Weighted Unequal Error Protection over a Rayleigh Fading Channel

1. Problem Statement

The paper addresses the problem of Weighted Unequal Error Protection (UEP) in channel coding over a quasi-static Rayleigh fading channel.

• Context: In modern communication (e.g., semantic or task-oriented communications), different parts of a message have different levels of importance. The goal is not necessarily to decode every bit perfectly, but to maximize the recovery of the most critical information.
• System Model:
  • A message bit string is divided into $K$ blocks (layers/packets), $B^m_1, \dots, B^m_K$, with lengths $m = k/K$.
  • Each block $i$ has an associated importance weight $d_i$, where $d_1 > d_2 > \dots > d_K > 0$ and $\sum d_i = 1$.
  • The channel is a Single-Input Single-Output (SISO) quasi-static Rayleigh fading channel: $Y = HX + Z$.
  • CSI Availability: Channel State Information (CSI) is available only at the receiver. The transmitter knows the distribution of the fading coefficient $H$ but not the instantaneous realization.
• Objective: Design a channel code to maximize the weighted sum of per-block decoding success probabilities:
  $$\text{Maximize } \sum_{i=1}^K P(B^m_i = \hat{B}^m_i) d_i$$
  Unlike traditional UEP where success is often nested (decoding layer $i$ requires success of $1 \dots i-1$), this objective rewards the marginal success of each block independently.

2. Methodology

The paper analyzes two distinct achievability schemes to solve the optimization problem under finite blocklength ($n$) and asymptotic ($n \to \infty$) regimes.

A. Power-Domain Superposition (PDS)

• Mechanism: Uses Superposition Coding where all $K$ blocks are transmitted simultaneously over the same time/frequency resources. The total power $P$ is split among blocks using fractions $\alpha_i$ ($\sum \alpha_i = 1$).
• Decoding: The receiver employs Successive Interference Cancellation (SIC). It decodes block 1, subtracts it from the received signal, then decodes block 2, and so on.
• Analysis: The interference from undecoded lower layers is treated as Gaussian noise (due to i.i.d. Gaussian codebooks). The Signal-to-Interference-plus-Noise Ratio (SINR) for layer $i$ depends on the fading state and the power allocation of subsequent layers.
• Optimization: The problem is transformed into finding the optimal power split vector $\vec{\alpha}$. The authors derive an asymptotic limit and provide algorithms to compute the optimal $\vec{\alpha}$ using the Lambert W function.

B. Orthogonal Resource Allocation (ORA) / Time-Sharing

• Mechanism: The channel resources (time or frequency) are partitioned orthogonally. Block $i$ is transmitted using a fraction $w_i$ of the total blocklength $n$ (where $\sum w_i = 1$).
• Decoding: Maximum Likelihood (ML) decoding is performed independently for each orthogonal segment.
• Analysis: Since resources are orthogonal, there is no interference between blocks. The performance depends on the resource allocation vector $\vec{w}$.
• Optimization: Similar to PDS, the authors derive the asymptotic limit and provide algorithms to find the optimal resource split $\vec{w}$.

C. Finite Blocklength Analysis

• The paper utilizes Error Exponent bounds and Normal Approximation bounds (based on channel dispersion) to analyze performance for finite $n$.
• The objective function involves an expectation over the exponential distribution of the channel gain $\gamma = |H|^2$.
• The authors define a parameter $\theta = \frac{2^R - 1}{P\sigma^2}$, representing the ratio of the threshold SNR to the average channel SNR.

3. Key Contributions

1. Formulation of Weighted UEP: The paper rigorously formulates the problem of maximizing a weighted sum of marginal success probabilities over a fading channel with CSIR (Channel State Information at Receiver) only.
2. Asymptotic Characterization:
   • Derives the first-order asymptotic performance limits for both PDS and ORA.
   • Proves structural properties of the optimal solutions (e.g., the number of active layers $\ell$ is bounded, and importance weights dictate the allocation).
3. Algorithm Development:
   • Algorithm 1 & 2 (PDS): Provides a globally optimal algorithm (Algorithm 1) and a computationally simpler locally optimal algorithm (Algorithm 2) to compute the optimal power split. These rely on the Lambert W function branches ($W_0$ and $W_{-1}$).
   • Algorithm 3 & 4 (ORA): Provides corresponding algorithms for optimal time/resource splitting.
   • The paper proves that the "local" algorithms (using only the principal branch of the Lambert function) yield results identical to the global search in numerical tests.
4. Upper Bounds on Active Blocks: Theorems 5 and 11 provide explicit upper bounds on the optimal number of blocks ($\ell$) to transmit as a function of the importance weights, coding rate, and channel SNR.
5. Finite Blocklength Approximation: The authors propose heuristic methods to approximate finite blocklength performance using the asymptotic optimal solutions, showing these approximations are tight.

4. Key Results

• PDS vs. ORA Performance:
  • PDS outperforms ORA, but the performance gap is negligible (less than 2%) in both asymptotic and finite blocklength regimes.
  • This suggests that the simpler ORA (time-sharing) scheme is a highly efficient alternative to the more complex PDS (superposition) scheme for weighted UEP.
• Finite vs. Asymptotic Gap:
  • For a blocklength of $n=1000$, the finite blocklength performance is within approximately 10% of the asymptotic limit.
  • For $n=5000$, the gap reduces to approximately 3%.
  • The gap increases as channel conditions worsen (lower SNR).
• Impact of Block Count ($K$):
  • Increasing the number of blocks $K$ (finer granularity of importance) yields performance improvements, particularly in poor channel conditions. However, there are diminishing marginal returns as $K$ increases.
• Optimal Structure:
  • The optimal strategy often involves transmitting only a subset of the most important blocks (setting $\alpha_i=0$ or $w_i=0$ for less important layers) when the channel is poor or the rate is high.
  • The optimal power/resource allocation is strictly decreasing with respect to the importance weights ($d_i$).

5. Significance

• Practicality for Modular Networks: The work supports the "Multi-Level Reliability Interface" concept, where application providers (source) and network providers (channel) can interact without full joint coding. It demonstrates how to optimize channel coding given pre-assigned importance levels.
• Simplicity vs. Performance: The finding that ORA (time-sharing) performs nearly as well as PDS (superposition) is significant. PDS requires complex SIC decoding and precise power control, while ORA is structurally simpler and easier to implement in practical systems.
• Robustness to Fading: The analysis provides concrete guidelines for resource allocation in delay-limited, non-ergodic fading channels where the transmitter cannot adapt to instantaneous channel states.
• Algorithmic Efficiency: The derivation of simplified algorithms (Algorithm 2 and 4) that avoid exhaustive searches over the secondary branches of the Lambert function makes the solution computationally feasible for real-time or near-real-time optimization in communication systems.

In conclusion, the paper provides a comprehensive theoretical and numerical framework for optimizing unequal error protection in fading channels, demonstrating that simple orthogonal resource allocation is a near-optimal, low-complexity solution for weighted semantic communication tasks.