Past Stochastic Analysis & Mathematical Finance Seminars

We develop a framework for showing that neural networks can overcome the curse of dimensionality in different high-dimensional approximation problems. Our approach is based on the notion of a catalog network, which is a generalization of a standard neural network in which the nonlinear activation functions can vary from layer to layer as long as they are chosen from a predefined catalog of functions. As such, catalog networks constitute a rich family of continuous functions. We show that under appropriate conditions on the catalog, catalog networks can efficiently be approximated with ReLU-type networks and provide precise estimates on the number of parameters needed for a given approximation accuracy. As special cases of the general results, we obtain different classes of functions that can be approximated with ReLU networks without the curse of dimensionality. 

A preprint is here: https://arxiv.org/abs/1912.04310

In this talk I will consider model-independent pricing problems in a stochastic interest rates framework. In this case the usual tools from Optimal Transport and Skorokhod embedding cannot be applied. I will show how some pricing problems in a fixed-income market can be reformulated as Weak Optimal Transport (WOT) problems as introduced by Gozlan et al. I will present a super-replication theorem that follows from an extension of WOT results to the case of non-convex cost functions.
This talk is based on joint work with M. Beiglboeck and G. Pammer.

The risk and return profiles of a broad class of dynamic trading strategies, including pairs trading and other statistical arbitrage strategies, may be characterized in terms of excursions of the market price of a portfolio away from a reference level. We propose a mathematical framework for the risk analysis of such strategies, based on a description in terms of price excursions, first in a pathwise setting, without probabilistic assumptions, then in a Markovian setting.

We introduce the notion of δ-excursion, defined as a path which deviates by δ from a reference level before returning to this level. We show that every continuous path has a unique decomposition into δ-excursions, which is useful for scenario analysis of dynamic trading strategies, leading to simple expressions for the number of trades, realized profit, maximum loss and drawdown. As δ is decreased to zero, properties of this decomposition relate to the local time of the path. When the underlying asset follows a Markov process, we combine these results with Ito's excursion theory to obtain a tractable decomposition of the process as a concatenation of independent δ-excursions, whose distribution is described in terms of Ito's excursion measure. We provide analytical results for linear diffusions and give new examples of stochastic processes for flexible and tractable modeling of excursions. Finally, we describe a non-parametric scenario simulation method for generating paths whose excursion properties match those observed in empirical data.

Joint work with Anna Ananova and Rama Cont: https://ssrn.com/abstract=3723980

Stochastic quantisation is, broadly speaking, the use of a stochastic differential equation to construct a given probability distribution. Usually this refers to Markovian Langevin evolution with given invariant measure. However we will show that it is possible to construct other kind of equations (elliptic stochastic partial differential equations) whose solutions have prescribed marginals. This connection was discovered in the '80 by Parisi and Sourlas in the context of dimensional reduction of statistical field theories in random external fields. This purely probabilistic results has a proof which depends on a supersymmetric formulation of the problem, i.e. a formulation involving a non-commutative random field defined on a non-commutative space. This talk is based on joint work with S. Albeverio and F. C. de Vecchi.

It is one of the most challenging issues in applied mathematics to approximately solve high-dimensional partial differential equations (PDEs) and most of the numerical approximation methods for PDEs in the scientific literature suffer from the so-called curse of dimensionality (CoD) in the sense that the number of computational operations employed in the corresponding approximation scheme to obtain an  approximation precision $\varepsilon >0$ grows exponentially in the PDE dimension and/or the reciprocal of $\varepsilon$. Recently, certain deep learning based approximation methods for PDEs have been proposed  and various numerical simulations for such methods suggest that deep neural network (DNN) approximations might have the capacity to indeed overcome the CoD in the sense that  the number of real parameters used to describe the approximating DNNs  grows at most polynomially in both the PDE dimension $d \in  \N$ and the reciprocal of the prescribed approximation accuracy $\varepsilon >0$. There are now also a few rigorous mathematical results in the scientific literature which  substantiate this conjecture by proving that  DNNs overcome the CoD in approximating solutions of PDEs.  Each of these results establishes that DNNs overcome the CoD in approximating suitable PDE solutions  at a fixed time point $T >0$ and on a compact cube $[a, b]^d$ but none of these results provides an answer to the question whether the entire PDE solution on $[0, T] \times [a, b]^d$ can be approximated by DNNs without the CoD. 
In this talk we show that for every $a \in \R$, $ b \in (a, \infty)$ solutions of  suitable  Kolmogorov PDEs can be approximated by DNNs on the space-time region $[0, T] \times [a, b]^d$ without the CoD. 

The Ricci flow on a surface is an intrinsic evolution of the metric converging to a constant curvature metric within the conformal class. It can be seen as an infinite-dimensional gradient flow. We introduce a natural 'Langevin' version of that flow, thus constructing an SPDE with invariant measure expressed in terms of Liouville Conformal Field Theory.
Joint work with Hao Shen (Wisconsin).

Trading of financial instruments has largely moved away from floor trading and onto electronic exchanges. Orders to buy and sell are queued at these exchanges in a limit-order book. While a full analysis of the dynamics of a limit-order book requires an understanding of strategic play among multiple agents, and is thus extremely complex, so-called zero-intelligence Poisson models have been shown to capture many of the statistical features of limit-order book evolution. These models can be addressed by traditional queueing theory techniques, including Laplace transform analysis. In this work, we demonstrate in a simple setting that another queueing theory technique, approximating the Poisson model by a diffusion model identified as the limit of a sequence of scaled Poisson models, can also be implemented. We identify the diffusion limit, find an embedded semi-Markov model in the limit, and determine the statistics of the embedded semi-Markov model. Along the way, we introduce and study a new type of process, a generalization of skew Brownian motion that we call two-speed Brownian motion.

Abstract: A recent paradigm views deep neural networks as discretizations of certain controlled ordinary differential equations, sometimes called neural ordinary differential equations. We make use of this perspective to link expressiveness of deep networks to the notion of controllability of dynamical systems. Using this connection, we study an expressiveness property that we call universal interpolation, and show that it is generic in a certain sense. The universal interpolation property is slightly weaker than universal approximation, and disentangles supervised learning on finite training sets from generalization properties. We also show that universal interpolation holds for certain deep neural networks even if large numbers of parameters are left untrained, and are instead chosen randomly. This lends theoretical support to the observation that training with random initialization can be successful even when most parameters are largely unchanged through the training. Our results also explore what a minimal amount of trainable parameters in neural ordinary differential equations could be without giving up on expressiveness.

Joint work with Martin Larsson, Josef Teichmann.

We revisit the variational characterization of conservative diffusion as entropic gradient flow and provide for it a probabilistic interpretation based on stochastic calculus. It was shown by Jordan, Kinderlehrer, and Otto that, for diffusions of Langevin–Smoluchowski type, the Fokker–Planck probability density flow maximizes the rate of relative entropy dissipation, as measured by the distance traveled in the ambient space of probability measures with finite second moments, in terms of the quadratic Wasserstein metric. We obtain novel, stochastic-process versions of these features, valid along almost every trajectory of the diffusive motion in the backward direction of time, using a very direct perturbation analysis. By averaging our trajectorial results with respect to the underlying measure on path space, we establish the maximal rate of entropy dissipation along the Fokker–Planck flow and measure exactly the deviation from this maximum that corresponds to any given perturbation. As a bonus of our trajectorial approach we derive the HWI inequality relating relative entropy (H), Wasserstein distance (W) and relative Fisher information (I).

Most of the basic results on martingale problems extend to the setting in which the generator depends on a control.  The “control” could represent a random environment, or the generator could specify a classical stochastic control problem.  The equivalence between the martingale problem and forward equation (obtained by taking expectations of the martingales) provides the tools for extending linear programming methods introduced by Manne in the context of controlled finite Markov chains to general Markov stochastic control problems.  The controlled martingale problem can also be applied to the study of constrained Markov processes (e.g., reflecting diffusions), the boundary process being treated as a control.  The talk includes joint work with Richard Stockbridge and with Cristina Costantini. 

Abstract: The calibration problem for local stochastic volatility models leads to two-dimensional stochastic differential equations of McKean-Vlasov type. In these equations, the conditional distribution of the second component of the solution given the first enters the equation for the first component of the solution. While such equations enjoy frequent application in the financial industry, their mathematical analysis poses a major challenge. I will explain how to prove the strong existence of stationary solutions for these equations, as well as the strong uniqueness in an important special case. Based on joint work with Daniel Lacker and Jiacheng Zhang.  
 

We study dynamic backward problems, with the computation of conditional expectations as a special objective, in a framework where the (forward) state process satisfies a Volterra type SDE, with fractional Brownian motion as a typical example. Such processes are neither Markov processes nor semimartingales, and most notably, they feature a certain time inconsistency which makes any direct application of Markovian ideas, such as flow properties, impossible without passing to a path-dependent framework. Our main result is a functional Itô formula, extending the Functional Ito calculus to our more general framework. In particular, unlike in the Functional Ito calculus, where one needs only to consider stopped paths, here we need to concatenate the observed path up to the current time with a certain smooth observable curve derived from the distribution of the future paths.  We then derive the path dependent PDEs for the backward problems. Finally, an application to option pricing and hedging in a financial market with rough volatility is presented.

Joint work with JianFeng Zhang (USC).

Abstract: 

In this talk I will discuss regularization by noise from a pathwise perspective using non-linear Young integration, and discuss the relations with occupation measures and local times. This methodology of pathwise regularization by noise was originally proposed by Gubinelli and Catellier (2016), who use the concept of averaging operators and non-linear Young integration to give meaning to certain ill posed SDEs. 

In a recent work together with   Nicolas Perkowski we show that there exists a class of paths with exceptional regularizing effects on ODEs, using the framework of Gubinelli and Catellier. In particular we prove existence and uniqueness of ODEs perturbed by such a path, even when the drift is given as a Scwartz distribution. Moreover, the flow associated to such ODEs are proven to be infinitely differentiable. Our analysis can be seen as purely pathwise, and is only depending on the existence of a sufficiently regular occupation measure associated to the path added to the ODE. 

As an example, we show that a certain type of Gaussian processes has infinitely differentiable local times, whose paths then can be used to obtain the infinitely regularizing effect on ODEs. This gives insight into the powerful effect that noise may have on certain equations. I will also discuss an ongoing extension of these results towards regularization of certain PDE/SPDEs by noise.​

Let ?={??}?∈ℤ be zero-mean stationary Gaussian sequence of random variables with covariance function ρ satisfying ρ(0)=1. Let φ:R→R be a function such that ?[?(?_0)2]<∞ and assume that φ has Hermite rank d≥1. The celebrated Breuer–Major theorem asserts that, if ∑|?(?)|^?<∞ then
the finite dimensional distributions of the normalized sum of ?(??) converge to those of ?? where W is
a standard Brownian motion and σ is some (explicit) constant. Surprisingly, and despite the fact this theorem has become over the years a prominent tool in a bunch of different areas, a necessary and sufficient condition implying the weak convergence in the
space ?([0,1]) of càdlàg functions endowed with the Skorohod topology is still missing. Our main goal in this paper is to fill this gap. More precisely, by using suitable boundedness properties satisfied by the generator of the Ornstein–Uhlenbeck semigroup,
we show that tightness holds under the sufficient (and almost necessary) natural condition that E[|φ(X0)|p]<∞ for some p>2.

Joint work with D Nualart
 

Many classical fractal functions, such as the Weierstrass and Takagi-van der Waerden functions, admit a finite p-th variation along a natural sequence of partitions. They can thus serve as integrators in pathwise Itô calculus. Motivated by this observation, we
introduce a new class of stochastic processes, which we call Weierstrass bridges. They have continuous sample paths and arbitrarily low regularity and so provide a new example class of “rough” stochastic processes. We study some of their sample path properties
including p-th variation and moduli of continuity. This talk includes joint work with Xiyue Han and Zhenyuan Zhang.

Generative Adversarial Networks (GANs) have celebrated great empirical success, especially in image generation and processing. Meanwhile, Mean-Field Games (MFGs),  as analytically feasible approximations for N-player games, have experienced rapid growth in theory of controls. In this talk, we will discuss a new theoretical connections between GANs and MFGs. Interpreting MFGs as GANs, on one hand, allows us to devise GANs-based algorithm to solve MFGs. Interpreting GANs as MFGs, on the other hand, provides a new and probabilistic foundation for GANs. Moreover, this interpretation helps establish an analytical connection between GANs and Optimal Transport (OT) problems, the connection previously understood mostly from the geometric perspective. We will illustrate by numerical examples of using GANs to solve high dimensional MFGs, demonstrating its superior performance over existing methodology.

For the problem of prediction with expert advice in the adversarial setting with geometric stopping, we compute the exact leading order expansion for the long time behavior of the value function using techniques from stochastic analysis and PDEs. Then, we use this expansion to prove that as conjectured in Gravin, Peres and Sivan the comb strategies are indeed asymptotically optimal for the adversary in the case of 4 experts.
 

We provide a detailed, probabilistic interpretation, based on stochastic calculus, for the variational characterization of conservative diffusion as entropic gradient flux. Jordan, Kinderlehrer, and Otto showed in 1998 that, for diffusions of Langevin-Smoluchowski type, the Fokker-Planck probability density flow minimizes the rate of relative entropy dissipation, as measured by the distance traveled in terms of the quadratic Wasserstein metric in the ambient space of configurations. Using a very direct perturbation analysis we obtain novel, stochastic-process versions of such features. These are valid along almost every trajectory of the diffusive motion in both the forward and, most transparently, the backward, directions of time. The original results follow then simply by taking expectations. As a bonus, we obtain the HWI inequality of Otto and Villani relating relative entropy, Fisher information and Wasserstein distance; and from it the celebrated log-Sobolev, Talagrand and Poincare inequalities of functional analysis. (Joint work with W. Schachermayer and B. Tschiderer, from the University of Vienna.)

 

In practice, it is standard to initialize Artificial Neural Networks (ANN) with random parameters. We will see that this allows to describe, in the functional space, the limit of the evolution of (fully connected) ANN when their width tends towards infinity. Within this limit, an ANN is initially a Gaussian process and follows, during learning, a gradient descent convoluted by a kernel called the Neural Tangent Kernel. 

This description allows a better understanding of the convergence properties of neural networks, of how they generalize to examples during learning and has 

practical implications on the training of wide ANNs. 

During this talk we will be interested in a one-dimensional exclusion process subject to strong kinetic constraints, which belongs to the class of cooperative kinetically constrained lattice gases. More precisely, its stochastic short range interaction exhibits a continuous phase transition to an absorbing state at a critical value of the particle density. We will see that the macroscopic behavior of this microscopic dynamics, under periodic boundary conditions and diffusive time scaling, is ruled by a non-linear PDE belonging to free boundary problems (or Stefan problems). One of the ingredients is to show that the system typically reaches an ergodic component in subdiffusive time.

Based on joint works with O. Blondel, C. Erignoux and M. Sasada

The deep neural network has achieved impressive results in various applications, and is involved in more and more branches of science. However, there are still few theories supporting its empirical success. In particular, we miss the mathematical tool to explain the advantage of certain structures of the network, and to have quantitive error bounds. In our recent work, we used a regularised relaxed control problem to model the deep neural network.  We managed to characterise its optimal control by the invariant measure of a mean-field Langevin system, which can be approximated by the marginal laws. Through this study we understand the importance of the pooling for the deep nets, and are capable of computing an exponential convergence rate for the (stochastic) gradient descent algorithm.

Consider a collection of particles whose state evolution is described through a system of interacting diffusions in which each particle
is driven by an independent individual source of noise and also by a small amount of noise that is common to all particles. The interaction between the particles is due to the common noise and also through the drift and diffusion coefficients that depend on the state empirical measure. We study large deviation behavior of the empirical measure process which is governed by two types of scaling, one corresponding to mean field asymptotics and the other to the Freidlin-Wentzell small noise asymptotics. 
Different levels of intensity of the small common noise lead to different types of large deviation behavior, and we provide a precise characterization of the various regimes. We also study large deviation behavior of  interacting particle systems approximating various types of Feynman-Kac functionals. Proofs are based on stochastic control representations for exponential functionals of Brownian motions and on uniqueness results for weak solutions of stochastic differential equations associated with controlled nonlinear Markov processes. 

We will talk about some stochastic parabolic and hyperbolic partial differential equations (SPDEs), which arise naturally in the context of Liouville quantum gravity. These dynamics are proposed to preserve the Liouville measure, which has been constructed recently in the series of works by David-Kupiainen-Rhodes-Vargas. We construct global solutions to these equations under some conditions and then show the invariance of the Liouville measure under the resulting dynamics. As a by-product, we also answer an open problem proposed by Sun-Tzvetkov recently.
 

Stochastic processes subject to weak noise often show a metastable
behaviour, meaning that they converge to equilibrium extremely slowly;
typically, the convergence time is exponentially large in the inverse
of the variance of the noise (Arrhenius law).
  
In the case of finite-dimensional Ito stochastic differential
equations, the large-deviation theory developed in the 1970s by
Freidlin and Wentzell allows to prove such Arrhenius laws and compute
their exponent. Sharper asymptotics for relaxation times, including the
prefactor of the exponential term (Eyring–Kramers laws) are known, for
instance, if the stochastic differential equation involves a gradient
drift term and homogeneous noise. One approach that has been very
successful in proving Eyring–Kramers laws, developed by Bovier,
Eckhoff, Gayrard and Klein around 2005, relies on potential theory.
  
I will describe Eyring–Kramers laws for some parabolic stochastic PDEs
such as the Allen–Cahn equation on the torus. In dimension 1, an
Arrhenius law was obtained in the 1980s by Faris and Jona-Lasinio,
using a large-deviation principle. The potential-theoretic approach
allows us to compute the prefactor, which turns out to involve a
Fredholm determinant. In dimensions 2 and 3, the equation needs to be
renormalized, which turns the Fredholm determinant into a
Carleman–Fredholm determinant.
  
Based on joint work with Barbara Gentz (Bielefeld), and with Ajay
Chandra (Imperial College), Giacomo Di Gesù (Vienna) and Hendrik Weber
(Warwick). 

References: 
https://dx.doi.org/10.1214/EJP.v18-1802
https://dx.doi.org/10.1214/17-EJP60

The optimal matching problem is about the rate of convergence
in Wasserstein distance of the empirical measure of iid uniform points
to the Lebesgue measure. We will start by reviewing the macroscopic
behaviour of the matching problem and will then report on recent results
on the mesoscopic behaviour in the thermodynamic regime. These results
rely on a quantitative large-scale linearization of the Monge-Ampere
equation through the Poisson equation. This is based on joint work with
Michael Goldman and Felix Otto.