Filtering methods for coupled inverse problems

Michael Herty,Elisa Iacomini

DOI: https://doi.org/10.48550/arXiv.2203.09841

2022-03-18

Abstract:We are interested in ensemble methods to solve multi-objective optimization problems. An ensemble Kalman method is proposed to solve a formulation of the nonlinear problem using a weighted function approach. An analysis of the mean field limit of the ensemble method yields an explicit update formula for the weights. Numerical examples show the improved performance of the proposed method.

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in multi - objective optimization problems, how to effectively solve coupled inverse problems. Specifically, the author focuses on how to determine the unknown parameter \(u\) given the observed data and the mathematical model, so that these parameters can simultaneously approximate multiple observed data. Such problems are widely present in geophysics, petroleum and water resource management and other fields.

The main contribution of the paper is to propose a method based on the Ensemble Kalman Filter (EnKF) to solve these problems. By introducing the method of weight functions, the multi - objective optimization problem is transformed into a weighted objective function, and an explicit formula for weight update is derived using the mean - field limit. This method not only improves the solution efficiency, but also can show better performance in numerical experiments.

The following are the key formulas in the paper:

1. **Mathematical model of the inverse problem**:
\[
y = G(u)+\eta
\]
where \(G\) is a nonlinear forward operator from the finite - dimensional space \(X = \mathbb{R}^d\) to \(Y=\mathbb{R}^k\), \(u\in X\) is the unknown parameter, \(y\in Y\) is the observed value, \(\eta\sim N(0,\Gamma)\) is the observation noise, and \(\Gamma\) is a known covariance matrix.

2. **Single - objective optimization problem**:
\[
u^*=\arg\min_{u\in X}\frac{1}{2}\left\|\Gamma^{-\frac{1}{2}}(y - G(u))\right\|^2
\]

3. **Multi - objective optimization problem**:
\[
\min_{u\in X}\left(\left\|\Gamma^{-\frac{1}{2}}(y_1 - G_1(u))\right\|,\ldots,\left\|\Gamma^{-\frac{1}{2}}(y_l - G_l(u))\right\|\right)
\]

4. **Weighted objective function**:
\[
G(u,\lambda):=\sum_{i = 1}^l\lambda_iG_i(u):X\times\Lambda\rightarrow Y
\]
where \(\Lambda:=\left\{\lambda\in\mathbb{R}_+^l:\lambda\cdot\mathbf{1} = 1\right\}\) and \(\mathbf{1}=(1,\ldots,1)^T\).

5. **Pareto optimal solution**:
\[
u^*(\lambda)=\arg\min_{u\in X}\Phi(u,\lambda),\quad\Phi(u,y,\lambda)=\frac{1}{2}\left\|\Gamma^{-\frac{1}{2}}\sum_{i = 1}^l\lambda_i(y_i - G_i(u))\right\|^2
\]

6. **EnKF update formula**:
\[
u_j^{n + 1}=u_j^n+C(U_n)\left(D(U_n)+\frac{1}{\Delta t}\Gamma^{-1}\right)^{-1}\left[y - G(u_n,\lambda)\right]
\]
where
\[
C(U_n)=\frac{1}{J}