Modern data analysis requires the study of tensors, or multi-way arrays. We consider the case where the dimension d is large and the order p is fixed. For dimension reduction and for interpretability, one considers tensor decompositions, where a tensor T can be decomposed into a sum of rank one tensors. In this thesis, I will describe some recent work that illustrate why and how to use decompositions for orthogonally decomposable tensors. Our developments are motivated by statistical applications where the data dimension is large. The estimation procedures will therefore aim to be computationally tractable while providing error rates that depend optimally on the dimension. A tensor is said to be orthogonally decomposable if it can be decomposed into rank one tensors whose component vectors are orthogonal. A number of data analysis tasks can be recast as the problem of estimating the component vectors from a noisy observation of an orthogonally decomposable tensor. In our first set of results, we study this decompositionproblem and derive perturbation bounds. For any two orthogonally decomposable tensors which are ε-perturbations of one another, we derive sharp upper bounds on the distances between their component vectors. While this is motivated by the extensive literature on bounds for perturbation of singular value decomposition, our work shows fundamental differences and requires new techniques. We show that tensor perturbation bounds have no dependence on eigengap, a quantity which is inevitable for matrices. Moreover, our perturbation bounds depend on the tensor spectral norm of the noise, and we provide examples to show that this leads to optimal error rates in several high dimensional statistical learning problems. Our results imply that matricizing a tensor is sub-optimal in terms of dimension dependence. The tensor perturbation bounds derived so far are universal, in that they depend only on the spectral norm of the perturbation. In subsequent chapters, we show that one can extract further information from how a noise is generated, and thus improve over tensor perturbation bounds both statistically and computationally. We demonstrate this approach for two different problems: first, in estimating a rank one spiked tensor perturbed by independent heavy-tailed noise entries; and secondly, in performing inference from moment tensors in independent component analysis. We find that an estimator benefits immensely— both in terms of statistical accuracy and computational feasibility — from additional information about the structure of the noise. In one chapter, we consider independent noise elements, and in the next, the noise arises as a difference of sample and population fourth moments. In both cases, our estimation procedures are determined so as to avoid accumulating the errors from different sources. In a departure from the tensor perturbation bounds, we also find that the spectral norm of the error tensor does not lead to the sharpest estimation error rates in these cases. The error rates of estimating the component vectors are affected only by the noise projected in certain directions, and due to the orthogonality of the signal tensor, the projected errors do not accumulate, and can be controlled more easily.

In this paper the Clifford Algebra is introduced and proposed as analternative to Gibbs' vector algebra as a unifying language for geometricoperations on vectors. Firstly, the algebra is constructed using a quotientof the tensor algebra and then its most important properties are proved,including how it enables division between vectors and how it is connected tothe exterior algebra. Further, the Clifford algebra is shown to naturallyembody the complex numbers and quaternions, whereupon its strength indescribing rotations is highlighted. Moreover, the wedge product, is shown asa way to generalize the cross product and reveal the true nature ofpseudovectors as bivectors. Lastly, we show how replacing the cross productwith the wedge product, within the Clifford algebra, naturally leads tosimplifying Maxwell's equations to a single equation.

The quadrupole coupling tensors at the B11 and Na23 sites in borax and tincalconite, two members of the hydrated sodium tetraborate family have been completely evaluated using broadline nuclear magnetic resonance (n.m.r.) techniques. The results have enabled deductions to be made concerning the crystal chemistry of boron, the coordinations of the sodium atoms and the crystallographic symmetries. Changes in the Li7 n.m.r. spectrum of ferroelectric lithium hydrazinium sulphate as a function of temperature were followed in detail. A previously undetected second order phase transition leading to a high temperature polymorph is complete at 164°C, The quadrupole coupling tensors at the Li7 sites have been completely evaluated in the two phases. The second moments or the proton n.m.r. spectrum from powdered lithium hydrazinium sulphate have been determined within the temperature range -183°C to 225°C, and single crystal data on the protons have been obtained at room temperature. It is shown that the nitrogen and hydrogen atoms exist in the structure as the hydrazinium ion NH2 and-NH3+. Detailed information is derived concerning the proton-proton vectors and the modes of reorientation of the NH2 and-NH3+ groups at various temperatures. Mechanisms for the ferro-electric switching at room temperature and for the transition to the high temperature polymorph are proposed on the basis of the experimental evidence. / Thesis / Doctor of Philosophy (PhD)

<p> An investigation of general Materials of Differential Type [MDT], and Motions With Constant Stretch History [MCSH] is presented. Rivlin-Ericksen tensors An are shown to result from a Taylor series expansion of the relative strain tensor Ct(T). Internal constraint in MDT is discussed. General Solutions of Motions of Differential Type are worked out. Dynamically possible stresses are found for certain irrotational motions. Theorems regarding necessary and sufficient conditions for MCSH are proved. A class of MCSH is introduced, and an approximate MCSH is suggested. Necessary equations regarding gradients of a scalar-valued tensor function are derived. </p> / Thesis / Master of Engineering (MEngr)

No description available.

In this thesis we consider the problem of clustering the data lying in a union of subspaces using spectral methods. Though the data generated may have high dimensionality, in many of the applications, such as motion segmentation and illumination invariant face clustering, the data resides in a union of subspaces having small dimensions. Furthermore, for a number of classification and inference problems, it is often useful to identify these subspaces and work with data in this smaller dimensional manifold. If the observations in each cluster were to be distributed around a centric, applying spectral clustering on an a nifty matrix built using distance based similarity measures between the data points have been used successfully to solve the problem. But it has been observed that using such pair-wise distance based measure between the data points to construct a similarity matrix is not sufficient to solve the subspace clustering problem. Hence, a major challenge is to end a similarity measure that can capture the information of the subspace the data lies in. This is the motivation to develop methods that use an affinity tensor by calculating similarity between multiple data points. One can then use spectral methods on these tensors to solve the subspace clustering problem. In order to keep the algorithm computationally feasible, one can employ column sampling strategies. However, the computational costs for performing the tensor factorization increases very quickly with increase in sampling rate. Fortunately, the advances in GPU computing has made it possible to perform many linear algebra operations several order of magnitudes faster than traditional CPU and multicourse computing. In this work, we develop parallel algorithms for subspace clustering on a GPU com-putting environment. We show that this gives us a significant speedup over the implementations on the CPU, which allows us to sample a larger fraction of the tensor and thereby achieve better accuracies. We empirically analyze the performance of these algorithms on a number of synthetically generated subspaces con gyrations. We ally demonstrate the effectiveness of these algorithms on the motion segmentation, handwritten digit clustering and illumination invariant face clustering and show that the performance of these algorithms are comparable with the state of the art approaches.

Herein we consider material tensors that pertain to thin sheets or thin films, which we model as two-dimensional objects. We assume that the thin sheet in question carries a crystallographic texture characterized by an orientation distribution function defined on the rotation group SO(3), which is almost transversely-isotropic about the sheet normal so that mechanical and physical properties of the thin sheet have weak planar-anisotropy. We present a procedure by which a special orthonormal basis can be determined in each tensor subspace invariant under the action of the orthogonal group O(2). We call members of such special bases irreducible basis tensors under O(2). For the class of thin sheets in question, we derive a representation formula in which each tensor in any given tensor subspace Z is written as the sum of a transversely-isotropic term and a linear combination of orthonormal irreducible basis tensors in Z, where the coefficients are given explicitly in terms of texture coefficients and undetermined material parameters. In addition to the general representation formula, we present also the specialized form for subspaces of tensor products of second-order symmetric tensors, a type commonly found in mechanics of materials.

Le suivi des zones côtières nécessite à la fois une bonne résolution spatiale, une bonne résolution spectraleassociée à un bon rapport signal sur bruit et enfin une bonne résolution temporelle pour visualiser deschangements rapides de couleur de l’eau.Les capteurs disponibles actuellement, et même ceux prévus prochainement, n’apportent pas à la fois unebonne résolution spatiale, spectrale ET temporelle. Dans cette étude, nous nous intéressons à la fusion de 2futurs capteurs qui s’inscrivent tous deux dans le programme Copernicus de l’agence spatiale européenne:MSI sur Sentinel-2 et OLCI sur Sentinel-3.Comme les capteurs MSI et OLCI ne fournissent pas encore d’images, il a fallu les simuler. Pour cela nousavons eu recours aux images hyperspectrales du capteur HICO. Nous avons alors proposé 3 méthodes : uneadaptation de la méthode ARSIS à la fusion d’images multispectrales (ARSIS), une méthode de fusion baséesur la factorisation de tenseurs non-négatifs (Tenseur) et une méthode de fusion basée sur l’inversion dematrices (Inversion)Ces 3 méthodes ont tout d’abord été évaluées à l’aide de paramètres statistiques entre les images obtenuespar fusion et l’image « parfaite » ainsi que sur les résultats d’estimation de paramètres biophysiques obtenuspar minimisation du modèle de transfert radiatif dans l’eau. / Monitoring coastal areas requires both a good spatial resolution, good spectral resolution associated with agood signal to noise ratio and finally a good temporal resolution to visualize rapid changes in water color.Available now, and even those planed soon, sensors do not provide both a good spatial, spectral ANDtemporal resolution. In this study, we are interested in the image fusion of two future sensors which are bothpart of the Copernicus program of the European Space Agency: MSI on Sentinel-2 and OLCI on Sentinel-3.Such as MSI and OLCI do not provide image yet, it was necessary to simulate them. We then used thehyperspectral imager HICO and we then proposed three methods: an adaptation of the method ARSIS fusionof multispectral images (ARSIS), a fusion method based on the non-negative factorization tensors (Tensor)and a fusion method based on the inversion de matrices (Inversion).These three methods were first evaluated using statistical parameters between images obtained by fusionand the "perfect" image as well as the estimation results of biophysical parameters obtained by minimizingthe radiative transfer model in water.

The principal contribution of this dissertation is a theory of Strongly Orthotropic Continuum Mechanics that is derived entirely from an assertion of geometric strain indeterminacy. Implementable into the finite element method, it can resolve widespread kinematic misrepresentations and offer unique and purportedly exact strain-induced energies by removing the assumptions of strain tensor symmetry. This continuum theory births the proposal of a new class of physical tensors described as the Intrinsic Field Tensors capable of generalising the response of most classical mechanical metrics, a number of specialised formulations and the solutions shown to be kinematically intermediate. A series of numerical examples demonstrate Euclidean objectivity, material frame-indifference, patch test satisfaction, and agreement between the subsequent Material Principal Co-rotation and P??I??C decomposition methods that produce the intermediary stress/strain fields. The encompassing theory has wide applicability owing to its fundamental divergence from conventional mechanics, it offers non-trivial outcomes when applied to even very simple problems and its use of not the Eulerian, Lagrangian but the Intrinsic Frame generates previously unreported results in strongly orthotropic continua.

During the last decade, my main research topic was on medical image analysis, and more particularly on image registration. However, I was also following in background a more theoretical research track on statistical computing on manifolds. With the recent emergence of computational anatomy, this topic gained a lot of importance in the medical image analysis community. During the writing of this habilitation manuscript, I felt that it was time to present a more simple and uni ed view of how it works and why it can be important. This is why the usual short synthesis of the habilitation became a hundred pages long text where I tried to synthesizes the main notions of statistical computing on manifolds with application in registration and computational anatomy. Of course, this synthesis is centered on and illustrated by my personal research work.