Inverse Design For Hamilton-Jacobi Equations

By Carlos Esteve, Enrique Zuazua

In many evolution models, the reconstruction of the initial state given an observation of the system at time T>0 represents a major challenge in mathematical modelling.

Especially if it involves irreversible processes, where sometimes, different initial conditions can lead the system to the same state at time T>0. In these cases, the goal would be to describe all the possible initial states that agree with the given observation.

Here we consider the following initial-value problem for a time-dependent Hamilton-Jacobi equation:

\left\{ \begin{array}{ll} \partial_t u + H(\nabla_x u) = 0, & \text{in} \ [0,T]\times \mathbb{R}^n, \\ u(0,x) = u_0(x), & \text{in} \ \mathbb{R}^n, \end{array}\right.

where u_0\in C^{0,1}(\mathbb{R}^n) and H: \mathbb{R}^n\rightarrow\mathbb{R} is a superlinear convex Hamiltonian. From now on, we will denote by S_T^+ u_0 the solution to (1) at time t=T.

In this case, due to the presence of the nonlinear term H(\nabla_x u) in the equation, one cannot in general expect the existence of a regular C^1 solution, even if the initial condition u_0 is assumed to be very smooth. As we shall see, this loss of regularity results in a loss of information of the initial condition and the impossibility of uniquely determining u_0 from the observed solution at time T.

Our goal is, for a given observation u_T\in C^{0,1}(\mathbb{R}^n), determine all the initial conditions u_0 for which the solution u to (1) satisfies u(T,\cdot)= u_T, or equivalently, S_T^+u_0 =u_T.

 

We split the problem in three steps:

Step 1: We first determine if the observation u_T is admissible, i.e. if there exists at least one u_0 satisfying S_T^+ u_0=u_T.

The natural candidate is the one obtained by reversing the time in the equation (1), considering u_T as terminal condition.

We denote by S_T^-u_T the backward in time solution at time t=0. In fact, it is proved in [1] that u_T is admissible if and only if the initial condition \tilde{u}_0 = S_T^- u_T satisfies S_T^+ \tilde{u}_T =u_T.

 

Step 2: Secondly, if the observation u_T is not admissible (it might be a consequence of noise effects or errors in the measurements) we compute the following projection of u_T on the set of admissible observations:

u_T^\ast : = S_T^+(S_T^- u_T).

Observe that, for any u_T, the projection u_T^\ast is, by definition, admissible. In [1], it is proven that, under suitable hypotheses on H, the function u_T^\ast is the unique viscosity solution to the nonlinear obstacle problem.

\min \left\{ v - u_T, \ -\lambda_n \left[ D^2 v - \dfrac{[H_{pp}(D v)]^{-1}}{T}\right] \right\} = 0.

Here, for a symmetric n\times n matrix, \lambda_n[X] represents its greatest eigenvalue. The projection u_T^\ast is, in fact, the smallest admissible observation bounded from below by u_T.

 

Step 3: Finally, given an admissible observation u_T, or its projection if it were not admissible, we construct all the initial conditions u_0 satisfying S_T^+ u_0=u_T.

This construction is done in [1], and needs of only two ingredients:

(1) The backward solution \tilde{u}_0 = S_T^- u_T, obtained by reversing the time in (1);

(2) and the following subset of \mathbb{R}^n:

X_T(u_T) : = \left\{ z- T\, H_p (\nabla u_T(z)); \ \forall z\in\mathbb{R}^n \ \text{such that} \ u_T(\cdot) \ \text{is differentiable at} \ z\right\},

obtained from the differentiability points of u_T.

Once these two elements are obtained, one can characterize the initial conditions u_0 satisfying S_T^+ u_0 =u_T as the functions that are bounded from below by \tilde{u}_0 in all \mathbb{R}^n and are equal to \tilde{u}_0 in X_T(u_T).

Click to play

If, for an admissible observation u_T, we denote

I_T(u_T):= \left\{ u_0 \in C^{0,1}(\mathbb{R}^n)\, ; \ S_T^+ u_0 = u_T\right\},

we can characterize this set of initial conditions as follows:

I_T(u_T) = \left\{ \tilde{u}_0 + \varphi \, ; \, \varphi\in C^{0,1}(\mathbb{R}^n)\ \text{such that} \ \varphi\geq 0\ \text{and}\ \text{supp} (\varphi) \subset \mathbb{R}^n\setminus X_T(u_T)\right\}.

 
Inverse Design for Hamilton-Jacobi Equation

We point out that, given an admissible observation u_T, the initial state u_0 can be uniquely determined if and only if X_T(u_T)=\mathbb{R}^n.

This is the case when the solution is C^1 in [0,T], however, as soon as the solution fails to be C^1 in [0,T], the set X_T(u_T) is a proper subset of \mathbb{R}^n and then the initial condition can no longer be uniquely determined in the region \mathbb{R}^n\setminus X_T(u_T) (white region in the right-hand-side plot).

 

References
[1] C. Esteve, E. Zuazua, The inverse problem for Hamilton-Jacobi equations and semiconcave envelopes, 2020 preprint arXiv:2003.06914

Like this?

You might be interested in
[content-slider id=”855,971″ number=”2″ format=”carousel”]

Tags:

Don't miss out our posts on Math & Research!

Probabilistic Constrained Optimization on Flow Networks By Michael Schuster Uncertainty often plays an important role in the context of flow problems. We analyze a stationary and a dynamic flow model with uncertain boundary data and we consider optimization problems with probabilistic constraints in this context. The […]
Perceptrons, Neural Networks and Dynamical Systems By Sergi Andreu // This post is last part of the “Deep Learning and Paradigms” post Binary classification with Neural Networks When dealing with data classification, it is very useful to just assign a color/shape to every label, and so be able to visualize […]
Deep Learning and Paradigms By Sergi Andreu // This post is the 2nd. part of the “Opening the black box of Deep Learning” post Deep Learning Now that we have some intuition about the data, it’s time to focus on how to approximate the functions that […]
Our last Publications
© 2019-2021 Chair for Dynamics, Control and Numerics - Alexander von Humboldt Professorship at FAU Erlangen-Nürnberg, Germany | Imprint