In communication systems, signals are rarely perfectly stationary. Instead, they often exhibit cyclostationarity—their statistical properties, such as mean and autocorrelation, vary periodically over time. This periodicity is a natural byproduct of human-engineered systems, arising from modulation schemes, carrier waves, and hardware clocks.
As Alon Kipnis notes in The Distortion Rate Function of Cyclostationary Gaussian Processes, while distortion-rate functions (DRFs) are well-understood for stationary processes, analyzing cyclostationary processes is historically difficult. Their time-varying statistics deviate from standard information-theoretic tools, requiring specialized approaches like polyphase decomposition to leverage their periodic nature.
However, a much deeper theoretical challenge emerges when these continuous-time signals are converted into discrete-time data via digital sampling.
The Breakdown of Information Stability
If a receiver samples a cyclostationary signal in perfect synchronization with the signal's internal period, the resulting discrete-time signal retains a clean, predictable periodic structure. But in real-world systems, sampling is frequently asynchronous. The receiver's sampling clock and the incoming signal's statistical period drift out of alignment.
This lack of synchronization transforms a wide-sense cyclostationary (WSCS) signal into a wide-sense almost cyclostationary (WSACS) process. From an information-theory perspective, this transformation causes a fundamental breakdown: the resulting process is non-information-stable.
Standard Shannon theory relies heavily on the Asymptotic Equipartition Property (AEP). The AEP guarantees that as block lengths grow toward infinity, the sample averages of information quantities (like log-probabilities) converge cleanly to their expected values (entropy). This convergence requires information stability.
As explained by Zikun Tan et al. in On the Rate-Distortion Function for Sampled Cyclostationary Gaussian Processes with Memory, asynchronous sampling prevents the direct application of the AEP. Because the statistical alignment shifts continuously across time, the information density does not converge to a single steady-state average. This bottleneck invalidates traditional formulas for both data compression (the rate-distortion function) and channel capacity.
The Mechanism: Bridging the Gap via Limit Superior
To determine the fundamental limits of non-information-stable processes, researchers must abandon the AEP and turn to the information-spectrum framework. Instead of looking at simple expectations, the information spectrum analyzes the asymptotic behavior of the entire probability distribution of the information density.
The core intuition for solving this problem lies in a clever approximation strategy: bridging the intractable asynchronous scenario using a sequence of tractable synchronous ones.
Asynchronous Sampling (WSACS) ──> Non-Information-Stable (AEP Fails)
▲
│ Limit Superior (lim sup)
│
Synchronous Approximations ──> Information-Stable Sequences
In On the Rate-Distortion Function of Sampled Cyclostationary Gaussian Processes, Emeka Abakasanga et al. demonstrate that the rate-distortion function (RDF) for an asynchronously sampled cyclostationary signal can be rigorously derived by taking the limit superior (lim sup) of the RDFs of a sequence of synchronously sampled approximations.
Why does this work? While the asynchronous process never perfectly repeats, it can be mathematically "sandwiched" by a sequence of periodic, synchronous processes with increasingly long periods. By calculating the rate-distortion bounds for these stable, synchronous approximations and taking the upper limit of that sequence, the authors pin down the exact compression limits of the unstable, asynchronous process in the low-distortion regime.
This same underlying philosophy applies to channel capacity. When a channel is corrupted by noise originating from an asynchronously sampled cyclostationary process, the noise is similarly non-information-stable. Nir Shlezinger et al. utilized the information-spectrum framework in The Capacity of Memoryless Channels With Sampled Cyclostationary Gaussian Noise to overcome this instability, proving that the lack of synchronization significantly impacts the maximum achievable data rate.
Accounting for Memory and Practical Delay
While memoryless models offer elegant mathematical tractability, physical communication signals possess memory—where past values influence future states. Incorporating memory into an already non-information-stable WSACS process compounds the analytical complexity.
In On the Rate-Distortion Function for Sampled Cyclostationary Gaussian Processes with Memory, the authors successfully extended the information-spectrum derivation to account for both memory and non-information-stability. By tracking how the spectral properties of the process evolve over infinite horizons, they established the true theoretical lower bound for lossy compression in these complex systems.
However, an infinite-horizon calculation is often uncomputable in practice. To make this framework useful for designing real-world hardware, Zikun Tan et al. introduced a realistic physical constraint in The Rate-Distortion Function for Sampled Cyclostationary Gaussian Processes with Memory and with Bounded Processing Delay: a bounded processing delay.
Enforcing a finite time window for data processing tames the infinite-horizon dependencies of the process's memory. This constraint simplifies the mathematical structure of the information spectrum, yielding a highly computable characterization of the rate-distortion function while preserving the physical realities of imperfectly synchronized digital systems.
Key Foundational Work
On the Rate-Distortion Function of Sampled Cyclostationary Gaussian Processes (2020) — contributed 8 concept elements to this direction.
The Capacity of Memoryless Channels With Sampled Cyclostationary Gaussian Noise (2020) — contributed 2 concept elements to this direction.
The Distortion Rate Function of Cyclostationary Gaussian Processes (2018) — contributed 2 concept elements to this direction.