Abstract
In wideband communications, sampling at the Nyquist rate is often hardware-prohibitive. While sparse signals can theoretically be sampled at the lower Landau rate, practical implementations usually require knowledge of which frequency bands are active to avoid destructive aliasing. This paper quantifies the "blindness penalty"—the information rate loss incurred when sampling a multiband signal below the Nyquist rate without knowing the occupancy pattern. Using a minimax framework and non-asymptotic random matrix theory, the authors prove that a randomized sampling strategy using random periodic modulation is minimax-optimal, effectively minimizing the worst-case capacity loss across all possible signal states.
Introduction
The rapid expansion of wideband communication systems has pushed the limits of analog-to-digital conversion. Sampling at the Nyquist rate for GHz-wide signals is often impossible or excessively power-hungry. Fortunately, many wideband signals are "sparse" in the frequency domain, meaning only a small fraction of the total bandwidth is occupied at any given time. According to Landau’s theorem, such signals can be perfectly reconstructed if sampled at a rate proportional to their actual occupancy rather than their total bandwidth.
However, a significant practical hurdle remains: most sub-Nyquist sampling architectures are state-dependent. They require the sampler to know exactly which subbands are active to configure the sampling kernels correctly. In dynamic environments where channel occupancy changes rapidly, this "blindness" to the signal state leads to aliasing and information loss. This paper investigates universal sampling: the design of a fixed sampling strategy that performs well regardless of the specific subband occupancy pattern.
System Model and the Blindness Penalty
The authors model the communication system as a multiband linear time-invariant (LTI) Gaussian channel divided into $n$ subbands, with $k$ bands active at any time. The sampling architecture consists of a bank of periodic modulators followed by low-pass filters.
The core problem is to quantify the capacity gap between an ideal, occupancy-aware sampler and a universal sampler. This gap is defined as the information rate loss. Because the occupancy state $s$ is unknown to the sampler, the goal is to design a sampling matrix $Q$ that minimizes this loss across the worst-case scenarios.
The Minimax Framework
To solve the universal design problem, the authors employ a minimax framework. Instead of maximizing absolute capacity for a specific state, they minimize the maximum "capacity loss"—the difference between the capacity achieved by an ideal sampler (with full state knowledge) and the proposed universal sampler.
The mathematical analysis utilizes non-asymptotic random matrix theory. Specifically, the authors analyze the concentration of log-determinant functions of large random matrices to prove that certain sampling structures behave predictably and optimally as the number of subbands $n$ grows large.
Key Insights: The Power of Randomness
The most striking finding is that "Gaussian sampling"—where the coefficients of the periodic modulation are chosen randomly—is nearly perfect. Because random matrices spread signal energy evenly across the sampling branches, they ensure that no specific occupancy pattern is significantly more disadvantaged than others. This property, rooted in the "sharp concentration of spectral measures," ensures that in large systems, the information loss is almost identical for every possible set of active subbands.
Furthermore, the authors characterize the minimax capacity loss primarily through two factors: the sparsity factor ($\beta = k/n$) and the undersampling factor ($\alpha = m/n$). The loss is essentially a function of the binary entropy of these ratios. Interestingly, as the Signal-to-Noise Ratio (SNR) increases, the ratio of the capacity loss to the total capacity vanishes. This implies that at high SNR, being "blind" to the channel occupancy costs very little in relative terms.
Quantitative Results and Limits
The paper establishes both a fundamental lower limit on the minimax information rate loss and an achievability result. They prove that Gaussian random sampling approaches this lower bound with exponentially high probability as $n$ increases. For the regime where the sampling rate is above the Landau rate ($\alpha \geq \beta$), the minimax loss per unit bandwidth is approximately: $$L \approx \frac{1}{2}H(\beta) - \frac{1}{2}\alpha H\left(\frac{\beta}{\alpha}\right)$$ where $H(\cdot)$ is the binary entropy function. This formula is consistent with classical sampling theory; as the sampling rate approaches the Nyquist rate ($\alpha \to 1$), the information loss drops to zero.
Conclusion and Future Work
This work provides a rigorous information-theoretic foundation for universal sub-Nyquist sampling. While the current derivations assume a frequency-flat channel, the framework can be extended to frequency-selective channels. Future research directions include applying this minimax universal sampling framework to other signal structures (such as wavelets) and multi-user scenarios like Multiple Access Channels (MAC).
Summary updated