On L∞ stability for wave propagation and for linear inverse problems

Rima Alaifari, Giovanni S. Alberti, Tandri Gauksson

TL;DR

This paper tackles the challenge of $L^\infty$ stability for wave propagation and inverse problems, which is crucial for understanding localized phenomena but often not guaranteed by standard methods. The authors introduce a novel regularization technique based on Fourier multipliers that achieves $L^\infty$ stability, extending these ideas to inverse problems and offering insights into the robustness of deep neural networks.

What Problem Does This Paper Address?

In many physical systems and mathematical models, particularly those described by partial differential equations (PDEs) like the wave equation, stability is a fundamental property. This means that small changes in initial conditions lead to small changes in the solution. While $L^2$ stability (related to energy conservation) is well-understood and often achieved, it doesn't always capture the behavior of localized features or "spikes" in a signal. The $L^\infty$ norm, which measures the maximum value of a function, is more relevant for detecting such localized phenomena. However, the linear wave equation, for instance, is notoriously unstable in the $L^\infty$ norm, meaning small initial perturbations can lead to arbitrarily large solutions. This instability is also a key concern in deep learning, where it manifests as adversarial examples – imperceptible input perturbations that cause drastic changes in output. This paper aims to bridge this gap by developing methods that ensure $L^\infty$ stability for wave propagation and inverse problems, thereby improving robustness and providing a theoretical foundation for understanding deep learning stability.

Methods

The core of the paper's contribution lies in developing a regularization technique to achieve $L^\infty$ stability. For the linear wave equation, they introduce a regularized solution operator that is constructed by applying a double filter in both space and frequency to the Fourier multiplier. This approach effectively smooths out the problematic high-frequency components that lead to $L^\infty$ instability.

These ideas are then extended to linear inverse problems. The authors leverage the singular value decomposition (SVD) of the operator and design a nonstandard spectral filtering method. This involves applying carefully chosen weights to the singular values to regularize the problem in a way that is stable in the $L^\infty$ norm.

The paper also connects these concepts to deep neural networks (DNNs). It frames the behavior of certain DNNs (like residual networks) as discretizations of hyperbolic PDEs. The instability of DNNs, particularly the existence of adversarial examples, is linked to the lack of $L^\infty$ stability in these underlying continuous models. The proposed regularization methods offer a potential pathway to enhance the robustness of DNNs against such adversarial attacks.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis and numerical examples using synthetic data generated from the wave equation and inverse problem models.

Key Insights and Ideas

  • $L^\infty$ instability is a fundamental challenge: The paper highlights that while $L^2$ stability is common, $L^\infty$ stability is often lacking, especially for wave-like phenomena and in inverse problems. This has direct implications for understanding and mitigating adversarial attacks in machine learning.
  • Filtering as a solution: The core idea is that by carefully filtering in both space and frequency (or equivalently, by spectrally filtering the SVD), one can suppress the problematic components that lead to $L^\infty$ instability.
  • Bridging PDEs and DNNs: The paper establishes a strong theoretical link between the stability properties of hyperbolic PDEs and the robustness of deep neural networks, suggesting that insights from PDE theory can inform the design of more stable neural architectures.
  • Adversarial sequences are key indicators: The concept of "adversarial sequences" is introduced to formally identify inputs that, while bounded in standard norms, lead to unbounded outputs in the target norm ($L^\infty$). The existence of such sequences signals instability.

Key Results

  • $L^\infty$ stable regularization for the wave equation: The authors demonstrate that their proposed regularized solution operator is stable in the $L^\infty$ norm, effectively overcoming the inherent instability of the linear wave equation.
  • $L^\infty$ stable regularization for inverse problems: They develop a spectral filtering method for inverting compact operators that is stable in the $L^\infty$ norm, which is crucial for applications where the output is expected to be well-behaved.
  • Theoretical connection to DNN robustness: The paper provides a theoretical framework that explains why DNNs can be susceptible to adversarial examples, linking it to the $L^\infty$ instability of underlying continuous models.
  • Numerical examples: Figures are provided illustrating the effectiveness of the proposed regularization methods in suppressing adversarial artifacts for the wave equation and in inverse problems, showing improved reconstructions compared to standard Tikhonov regularization.

Limitations and Future Work

  • Focus on linear models: The current work primarily focuses on linear wave equations and linear inverse problems. The authors acknowledge that extending these ideas to more complex nonlinear PDEs and nonlinear inverse problems is a significant and interesting avenue for future research.
  • Specific filters and conditions: While the paper proposes a general framework, the effectiveness of the regularization depends on the choice of filters and specific conditions on the operator's singular values and singular vectors.
  • Deep neural networks: The connection to DNNs is made through simplified models. Capturing the full complexity of modern deep learning architectures, especially their nonlinearities, requires further investigation.
  • Generalization to other norms: While the paper focuses on $L^\infty$ stability, exploring stability with respect to other norms or combinations of norms could be a valuable extension.

Summary updated

Abstract

Stability is a key property of both forward models and inverse problems, and depends on the norms considered in the relevant function spaces. For instance, stability estimates for hyperbolic partial differential equations are often based on energy conservation principles, and are therefore expressed in terms of [Formula: see text] norms. The focus of this paper is on stability with respect to the [Formula: see text] norm, which is more relevant to detect localized phenomena. The linear wave equation is not stable in [Formula: see text], and we design an alternative solution method based on the regularization of Fourier multipliers, which is stable in [Formula: see text]. Furthermore, we show how these ideas can be extended to inverse problems, and design a regularization method for the inversion of compact operators that is stable in [Formula: see text]. We also discuss the connection with the stability of deep neural networks modeled by hyperbolic PDEs.
BibTeX
@article{alaifari2026l,
  author = {Rima Alaifari and Giovanni S. Alberti and Tandri Gauksson},
  title = {{On L∞ stability for wave propagation and for linear inverse problems}},
  journal = {Journal of Hyperbolic Differential Equations},
  year = {2026},
  doi = {10.1142/s0219891626400011},
}
Figure