In classical signal processing, the Nyquist-Shannon sampling theorem provides a comforting guarantee: if a signal is band-limited and has finite energy (residing in the $L^2$ space), we can perfectly reconstruct it from discrete samples. We often assume that if we simply increase the sampling rate or choose our points more cleverly, the reconstruction error will vanish.
However, research into integrable function spaces—such as the Paley-Wiener space $\mathcal{PW}^1_\pi$ and the Hardy space $\mathcal{H}^1$—reveals a much more volatile reality. In these domains, standard reconstruction procedures don't just fail to converge; they can strongly diverge, causing the reconstruction error to blow up to infinity even as we add more data.
The Collapse of Point-Based Sampling
The fundamental problem lies in the nature of point evaluations. In many mathematical spaces used to model real-world signals, a single point measurement does not carry enough "information" to constrain the behavior of the function between samples.
Brendan Farrell demonstrates in Strong Divergence of Reconstruction Procedures for the Paley-Wiener Space $\mathcal{PW}^1_π$ and the Hardy Space $\mathcal{H}^1$ that for certain signals, the reconstruction error doesn't just fluctuate—it oscillates wildly toward infinity. This "strong divergence" is unavoidable; no matter how you choose your sampling points or how high your sampling rate is, there exist signals within these spaces that will cause the reconstruction to fail catastrophically.
This isn't just a theoretical curiosity. Volker Pohl and Holger Boche argue in System Approximations and Generalized Measurements in Modern Sampling Theory that this represents a "digital-analog gap." They show that certain stable analog linear time-invariant (LTI) systems simply cannot be implemented digitally if the process relies solely on point samples. Even with infinite oversampling, the discrete points fail to capture the necessary dynamics of the system.
Why Reconstruction Blows Up
To understand why this happens, consider the geometry of the functions involved. In $L^2$ spaces, the "energy" of the signal is spread out in a way that allows sinc-function interpolation to balance itself out. But in $L^1$ spaces (integrable signals), functions can exhibit "peaking" behavior.
This is seen clearly in spherical geometries. Holger Rauhut and Rachel Ward note in Sparse recovery for spherical harmonic expansions that basis functions like spherical harmonics often peak at the poles. This lack of uniform boundedness makes it difficult for standard compressed sensing tools to provide stable recovery. If the signal "hides" its complexity in these peaks, a point-based sample is likely to miss it or misrepresent it so severely that the resulting reconstruction overcompensates, leading to divergence.
Turning Instability into a Tool
While divergence is usually a hurdle for engineers, it can be a feature for security. Loukas Grafakos and colleagues exploit this mathematical fragility in Signal Encryption Strategy based on Domain change of the Fractional Fourier Transform.
By mapping a signal into a non-integrable functional space using the Fractional Fourier Transform (FRFT), they create a version of the signal where standard inverse transforms fail to converge. Only a receiver who knows the exact "approximate identities" (specific summability methods) can stabilize the reconstruction. To anyone else, the signal appears as a divergent, unrecoverable mess—effectively using the inherent instability of the function space as a layer of encryption.
Beyond Points: Generalized Measurements
The recurring lesson across these papers is that the classical reliance on point-wise values is insufficient for complex signal spaces. To achieve stability, the field is moving toward generalized measurements.
Instead of looking at a signal at a single instant, generalized measurements might involve local averages or weighted integrals. As Volker Pohl suggests, bridging the gap between analog systems and digital hardware requires moving away from the "point evaluation" paradigm and toward measurement strategies that are robust against the strong divergence inherent in integrable spaces.