Regularized Multiwavelet Collocation Methods for Ill-Posed Fredholm Integral Equations of the First Kind

Technical Summary: Regularized Multiwavelet Collocation Methods for Ill-Posed Fredholm Integral Equations of the First Kind

Problem Statement
Fredholm integral equations of the first kind arise naturally in inverse problems where measurements are indirect, smoothed, or contaminated by noise. Because the underlying integral operators are compact, direct discretization of these equations typically yields unstable reconstructions; small perturbations in the data can cause arbitrarily large changes in the solution. This ill-posedness necessitates a numerical strategy that combines approximation with regularization. While wavelet and polynomial bases have been extensively used to discretize such equations, the specific influence of basis selection on the stabilization behavior of the inverse problem—beyond mere approximation accuracy—remains a diagnostic gap.

Methodology
The paper develops a unified regularized multiwavelet-collocation framework to solve the generic first-kind Fredholm equation $Kf = g$. The methodology proceeds through the following steps:

1. Discretization via Wavelet Bases: The unknown function $f$, the data $g$, and the kernel $K$ are approximated using five distinct compactly supported bases:

   • Linear Legendre Multiwavelets (LLMW)
   • Legendre Wavelets
   • Haar Wavelets
   • B-spline Wavelets
   • Hybrid Wavelets (combining block-pulse and Legendre polynomials)
     The integral equation is transformed into a linear algebraic system $A_N C \approx G_N$, where the matrix $A_N$ depends on the specific quadrature rules and orthogonality properties of the chosen basis.

2. Tikhonov Regularization: To stabilize the ill-conditioned system, the authors employ Tikhonov regularization. The unstable first-kind formulation is converted into a stabilized problem by minimizing the functional $J_\alpha(f) = \|Kf - g_\delta\|^2 + \alpha\|f\|^2$. This leads to the solution of the normal equations $(A_N^T A_N + \alpha I)C = A_N^T G_\delta$.

3. Parameter Selection: The regularization parameter $\alpha$ is selected a posteriori using Morozov's discrepancy principle. Specifically, the study utilizes a cubic-convergence procedure based on the damped discrepancy principle to ensure the residual $\|A_N C - G_\delta\|$ matches the noise level $\delta$.

4. Diagnostic Protocol: A critical aspect of the methodology is the use of a fixed computational protocol for all bases. This includes identical discretization sizes ($N=32$), collocation grids, noise realizations (using a fixed random seed), and noise levels ($\delta = 10^{-4}$ to $5 \times 10^{-2}$). This ensures that observed differences are attributable to the basis properties rather than implementation variances.

Key Contributions
The primary contribution of this work is a systematic, reproducible diagnostic comparison of the five wavelet bases under a single Tikhonov–Morozov regularization framework. The paper moves beyond standard approximation error tables to evaluate the following metrics for each basis:

• Solution Error ($E_{sol}$): Accuracy of the recovered function.
• Residual Error ($E_{res}$): How well the regularized solution satisfies the noisy algebraic system.
• Conditioning: Both the condition number of the discrete operator $\kappa(A_N)$ and the regularized normal equations $\kappa(A_N^T A_N + \alpha I)$.
• Parameter Behavior: The magnitude of the discrepancy-selected $\alpha$ across different noise levels.
• Computational Cost: Assembly and solve times.

Results
The numerical experiments, conducted on benchmark integration problems and 1D/2D Fredholm equations (including Gaussian smoothing kernels with both smooth and nonsmooth targets), yield the following findings:

• Basis Selection as Regularization: The choice of basis significantly alters the stabilization behavior. Bases that perform similarly in noise-free approximation may behave drastically differently when noise and regularization are introduced.
• Smooth vs. Nonsmooth Targets:
  • For smooth solutions, higher-order local bases (Linear Legendre Multiwavelets, Legendre wavelets, and B-splines) generally provide the best reconstruction accuracy and favorable conditioning after regularization.
  • For nonsmooth solutions (e.g., functions with corners), the advantage of high-order polynomial bases diminishes. In these cases, Haar and hybrid bases, which offer compact support and locality, remain competitive and robust.
• Conditioning and Stability: Unregularized collocation matrices exhibit extreme ill-conditioning (order $10^{16}$). Tikhonov regularization reduces this by several orders of magnitude, but the residual condition number still varies by basis. The study demonstrates that a basis with low approximation error in a clean setting may require more aggressive regularization or exhibit higher sensitivity to noise than a simpler basis.
• Computational Trade-offs: While Haar wavelets are computationally efficient and simple to implement, they often yield less accurate smooth reconstructions. Conversely, LLMW and B-spline bases offer superior accuracy for smooth problems but incur higher assembly costs. The "best" basis is not determined by accuracy alone but by the balance of accuracy, conditioning, noise robustness, and cost.

Significance and Claims
The paper claims that for ill-posed first-kind Fredholm equations, basis selection should be viewed as an integral component of the regularization strategy itself, rather than solely as an approximation-theoretic choice.

The authors argue that the interaction between the discretization basis and the regularization mechanism is critical. The basis influences the spectrum of the discrete operator, the conditioning of the normal equations, the value of the regularization parameter selected by the discrepancy principle, and the propagation of noise. Consequently, selecting an approximation space requires balancing accuracy, stability, conditioning, and computational cost. The study provides practical guidelines:

• Use Linear Legendre Multiwavelets, Legendre wavelets, or B-splines for smooth inverse problems where high reconstruction accuracy and stability are paramount.
• Use Haar or Hybrid wavelets when computational simplicity, low cost, locality, or the recovery of nonsmooth features are the primary concerns.

This diagnostic perspective distinguishes the work from previous studies that focused primarily on wavelet approximation properties or isolated regularization techniques, offering a unified framework for evaluating the reliability of inverse recovery in noisy environments.