Cette thèse se compose de deux parties indépendantes qui portent sur le contrôle stochastique avec des espérances non linéaires et les équations stochastiques rétrogrades (EDSR), ainsi que sur les méthodes numériques de résolution de ces équations. Dans la première partie on étudie une nouvelle classe d'équations stochastiques rétrogrades, dont la particularité est que la condition terminale n'est pas fixée mais vérifie une contrainte non linéaire exprimée en termes de "f-espérances". Ce nouvel objet mathématique est étroitement lié aux problèmes de couverture approchée des options européennes où le risque de perte est quantifié en termes de mesures de risque dynamiques, induites par la solution d'une EDSR non linéaire. Dans le chapitre suivant on s'intéresse aux problèmes d'arrêt optimal pour les mesures de risque dynamiques avec sauts. Plus précisément, on caractérise dans un cadre markovien la mesure de risque minimale associée à une position financière comme l'unique solution de viscosité d'un problème d'obstacle pour une équation intégro-différentielle. Dans le troisième chapitre, on établit un principe de programmation dynamique faible pour un problème mixte de contrôle stochastique et d'arrêt optimal avec des espérances non linéaires, qui est utilisé pour obtenir les EDP associées.La spécificité de ce travail réside dans le fait que la fonction de gain terminal ne satisfait aucune condition de régularité (elle est seulement considérée mesurable), ce qui n'a pas été le cas dans la littérature précédente. Dans le chapitre suivant, on introduit un nouveau problème de jeux stochastiques, qui peut être vu comme un jeu de Dynkin généralisé (avec des espérances non linéaires). On montre que ce jeu admet une fonction valeur et on obtient des conditions suffisantes pour l'existence d'un point selle. On prouve que la fonction valeur correspond à l'unique solution d'une équation stochastique rétrograde doublement réfléchie avec un générateur non linéaire général. Cette caractérisation permet d'obtenir de nouveaux résultats sur les EDSR doublement réfléchies avec sauts. Le problème de jeu de Dynkin généralisé est ensuite étudié dans un cadre markovien.Dans la deuxième partie, on s'intéresse aux méthodes numériques pour les équations stochastiques rétrogrades doublement réfléchies avec sauts et barrières irrégulières, admettant des sauts prévisibles et totalement inaccessibles. Dans un premier chapitre, on propose un schéma numérique qui repose sur la méthode de pénalisation et l'approximation de la solution d'une EDSR par une suite d'EDSR discrètes dirigées par deux arbres binomiaux indépendants (un qui approxime le mouvement brownien et l'autre le processus de Poisson composé). Dans le deuxième chapitre, on construit un schéma en discrétisant directement l'équation stochastique rétrograde doublement réfléchie, schéma qui présente l'avantage de ne plus dépendre du paramètre de pénalisation. On prouve la convergence des deux schémas numériques et on illustre avec des exemples numériques les résultats théoriques. / This thesis consists of two independent parts which deal with stochastic control with nonlinear expectations and backward stochastic differential equations (BSDE), as well as with the numerical methods for solving these equations.We begin the first part by introducing and studying a new class of backward stochastic differential equations, whose characteristic is that the terminal condition is not fixed, but only satisfies a nonlinear constraint expressed in terms of "f - expectations". This new mathematical object is closely related to the approximative hedging of an European option, when the shortfall risk is quantified in terms of dynamic risk measures, induced by the solution of a nonlinear BSDE. In the next chapter we study an optimal stopping problem for dynamic risk measures with jumps.More precisely, we characterize in a Markovian framework the minimal risk measure associated to a financial position as the unique viscosity solution of an obstacle problem for partial integrodifferential equations. In the third chapter, we establish a weak dynamic programming principle for a mixed stochastic control problem / optimal stopping with nonlinear expectations, which is used to derive the associated PDE. The specificity of this work consists in the fact that the terminal reward does not satisfy any regularity condition (it is considered only measurable), which was not the case in the previous literature. In the next chapter, we introduce a new game problem, which can be seen as a generalized Dynkin game (with nonlinear expectations ). We show that this game admits a value function and establish sufficient conditions ensuring the existence of a saddle point . We prove that the value function corresponds to the unique solution of a doubly reected backward stochastic equation (DRBSDE) with a nonlinear general driver. This characterization allows us to obtain new results on DRBSDEs with jumps. The generalized Dynkin game is finally addressed in a Markovian framework.In the second part, we are interested in numerical methods for doubly reected BSDEs with jumps and irregular barriers, admitting both predictable and totally inaccesibles jumps. In the first chapter we provide a numerical scheme based on the penalisation method and the approximation of the solution of a BSDE by a sequence of discrete BSDEs driven by two independent random walks (one approximates the Brownian motion and the other one the compensated Poisson process). In the second chapter, we construct an alternative scheme based on the direct discretisation of the DRBSDE, scheme which presents the advantage of not depending anymore on the penalization parameter. We prove the convergence of the two schemes and illustrate the theoretical results with some numerical examples.

Experimental studies have shown that the protein abundance in living cells varies from few tens to several thousands molecules per species. Molecular fluctuations roughly scale as the inverse square root of the number of molecules due to the random timing of reactions. It is hence expected that intrinsic noise plays an important role in the dynamics of biochemical networks. The Chemical Master Equation is the accepted description of these systems under well-mixed conditions. Because analytical solutions to this equation are available only for simple systems, one often has to resort to approximation methods. A popular technique is an expansion in the inverse volume to which the reactants are confined, called van Kampen's system size expansion. Its leading order terms are given by the phenomenological rate equations and the linear noise approximation that quantify the mean concentrations and the Gaussian fluctuations about them, respectively. While these approximations are valid in the limit of large molecule numbers, it is known that physiological conditions often imply low molecule numbers. We here develop systematic approximation methods based on higher terms in the system size expansion for general biochemical networks. We present an asymptotic series for the moments of the Chemical Master Equation that can be computed to arbitrary precision in the system size expansion. We then derive an analytical approximation of the corresponding time-dependent probability distribution. Finally, we devise a diagrammatic technique based on the path-integral method that allows to compute time-correlation functions. We show through the use of biological examples that the first few terms of the expansion yield accurate approximations even for low number of molecules. The theory is hence expected to closely resemble the outcomes of single cell experiments.

Stochastic differential equations are becoming a popular tool for modeling the transport and acceleration of cosmic rays in the heliosphere. In diffusive shock acceleration, cosmic rays diffuse across a region of discontinuity where the up- stream diffusion coefficient abruptly changes to the downstream value. Because the method of stochastic integration has not yet been developed to handle these types of discontinuities, I utilize methods and ideas from probability theory to develop a conceptual framework for the treatment of such discontinuities. Using this framework, I then produce some simple numerical algorithms that allow one to incorporate and simulate a variety of discontinuities (or boundary conditions) using stochastic integration. These algorithms were then modified to create a new algorithm which incorporates the discontinuous change in diffusion coefficient found in shock acceleration (known as Skew Brownian Motion). The originality of this algorithm lies in the fact that it is the first of its kind to be statistically exact, so that one obtains accuracy without the use of approximations (other than the machine precision error). I then apply this algorithm to model the problem of diffusive shock acceleration, modifying it to incorporate the additional effect of the discontinuous flow speed profile found at the shock. A steady-state solution is obtained that accurately simulates this phenomenon. This result represents a significant improvement over previous approximation algorithms, and will be useful for the simulation of discontinuous diffusion processes in other fields, such as biology and finance.

We introduce computationally efficient Monte Carlo methods for studying the statistics of stochastic differential equations in two distinct settings. In the first, we derive importance sampling methods for data assimilation when the noise in the model and observations are small. The methods are formulated in discrete time, where the "posterior" distribution we want to sample from can be analyzed in an accessible small noise expansion. We show that a "symmetrization" procedure akin to antithetic coupling can improve the order of accuracy of the sampling methods, which is illustrated with numerical examples. In the second setting, we develop "stochastic continuation" methods to estimate level sets for statistics of stochastic differential equations with respect to their parameters. We adapt Keller's Pseudo-Arclength continuation method to this setting using stochastic approximation, and generalized least squares regression. Furthermore, we show that the methods can be improved through the use of coupling methods to reduce the variance of the derivative estimates that are involved.

Orientador: Maria Cristina de Castro Cunha / Tese (doutorado) - Universidade Estadual de Campinas, Instituto de Matemática Estatística e Computação Científica / Made available in DSpace on 2018-08-26T19:33:01Z (GMT). No. of bitstreams: 1 Campos_FabioAntonioAraujode_D.pdf: 1374668 bytes, checksum: 6318414d486cf4810705b84e0d722e77 (MD5) Previous issue date: 2015 / Resumo: Neste trabalho estudamos o sistema de equações diferenciais estocásticas obtido na linearização do modelo de propagação de ondas acústicas. Mais especificamente, analisamos métodos para solução do sistema de equações diferenciais usado na acústica linear, onde a matriz com dados aleatórios e um vetor de funções aleatórias que define as condições iniciais. Além do tradicional Método de Monte Carlo aplicamos o Método de Transformações de Variáveis Aleatórias e o Método de Galerkin Estocástico. Apresentamos resultados obtidos usando diferentes distribuições de probabilidades dos dados do problema. Também comparamos os métodos através da distribuição de probabilidade e momentos estatísticos da solução / Abstract: On the present work we study the system of stochastic differential equations obtained from the linearization of the propagation model of acoustic waves. More specifically we analyze methods for the solution of the system of differential equations used in the linear acoustics, where the matrix with random data and a vector of random functions defining initial conditions. In addition to the traditional Monte Carlo Method we apply the Variable Transformations of Random Method and the Galerkin Stochastic Method. We present results obtained using different probability distributions of problem data. We also compared the methods through the distribution of probabilities and statistical moments of the solution / Doutorado / Matematica Aplicada / Doutor em Matemática Aplicada

In the thesis, multivariate fractional Brownian motions with possibly different Hurst indices in different coordinates are considered and a Girsanov-type theo- rem for these processes is shown. Two applications of this theorem to stochastic differential equations driven by multivariate fractional Brownian motions (SDEs) are given. Firstly, the existence of a weak solution to an SDE with a drift coeffi- cient that can be written as a sum of a regular and a singular part and a diffusion coefficient that is dependent on time and satisfies suitable conditions is shown. The results are applied for the proof of existence of a weak solution of an equation describing stochastic harmonic oscillator. Secondly, the Girsanov-type theorem is used to find the maximum likelihood scalar estimator that appears in the drift of an SDE with additive noise. 1

The use of the non-linear SPDEs are inevitable in both physics and applied mathematics since many of the physical phenomena in nature can be effectively modeled in random and non-linear way. The Kardar-Parisi-Zhang (KPZ) equation is well-known for its applications in describing various statistical mechanical models including randomly growing surfaces, directed polymers and interacting particle systems. We consider the upper and lower tail probabilities for the centered (by time$/24$) and scaled (according to KPZ time$^{1/3}$ scaling) one-point distribution of the Cole-Hopf solution of the KPZ equation. We provide the first tight bounds on the lower tail probability of the one point distribution of the KPZ equation with narrow wedge initial data. Our bounds hold for all sufficiently large times $T$ and demonstrates a crossover between super-exponential decay with exponent $\tfrac{5}{2}$ (and leading pre-factor $\frac{4}{15\pi} T^{1/3}$) for tail depth greater than $T^{2/3}$ (deep tail), and exponent $3$ (with leading pre-factor at least $\frac{1}{12}$) for tail depth less than $T^{2/3}$ (shallow tail). We also consider the case when the initial data is drawn from a very general class. For the lower tail, we prove an upper bound which demonstrates a crossover from super-exponential decay with exponent $3$ in the shallow tail to an exponent $\frac{5}{2}$ in the deep tail. For the upper tail, we prove super-exponential decay bounds with exponent $\frac{3}{2}$ at all depths in the tail. We study the correlation of fluctuations of the narrow wedge solution to the KPZ equation at two different times. We show that when the times are close to each other, the correlation approaches one at a power-law rate with exponent $\frac{2}{3}$, while when the two times are remote from each other, the correlation tends to zero at a power-law rate with exponent $-\frac{1}{3}$.

NR 20140805

Turbulent collapse and recovery are both observed to occur abruptly in the atmospheric stable boundary layer (SBL). The understanding and predictability of turbulent recovery remains limited, reducing numerical weather prediction accuracy. Previous studies have shown that regime occupation is the result of the net effect of highly variable processes, from turbulent to synoptic scales, making stochastic methods a compelling approach. Idealized stable boundary layer models have shown that under some circumstances, regimes can be related to the stable branches of model equilibria, and an additional unstable equilibrium is predicted. This work seeks to determine the extent to which the SBL regime occupation can be explained using a one-dimensional stochastic differential equation (SDE). The drift and diffusion coefficients of the SDE of an input time series are approximated from the statistics of its averaged time tendencies. These approximated coefficients are fit using Gaussian Process Regression. Probabilistic estimates of the system's equilibrium points are then found and used to create an empirical bifurcation diagram without making any prior assumptions on the dynamical form of the system. This data driven bifurcation diagram is compared to modelled predictions. The analysis is repeated on several meteorological towers around the world to assess the influence of local meteorological settings. This work provides empirical insights into the nature of regime dynamics and the extent to which the SBL displays hysteresis. / Graduate

In this thesis, we center our research around the analytical approximation of American put options with the Multilevel Monte Carlo simulation approach. The focus lies on reducing the computational complexity of estimating an expected value arising from a stochastic differential equation. Numerical results showcase that the simulations are consistent with the theoretical order of convergence of Monte Carlo simulations. The approximations are accurate and considerately more computationally efficient than the standard Monte Carlo simulation method.