This dissertation addresses the task of learning to rank, both in the supervised and unsupervised settings, by exploiting the interplay of convex functions, monotonic mappings and their fixed points. In the supervised setting of learning to rank, one wishes to learn from examples of correctly ordered items whereas in the unsupervised setting, one tries to maximize some quantitatively defined characteristic of a "good" ranking. A ranking method selects one permutation from among the combinatorially many permutations defined on the items to rank. Accomplishing this optimally in the supervised setting, with minimal loss in generality, if any, is challenging. In this dissertation this problem is addressed by optimizing, globally and efficiently, a statistically consistent loss functional over the class of compositions of a linear function by an arbitrary, strictly monotonic, separable mapping with large margins. This capability also enables learning the parameters of a generalized linear model with an unknown link function. The method can handle infinite dimensional feature spaces if the corresponding kernel function is known. In the unsupervised setting, a popular ranking approach is is link analysis over a graph of recommendations, as exemplified by pagerank. This dissertation shows that pagerank may be viewed as an instance of an unsupervised consensus optimization problem. The dissertation then solves a more general problem of unsupervised consensus over noisy, directed recommendation graphs that have uncertainty over the set of "out" edges that emanate from a vertex. The proposed consensus rank is essentially the pagerank over the expected edge-set, where the expectation is computed over the distribution that achieves the most agreeable consensus. This consensus is measured geometrically by a suitable Bregman divergence between the consensus rank and the ranks induced by item specific distributions Real world deployed ranking methods need to be resistant to spam, a particularly sophisticated type of which is link-spam. A popular class of countermeasures "de-spam" the corrupted webgraph by removing abusive pages identified by supervised learning. Since exhaustive detection and neutralization is infeasible, there is a need for ranking functions that can, on one hand, attenuate the effects of link-spam without supervision and on the other hand, counter spam more aggressively when supervision is available. A family of non-linear, iteratively defined monotonic functions is proposed that propagates "rank" and "trust" scores through the webgraph. It relies on non-linearity, monotonicity and Schurconvexity to provide the resistance against spam. / text

published_or_final_version / Mathematics / Doctoral / Doctor of Philosophy

A norm ||⋅|| on a Banach space X is Fréchet differentiable at x ∈ X if there is a functional ∫∈ X* such that [Formula Omitted] This concept reflects the smoothness characteristics of X. A dual Banach space X* has the Radon-Nikodym Property (RNP) if whenever C ⊂ X* is weak*-compact and convex, and ∈ > 0, there is an x ∈ X and an ⍺ > 0 such that diameter [Formula Omitted] this property reflects the convexity characteristics of X*. Culminating several years of work by many researchers, the following theorem established a strong connection between the smoothness of X and the convexity of X*: Every equivalent norm on X is Fréchet differentiable on a dense set if and only if X* has the RNP. A more general measure of convexity has been recently receiving a great deal of attention: A dual Banach space X* has the weak* Convex Point-of-Continuity Property (C*PCP) if whenever ɸ ≠ C ⊂ X* is weak*-compact and convex, and ∈ > 0, there is a weak*-open set V such that V ⋂ C ≠ ɸ and diam V ⋂ C < ∈. In this thesis, we develop the corresponding smoothness properties of X which are dual to C*PCP. For this, a new type of differentiability, called cofinite Fréchet differentiability, is introduced, and we establish the following theorem: Every equivalent norm on X is cofinitely Fréchet differentiable everywhere if and only if X* has the C*PCP. Representing joint work with R. Deville, G. Godefroy and V. Zizler, an alternate approach is developed in the case when X is separable. We show that if X is separable, then every equivalent norm on X which has a strictly convex dual is Fréchet differentiable on a dense set if and only if X* has the C*PCP, if and only if every equivalent norm on X which is Gâteaux differentiable (everywhere) is Fréchet differentiable on a dense set. This result is used to show that if X* does not have the C*PCP, then there is a subspace Y of X such that neither Y* nor (X/Y)* have the C*PCP, yet both Y and X/Y have finite dimensional Schauder decompositions. The corresponding result for spaces X* failing the RNP remains open. / Science, Faculty of / Mathematics, Department of / Graduate

El texto completo de este trabajo no está disponible en el Repositorio Académico UPC por restricciones de la casa editorial donde ha sido publicado. / The present investigation focuses on the mathematical concept of convexity, as the main tool for the graphic construction of bounded surfaces explicitly and implicitly described, as well as the construction of unconventional solids using GeoGebra. Two cases are presented in which the importance of the concept of convexity is highlighted, in the first situation the convexity is used in the argument of the surface command together with the curves that delimit it to graph a bounded surface, while in the second situation the convexity is evidenced by expressing the coordinates of the surface in parametric form. On the other hand, the 3D graphic view combined with the GeoGebra AR tool allows one to visualize, manipulate, understand and improve the abstraction of mathematical objects that are built in three-dimensional space in a dynamic and friendly environment. These constructions in three-dimensional space that are complex when sketching them with pencil and paper are easier when linking the mathematical definitions with free software such as GeoGebra. / Revisón por pares

Dans de nombreuses applications, nous souhaitons interpoler ou approcher une fonction de plusieurs variables possédant certaines propriétés ou “formes” géométriques, telles que la régularité, la monotonie, la convexité ou la non-négativité. Ces propriétés sont importantes pourdes applications en physique (par exemple, la courbe pression-volume doit avoir une dérivée non négative), aussi bien où le problème de l’interpolation conservant la forme est essentiel dans divers problèmes de l’industrie (par exemple, modélisation automobile, construction de la surface dumasque). Par conséquent, une question importante se pose : comment calculer la meilleure approximation possible à une fonction donnée f lorsque certaines de ses propriétés caractéristiques supplémentaires sont connues ?Cette thèse présente plusieurs nouvelles techniques pour trouver une bonne approximation des fonctions de plusieurs variables par des opérateurs linéaires dont l’erreur d’approximation A( f ) - f garde un signe constant pour toute fonction f satisfaisant une certaine convexité généralisée. Nous nous concentrons dans cette thèse sur la classe des fonctions convexesou fortement convexes. Nous décrirons comment la connaissance a priori de cette information peut être utilisée pour déterminer une bonne majoration de l’erreur pour des fonctions continuellement différentiables avec des gradients Lipschitz continus. Plus précisément, nous montrons que les estimations d’erreur basées sur ces opérateurs sont toujours contrôléespar les constantes de Lipschitz des gradients, le paramètre de la convexité forte ainsi que l’erreur commise associée à l’utilisation de la fonction quadratique. En supposant en plus que la fonction que nous voulons approcher est également fortement convexe, nous établissons de meilleures bornes inférieures et supérieures pour les estimations d’erreur de l’approximation. Lesméthodes de quadrature multidimensionnelle jouent un rôle important, voire fondamental, en analyse numérique. Une analyse satisfaisante des erreurs provenant de l’utilisationdes formules de quadrature multidimensionnelle est bien moins étudiée que dans le cas d’une variable. Nous proposons une méthode d’approximation de l’intégrale d’une fonction réelle donnée à plusieurs variables par des formules de quadrature, qui conduisent à des valeurs approchées par excès (respectivement par défaut) des intégrales des fonctions ayantun certain type de convexité. Nous verrons aussi, comme nous l’avons fait pour l’approximation des fonctions, que pour de telles formules d’intégration, on peut établir un résultat de caractérisation en termes d’estimations d’erreur. En outre, nous avons étudié le problèmede l’approximation d’une intégrale définie d’une fonction donnée quand un certain nombre d’intégrales de cette fonction sur certaines sections hyperplanes d’un l’hyper-rectangle sont seulement disponibles.La motivation derrière ce type de problème est multiple. Il se pose dans de nombreuses applications, en particulier en physique expérimentale et en ingénierie, où les valeurs standards des échantillons discrets des fonctions ne sont pas disponibles, mais où seulement leurs valeurs moyennes sont accessibles. Par exemple, ce type de données apparaît naturellement dans la tomographie par ordinateur avec ses nombreuses applications en médecine, radiologie, géologie, entre autres. / In many applications, we may wish to interpolate or approximate a multivariate function possessing certain geometric properties or “shapes” such as smoothness, monotonicity, convexityor nonnegativity. These properties may be desirable for physical (e.g., a volume-pressure curve should have a nonnegative derivative) or practical reasons where the problem of shape preserving interpolation is important in various problems occurring in industry (e.g., car modelling, construction of mask surface). Hence, an important question arises: How can we compute the best possible approximation to a given function f when some of its additional characteristic properties are known?This thesis presents several new techniques to find a good approximation of multivariate functions by a new kind of linear operators, which approximate from above (or, respectively, from below) all functions having certain generalized convexity. We focus on the class of convex and strongly convex functions. We would wish to use this additional informationin order to get a good approximation of f . We will describe how this additional condition can be used to derive sharp error estimates for continuously differentiable functions with Lipschitz continuous gradients. More precisely we show that the error estimates based on such operators are always controlled by the Lipschitz constants of the gradients, the convexity parameter of the strong convexity and the error associated with using the quadratic function. Assuming, in addition, that the function, we want to approximate, is also strongly convex, we establish sharp upper as well as lower refined bounds for the error estimates.Approximation of integrals of multivariate functions is a notoriously difficult tasks and satisfactory error analysis is far less well studied than in the univariate case. We propose a methodto approximate the integral of a given multivariate function by cubature formulas (numerical integration), which approximate from above (or from below) all functions having a certain type of convexity. We shall also see, as we did for for approximation of functions, that for such integration formulas, we can establish a characterization result in terms of sharp error estimates. Also, we investigated the problem of approximating a definite integral of a given function when a number of integrals of this function over certain hyperplane sections of d-dimensional hyper-rectangle are only available rather than its values at some points.The motivation for this problem is multifold. It arises in many applications, especially in experimental physics and engineering, where the standard discrete sample values fromfunctions are not available, but only their mean values are accessible. For instance, this data type appears naturally in computer tomography with its many applications inmedicine, radiology, geology, amongst others.

This thesis consists of two parts: digital geometry and discrete optimization. In the first part we study the structure of digital straight line segments. We also study digital curves from a combinatorial point of view. In Paper I we study the straightness in the 8-connected plane and in the Khalimsky plane by considering vertical distances and unions of two segments. We show that we can investigate the straightness of Khalimsky arcs by using our knowledge from the 8-connected plane. In Paper II we determine the number of Khalimsky-continuous functions with 2, 3 and 4 points in their codomain. These enumerations yield examples of known sequences as well as new ones. We also study the asymptotic behavior of each of them. In Paper III we study the number of Khalimsky-continuous functions with codomain Z and N. This gives us examples of Schröder and Delannoy numbers. As a byproduct we get some relations between these numbers. In Paper IV we study the number of Khalimsky-continuous functions between two points in a rectangle. Using a generating function we get a recurrence formula yielding this numbers.   In the second part we study an analogue of discrete convexity, namely lateral convexity. In Paper V we define by means of difference operators the class of lateral convexity. The functions have plus infinity in their codomain. For the real-valued functions we need to check the difference operators for a smaller number of points. We study the relation between this class and integral convexity. In Paper VI we study the marginal function of real-valued functions in this class and its generalization. We show that for two points with a certain distance we have a Lipschitz property for the points where the infimum is attained. We show that if a function is in this class, the marginal function is also in the same class. / At the time of the doctoral defense, the following papers were unpublished and had a status as follows: Paper 4: Submitted. Paper 5: Manuscript. Paper 6: Manuscript.

Orthographically-projected biological motion point-light displays generally contain no information about their in-depth orientation, yet observers consistently prefer the facing-the-viewer (FTV) interpretation (Vanrie, Dekeyser and Verfaillie, 2004). This bias has been attributed to the social relevance of such stimuli (Brooks et al., 2008) although local stimulus properties appear to influence the bias (Schouten, Troje and Verfaillie, 2011). In the present study we investigated the cause of the FTV bias. In Experiment 1 we compared FTV bias for various configurations of stick-figures and depth ambiguous human silhouettes. The FTV bias was not present for silhouettes, but was strongly elicited for most stick-figures. We concluded that local attitude assignments for intrinsic structures of stick-figures are subject to inferences about the flexion of body surfaces, and that a visual bias that assumes surfaces to be convex drives the FTV bias. In Experiment 2 we manipulated silhouettes to permit local attitude assignments by using point-lights on emphasized flexion points. As predicted, the inclusion of intrinsic structures produced FTV bias for silhouettes. The results help to unify various findings regarding the FTV bias. We conclude that the FTV bias emerges during the 2 ½-D sketch stage of visual processing (Marr and Nishihara, 1978). / Thesis (Master, Psychology) -- Queen's University, 2013-08-08 19:06:06.84

The purpose of this study was to investigate the social perceptions of subjects with differing lip position and facial convexity in three dimensions. A 3dMD camera (3dMD, Atlanta, GA) was used to capture 3D images of 9 subjects’ faces. The images were altered to have ideal lip position and ideal convexity, ideal lip position and Class II convexity, Class II lip position and ideal convexity, and Class II lip position and Class II convexity. 400 laypersons rated their perceptions of the subjects’ athletic ability, popularity, leadership, and intelligence on a VAS scale. Subjects with ideal lip position relative to the E-line were rated significantly higher for leadership and intelligence. Males with ideal facial convexity were judged to be better leaders and more intelligent than those with Class II convexity. Subjects with ideal lip position were given the highest mean VAS scores for all four social attributes. The perception of differences related to facial convexity was inconsistent.

Polyconvexity is a standard assumption on hyperelastic stored energy densities which, together with some growth conditions, ensures the weak lower semicontinuity of the respective energy functional. The present work first reviews known results about gradient polyconvexity, introduced by Benešová, Kružík and Schlömerkemper in 2017. It is an alternative property to polyconvexity, better-suited e.g. for the modelling of shape-memory alloys. The principal result of this thesis is the extension of an elastic material model with gradient polyconvex energy functional to an elastoplastic body and proving the existence of an energetic solution to an associated rate- independent evolution problem, proceeding from previous work of Mielke, Francfort and Mainik. 1

Thesis advisor: Robert Murphey / This paper demonstrates that a linear Phillips Curve has neither theoretical nor empirical justification. I first alter the traditional linear model specification to allow for non-linearity between inflation and unemployment. I show that these non-linear models produce greater R2’s than similar linear versions. I provide theoretical justification for the non-linear models and demonstrate why the theoretical reasoning for linear models is flawed. Finally, by introducing the natural rate of unemployment as a separate independent variable, I increase the explanatory power of the model. I allow the natural rate’s marginal effect on inflation to vary with time and suggest a theoretical framework that supports this final model. I conclude that non-linearity and therefore convexity between inflation and unemployment is the correct framework under any time period for Phillips Curve analysis and application. / Thesis (BA) — Boston College, 2017. / Submitted to: Boston College. College of Arts and Sciences. / Discipline: Departmental Honors. / Discipline: Economics.