Optimal Diffusion Processes

Saber Jafarizadeh · 2024

Read Optimal Diffusion Processes by Saber Jafarizadeh (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-20934.txt.

Description

Abstract Of stochastic differential equations, diffusion processes have been adopted in numerous applications, as more relevant and flexible models. This paper studies diffusion processes in a different setting, where for a given stationary distribution and average variance, it seeks the diffusion process with optimal convergence rate.

Catalog record

CREATORSaber Jafarizadeh
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-20934

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Optimal Diffusion Processes Saber Jafarizadeh Abstract Of stochastic differential equations, diffusion processes have been adopted in numerous applications, as more relevant and flexible models. This paper studies diffusion processes in a different setting, where for a given stationary distribution and average variance, it seeks the diffusion process with optimal convergence rate. It is shown that the optimal drift function is a linear function and the convergence rate of the stochastic process is bounded by the ratio of the average variance to the variance of the stationary distribution. Furthermore, the concavity of the optimal relaxation time as a function of the stationary distribution has been proven, and it is shown that all Pearson diffusion processes of the Hypergeometric type with polynomial functions of at most degree two as the variance functions are optimal. Keywords: Stochastic differential equation, Diffusion process, Fokker-Planck equation, Optimization 1 Introduction Due to random nature of any real signal, propagation media, interference and even information itself, the stochastic approach has been one of the popular choices towards problems associated with information transmission, as well as other applications in physics, mathematical biology, queueing, electrical engineering, economic and finance [ 18 , 1 , 14 , 16 , 20 , 7 , 8 ] . A particularly effective application of the stochastic approach is the modeling of the phenomena as random processes which are the solution of an appropriate Stochastic Differential Equation (SDE), i.e., an ordinary differential equation with random excitation, due to the presence of noise in the dynamics of the process [ 10 , 3 ] . The basic idea behind this approach is that the state of the studied process is the solution of an SDE with a priori information, such as the stationary distribution. SDEs are generally identified by two main parameters, the drift and the variance functions. Based on dependency of the drift and variance functions on the previous values of the process as well as the present values, the SDEs are categorized into Itô and diffusion processes. In diffusion processes (which are the focus of studies in this paper), the drift and variance functions depend only on the present value of the process and their solution satisfy the Markov property [ 4 ] , which results in a unique stationary distribution for the diffusion process. Diffusion processes with invariant equilibrium distributions have found applications in approximating the process of sampling from high dimensional probability distributions [ 21 ] . One of the important factors in such applications is the convergence rate of the diffusion process and its optimizations. [ 9 ] is regarded as one of the pioneering works that has studied the optimization of convergence rate of the diffusion processes. In [ 9 , 21 ] (and the following works of [ 9 ] ), optimization of the convergence rate of the non-reversible diffusion processes have been addressed, but at the expense of limiting the variance function to be a constant function. This is unlike the studied presented in this paper, where no such restrictions have been applied on the variance function. In a nutshell, in this paper, for different types of diffusion processes with given average variance function and stationary distribution, we have provided the diffusion process with the optimal convergence rate, which has been possible by adjust the variance function. In detail, for a given distribution \pi(x) , our goal is to construct the diffusion process with optimal convergence rate whose stationary distribution is \pi(x) . We have shown that the optimal drift function is a linear function and the optimal convergence rate of the diffusion process is equal to the ratio of the average variance to the variance of the stationary distribution. Therefore, for all stochastic processes with given stationary distribution \pi(x) and average variance of \widehat{\sigma}^{2}/2 , the convergence rate of the stochastic process is bounded by \frac{\widehat{\sigma}^{2}/2}{m_{2}-m_{1}^{2}} , where m_{2}-m_{1}^{2} is the variance of the stationary distribution \pi(x) . Furthermore, we have proved the concavity of the optimal relaxation time (\tau) as a function of the stationary distribution (\pi(x)) . We have shown that all Pearson diffusion processes of the Hypergeometric type with polynomial functions of at most degree two as the variance functions are optimal. In a table, we have provided the optimal convergence rate along with other attributes for the Pearson diffusion processes. Pearson diffusion processes are diffusion processes whose drift and variance functions are polynomials of degrees at most one and two, respectively. Also, we have provided the optimal results for two other types of diffusion processes, namely (i) diffusion processes with polynomial functions of degree higher than two as the variance function, and (ii) diffusion processes with hyperexponential distribution as the stationary distribution, which is an example of a diffusion process with a non-polynomial function as its variance function. The rest of the paper is organized as follows. In Section 2 , Preliminaries on Fokker-Plank equations are presented. Our general approach towards optimization of the diffusion processes has been provided in Section 3 . In Section 4 , main results of the paper have been presented, and in Section 5 , the optimal results for three different types of diffusion processes have been reported. and Section 6 concludes the paper. 2 Preliminaries 2.1 Fokker-Plank Equation Let’s consider the following time-homogeneous diffusion process (also known as Itô diffusion) X(t) , \begin{gathered}\mathrm{d}X(t)=\mu(X(t))\,\mathrm{d}t+\sigma(X(t))\,\mathrm{d}W(t),\end{gathered} (1) where W(t) is a Brownian motion in \mathbb{R} , The function \mu is the drift coefficient of X and \sigma is known as the diffusion coefficient of X and it can be interpreted as a variance. Imposing any control term to either the drift coefficient or the diffusion coefficient will transform the diffusion process ( 1 ) into a controlled diffusion process [ 12 , 11 ] . The domain of X(t) is denoted by I=(x_{1},x_{2}) where -\infty\leq x_{1}<x_{2}\leq\infty . The process X(t) is Markov, and its transition density p(t;x,y) is defined as \begin{gathered}P\left(X(t)\in A|X_{0}=x\right)=\int_{A}p(t;x,y)dy,\end{gathered} (2) for t\geq 0 and any Borel set A\subseteq I . Under regularity conditions, the transition density solves the backward Kolmogorov equation, \begin{gathered}\frac{\partial p(t;x,y)}{\partial t}=\mathcal{A}\left[\mu,\sigma^{2}\right]p(t;x,y),\quad t>0,\quad(x,y)\in I,\end{gathered} (3) where \mathcal{A} is an infinitesimal generator defined as below, \begin{gathered}\mathcal{A}\left[\mu,\sigma^{2}\right]p(t;x,y)=\mu(x)\frac{\partial p(t;x,y)}{\partial x}+\frac{1}{2}\sigma^{2}(x)\frac{\partial^{2}p(t;x,y)}{\partial x^{2}}.\end{gathered} (4) p(t;x,y) can be completely characterised by the generator \mathcal{A}\left[\mu,\sigma^{2}\right] . Therefore, it is a functional of \mu and \sigma^{2} and it can be written as p(t;x,y)=p\left(t;x,y;\mu,\sigma\right) . Suppose that \{X(t)\} is strictly stationary and ergodic, in which case it has a stationary marginal density which we denote \pi(x) , P\left(X(t)\in A\right)=\int_{A}\pi(x)dx , for t\geq 0 and any Borel set A\subseteq I . Note that \pi is invariant, i.e. \begin{gathered}\pi(y)=\int_{I}p(t;x,y)\pi(x)dx,\quad\text{for}\quad t\geq 0.\end{gathered} (5) Alternatively, the characterization of the transition density can be obtained through the forward Kolmogorov equation, \begin{gathered}\frac{\partial p(t;x,y)}{\partial t}=-\frac{\partial\left(\mu(y)p(t;x,y)\right)}{\partial y}+\frac{1}{2}\frac{\partial^{2}\left(\sigma^{2}(y)p(t;x,y)\right)}{\partial y^{2}},\end{gathered} (6) for t>0 , x\in I with initial value p(0;x,y)=\delta(y-x) and the following boundary conditions, which are corresponding to reflecting barriers, \begin{gathered}\left.\frac{1}{2}\frac{\partial}{\partial y}\left(\sigma^{2}(y)p(t;x,y)\right)-\mu(y)p(t;x,y)\right|_{y=y_{1},y_{2}}=0.\end{gathered} (7) Multiplying ( 6 ) with \pi(x) and integrating over x and using ( 5 ) the following is obtained, \begin{gathered}\frac{d}{dy}\left(-\mu(y)\pi(y)+\frac{1}{2}\frac{d}{dy}\left(\sigma^{2}(y)\pi(y)\right)\right)=0.\end{gathered} (8) If the detailed balance condition is satisfied, i.e. \begin{gathered}-\mu(y)\pi(y)+\frac{1}{2}\frac{d}{dy}\left(\sigma^{2}(y)\pi(y)\right)=0,\end{gathered} (9) then the process X(t) is reversible wrt \pi(x) . Using the detailed balance condition ( 9 ), we can conclude \begin{gathered}\mu(x)=\frac{1}{2\pi(x)}\frac{\partial}{\partial x}\left(\sigma^{2}(x)\pi(x)\right).\end{gathered} (10) All the control effort will be

Read the complete text