Source-linked AI summary
Error Analysis of the Inverse Conductivity Problem with Scattered Measurements
Bangti Jin, Qimeng Quan, Wenlong Zhang
TL;DR
The paper studies recovering an elliptic equation’s conductivity from finitely many noisy measurements, where existing convergence analysis is limited and low-regularity conductivity weakens standard regularization analysis. It uses a W 1,4(Ω)-penalized least-squares formulation with Galerkin finite elements and establishes high-probability error bounds for both regularized and discrete solutions. Numerical experiments illustrate the theoretical findings and show convergence as sampling increases.
Problem
Conductivity must be recovered from finitely many noisy measurements, while convergence analysis for scattered-measurement PDE parameter identification remains limited.
Method
The paper uses a W 1,4(Ω)-penalized regularized least-squares formulation and discretizes it with Galerkin finite elements using continuous piecewise linear elements.
Results
High-probability L2(Ω) error bounds are established for the regularized solution and its Galerkin approximation, with dependencies on γ, n, and h.
Takeaways & Limitations
Numerical experiments support the analysis: eq and eu converge steadily to zero as the number n of sampling points increases, and the a priori parameter choice is effective.
Takeaways & Limitations
The analysis is constrained by limited smoothing for low-regularity conductivity and insufficient Sobolev regularity of the discrete state.
Abstract
from arXiv · showhide
In this work, we investigate the inverse problem of recovering the conductivity coefficient in an elliptic equation from noisy measurements collected at finitely many deterministic scattered points in the domain $Ω$, and corrupted by random noise. Inspired by the regularity analysis, we propose a numerical scheme based on the regularized least-squares formulation with a $W^{1,4}(Ω)$ penalty, and discretize the regularized problem using the Galerkin finite element method with continuous piecewise linear elements. Under suitable assumptions on the problem data, we provide an error analysis of the regularized solution and its Galerkin approximation. We establish $L^2(Ω)$ error bounds in a high-probability sense, which depend explicitly on the regularization parameter $γ$, the number $n$ of data points and the mesh size $h$. We also present numerical experiments to illustrate the theoretical findings.
1 Introduction
The paper addresses conductivity recovery from finitely many noisy point measurements using a W 1,4(Ω)-regularized least-squares formulation and Galerkin finite elements. It develops high-probability L2(Ω) error analysis for regularized and discrete solutions, while accounting for limited smoothing and discrete-state regularity.
- Problem setting: The inverse conductivity problem estimates q† from finitely many noisy measurements at deterministic locations, reflecting fixed sensors with limited precision.The noise is modeled as i.i.d. zero-mean sub-Gaussian with noise level σ.
- Motivation: Existing convergence analysis for scattered-measurement PDE parameter identification is limited, and standard H1(Ω) penalization is insufficient for low-regularity conductivity.Because conductivity enters the leading elliptic-operator term, the forward solution operator has limited smoothing.
- Method: The proposed reconstruction uses a W 1,4(Ω) penalty and Galerkin finite elements to discretize the regularized conductivity problem.The paper first establishes well-posedness and Lipschitz stability for the regularized formulation.
- Theoretical analysis: High-probability bounds are established for the regularized solution, including its W 1,4(Ω) norm and the semi-discrete state approximation.The analysis combines probabilistic bounds with a conditional stability estimate.
- Theoretical analysis: The Galerkin approximation with continuous piecewise linear elements satisfies a high-probability a priori stochastic L2(Ω) error bound.The discrete argument differs substantially because the discrete state lacks sufficient Sobolev regularity.
- Scope: The paper analyzes dependence on the regularization parameter γ, observation count n, and mesh size h, and includes numerical illustrations.The paper is organized around formulation, proofs, numerical experiments, and supporting preliminaries.
2 Main results and discussions
The paper formulates conductivity recovery from noisy scattered measurements using W^{1,4}(Ω)-regularized least squares and continuous piecewise-linear Galerkin FEM. It establishes high-probability L2(Ω) error bounds for both continuous and discrete regularized solutions, with explicit dependence on γ, n, and h.
- Regularized formulation: The conductivity is recovered by minimizing a least-squares functional with a W^{1,4}(Ω) penalty.The penalty supports regularity and Lipschitz stability properties of the forward map used in the analysis.
- FEM approximation: The Galerkin discretization uses continuous piecewise-linear finite elements on a quasi-uniform simplicial triangulation.The spaces V_h and X_h discretize the conductivity and state, respectively.
- Data assumptions: The sampling points are scattered but quasi-uniformly distributed, making the discrete sampling semi-norm comparable with the L2(Ω) norm.The distribution condition requires d_max ≤ b d_min for all sufficiently large n.
- Discrete error estimate: Theorem 2.6 establishes a high-probability stochastic L2(Ω) error bound for the Galerkin approximation q*_h.The bound depends on the regularization parameter γ, the mesh size h, the number of data points n, and quantities determined by the assumptions.
- Parameter guidance: Choosing h = O(γ^{1/4}) provides guidelines for balancing mesh size and penalty parameter, while the discrete convergence rate agrees with the continuous case.The discrete estimate has a larger exponent on ρ0 because FEM analysis repeatedly uses H2(Ω) regularity.
3 Proof of Theorem 2.4
The proof of Theorem 2.4 combines forward-solution regularity, stability, sampling-norm estimates, and stochastic bounds to control the regularized conductivity. Existence and continuous dependence of a minimizer support the high-probability L2(Ω) error analysis.
- Forward-map regularity: Sobolev embeddings and iterative elliptic regularity establish H2(Ω) regularity for the state u(q).For q, q̃ ∈ W^{1,4}(Ω) ∩ A and f ∈ L2(Ω), the state also satisfies a Lipschitz stability estimate with respect to q.
- Forward-map regularity: Energy estimates and Poincaré’s inequality bound the state in H1(Ω) using the L2(Ω) norm of f.The proof selects dimension-dependent Hölder exponents to control nonlinear terms.
- Minimizer properties: A global minimizer q* exists almost surely and depends continuously on the data perturbation.The existence proof uses pointwise evaluation enabled by H2(Ω) embedding into C(Ω), weak lower semicontinuity, and compactness.
- Stochastic control: Sub-Gaussian and ψ2-Orlicz bounds control the sampling discrepancy and the W^{1,4}(Ω) norm of q* with high probability.The stochastic estimates are combined with the minimizing property of q* and entropy bounds for function classes.
- Sampling-norm comparison: The proof connects the discrete sampling semi-norm to the standard L2(Ω) norm for sufficiently regular functions.Under quasi-uniform sampling, the comparison is used to transfer stochastic measurement estimates into spatial error control.
4 Proof of Theorem 2.6
The proof of Theorem 2.6 adds finite-element approximation and stability estimates to the continuous stochastic analysis. It bounds state and conductivity discretization errors while retaining high-probability control.
- FEM approximation: Interpolation, projection, and inverse estimates quantify approximation errors in the continuous piecewise-linear FEM spaces.The estimates apply for Sobolev indices satisfying sp > d and support the subsequent state-error analysis.
- State approximation: Céa’s lemma and Nitsche’s trick provide energy- and L2(Ω)-error estimates for the finite-element state approximation.The duality argument requires an H2(Ω)-conforming auxiliary space and explains the role of additional regularity.
- Discrete stochastic control: The discrete analogue of the stochastic lemma supplies ψ2-Orlicz bounds for the discrete minimizer q*_h.These bounds incorporate the mesh size h and are combined with estimates for the discrete state and interpolated exact coefficient.
- Discrete stochastic control: Sub-Gaussian process estimates control the stochastic terms over the finite-element spaces using their dimension and approximation properties.The proof decomposes the error into five events and bounds them separately.
5 Numerical experiments and discussions
The numerical experiments evaluate regularization and discretization behavior under varying noise, sampling density, and conductivity profiles. Results support the theoretical parameter choice and show decreasing errors with more observations.
- Numerical method: The reconstruction problems are solved with steepest descent, while the nonlinear 4-Laplace subproblem uses Newton iteration initialized by a Poisson solution.The discrete approximation is examined using defined errors, with γ selected according to the theoretical guideline.
- Regularization selection: The optimal γ values are 10^-8.5 for σ = 5.0% and 10^-9 for σ = 1.0%, close to the a priori choices 3.32e-9 and 3.89e-10, respectively.The reported observation confirms the effectiveness of the a priori choice (5.3).
- Sampling-density convergence: For both σ = 5.0% and σ = 1.0%, the metrics eq and eu steadily converge to zero as n increases from 512 to 4012.The table reports eq values 3.13e-2 to 9.77e-3 and eu values 2.75e-4 to 1.32e-4 for σ = 5.0%, while the lower-noise case reports eq values 2.07e-2 to 1.02e-2 and eu values 1.01e-4 to 6.96e-5.
- Conductivity reconstructions: The experiments include exact conductivities with either one bump or multiple bumps, and the latter reconstruction captures the multiple-bump contours well.Pointwise errors concentrate near the bumps of the exact state u† in both conductivity profiles.
A Preliminaries on sub-Gaussian random variables
The appendix introduces sub-Gaussian random variables, Orlicz norms, empirical-process tools, and covering-number bounds used in the stochastic analysis. These preliminaries support concentration and complexity estimates for Sobolev and finite-dimensional classes.
- Sub-Gaussian variables: A sub-Gaussian random variable is defined through a parameter σ and has an exponentially decaying probability tail.The appendix also records converse relationships between tail bounds and sub-Gaussian estimates.
- Orlicz norms: The appendix defines the Orlicz norm using a monotonically increasing convex function ψ and specializes to ψ2(t) = exp(t^2) − 1.This norm is used to state bounds for sub-Gaussian variables.
- Random processes: A sub-Gaussian random process is defined on a semimetric space, and a maximal inequality is stated for separable processes indexed by T.These tools provide stochastic convergence estimates through process-level control.
- Covering numbers: The appendix defines ε-covering numbers as the minimum cardinality of an ε-cover under a semimetric d.An ε-cover requires every t ∈ T to lie within ε of at least one selected point.
- Function-class complexity: Covering-number bounds are supplied for Sobolev subsets and bounded finite-dimensional subsets, with conditions involving s, p, q, and the dimension d.These bounds quantify the complexity of function classes used in the analysis.