Expected Supremum of a Random Linear Combination of Shifted Kernels

Holger Boche, Brendan Farrell, Michel Ledoux, Moritz WiesearXiv 2012 · arXiv

Abstract

In signal processing and communications engineering, the Peak-to-Average Power Ratio (PAPR) is a critical metric influencing hardware efficiency and signal integrity. This paper analyzes the expected supremum of signals constructed from random linear combinations of shifted sinc kernels. Unlike standard multi-carrier systems where peaks grow at a rate of $\sqrt{n \log n}$, we demonstrate that sinc-based signals exhibit significantly slower growth. Specifically, for Gaussian weights, the expected peak grows as $\sqrt{\log n}$, while for bounded weights, it grows at the remarkably slow rate of $\log \log n$. These findings provide a theoretical foundation for the superior PAPR characteristics observed in single-carrier communication systems.

Introduction

The "sinc" kernel, defined as $\text{sinc}(t) = \frac{\sin \pi t}{\pi t}$, serves as the mathematical cornerstone of sampling theory. In modern communications, signals are often constructed by summing shifted versions of a fundamental kernel weighted by random coefficients. A persistent challenge in this domain is the Peak-to-Average Power Ratio (PAPR). High PAPR necessitates power amplifiers with significant "headroom" to prevent non-linear distortion, leading to increased power consumption and heat generation.

While traditional analysis focuses on orthonormal bases over fixed intervals (such as the Fourier basis in OFDM), this research explores the behavior of signals on the real line using the sinc kernel. The central question is: how does the maximum value (supremum) of such a signal scale as the number of components $n$ increases?

Mathematical Framework

The study investigates the expected supremum $E[S_n]$ of the random process: $$S_n = \sup_{t \in \mathbb{R}} \left| \sum_{k=1}^n a_k \text{sinc}(t-k) \right|$$ where $a_k$ are independent random variables.

To derive bounds for $S_n$, the authors employ several sophisticated tools from probability theory and harmonic analysis:

  1. Discretization: The continuous supremum over $t \in \mathbb{R}$ is bounded by analyzing the maximum over a sufficiently dense finite set of points.
  2. Concentration Inequalities: Hoeffding’s inequality is utilized to bound the probability of large deviations from the mean.
  3. Gaussian Process Theory: For Gaussian coefficients, Slepian’s and Sudakov’s comparison theorems are applied to relate the supremum of the sinc process to simpler Gaussian processes.
  4. Combinatorial Analysis: For bounded coefficients (e.g., Bernoulli $\pm 1$), a "law of large numbers" approach is applied to sub-intervals to establish the asymptotic growth rate.

Key Results and Insights

The research reveals a striking "dichotomy" in peak growth based on the distribution of the coefficients $a_k$.

1. Gaussian Coefficients

When the weights $a_k$ are Gaussian, the expected supremum grows as: $$E[S_n] \sim \sqrt{\log n}$$ This is already a significant improvement over the $\sqrt{n \log n}$ growth seen in standard orthonormal systems.

2. Bounded/Bernoulli Coefficients

If the weights are bounded (e.g., $a_k \in \{-1, 1\}$), the growth rate drops even further: $$E[S_n] \sim \log \log n$$ To illustrate the magnitude of this difference, consider a system with $n = 10^6$ components. In a standard OFDM system, the peak might scale with a factor of approximately 3,700. In contrast, a sinc-based system with bounded weights would see a peak scaling factor of approximately $\log(\log(10^6)) \approx 2.6$.

The Role of Kernel Decay

The slow growth of the supremum is intrinsically linked to the decay properties of the sinc kernel. Because the sinc kernel is not absolutely integrable (its tails decay as $1/t$), the peak grows with $n$. The authors note that if a faster-decaying kernel were used (such as a root-raised cosine kernel, which is integrable), the expected supremum would remain bounded by a constant, independent of $n$.

Discussion and Implications

These results provide a rigorous theoretical explanation for why single-carrier systems, which closely resemble the shifted-kernel model, typically exhibit much lower PAPR than multi-carrier systems like OFDM.

However, several practical considerations remain:

  • Implementation: The sinc kernel is non-causal and extends infinitely in time, making it impossible to implement perfectly in real-time hardware. Practical systems must use truncated or faster-decaying approximations.
  • Worst-Case Analysis: While the expected value is a vital metric for average performance, engineering specifications often require bounds on the "worst-case" peak or high-percentile behavior (the "tails" of the distribution).
  • Peak Suppression: Modern systems often use active peak reduction techniques, where coefficients are carefully chosen rather than purely random. Extending this theoretical framework to include such deterministic constraints is a promising area for future research.

Conclusion

By shifting the focus from fixed-interval orthonormal bases to shifted kernels on the real line, this work demonstrates that the growth of signal peaks is far more manageable than previously suggested by standard models. The discovery of the $\log \log n$ growth rate for bounded coefficients highlights the inherent efficiency of sinc-based signaling in managing power dynamics.

Summary updated

BibTeX
@misc{boche2012expected,
  author = {Holger Boche and Brendan Farrell and Michel Ledoux and Moritz Wiese},
  title = {{Expected Supremum of a Random Linear Combination of Shifted Kernels}},
  year = {2012},
  eprint = {1208.5281},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1208.5281},
}
Figure