Adaptive Causal RLNC

• AC-RLNC is a feedback-driven network coding technique that adapts coding and retransmission strategies based on channel conditions and ACK/NACK feedback.
• It employs both a priori and a posteriori FEC mechanisms to proactively and reactively compensate for packet erasures, optimizing throughput and delay.
• Its layered extension, LAC-RLNC, integrates base and enhancement layers to deliver ultra-reliable low-latency streaming across single-path and multipath networks.

Adaptive Causal Random Linear Network Coding (AC-RLNC) is a family of feedback-driven network coding techniques for streaming data over packet erasure channels. AC-RLNC and its layered extension, LAC-RLNC, dynamically adapt coding and retransmission strategies in response to channel conditions and feedback, optimizing the throughput–delay–efficiency trade-offs under diverse network scenarios. These approaches have found applications in both single-path and multipath multi-hop networks, and enable efficient, ultra-reliable low-latency communications with provable performance bounds (Cohen et al., 2019, Cohen et al., 2019, Cohen et al., 2022, Waxman et al., 17 Feb 2025).

1. Fundamental Mechanisms of AC-RLNC

AC-RLNC operates over time-slotted packet erasure channels, where feedback about packet delivery (ACK/NACK) is returned after a fixed round-trip time (RTT). At each slot, the sender transmits an RLNC-coded packet, encoding a sliding window of source packets. Crucially, encoding decisions are causal: every coded packet depends only on information and feedback available up to $t - \mathrm{RTT}$.

Adaptivity in AC-RLNC is realized through two complementary forward error correction (FEC) mechanisms:

• A priori FEC: Following every batch of $n$ new transmissions, the sender injects $\lceil \hat{\epsilon}n \rceil$ additional RLNC-coded packets to proactively compensate for anticipated erasures, where $\hat{\epsilon}$ is the current estimated erasure probability.
• A posteriori (feedback-based) FEC: After each feedback event (ACK/NACK), the sender computes the "DoF rate gap"

$\Delta_t = r_{t-\mathrm{RTT}} - d_t - th,$

where $r_{t-\mathrm{RTT}}$ is the estimated channel rate, $d_t$ is the observed ratio of erased to added degrees of freedom for the current window, and $th$ is a tunable threshold. If $\Delta_t > 0$, indicating a lag in reliable delivery, the sender retransmits a previously sent degree of freedom (with fresh coefficients).

The combination of these mechanisms ensures that the code is both adaptive (to instantaneous channel conditions) and causal (never using future information or unresolved events in encoding), yielding zero-error packet delivery and bounded in-order delays under bursty and time-varying erasure patterns (Cohen et al., 2019, Cohen et al., 2019).

2. Layered Extension: LAC-RLNC for Streaming with Mixed Delay

LAC-RLNC generalizes AC-RLNC by embedding the latter in a two-layer broadcast architecture for streaming under mixed delay constraints (Cohen et al., 2022). Each data frame consists of $n$ coded packets encoding both a base layer ($k_1$ packets) and an enhancement layer ($k_2$ packets, $k_1+k_2 \le n$):

• The base layer is allocated sufficient redundancy (via FEC) to guarantee ultra-low delay delivery, meeting strict deadlines for service continuity.
• The enhancement layer contains additional data for quality augmentation, subject to looser delay constraints, enabling adaptation of the throughput–delay trade-off.

Layer sizes are chosen such that the receiver can decode the base layer after receiving any $k_1$ erasure-free packets from a frame, ensuring robustness up to a prescribed $\epsilon^{\max}$. The coding uses nested generator matrices $G_1$ and $G_2$:

$$[p_1^n] = [\ell_1^{k_1} \;\;\ell_2^{k_2}] \begin{bmatrix} G_1 \ G_2 \end{bmatrix}$$

where $\ell_1$ is the base layer and $\ell_2$ the enhancement layer content.

After transmitting a frame, additional FEC packets are sent as:

$$m_t = \bigl\lceil \epsilon^{\max}_{t-\mathrm{RTT}} n \bigr\rceil$$

to push the base layer over worst-case erasure events.

This approach yields dramatic improvements in delay for the base layer—experimentally, $3\times$ reductions (mean and max) versus non-layered AC-RLNC, approaching the lower bound $D_{\min} \sim \mathrm{RTT}/2$, while maintaining throughput close to $1-\epsilon^{\mathrm{mean}}$ (i.e., the channel capacity under zero packet loss) (Cohen et al., 2022).

3. Multi-Path and Multi-Hop AC-RLNC

The original AC-RLNC framework naturally extends to multipath and multi-hop networks (Cohen et al., 2019, Cohen et al., 2020, Waxman et al., 17 Feb 2025). Here, each path (or hop) is modeled as a BEC with individual erasure rates. The sender maintains separate coding windows for each path and assigns new vs. retransmission DoFs adaptively, leveraging feedback from all available links.

Multipath Scheduling and Bit-Filling

A bit-filling (discrete water-filling) algorithm is used to allocate paths between new degrees of freedom and retransmissions. The sender maximizes aggregate new-DoF throughput, subject to ensuring that the total retransmission rate over remaining paths meets the necessary FEC quota, dictated by the observed DoF gap (Cohen et al., 2019, Cohen et al., 2020). The optimization:

$$\hat{\mathcal{P}}_{\xi} = \arg\max_{\mathcal{P}_{\xi} \subset \mathcal{P}} \sum_{i \in \mathcal{P}_{\xi}} r_i \quad \text{subject to} \quad \sum_{j \notin \mathcal{P}_{\xi}} r_j \geq \Delta$$

assigns paths for new or repeat transmissions according to current channel state.

Decentralized Multi-Hop Balancing

In chains of $H$ hops, each node matches its best incoming and outgoing links ("natural matching") to maximize the end-to-end min-cut:

$$\eta_{\max} = \sum_{p=1}^P \min_{h=1..H} r_{G(p,h),h}$$

This decentralized balancing is shown to be globally optimal and enables per-path throughput guarantees without increased delay, provided that intermediate nodes implement "selective recoding" that preserves the DoF/FEC structure established end-to-end (Cohen et al., 2019, Cohen et al., 2020).

Selective Recoding

To preserve the feedback-driven FEC structure of AC-RLNC, intermediate nodes must distinguish between "new-DoF" and "FEC-repeat" packets:

• New-DoF recoding: mix only new degrees of freedom for further forwarding.
• FEC-repeat recoding: mix only repeat transmissions of the same DoF.

This restriction ensures that retransmission counts remain valid and that the window-based FEC remains effective, even after multiple hops (Cohen et al., 2019).

4. Advanced Scheduling: Blank-Space AC-RLNC and Efficiency

Blank-Space AC-RLNC (BS AC-RLNC) extends standard AC-RLNC to optimize the throughput–delay–efficiency trilemma in multi-hop streaming (Waxman et al., 17 Feb 2025). BS introduces lightweight scheduling algorithms at intermediate nodes:

• Blank Space Period (BSP): Nodes dynamically suspend transmissions for durations based on the bottleneck channel downstream, as determined by real-time RTT and erasure estimates. This prevents futile transmission when the bottleneck cannot accept new DoFs, reducing bandwidth waste.
• No-New No-FEC Rule: If neither new data nor necessary FEC is available, the node pauses transmission, exploiting local buffer emptiness for resource savings.

Experimental results show up to 20%–40% reduction in channel usage without sacrificing throughput or delay, confirming the benefit of exploiting bottleneck awareness and adaptive suspension rules in distributed coding (Waxman et al., 17 Feb 2025).

5. Performance Analysis and Bounds

For both single and multipath/multihop settings, AC-RLNC achieves provably near-capacity performance:

• Throughput: Upper and lower bounds based on the sender's rate estimation and Bhattacharyya distance between sender's channel estimate and the true channel realization:

$$\eta \leq \sum_{p=1}^P [r_p(t^-) - l(r_p(t), r_p(t^-))]$$

with the loss term $l(\cdot)$ vanishing for slowly varying channels.

• Mean and Maximum Delay: Closed-form (or closed-bound) expressions in terms of window size $\bar{o}$, RTT, and mean erasure probability $\bar{\epsilon}$; max delays scale at most linearly with $\bar{o}$ and can be made within a small multiplicative factor of the theoretical minimum.
• Empirical Validation: In BEC and Gilbert–Elliott channels, AC-RLNC delivers up to $2\times$ throughput and $3\times$ lower mean delay compared to SR-ARQ, maintaining $>90\%$ of channel capacity even in bursty or rapidly varying conditions (Cohen et al., 2019, Cohen et al., 2019).

6. Implementation: Sender and Network Algorithms

The sender-side algorithm maintains window state $(w_{\min}, w_{\max})$, monitors feedback history, updates rate and erasure estimates, and applies decision logic:

• If the feedback-based DoF gap warrants, send FEC/repeat DoF.
• Otherwise, inject new information by expanding the window.
• Upon frame completion or window overflow, shift windows based on cumulative ACKs.
• In LAC-RLNC, invoke layered coding, scheduling strict FEC for base layer and relaxed FEC for enhancement data (Cohen et al., 2022).

Intermediate nodes in multi-hop scenarios maintain their own erasure statistics and use the lightweight NET algorithm to apply the same local causal adaptation and blank-space scheduling (Waxman et al., 17 Feb 2025).

7. Extensions, Limitations, and Research Directions

• Layered and Heterogeneous Receivers: LAC-RLNC supports more than two layers and could target heterogeneous client delay/quality profiles (Cohen et al., 2022).
• General Network Topologies: SDN-based modular architectures distribute AC-RLNC primitives, supporting highly meshed, heterogeneous, multi-source/multi-destination deployments (Cohen et al., 2020).
• Security and Distributed Computation: AC-RLNC primitives can be integrated with cryptographic overlays, storage, and coded computation frameworks.
• Limitations: Analysis is largely empirical for LAC-RLNC; rigorous theoretical bounds under arbitrary erasure dynamics remain open. Bottleneck-awareness in BS assumes well-characterized RTTs; adapting to variable RTTs, noisy feedback, or general topologies remains an open challenge (Waxman et al., 17 Feb 2025, Cohen et al., 2022).
• Practical Impact: AC-RLNC and its extensions have demonstrated near-optimal delay and throughput with reduced resource consumption, and are of direct relevance for ultra-reliable low-latency communications and streaming in modern networked systems (Cohen et al., 2019, Cohen et al., 2019, Cohen et al., 2022, Waxman et al., 17 Feb 2025).