🔢 mathematics

Variational optimization approach for reconstruction of dielectric permittivity and conductivity functions using partial boundary measurements

This paper presents a variational optimization framework based on a weak Lagrangian formulation to simultaneously reconstruct dielectric permittivity and conductivity in time-dependent Maxwell's equations from partial boundary measurements, supported by theoretical stability analysis and validated through 2D and 3D numerical experiments.

Original authors: Eric Lindström, Larisa Beilina

Published 2026-02-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eric Lindström, Larisa Beilina

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out what's inside a sealed, opaque box without opening it. You can't see inside, but you can tap on the outside, listen to the echoes, and measure how the vibrations change. Based on those echoes, you want to build a 3D map of the hidden objects inside, knowing exactly what they are made of and where they are.

This is essentially what the paper is about, but instead of a box and tapping, it's about microwaves and human tissue.

Here is a breakdown of the paper's concepts using simple analogies:

1. The Big Problem: The "Blind" Doctor

The researchers are trying to solve a Coefficient Inverse Problem. In plain English, this means: "We know the rules of how waves move (Maxwell's equations), and we can measure the waves bouncing off the outside of a patient. But we don't know the properties of the tissue inside (how conductive it is or how it slows down waves)."

The Goal: Reconstruct a 3D image of a tumor (specifically malignant melanoma) inside the skin.
The Challenge: This is like trying to guess the shape and material of a hidden object just by listening to the echo of a shout in a cave. It's a "messy" problem because small errors in your measurements can lead to huge errors in your guess. This is called an ill-posed problem.

2. The Solution: The "Smart Guessing" Game

To solve this, the authors use a Variational Optimization Approach. Think of this as a game of "Hot and Cold," but played by a super-smart computer.

The Starting Point: The computer starts with a "blank slate" guess (e.g., "The whole body is just normal skin").
The Simulation: It simulates sending microwaves into this "blank" body and calculates what the echoes should look like.
The Comparison: It compares its simulated echoes with the real echoes measured from the patient.
The Correction: If the echoes don't match, the computer knows its guess is wrong. It needs to tweak the "ingredients" (dielectric permittivity and conductivity) of the tissue in its simulation to make the echoes match better.

3. The Secret Weapon: The "Shadow Puppet" (The Adjoint Problem)

How does the computer know exactly where to tweak the ingredients? It uses a mathematical trick called a Lagrangian approach involving an Adjoint Problem.

The Analogy: Imagine you are trying to find a leak in a complex plumbing system. Instead of checking every pipe one by one, you send a "reverse wave" of water from the point where the leak is suspected, flowing backward through the pipes.
In the Paper: The computer runs a "reverse simulation" (the Adjoint Problem). This reverse wave travels backward in time from the sensors on the skin. Where this reverse wave meets the forward wave, it highlights exactly which parts of the tissue need to be changed to fix the error. This tells the computer the most efficient path to the correct answer.

4. The Map Maker: Adaptive Mesh Refinement

The paper introduces a method called Adaptive Conjugate Gradient (ACGA).

The Analogy: Imagine you are drawing a map of a city.
- Old Way: You draw a grid of squares over the whole city. Some squares cover the ocean (empty space), and some cover the dense downtown. You spend the same amount of effort drawing both.
- New Way (ACGA): You start with a rough grid. Then, the computer looks at the "downtown" area (where the tumor is) and says, "This area is too blurry; let's zoom in and draw smaller, more detailed squares here." It ignores the empty ocean areas.
Why it matters: This saves massive amounts of computing power and makes the image of the tumor much sharper and more accurate, especially for deep or complex shapes.

5. The Real-World Application: Finding Cancer

The ultimate goal is Microwave Medical Imaging.

Why Microwaves? Unlike X-rays (which use radiation) or MRIs (which are huge and expensive), microwaves are non-invasive and safe.
The Difference: Cancerous tissue (malignant melanoma) holds more water and has different electrical properties than healthy skin. It acts like a different "material" to the microwaves.
The Result: The paper shows that their math can successfully reconstruct 3D images of tumors, distinguishing them from healthy tissue, even when the data is noisy (like static on a radio).

Summary

The authors have built a sophisticated mathematical "flashlight."

They shine microwaves at a patient.
They catch the messy, noisy echoes.
They use a reverse-wave algorithm (Lagrangian/Adjoint) to figure out exactly what caused the mess.
They use smart zooming (Adaptive Mesh) to focus their computational power only where the tumor is.
The result is a clear, 3D picture of a tumor inside the skin, helping doctors detect cancer earlier and more accurately without radiation.

It's a blend of advanced calculus, physics, and computer science designed to save lives by "seeing the invisible."

1. Problem Statement

The paper addresses a Coefficient Inverse Problem (CIP) within the context of time-dependent Maxwell's equations. The objective is to simultaneously reconstruct two spatially distributed coefficients:

Dielectric permittivity ( $\varepsilon$ )
Electrical conductivity ( $\sigma$ )

Key Constraints and Challenges:

Data Limitation: Reconstruction relies on partial boundary measurements of the scattered electric field (backscattered data), often containing noise.
Ill-posedness: CIPs are inherently ill-posed, meaning solutions may not exist, may not be unique, or may not depend continuously on the data.
Medium: The study focuses on conductive media (unlike previous works that often assumed non-conductive media or used scalar wave approximations).
Application: The primary motivation is microwave medical imaging, specifically the detection of Malignant Melanoma (MM) and Improvised Explosive Devices (IEDs), where distinguishing tissue properties based on $\varepsilon$ and $\sigma$ is critical.

2. Mathematical Model

The authors utilize a stabilized Maxwell system for a non-magnetic object ( $\mu_r \equiv 1$ ). The forward problem is modeled by a second-order hyperbolic equation for the electric field $E$ :
$\varepsilon \partial_{tt}E + \sigma \partial_t E - \Delta E - \nabla \nabla \cdot (\varepsilon - 1)E = -\partial_t J$
subject to initial conditions and Neumann boundary conditions. The term $\nabla \nabla \cdot (\varepsilon - 1)E$ acts as a Coulomb gauge-type stabilization term to ensure the well-posedness of the system.

3. Methodology

The authors propose a Variational Optimization Approach based on the minimization of a regularized Tikhonov functional using a Lagrangian formulation.

A. Tikhonov Functional

The inverse problem is transformed into an optimization problem by minimizing:
$F(\varepsilon, \sigma) = \frac{1}{2} \| E(\varepsilon, \sigma) - \tilde{E}_{obs} \|^2_{\Gamma_T} + \frac{\gamma_\varepsilon}{2} \| \varepsilon - \varepsilon_0 \|^2 + \frac{\gamma_\sigma}{2} \| \sigma - \sigma_0 \|^2$
Where:

$\tilde{E}_{obs}$ is the noisy boundary measurement.
$\varepsilon_0, \sigma_0$ are initial guesses.
$\gamma_\varepsilon, \gamma_\sigma$ are regularization parameters.

B. Lagrangian and Weak Formulation

To handle the implicit dependence of $E$ on $\varepsilon$ and $\sigma$ , a Lagrangian $L(E, \lambda, \varepsilon, \sigma)$ is constructed. This involves:

Defining an adjoint state $\lambda$ (solving a backward-in-time problem).
Deriving the weak form of the Lagrangian using integration by parts to lower regularity requirements on the solution spaces ( $H^1$ ).
Establishing optimality conditions by setting the Fréchet derivatives of the Lagrangian with respect to $E$ , $\lambda$ , $\varepsilon$ , and $\sigma$ to zero.

C. Theoretical Analysis

The paper provides rigorous mathematical proofs for:

Stability Estimates: Energy estimates for both the forward and adjoint problems, proving that the solution depends continuously on the data.
Fréchet Differentiability: Proof that the Tikhonov functional is Fréchet differentiable, allowing the use of gradient-based methods.
Existence of Minimizers: Proof that unique minimizers exist for both the non-regularized and regularized functionals under specific conditions.

D. Numerical Algorithms

Two gradient-based algorithms are implemented:

Conjugate Gradient Algorithm (CGA): Standard gradient descent using the computed gradients of the functional.
Adaptive Conjugate Gradient Algorithm (ACGA): Enhances CGA by employing adaptive local mesh refinement. The mesh is refined in regions where the gradients of the coefficients (or error indicators) are large, significantly improving reconstruction accuracy without globally increasing computational cost.
- The method uses a Domain Decomposition strategy, combining Finite Element Methods (FEM) in the region of interest and Finite Difference Methods (FDM) in the surrounding domain.

4. Key Contributions

Simultaneous Reconstruction: A novel framework for reconstructing both permittivity and conductivity in conductive media, moving beyond scalar approximations.
Rigorous Theory: Complete derivation of the weak Lagrangian, optimality conditions, stability estimates for the adjoint problem, and proofs of Fréchet differentiability and minimizer existence.
Adaptive Strategy: Integration of a posteriori error indicators to drive local mesh refinement (ACGA), which is crucial for resolving sharp contrasts in medical imaging.
3D Medical Imaging Application: Successful application to 3D models of malignant melanoma, demonstrating the feasibility of detecting tumors based on dielectric contrasts.

5. Numerical Results

The authors validated the approach using 2D and 3D numerical experiments:

2D Tests:
- Demonstrated convergence of the CGA algorithm.
- Showed that starting with initial guesses closer to the true values avoids local minima and ensures consistent error decay.
- Verified the decay of adjoint norms and gradient norms as expected by theory.
3D Tests (Malignant Melanoma Detection):
- Setup: Simulated backscattered data from 3D models of skin with embedded melanoma tumors (at different growth stages) at 6 GHz. Data included 10% Gaussian noise.
- Performance:
  - CGA on coarse meshes: Could identify the location and approximate shape but struggled with sharp boundaries and conductivity contrast.
  - ACGA: Significantly improved reconstruction quality. As the mesh was adaptively refined (up to 4 levels), the reconstructed shapes of the tumors became more accurate, and the contrast in $\sigma$ improved substantially.
- Metrics: Relative $L^2$ errors for permittivity and conductivity decreased as the mesh was refined. The method successfully reconstructed the conical shape of the tumor at a depth of ~4 mm.

6. Significance and Future Work

Significance: This work bridges the gap between rigorous mathematical theory and practical 3D imaging applications. It proves that variational optimization with Lagrangian approaches is a viable, stable, and accurate method for solving complex Maxwell CIPs in conductive media.
Impact: The ability to non-invasively detect malignant melanoma using microwaves offers a potential alternative or complement to X-ray mammography and MRI, particularly for patients where radiation or contrast agents are undesirable.
Future Directions:
- Reconstructing unweighted (real) dielectric properties rather than the weighted values used in the current tests.
- Developing a new adaptive DDFE/FDM method specifically for finite difference domains to replace the acoustic model with a damped wave equation for higher fidelity.
- Validating the approach with experimentally acquired data.