Signal and System Approximation from General Measurements

Holger Boche, Ullrich J. MönichApplied and Numerical Harmonic Analysis 2014 · DOI · OpenAlex

Abstract

This paper provides a rigorous mathematical analysis of the limits inherent in digital signal processing. Authored by Holger Boche and Ullrich J. Mönich, the work challenges the fundamental assumption that any stable analog system can be perfectly replicated in the digital domain through standard sampling techniques. Focusing on the Paley–Wiener space $PW^1_\pi$, the authors demonstrate that pointwise sampling can lead to divergence in digital systems, even when exceeding the Nyquist rate. They propose a solution involving "general measurements" and oversampling to guarantee mathematical stability in digital approximations.

The Challenge of Digital Approximation

In the modern digital era, the seamless transition between analog and digital domains is often taken for granted. The standard engineering workflow involves sampling an analog signal, processing the resulting discrete data in a computer, and reconstructing the output to return to the analog world. This process assumes that the digital representation is a stable and faithful proxy for the continuous-time signal.

Conceptual framework comparing analog and digital system implementations.

However, the authors ask a deep theoretical question: Is this process always stable? Specifically, if we have a stable analog system (such as a filter or a Hilbert transform), can we always build a digital version that uses only a finite number of samples and produces an output that converges to the true analog output? The paper focuses on signals in the Paley–Wiener space $PW^1_\pi$ (signals whose Fourier transforms are integrable). While this might seem like a niche mathematical detail, it represents a broad class of practically relevant signals. The "gap" identified is that for this specific space, traditional Shannon sampling is not robust enough to guarantee that digital processing will not fail.

Theoretical Framework and Methodology

The study is rooted in functional analysis and operator theory. The authors treat Linear Time-Invariant (LTI) systems as operators and analyze their norms to determine stability. A key tool employed is the Banach-Steinhaus Theorem, which is used to prove that if a sequence of operators is not uniformly bounded, there must exist some "pathological" input signal that causes the system to diverge.

Beyond standard pointwise samples $f(t_k)$, the authors introduce "general measurement functionals" $c_k(f)$. These functionals represent more complex ways of measuring a signal, such as taking a weighted average over a small window of time. This approach acknowledges that physical sensors often integrate signal values over a finite duration rather than capturing an infinitely thin point in time. Additionally, the paper explores the use of Walsh functions—square-wave-like basis functions—to create measurement sequences that may exhibit better convergence properties than standard trigonometric bases.

The Fragility of Pointwise Sampling

The most striking insight of the research is that pointwise sampling is inherently fragile. For signals in $PW^1_\pi$, a single point value does not carry enough "weight" to ensure the stability of the overall system approximation. The authors prove that for any sampling sequence, one can always find a stable analog system and a valid signal that, when combined, cause the digital approximation to diverge to infinity.

Crucially, the common engineering intuition that oversampling (sampling faster than the Nyquist rate) solves all reconstruction issues is challenged. This paper shows that for pointwise sampling in $PW^1_\pi$, oversampling still fails to prevent divergence. This suggests a fundamental limitation in how we interface the analog and digital worlds using discrete points.

Stability through General Measurements

The path to stability lies in changing how signals are measured. The authors demonstrate that stability is recovered if "general measurement functionals" are used in conjunction with oversampling. By looking at a small neighborhood of the signal rather than a single point, the digital implementation can be made robust.

Furthermore, the research identifies interesting behavior regarding subsequences. Even without oversampling, convergence can be achieved if one is selective about which samples are retained. By utilizing specific subsequences—such as doubling the number of measurements at each step using Walsh functions—stability can be maintained even at the Nyquist rate.

Key Mathematical Results

The paper formalizes these findings through several landmark theorems:

  • Theorem 2: Proves that for any ordered complete interpolating sequence, there exists a stable LTI system and a signal where the digital approximation diverges pointwise.
  • Theorem 4: Establishes that stable approximation is mathematically guaranteed if one employs general measurement functionals and oversampling.
  • Theorem 6: Shows that without oversampling, even general measurement procedures will fail for a large class of practically relevant systems.
  • Theorem 7: Demonstrates a specific "universal" subsequence of measurements based on Walsh functions that converges for all signals in the space without requiring oversampling.

Conclusion and Future Directions

While this work provides a rigorous foundation for understanding the limits of digital approximation, it also opens new avenues for research. The study is currently limited to Linear Time-Invariant systems; non-linear systems, which are ubiquitous in modern signal processing, likely present even greater stability challenges. Additionally, while the mathematical necessity of general measurements has been proven, a corresponding "implementation theory" for hardware—accounting for energy consumption, transistor limits, and quantization—remains to be developed. Finally, the authors leave "Conjecture 3" open: whether the divergence seen in pointwise sampling also holds for every possible subsequence of those samples.

Summary updated

BibTeX
@inproceedings{boche2014signal,
  author = {Holger Boche and Ullrich J. Mönich},
  title = {{Signal and System Approximation from General Measurements}},
  booktitle = {Applied and Numerical Harmonic Analysis},
  year = {2014},
  doi = {10.1007/978-3-319-08801-3\_6},
}

Loading related content...

Figure