京都大学 / 新制・課程博士 / 博士(情報学) / 甲第25439号 / 情博第877号 / 京都大学大学院情報学研究科数理工学専攻 / (主査)教授 山下 信雄, 教授 梅野 健, 准教授 加嶋 健司 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Στην παρούσα διατριβή μελετάμε το πρόβλημα της βελτιστοποίησης μη γραμμικών συναρτήσεων πολλών μεταβλητών, όπου η αντικειμενική συνάρτηση είναι συνεχώς διαφορίσιμη σε ένα ανοιχτό υποσύνολο του Rn. Αναπτύσσουμε μαθηματικές μεθόδους βελτιστοποίησης αποσκοπώντας στην επίλυση προβλημάτων μεγάλης κλίμακας, δηλαδή προβλημάτων των οποίων οι μεταβλητές είναι πολλές χιλιάδες, ακόμα και εκατομμύρια. Η βασική ιδέα των μεθόδων που αναπτύσσουμε έγκειται στη θεωρητική μελέτη των χαρακτηριστικών μεγεθών των Quasi-Newton ενημερώσεων ελάχιστης και μικρής μνήμης. Διατυπώνουμε θεωρήματα αναφορικά με το χαρακτηριστικό πολυώνυμο, τον αριθμό των διακριτών ιδιοτιμών και των αντίστοιχων ιδιοδιανυσμάτων. Εξάγουμε κλειστούς τύπους για τον υπολογισμό των ανωτέρω ποσοτήτων, αποφεύγοντας τόσο την αποθήκευση όσο και την παραγοντοποίηση πινάκων. Τα νέα θεωρητικά απoτελέσματα εφαρμόζονται αφενός μεν στην επίλυση μεγάλης κλίμακας υποπροβλημάτων περιοχής εμπιστοσύνης, χρησιμοποιώντας τη μέθοδο της σχεδόν ακριβούς λύσης, αφετέρου δε, στην καμπυλόγραμμη αναζήτηση, η οποία χρησιμοποιεί ένα ζεύγος κατευθύνσεων μείωσης, την Quasi-Newton κατεύθυνση και την κατεύθυνση αρνητικής καμπυλότητας. Η νέα μέθοδος μειώνει δραστικά τη χωρική πολυπλοκότητα των γνωστών αλγορίθμων του μη γραμμικού προγραμματισμού, διατηρώντας παράλληλα τις καλές ιδιότητες σύγκλισής τους. Ως αποτέλεσμα, οι προκύπτοντες νέοι αλγόριθμοι έχουν χωρική πολυπλοκότητα Θ(n). Τα αριθμητικά αποτελέσματα δείχνουν ότι οι νέοι αλγόριθμοι είναι αποδοτικοί, γρήγοροι και πολύ αποτελεσματικοί όταν χρησιμοποιούνται στην επίλυση προβλημάτων με πολλές μεταβλητές. / In this thesis we study the problem of minimizing nonlinear functions of several variables, where the objective function is continuously differentiable on an open subset of Rn. We develop mathematical optimization methods for solving large scale problems, i.e., problems whose variables are many thousands, even millions. The proposed method is based on the theoretical study of the properties of minimal and low memory Quasi-Newton updates. We establish theorems concerning the characteristic polynomial, the number of distinct eigenvalues and corresponding eigenvectors. We derive closed formulas for calculating these quantities, avoiding both the storage and factorization of matrices. The new theoretical results are applied in the large scale trust region subproblem for calculating nearly exact solutions as well as in a curvilinear search that uses a Quasi-Newton and a negative curvature direction. The new method is drastically reducing the spatial complexity of known algorithms of nonlinear programming. As a result, the new algorithms have spatial complexity Θ(n), while they are maintaining good convergence properties. The numerical results show that the proposed algorithms are efficient, fast and very effective when used in solving large scale problems.

This thesis proposes a new class of bivariate autoregressive conditional median duration models for matched high-frequency data and develops some inferential methods for an existing univariate model as well as the bivariate models introduced here to facilitate model fitting and forecasting. During the last two decades, the autoregressive conditional mean duration (ACD) model has been playing a dominant role in analyzing irregularly spaced high-frequency financial data. Univariate ACD models have been extensively discussed in the literature. However, some major challenges remain. The existing ACD models do not provide a good distributional fit to financial durations, which are right-skewed and often exhibit unimodal hazard rates. Birnbaum-Saunders (BS) distribution is capable of modeling a wide variety of positively skewed data. Median is not only a robust measure of central tendency, but also a natural scale parameter of the BS distribution. A class of conditional median duration models, the BS-ACD and the scale-mixture BS ACD models based on the BS, BS power-exponential and Student-t BS (BSt) distributions, have been suggested in the literature to improve the quality of the model fit. The BSt-ACD model is more flexible than the BS-ACD model in terms of kurtosis and skewness. In Chapter 2, we develop the maximum likelihood estimation method for the BSt-ACD model. The estimation is performed by utilizing a hybrid of optimization algorithms. The performance of the estimates is then examined through an extensive Monte Carlo simulation study. We also carry out model discrimination using both likelihood-based method and information-based criterion. Applications to real trade durations and comparison with existing alternatives are then made. The bivariate version of the ACD model has not received attention due to non-synchronicity. Although some bivariate generalizations of the ACD model have been introduced, they do not possess enough flexibility in modeling durations since they are conditional mean-based and do not account for non-monotonic hazard rates. Recently, the bivariate BS (BVBS) distribution has been developed with many desirable properties and characteristics. It allows for unimodal shapes of marginal hazard functions. In Chapter 3, upon using this bivariate BS distribution, we propose the BVBS-ACD model as a natural bivariate extension of the BS-ACD model. It enables us to jointly analyze matched duration series, and also capture the dependence between the two series. The maximum likelihood estimation of the model parameters and associated inferential methods have been developed. A Monte Carlo simulation study is then carried out to examine the performance of the proposed inferential methods. The goodness-of-fit and predictive performance of the model are also discussed. A real bivariate duration data analysis is provided to illustrate the developed methodology. The bivariate Student-t BS (BVBSt) distribution has been introduced in the literature as a robust extension of the BVBS distribution. It provides greater flexibility in terms of the kurtosis and skewness through the inclusion of an additional shape parameter. In Chapter 4, we propose the BVBSt-ACD model as a natural extension of the BSt-ACD model to the bivariate case. We then discuss the maximum likelihood estimation of the model parameters. A simulation study is carried out to investigate the performance of these estimators. Model discrimination is then done by using information-based criterion. Methods for evaluating the goodness-of-fit and predictive ability of the model are also discussed. A simulated data example is used to illustrate the proposed model as compared to the BVBS-ACD model. Finally, in Chapter 5, some concluding comments are made and also some problems for future research are mentioned. / Thesis / Master of Science (MSc)