Large Deviations Results for Non-Gradient Systems Christian Cofoid Boston College cofoid@bc.edu September 16, 2015 Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 1/9 Overview 1 Introduction 2 The Problem 3 Summer Work 4 Future Work Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 2/9 Introduction Introduction 2 Consider the one dimensonal potential well given by V (x) = x 2 − 1 , and the ODE dV ẋ = − . dx This ODE has two asymptotically stable fixed points at x = ±1 and a saddle point at x = 0. Add dampened ”white noise” to the ODE, giving the one-dimensional SDE: dx(t) = −∇V (x)dt + εdW (t), where ε 1, and W (t) is a canonical Weiner process. More generally, we will consider the n-dimensional SDE dX (t) = −∇V (X )dt + εdW (t). Christian Cofoid (Boston College) Brown REU 2015 (1) September 16, 2015 3/9 Introduction Plot and Contour for n = 2. 2 Two Dimensional Gradient System with Noise: epsilon = 0.5 1.5 1 y 0.5 0 -0.5 -1 -1.5 -2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 x Figure: V (x, y ) = x 4 − 4xy + 2y 2 Christian Cofoid (Boston College) Figure: 50 paths following dx(t) = −∇V (x)dt + εdW (t), ε = 0.5. Brown REU 2015 September 16, 2015 4/9 Introduction Current Knowledge We can compute the expected time of escape τΩ from a domain of attraction Ω centered about a stable fixed point. Let x ∗ ∈ Rn be the stable fixed point and z ∗ ∈ Rn be the saddle point. Then, the expected time of escape is approximated by the Eyring-Kramers law [Berg11]; s | det (∇2 V (z ∗ )) | 2 (V (z ∗ ) − V (x ∗ )) 2π exp , E [τΩ ] ' |λ1 (z ∗ ) | det (∇2 V (x ∗ )) ε2 where λ1 (z ∗ ) is the negative eigenvalue of the Hessian ∇2 V (z ∗ ). Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 5/9 Introduction Current Knowledge (cont’d) A large deviation principle is a relation stating that for sufficiently small ε, the probability of the sample paths xt being close to a function ϕ(t) is given by P ((xt ' ϕ(t), 0 ≤ t ≤ T ) ' e −I (ϕ)/2ε , where I (ϕ) is called the rate functional. It is known that 1 I (ϕ) = 2 Z T ||ϕ̇(t) − f (ϕ(t))|| dt. (2) 0 With some theoretical legwork, a large deviations principle tells us that if a particle is going to escape Ω, it will do so in a ”straight shot” along the heteroclinic orbit. Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 6/9 The Problem The Problem If we add a small perturbation µg (X ), µ 1, and g is non-gradient to equation (1), we obtain a non-gradient system dX (t) = −∇V (X )dt + µg (X )dt + εdW (t). (3) The question of interest is; what is the expected time and expected path of escape from Ω centered about an asymptotically stable fixed point p in a perturbed system such as this? Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 7/9 Summer Work Work Done This Summer 1 We have found a technique that seems promising, called importance sampling. Basically, we change the distribution from which we sample N (0, 1) from one normal distribution to one with a larger mean, increasing the probability of realizing large amounts of noise, and thus the time needed to escape Ω. We then correct for this bias to find the unbiased expected time of escape. Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 8/9 Future Work Future Work 1 We would like to conclude that our algorithm confirms the theoretical results for gradient systems. 2 We would then like to apply this to non-gradient systems, and conduct importance sampling on the expected path of escape. Christian Cofoid (Boston College) Brown REU 2015 September 16, 2015 9/9