DOCSLIB.ORG

Mean-Field Langevin Dynamics and Neural Networks

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Mean-Field Langevin Dynamics and Neural Networks Mean-Field Langevin Dynamics and Neural Networks Zhenjie Ren CEREMADE, Université Paris-Dauphine joint works with Giovanni Conforti, Kaitong Hu, Anna Kazeykina, David Siska, Lukasz Szpruch, Xiaolu Tan, Junjian Yang MATH-IMS Joint Applied Mathematics Colloquium Series August 28, 2020 Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 1 / 37 Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 ∗ ∗ − f (x) m (x; v) m (x) = Ce σ2 In particular, f does NOT need to be convex. Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 In particular, f does NOT need to be convex. Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 2 1 2 ∗ − 2 f (x) ∗ − f (x)+ jvj m (x) = Ce σ m (x; v) = Ce σ2 2 Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 In particular, f does NOT need to be convex. Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 2 1 2 ∗ − 2 f (x) ∗ − f (x)+ jvj m (x) = Ce σ ! δarg min f (x) m (x; v) = Ce σ2 2 Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 In particular, f does NOT need to be convex. Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 ∗ − f (x) 1 ∗ σ2 m (x; v) ! δ 2 m (x) = Ce ! δarg min f (x) arg min f (x)+ 2 jvj Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 In particular, f does NOT need to be convex. Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 ∗ ∗ − 2 f (x) m (x; v) ! δ(arg min f (x);0) m (x) = Ce σ ! δarg min f (x) Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 Classical Langevin dynamics and non-convex optimization The Langevin dynamics was first introduced in statistical physics to describe the motion of a particle with position X and velocity V in a potential field rx f subject to damping and random collision. Overdamped Langevin dynamics Underdamped Langevin dynamics dXt = −rx f (Xt )dt + σdWt dXt = Vt dt dVt = −rx f (Xt )−γVt dt+σdWt Under mild conditions, the two Markov diffusions admits unique invariant measures whose densities read: Overdamped Langevin dynamics Underdamped Langevin dynamics 2 ∗ ∗ − 2 f (x) m (x; v) ! δ(arg min f (x);0) m (x) = Ce σ ! δarg min f (x) In particular, f does NOT need to be convex. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 2 / 37 Taking into account the Brownian noise with constant σ, we may produce samplings of the invariant measures the overdamped Langevin , MCMC the underdamped Langevin , Hamiltonian MCMC The convergence rate of MCMC algorithm is in general dimension dependent! One may diminish σ # 0 along the simulation ) Simulation annealing. Relation with classical algorithms If we overlook the Brownian noise, then the overdamped process ) gradient descent algorithm the underdamped process ) Hamiltonian gradient descent algorithm But their convergence to the minimizer is ensured only for convex potential function f . Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 3 / 37 One may diminish σ # 0 along the simulation ) Simulation annealing. Relation with classical algorithms If we overlook the Brownian noise, then the overdamped process ) gradient descent algorithm the underdamped process ) Hamiltonian gradient descent algorithm But their convergence to the minimizer is ensured only for convex potential function f . Taking into account the Brownian noise with constant σ, we may produce samplings of the invariant measures the overdamped Langevin , MCMC the underdamped Langevin , Hamiltonian MCMC The convergence rate of MCMC algorithm is in general dimension dependent! Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 3 / 37 Relation with classical algorithms If we overlook the Brownian noise, then the overdamped process ) gradient descent algorithm the underdamped process ) Hamiltonian gradient descent algorithm But their convergence to the minimizer is ensured only for convex potential function f . Taking into account the Brownian noise with constant σ, we may produce samplings of the invariant measures the overdamped Langevin , MCMC the underdamped Langevin , Hamiltonian MCMC The convergence rate of MCMC algorithm is in general dimension dependent! One may diminish σ # 0 along the simulation ) Simulation annealing. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 3 / 37 It is natural to study this problem using Mean-field Langevin equations. Deep neural networks The deep neural networks have won and continue gaining impressive success in various applications. Mathematically speaking, we may approximate a given function f with the parametrized function: ni X i i i f (z) ≈ 'n ◦ · · · ◦ '1(z); where 'i (z) := ck '(Ak z + bk ) k=1 and ' is a given non-constant, bounded, continuous activation function. The expressiveness of the neural network is ensured by the universal representation theorem. However, the efficiency of such over-parametrized, non-convex optimization is still a mystery for mathematical analysis. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 4 / 37 Deep neural networks The deep neural networks have won and continue gaining impressive success in various applications. Mathematically speaking, we may approximate a given function f with the parametrized function: ni X i i i f (z) ≈ 'n ◦ · · · ◦ '1(z); where 'i (z) := ck '(Ak z + bk ) k=1 and ' is a given non-constant, bounded, continuous activation function. The expressiveness of the neural network is ensured by the universal representation theorem. However, the efficiency of such over-parametrized, non-convex optimization is still a mystery for mathematical analysis. It is natural to study this problem using Mean-field Langevin equations. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 4 / 37 Two-layer Network and Mean-field Langevin Equation Table of Contents 1 Two-layer Network and Mean-field Langevin Equation 2 Application to GAN 3 Deep neural network and MFL system 4 Game on random environment Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 5 / 37 Note that F is convex in ν. Take Ent(·), the relative entropy w.r.t. Lebesgue measure, as a regularizer, and note that Ent(·) is strictly convex. How to characterize the minimizer of a function of probabilities ? Two-layer Network and Mean-field Langevin Equation Two-layer neural network In the work with K. Hu, D. Siska, L. Szpruch ’19, we focused on the two-layer network, and aimed at minimizing n h X 2i inf E f (Z) − ck '(Ak Z + bk ) ; n;(c ;A ;b ) k k k k=1 where Z represents the data and E is the expectation under the law of the data. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 6 / 37 Note that F is convex in ν. Take Ent(·), the relative entropy w.r.t. Lebesgue measure, as a regularizer, and note that Ent(·) is strictly convex. How to characterize the minimizer of a function of probabilities ? Two-layer Network and Mean-field Langevin Equation Two-layer neural network In the work with K. Hu, D. Siska, L. Szpruch ’19, we focused on the two-layer network, and aimed at minimizing n h 1 X 2i inf E f (Z) − ck '(Ak Z + bk ) ; n;(c ;A ;b ) n k k k k=1 where Z represents the data and E is the expectation under the law of the data. Zhenjie Ren (CEREMADE) MF Langevin 28/08/2020 6 / 37 Note that F is convex in ν. Take Ent(·), the relative entropy w.r.t. Lebesgue measure, as a regularizer, and note that Ent(·) is strictly convex. How to characterize the minimizer of a function of probabilities ? Two-layer Network and Mean-field Langevin Equation Two-layer neural network In the work with K.
Load more

Recommended publications
  • Iterated Importance Sampling in Missing Data Problems
    Iterated importance sampling in missing data problems Gilles Celeux INRIA, FUTURS, Orsay, France Jean-Michel Marin ∗ INRIA, FUTURS, Orsay, France and CEREMADE, University Paris Dauphine, Paris, France Christian P. Robert CEREMADE, University Paris Dauphine and CREST, INSEE, Paris, France Abstract Missing variable models are typical benchmarks for new computational techniques in that the ill-posed nature of missing variable models offer a challenging testing ground for these techniques. This was the case for the EM algorithm and the Gibbs sampler, and this is also true for importance sampling schemes. A population Monte Carlo scheme taking advantage of the latent structure of the problem is proposed. The potential of this approach and its specifics in missing data problems are illustrated in settings of increasing difficulty, in comparison with existing approaches. The improvement brought by a general Rao–Blackwellisation technique is also discussed. Key words: Adaptive algorithms, Bayesian inference, latent variable models, population Monte Carlo, Rao–Blackwellisation, stochastic volatility model ∗ Corresponding author: CEREMADE, Place du Mar´echal De Lattre de Tassigny, 75775 Paris Cedex 16, France, [email protected] Preprint submitted to Computational Statistics and Data Analysis 25 July 2005 1 Introduction 1.1 Missing data models Missing data models, that is, structures such that the distribution of the data y can be represented via a marginal density Z f(y|θ) = g(y, z|θ)dz , Z where z ∈ Z denotes the so-called ”missing data”, have often been at the forefront of computational Statistics, both as a challenge to existing techniques and as a benchmark for incoming techniques. This is for instance the case with the EM algorithm (Dempster et al., 1977), which was purposely designed for missing data problems although it has since then been applied in a much wider setting.
    [Show full text]
  • Testing Adverse Selection and Moral Hazard on French Car Insurance Data
    TESTING ADVERSE SELECTION AND MORAL HAZARD ON FRENCH CAR INSURANCE DATA Guillaume CARLIER1 Université Paris Dauphine Michel GRUN-REHOMME2 Université Paris Panthéon Olga VASYECHKO3 Université Nationale d'Economie de Kyiv This paper is a modest contribution to the stream of research devoted to find empirical evidence of asymmetric information. Building upon Chiappori and Salanié's (2000) work, we propose two specific tests, one for adverse selection and one for moral hazard. We implement these tests on French car insurance data, circumventing the lack of dynamic data in our data base by a proxy of claim history. The first test suggests presence of adverse selection whereas the second one seems to contradict presence of pure moral hazard. Keywords: empirical evidence of moral hazard, adverse selection, car insurance. JEL Classification: C35, D82. 1 Université Paris Dauphine, CEREMADE, Pl. de Lattre de Tassigny, 75775 Paris Cedex 16, FRANCE, [email protected] 2 Université Paris Panthéon, ERMES, 12 Pl. du Panthéon, 75005 Paris, FRANCE, [email protected] 3 Université Nationale d'Economie de Kyiv, UKRAINE, [email protected] BULLETIN FRANÇAIS D’ACTUARIAT, Vol. 13, n° 25, janvier – juin 2013, pp. 117- 130 118 G. CARLIER – M. GRUN-REHOMME – O. VASYECHKO 1. INTRODUCTION There has been an intensive line of research in the last decades devoted to find empirical evidence of asymmetric information on insurance data. Under adverse selection, the insured has some private information about his type of risk, which the insurer cannot observe before the subscription of an insurance contract. The adverse selection assumption stipulates that the high-risks tend to choose more coverage than the low-risks (Rothschild and Stiglitz (1976), Wilson (1977), Spence (1978)).
    [Show full text]
  • Dirk Hundertmark Department of Mathematics, MC-382 University of Illinois at Urbana-Champaign Altgeld Hall 1409 W
    Dirk Hundertmark Department of Mathematics, MC-382 University of Illinois at Urbana-Champaign Altgeld Hall 1409 W. Green Street Urbana, IL 61801 +1 (217) 333-3350 (217) 333-9516 (fax) [email protected] 1401 W. Charles, Champaign, IL 61821 +1 (217) 419-1088 (home) http://www.math.uiuc.edu/∼dirk Personal Data German and American citizen, married, one daughter. Research interests Partial differential equations, analysis, variational methods, functional analysis, spectral theory, motivated by problems from Physics, especially quantum mechanics, and Engineering. Spectral theory of random operators and and its connection to statistical mechanics and some probabilistic problems from solid state physics. More recently, mathematical problems in non-linear fiber-optics, especially properties of dispersion managed solitons. Education 5/2003 Habilitation in Mathematics, Ludwig{Maximilians-Universit¨atM¨unchen. 7/1992 { 11/1996 Ph.D. (Dr. rer. nat., summa cum laude) in Mathematics, Ruhr- Universit¨atBochum, Germany. Thesis: \ On the theory of the magnetic Schr¨odingersemigroup." Advisor: Werner Kirsch 11/1985 { 2/1992 Study of Physics, Friedrich-Alexander-Universit¨atErlangen, Germany. Graduated with Diplom. Advisor: Hajo Leschke. Employment Since 9/2007 Member of the Institute of Condensed Matter Theory at UIUC. Since 8/2006 Associate Professor (tenured), Department of Mathematics, University of Illinois at Urbana-Champaign (on leave 8/2006 { 8/2007). 8/2006 {8/2007 Senior Lecturer, School of Mathematics, University of Birmingham, England. 1/2003 { 7/2006 Assistant Professor, Department of Mathematics, University of Illinois at Urbana-Champaign. 9 { 12/2002 Research fellow at the Institut Mittag-Leffler during the special program \Partial Differential Equations and Spectral Theory." 9/1999{8/2002 Olga Taussky-John Todd Instructor of Mathematics, Caltech.
    [Show full text]
  • Mckean–Vlasov Optimal Control: the Dynamic Programming Principle
    McKean–Vlasov optimal control: the dynamic programming principle∗ Mao Fabrice Djete† Dylan Possamaï ‡ Xiaolu Tan§ March 25, 2020 Abstract We study the McKean–Vlasov optimal control problem with common noise in various formulations, namely the strong and weak formulation, as well as the Markovian and non–Markovian formulations, and allowing for the law of the control process to appear in the state dynamics. By interpreting the controls as probability measures on an appropriate canonical space with two filtrations, we then develop the classical measurable selection, conditioning and concatenation arguments in this new context, and establish the dynamic programming principle under general conditions. 1 Introduction We propose in this paper to study the problem of optimal control of mean–field stochastic differential equations, also called Mckean–Vlasov stochastic differential equations in the literature. This problem is a stochastic control problem where the state process is governed by a stochastic differential equation (SDE for short), which has coefficients depending on the current time, the paths of the state process, but also its distribution (or conditional distribution in the case with common noise). Similarly, the reward functionals are allowed to be impacted by the distribution of the state process. The pioneering work on McKean–Vlasov equations is due to McKean [60] and Kac [49], who were interested in studying uncontrolled SDEs, and in establishing general propagation of chaos results. Let us also mention the illuminating notes of Snitzman [75], which give a precise and pedagogical insight into this specific equation. Though many authors have worked on this equation following these initial papers, there has been a drastic surge of interest in the topic in the past decade, due to the connection that it shares with the so–called mean–field game (MFG for short) theory, introduced independently and simultaneously on the one hand by Lasry and Lions in [53; 54; 55] and on the other hand by Huang, Caines, and Malhamé [44; 45; 46; 47; 48].
    [Show full text]
  • Time Averages for Kinetic Fokker-Planck Equations
    Time averages for kinetic Fokker-Planck equations Giovanni M. Brigatia,b,1 aCEREMADE, CNRS, UMR 7534, Universit´eParis-Dauphine, PSL University, Place du Marechal de Lattre de Tassigny, 75016 Paris, France bDipartimento di Matematica “F. Casorati”, Universit`adegli Studi di Pavia, Via Ferrata 5, 27100 Pavia, Italia Abstract We consider kinetic Fokker-Planck (or Vlasov-Fokker-Planck) equations on the torus with Maxwellian or fat tail local equilibria. Results based on weak norms have recently been achieved by S. Armstrong and J.-C. Mourrat in case of Maxwellian local equilibria. Using adapted Poincar´eand Lions-type inequali- ties, we develop an explicit and constructive method for estimating the decay rate of time averages of norms of the solutions, which covers various regimes corresponding to subexponential, exponential and superexponential (including Maxwellian) local equilibria. As a consequence, we also derive hypocoercivity estimates, which are compared to similar results obtained by other techniques. Keywords: Kinetic Fokker-Planck equation, Ornstein-Uhlenbeck equation, time average, local equilibria, Lions’ lemma, Poincar´einequalities, hypocoercivity. 2020 MSC: Primary: 82C40; Secondary: 35B40, 35H10, 47D06, 35K65. 1. Introduction Let us consider the kinetic Fokker-Planck equation α−2 ∂tf + v ⋅ ∇xf ∇v ⋅ ∇vf + α v vf , f 0, ⋅, ⋅ f0. (1) = ⟨ ⟩ ( ) = where f is a function of time t 0, position x, velocity v, and α is a positive arXiv:2106.12801v1 [math.AP] 24 Jun 2021 parameter. Here we use the notation≥ d v »1 + v 2, ∀ v R . ⟨ ⟩ = S S ∈ d We consider the spatial domain Q ∶ 0,L x, with periodic boundary con- = ( ) ∋ ditions, and define Ωt ∶ t,t + τ × Q, for some τ 0, t 0 and Ω Ω0.
    [Show full text]
  • Curriculum Vitae Béatrice De Tilière
    CURRICULUM VITAE BÉATRICE DE TILIÈRE University Paris Dauphine (CEREMADE) Place du Maréchal De Lattre De Tassigny Birth name: Fässler 75775 PARIS Cedex 16 Born on December 25, 1976 [email protected] Nationalities French and Swiss https://www.ceremade.dauphine.fr/ detiliere/ Married, 2 children ∼ 1. POSITIONS 2019 - Professor. CEREMADE. University Paris Dauphine. 2015 -2018 Professor. LAMA. University Paris-Est Créteil. 2010 -2015 Maître de conférences (Assistant Professor). LPMA. University Pierre et Marie Curie, Paris. 2007 - 2010 Maître-Assistante (Assistant Professor). University of Neuchâtel. 2005 - 2007 Swiss National Fund Post-doctoral position in the probability group of E. Bolthausen. University of Zürich. 2004 - 2005 Research and Teaching Position (ATER). University Paris Descartes. 2. SCIENTIFIC INTERESTS Statistical mechanics, probability theory, phase transitions, critical phenomena, discrete geometry, combinatorics, complex analysis. Dimer model, Ising model, spanning trees. 3. EDUCATION 2013 Habilitation à diriger des recherches en Mathématiques. (Accreditation to supervise research). University Pierre et Marie Curie. Title: “Exactly solvable models of 2-dimensional statistical mechanics: Ising model, dimer model and spanning trees.” Referees: G. Grimmett, C. Sabot, S. Smirnov. Jury: B. Duplantier, T. Duquesne, Y. Le Jan, C. Sabot, W. Werner. 2001 - 2004 Ph.D. in Mathematics. University Paris Sud, Orsay. Grade: “très honorable". Title: “Dimers on isoradial graphs and random interface model in dimension 2+2”. Advisor: R. Kenyon. Reporters: G. Ben Arous, D. Wilson. Referees: R. Kenyon, J.-F. Le Gall, Y. Le Jan, G. Schaeffer, W. Werner. 2000 - 2001 Graduate courses in Mathematics. UC Berkeley. Y. Peres: Probability on trees and networks; D. Aldous: Interacting particle systems; A. Weinstein: Differentiable manifolds; D.
    [Show full text]
  • Master in Mathematics and Applications
    MASTER IN MATHEMATICS AND APPLICATIONS The fruit of the collaboration between the Ecole normale supérieure, EHESS, MINES ParisTech, the Observatoire de Paris and Paris-Dauphine university, the Master’s degree in “Mathematics and applications” from Université PSL provides advanced training in fundamental and applied mathematics. Its ambition is to train specialists in mathematics in the broad sense, preparing for professions both in the academic sector and in those of the service and industrial sectors where modeling takes a predominant place. MAIN ASSETS LEARNING OUTCOMES — Unprecedented collaboration between some of The master’s degree offers in-depth training in the most prestigious schools in France in the field of mathematics during the first year, open to both funda- mathematics. mental mathematics and applied mathematics. — Immersion in an advanced international research eco- The second year offers a specialization that builds upon system, working closely with major laboratories and the theoretical knowledge acquired during the first scientific innovation. year, leading to tracks in applied or theoretical — A very wide range of fundamental teaching and tracks mathematics. combining the complementary expertise of various partners; opportunities for varied specializations deeply rooted in the major contemporary issues of mathema- tics (digital transition, artificial intelligence, modeling CAREER OPPORTUNITIES and analysis of complex systems, data sciences related According to the specialization track chosen in the 2nd to human sciences). year: — A framework of international studies, with part of the — Doctorate in theoretical or applied mathematics teaching given in English by researchers from foreign — Professions in the service sector, particularly related universities. to the evaluation of risk and decision-making (actuarial — A strong link with the industrial and socio-econo- sciences, insurance, finance, statistics and data mic world, a real springboard for a professional career, science).
    [Show full text]
  • Mathieu LEWIN
    Mathieu LEWIN CEREMADE Tel. (+33) 1 44 05 47 96 Universit´eParis-Dauphine [email protected] Place de Lattre de Tassigny http://www.ceremade.dauphine.fr/∼lewin F-75 016 PARIS N´ele 14 Novembre 1977 `aSenlis (France) Positions Oct. 2014 – Directeur de Recherche CNRS CEREMADE, Universit´eParis-Dauphine Sept. 2017– Professeurcharg´edecours Ecole´ Polyechnique, Palaiseau Oct. 2005 – Sept. 2014 Charg´ede Recherche CNRS Laboratoire de Math´ematiques, Universit´ede Cergy-Pontoise Mar. 2005 – Aoˆut 2005 Post-doc INRIA CERMICS, Ecole´ Nationale des Ponts et Chauss´ees, avec Eric´ Canc`es Aoˆut 2004 – F´ev. 2005 EU Post-doc University of Copenhagen (Danemark), avec Jan Philip Solovej Diplˆomes et formation Juin 2009 Habilitation `adiriger des recherches, Universit´ede Cergy-Pontoise Juin2004 Th`esededoctoratavec Eric´ S´er´e, Universit´eParis Dauphine Sept. 1998 – Aoˆut 2002 El`eve´ normalien `al’ENS Cachan Distinctions 2022 Conf´erencier invit´e au congr`es international de math´ematiques (ICM, Saint- Petersburg) 2017 – 2022 Consolidator Grant du European Research Council (ERC) 2015 Plenary speaker at the International Congress of Mathematical Physics 2012 EMSPrize 2010 – 2015 Starting Grant du European Research Council (ERC) Activit´es ´editoriales 2019 – Editeur´ pour Probability and Mathematical Physics 2014 – Editeur´ pour Letters in Mathematical Physics 2013 – Editeur´ pour Mathematical Models and Methods in Applied Sciences (M3AS) 1 Responsabilit´es administratives 2020 – Membre du comit´escientifique du labex CEMPI de Lille 2019 – Membre du comit´e´ethique de la recherche de l’universit´eParis-Dauphine 2015 – 2021 Membre ´elu du conseil d’administration de la SMAI 2015 – 2020 Membre ´elu du conseil ex´ecutif de l’association internationale de physique math´ematique (IAMP) Oct.
    [Show full text]
  • The Obstacle Version of the Geometric Dynamic Programming Principle : Application to the Pricing of American Options Under Constraints
    The obstacle version of the Geometric Dynamic Programming Principle : Application to the pricing of American options under constraints Bruno Bouchard Thanh Nam Vu CEREMADE, Universit´eParis Dauphine CEREMADE, Universit´eParis Dauphine and CREST-ENSAE [email protected] [email protected] First version : November 2008. This version : June 2009 Abstract We provide an obstacle version of the Geometric Dynamic Programming Principle of Soner and Touzi [23] for stochastic target problems. This opens the doors to a wide range of applications, particularly in risk control in finance and insurance, in which a controlled stochastic process has to be maintained in a given set on a time interval [0;T ]. As an example of application, we show how it can be used to provide a viscosity characterization of the super-heging cost of American options under portfolio constraints, without appealing to the standard dual formulation from mathematical finance. In particular, we allow for a degenerate volatility, a case which does not seem to have been studied so far in this context. Key words: Stochastic target problem, discontinuous viscosity solutions, American options. AMS 1991 subject classifications: Primary 49L25, 60J60; secondary 49L20, 35K55. 1 Introduction A stochastic target problem can be described as follows. Given a set of controls A, a family of d- ν ν d dimensional controlled processes Z0;z with initial condition Z0;z(0) = z and a Borel subset G of R , ν ν find the set V (0) of initial conditions z for Z0;z such that Z0;z(T ) 2 G P−a:s: for some ν 2 A.
    [Show full text]
  • Richard H. Day
    RICHARD H. DAY Richard H. Day is Professor of Economics, Emeritus of the University of Southern California. He studied General Science at Iowa State University and, as an Honor Scholar, received the Bachelor of Science in June 1955. He attended Harvard University 1955-1958, received the Ph.D. degree in Economics in June 1961. His dissertation, published by North-Holland in 1963,, presented a class of "recursive programming" models for explaining production, investment and technological change in agriculture and industrial sectors. This work, originally sponsored by the U.S. Department of Agriculture, was the prototype for a U.S. National Model in the 1960s and for a variety of similar studies in a number of countries. His work at this time also included the derivation of sufficient conditions for aggregation in production models, estimates of the distribution of field crop yields, and an analysis of the effect of technological change on the migration from rural to urban centers. During the period 1959–1962, Day served in the U.S.A.F. as a mathematician and systems analyst and subsequently spent a brief stint as a consultant to Richard Reuter who was then Director of the Food for Peace Program and Special Assistant to President Kennedy. After joining the University of Wisconsin in 1963, with the support of the National Science Foundation, conducted recursive programming studies of the U.S. Coal, Steel and Petroleum Industries and of agricultural development in India. This work showed how structural change can be explained in economic terms and how it can completely transform a region or industry in less than a generation.
    [Show full text]
  • Optimal Insurance for a Minimal Expected Retention: the Case of an Ambiguity-Seeking Insurer
    risks Article Optimal Insurance for a Minimal Expected Retention: The Case of an Ambiguity-Seeking Insurer Massimiliano Amarante 1 and Mario Ghossoub 2,* 1 Département de Sciences Économiques, Université de Montréal, C.P. 6128, succursale Centre-ville, Montréal, QC H3C 3J7, Canada; [email protected] 2 Imperial College London, South Kensington Campus, London SW7 2AZ, UK * Correspondence: [email protected]; Tel.: +44-020-7594-9130 Academic Editor: Alexander Szimayer Received: 22 December 2015; Accepted: 9 March 2016; Published: 21 March 2016 Abstract: In the classical expected utility framework, a problem of optimal insurance design with a premium constraint is equivalent to a problem of optimal insurance design with a minimum expected retention constraint. When the insurer has ambiguous beliefs represented by a non-additive probability measure, as in Schmeidler, this equivalence no longer holds. Recently, Amarante, Ghossoub and Phelps examined the problem of optimal insurance design with a premium constraint when the insurer has ambiguous beliefs. In particular, they showed that when the insurer is ambiguity-seeking, with a concave distortion of the insured’s probability measure, then the optimal indemnity schedule is a state-contingent deductible schedule, in which the deductible depends on the state of the world only through the insurer’s distortion function. In this paper, we examine the problem of optimal insurance design with a minimum expected retention constraint, in the case where the insurer is ambiguity-seeking. We obtain the aforementioned result of Amarante, Ghossoub and Phelps and the classical result of Arrow as special cases. Keywords: optimal insurance; deductible; minimum retention; ambiguity; Choquet integral; probability distortion JEL: G22 1.
    [Show full text]
  • Symbolic Data Analysis: Definitions and Examples
    SYMBOLIC DATA ANALYSIS: DEFINITIONS AND EXAMPLES L. Billard E. Diday Department of Statistics CEREMADE University of Georgia Universite de Paris 9 Dauphine Athens, GA 30602 -1952 USA 75775 Paris Cedex 16 France Abstract With the advent of computers, large, very large datasets have become routine. What is not so routine is how to analyse these data and/or how to glean useful information from within their massive confines. One approach is to summarize large data sets in such a way that the resulting summary dataset is of a manageable size. One consequence of this is that the data may no longer be formatted as single values such as is the case for classical data, but may be represented by lists, intervals, distributions and the like. These summarized data are examples of symbolic data. This paper looks at the concept of symbolic data in general, and then attempts to review the methods currently available to analyse such data. It quickly becomes clear that the range of methodologies available draws analogies with developments prior to 1900 which formed a foundation for the inferential statistics of the 1900’s, methods that are largely limited to small (by comparison) data sets and limited to classical data formats. The scarcity of available methodologies for symbolic data also becomes clear and so draws attention to an enormous need for the development of a vast catalogue (so to speak) of new symbolic methodologies along with rigorous mathematical foundational work for these methods. 1 Introduction With the advent of computers, large, very large datasets have become routine. What is not so routine is how to analyse the attendant data and/or how to glean useful information from within their massive confines.
    [Show full text]