Hölder-Logarithmic Stability and Convergence Rates for an Inverse Random Source Problem

1. Problem Formulation

The paper addresses the Inverse Random Source Problem (IRSP) for the time-harmonic acoustic wave equation (Helmholtz equation) in $\mathbb{R}^3$.

• Governing Equation: $\Delta u + \kappa^2 u = -Q$, where $Q$ is a random source supported in a bounded domain $D$.
• Source Model: The source $Q$ is modeled as an uncorrelated, centered Gaussian random process with zero mean ($E[Q]=0$). The unknown quantity to be recovered is the source power function $q(y)$, which defines the covariance operator of the source:
  $$\text{Cov}[Q](v)(y) = q(y)v(y)$$
• Data: The available data are correlation measurements (sample covariances) of the wave field $u$ measured on a boundary surface $M = \partial O$ surrounding the source. The forward operator maps the source strength $q$ to the covariance operator of the solution restricted to $M$.
• Context: This problem arises in applications like aeroacoustics (noise localization) and helioseismic holography. Unlike deterministic inverse source problems, IRSPs are uniquely solvable even with single-frequency data, but they remain ill-posed, requiring regularization for stable reconstruction.

2. Methodology

The authors employ a combination of Complex Geometrical Optics (CGO) solutions, Variational Source Conditions, and Spectral Regularization theory.

• Complex Geometrical Optics (CGO): The core analytical tool involves constructing CGO solutions to the homogeneous Helmholtz equation. These are exponentially growing solutions of the form $u_\xi(x) = e^{i\xi \cdot x}$ where $\xi \in \mathbb{C}^3$ and $\xi \cdot \xi = \kappa^2$.

  • The authors use these solutions to bound the Fourier coefficients of the difference between two source strengths ($q_1 - q_2$) in terms of the Hilbert-Schmidt norm of the difference in their covariance data.
  • This allows them to derive estimates on the low-frequency components of the source, which are crucial for stability analysis.

• Variational Source Conditions (VSC): Instead of deriving stability estimates directly, the authors establish a Variational Source Condition (a variational inequality involving the source term and the forward operator).

  • They verify this condition by splitting the problem into two parts: a generalized smoothness assumption (Sobolev regularity of the source) and an ill-posedness estimate (derived via the CGO bounds).
  • This approach allows them to leverage existing theory (specifically from [20]) to translate the VSC into both stability estimates and convergence rates for spectral regularization methods.

• Extension to Convected Helmholtz: The methodology is extended to the convected Helmholtz equation (relevant for aeroacoustics with background flow) using a Lorentz transformation. This maps the convected problem to a standard Helmholtz problem with a modified wave number, allowing the previous proofs to be adapted.

3. Key Contributions

• Conditional Hölder-Logarithmic Stability: The paper derives conditional stability estimates of the form:
  $$\|q_1 - q_2\|_{H^m} \leq C \left( \log(3 + \delta^{-2}) \right)^{-(s-m)}$$
  where $\delta$ is the noise level, $s$ is the Sobolev regularity of the source, and $m$ is the regularity of the reconstruction space.

  • Improvement over [26]: Previous work established logarithmic stability where constants deteriorated as the wave number $\kappa \to \infty$. This paper shows that if $\kappa$ grows sufficiently fast as $\delta \to 0$, the stability becomes Hölder-type (i.e., polynomial convergence).
  • Reduced Regularity: The minimal Sobolev regularity index required for the source is reduced from $s \geq 3$ (in [26]) to $s > 0$.

• Convergence Rates for Regularization: By establishing the VSC, the authors prove convergence rates for spectral regularization methods (including Tikhonov, Landweber iteration, and spectral cut-off) applied to noisy data. The rates match the stability estimates.

• Joint Limit Analysis: A significant theoretical contribution is the analysis of the joint limit where the noise level $\delta \to 0$ and the wave number $\kappa \to \infty$. The authors show that in this regime, the logarithmic stability estimate transitions to a Hölder-type estimate with an exponent that depends on the Sobolev smoothness $s$.

4. Results

• Theoretical Bounds:

  • Theorem 2.3: Establishes the conditional logarithmic stability estimate.
  • Theorem 2.4: Proves convergence rates for spectral regularization methods under the derived VSC.
  • Theorem 5.2: Extends these results to the convected Helmholtz equation, showing similar stability and convergence behavior dependent on the Mach number.

• Numerical Experiments:

  • The authors performed numerical simulations using Tikhonov regularization with synthetic data generated from random sources (modeled by B-splines of degree 1 and 3).
  • Findings:
    1. Convergence: The numerical error decreases as the noise level decreases, confirming the predicted logarithmic convergence rates.
    2. Smoothness: Higher Sobolev smoothness ($s$) leads to faster convergence (larger exponents in the error bound).
    3. Wave Number Dependence: Increasing the wave number $\kappa$ significantly improves reconstruction quality and shifts the onset of the asymptotic convergence regime to higher noise levels. This validates the theoretical finding that high-frequency data compensates for ill-posedness.
    4. Optimality: The observed numerical exponents were slightly larger (better) than the theoretical upper bounds, suggesting the theoretical bounds might not be sharp, though proving optimality remains an open challenge.

5. Significance

• Bridging Deterministic and Random Inverse Problems: The work clarifies the unique solvability and stability properties of random source problems, distinguishing them from their deterministic counterparts.
• High-Frequency Advantage: It provides a rigorous mathematical justification for the empirical observation in imaging (like seismic or aeroacoustic) that using higher frequencies improves resolution and stability, even in the presence of noise.
• Practical Applicability: The reduction of the required regularity ($s>0$) and the extension to convected flows make the results more applicable to real-world engineering problems where sources may be rough and flows present.
• Framework for Future Work: The use of Variational Source Conditions provides a robust framework for analyzing other inverse problems involving random data, opening avenues for statistical convergence analysis and handling correlated sources.

In summary, this paper provides a rigorous mathematical foundation for solving inverse random source problems, demonstrating that while the problem is ill-posed (logarithmic stability), the ill-posedness can be mitigated by high-frequency data and source smoothness, leading to Hölder-type convergence in specific regimes.