Introduction: The Hidden Fragility of Physical Systems
In physics and engineering, the stability of a system is traditionally measured using the $L^2$ norm, which corresponds to the total energy. A system is considered stable if a small change in input energy leads to a small change in output energy. However, this standard metric can be deceptive. $L^2$ stability allows for "localized spikes"—perturbations that, while low in total energy, possess extremely high peak values.
This paper explores the critical need for $L^\infty$ stability, which governs the maximum value (peak) of a signal. This distinction is vital for high-stakes applications like medical imaging and for the development of robust neural networks, which are often vulnerable to "adversarial attacks"—tiny, imperceptible $L^\infty$ perturbations that cause catastrophic model failure.
The Problem: When Waves Focus into Spikes
The authors demonstrate that the standard wave equation, a cornerstone for many physics-inspired neural network architectures (like Hyperbolic CNNs), is inherently unstable in $L^\infty$. Even if the total energy is conserved, a small initial ripple can focus over time into a massive, destructive spike at a single point.
As shown in the figure above, as the time $t$ progresses from 0 to 1, perturbations with different widths ($w$) converge toward the center. While the initial conditions at $t=0$ appear relatively similar and bounded, the resulting solutions at $t=1$ exhibit massive growth at the origin ($r=0$), particularly for smaller values of $w$. This "focusing" effect is a primary source of $L^\infty$ instability.
Proposed Regularization Strategies
To address these instabilities, the paper introduces two primary mathematical frameworks:
- Double Filtering for Forward Waves: For wave propagation, the authors propose a "double filter" method. By applying filters in both the frequency domain (to suppress high-frequency noise) and the spatial domain, they prove that the solution operator becomes bounded and stable in $L^\infty$.
- Spectral Filtering for Inverse Problems: In linear inverse problems (like deblurring), standard Tikhonov regularization often fails to prevent $L^\infty$ spikes. The authors propose a non-uniform weighting of the Singular Value Decomposition (SVD). This "Spectral Filtering" specifically identifies and suppresses the singular vectors most prone to creating peak-value instabilities.
Key Results and Insights
The theoretical findings are validated through synthetic numerical benchmarks, including 3D wave equations and periodic convolutions. The core insight is that energy conservation does not imply safety. A system can be perfectly stable in terms of energy ($L^2$) while being completely unreliable in terms of peak values ($L^\infty$).
The numerical experiments confirm that without specific $L^\infty$ regularization, even small perturbations lead to significant deviations from the "clean" solution. The proposed methods provide a rigorous path toward ensuring that reconstructions and wave simulations remain physically meaningful and robust against adversarial noise.
Conclusion
This work bridges the gap between partial differential equations (PDEs) and deep learning robustness. By providing a framework for $L^\infty$ stability, the authors offer a mathematical foundation for building neural networks that are not just energy-stable, but truly robust against localized artifacts. Future research will focus on extending these linear frameworks to the nonlinear operators common in modern AI architectures.
Summary updated