Forward Frame-Erasure Correction

• Forward frame-erasure correction is a technique that applies FEC to data units like frames, packets, or segments to repair losses without round-trip retransmissions.
• The method employs various code families such as Reed–Solomon, LDPC-Band, and convolutional codes, balancing decoder complexity, latency, and burst loss recovery.
• Adaptive protection and cross-layer designs enable dynamic adjustment of redundancy, improving system performance in applications like video streaming, deep-space, and wireless communications.

Forward frame-erasure correction denotes the use of forward error correction at the granularity of frames, packets, fragments, or segments so that losses can be repaired without costly round-trip exchanges and without waiting for retransmissions. In the cited literature, the protected unit ranges from RTP video frames and QUIC packets to LoRa fragments, DTN code matrices, and streaming segments, and the protected channel ranges from bursty Internet paths to deep-space links and wireless slices (Díaz et al., 2024, Michel et al., 2019, Singh et al., 14 Aug 2025, Chen et al., 19 Jun 2025, Vasudevan et al., 4 Jan 2026). The subject is therefore a design space rather than a single code family: it includes block codes, convolutional and random linear codes, unequal loss protection, hybrid iterative–maximum-likelihood decoding, and cross-layer adaptation.

1. Formal model and coding foundations

Forward frame-erasure correction is usually formulated over an erasure channel in which loss locations are known but payload contents are absent. In block-code form, a codeword has length $n$, dimension $k$, and rate $r = k/n$. For erasure-only Reed–Solomon decoding, the minimum distance is

$d_{\min} = n-k+1,$

and up to $e \le n-k$ erasures can be corrected (Sen et al., 2016).

Several systems in the literature express frame-erasure correction directly in terms of recoverable fragments. In LoRa-based federated learning, a compressed update of size $U$ bytes is divided into

$k = \left\lceil U/b \right\rceil$

source fragments of $b$ bytes each, then encoded into $n = k/r$ coded fragments using an MDS code; any set of $k$ coded fragments suffices for decoding, and an update is considered successfully delivered if at least $k$0 coded fragments are received (Singh et al., 14 Aug 2025). This fragment model is a direct frame-erasure abstraction: the receiver does not need bit-level error locations, only enough surviving coded units.

A more algebraic construction appears in LDPC-Band codes. The generator matrix is systematic,

$k$1

where $k$2 is the $k$3 identity matrix and $k$4 is a $k$5 banded binary matrix. The associated parity-check matrix is

$k$6

where $k$7 is sparse and $k$8 is a lower-triangular Toeplitz invertible binary matrix. From $k$9, the paper derives

$r = k/n$0

This construction jointly exposes two structures: sparsity for iterative decoding and bandedness for faster maximum-likelihood decoding on erasure channels (0901.3467).

2. Code families and decoding architectures

The literature spans simple parity schemes, algebraic block codes, sliding-window codes, and hybrid constructions. Their differences are not merely taxonomic; they determine recovery latency, decoder complexity, and robustness to burst losses.

Scheme	Structure	Stated property
XOR	Block parity	Repairs a single loss per block
Reed–Solomon	Systematic block code	Recovers up to $r = k/n$1 lost packets in a block
Convolutional RLC	Sliding window	Lower recovery latency
LDPC-Band	Sparse $r = k/n$2, banded $r = k/n$3	Iterative $r = k/n$4, ML $r = k/n$5

In QUIC implementations, XOR is described as simple and fast but insufficient when losses are bursty. Reed–Solomon is characterized as handling burst losses well at the cost of higher computation. Convolutional Random Linear Code interleaves redundancy with data and is more suited for streaming or steady flows because the time to recover from a lost packet is reduced compared to block codes (Michel et al., 2019, Michel et al., 2018).

LDPC-Band codes were designed so as to optimize a hybrid iterative-Maximum Likelihood decoding. Their sparse parity-check matrix enables belief propagation or message-passing iterative LDPC decoding with complexity $r = k/n$6, while the banded generator matrix enables optimized LU decomposition with ML decoding complexity $r = k/n$7. The paper states that $r = k/n$8 is typically chosen close to $r = k/n$9, and reports that when ML decoding is used, LDPC-Band is $d_{\min} = n-k+1,$0–$d_{\min} = n-k+1,$1 times faster than windowed or regular LDPC codes while preserving a very good trade-off between complexity and erasure correction capability (0901.3467).

Hardware realizations make the same trade-off explicit. An FPGA erasure-only RS decoder over $d_{\min} = n-k+1,$2 is built from parametric numbers of multiplier cores. In the reported Zynq7020 example for RS(200,136) correcting $d_{\min} = n-k+1,$3 erasures, $d_{\min} = n-k+1,$4 requires $d_{\min} = n-k+1,$5 LUTs and $d_{\min} = n-k+1,$6 FFs for $d_{\min} = n-k+1,$7, whereas $d_{\min} = n-k+1,$8 requires $d_{\min} = n-k+1,$9 LUTs and $e \le n-k$0 FFs for $e \le n-k$1. The paper further states that, for a fixed resource, throughput decreases with the square of the number of erasures (Sen et al., 2016).

3. Adaptive protection and cross-layer control

A central development in forward frame-erasure correction is the shift from static redundancy to application-aware and channel-aware adaptation. This is especially visible in video, transport, and delayed-feedback environments.

For RTP video streaming, the VA-ULP system performs frame-level unequal loss protection. Frame relevance is ranked by frame type, frame size, and distance to GOP end; channel losses are modeled with the simplified Gilbert–Elliot model; and a fixed bitrate budget is distributed across Decision Frame Sets, with I-DFS prioritized and some budget from PB-DFS reserved for forthcoming I-DFSs. Protection decisions minimize

$e \le n-k$2

where expected distortion combines frame importance and the probability that a frame cannot be recovered. Protection uses the 1-D interleaved parity FEC scheme based on XOR operations. Simulation results show higher average PSNR than Uniform Protection and Marx & Farah, especially at lower redundancy rates, while prioritizing and recovering significantly more I and P frames (Díaz et al., 2024).

In QUIC, FlEC exposes two principal hooks to the application, $e \le n-k$3 and $e \le n-k$4, so that reliability can range from pure retransmissions to various forms of FEC. The framework supports Random Linear Codes as the primary implementation, conceptually supports Reed-Solomon and XOR codes, and adapts to bulk transfer, buffer-limited transfer, and delay-constrained messages. For the bulk case, the representative condition is

$e \le n-k$5

For delay-constrained messaging, the evaluation reports that $e \le n-k$6–$e \le n-k$7 of video frames are delivered within deadline under loss and high RTT, whereas standard QUIC quickly falls below this threshold as RTT increases (Michel et al., 2022).

AC-RLNC addresses delayed feedback more explicitly. It learns the erasure pattern of the channel via feedback acknowledgments, inserts FEC both a priori and a posteriori, and transmits new information when the estimated channel rate $e \le n-k$8 exceeds the required DoF rate $e \le n-k$9 by a threshold $U$0: $U$1 The paper proves upper bounds on throughput and mean and maximum in-order delivery delay, and states that, in the non-asymptotic regime, the proposed code may achieve more than $U$2 of the channel capacity. On experimentally obtained commercial traces, AC-RLNC can double the throughput gains and triple the gain in terms of mean in-order delivery delay vis-à-vis SR-ARQ when the channel is bursty (Cohen et al., 2019).

Deep-space DTN introduces a different adaptive regime because feedback is enormously delayed. An RL-based adaptive FEC algorithm for LTP chooses the code rate $U$3 from $U$4 using a Dueling-DQN state comprising estimated packet loss rate, last-matrix decoding status, and persistence of the current channel state. In simulated Earth–Moon and Earth–Mars links, matrix decoding failures are reduced by at least $U$5, and the reported Earth–Mars case shows goodput $U$6 Mbps, delay $U$7 s, and $U$8 decoding failures for the RL-based method, compared with $U$9 Mbps, $k = \left\lceil U/b \right\rceil$0 s, and $k = \left\lceil U/b \right\rceil$1 failures for feedback-based adaptation (Chen et al., 19 Jun 2025).

A more recent streaming formulation is TAROT, which selects redundancy, block size, and symbolization on a per-segment basis. Its multi-objective score penalizes insufficient coverage, excessive overhead, and slow block completion, and it supports Reed-Solomon, RaptorQ, and XOR-based codes. Across Low-Latency Live and Video-on-Demand modes, diverse network traces, and multiple ABR algorithms, TAROT reduces FEC overhead by up to $k = \left\lceil U/b \right\rceil$2 while improving perceptual quality by $k = \left\lceil U/b \right\rceil$3 VMAF units with minimal rebuffering (Sidhu et al., 10 Feb 2026).

4. Protocol embedding and system architecture

Forward frame-erasure correction is often constrained less by coding theory than by where the coding layer sits and what metadata it can access.

In RTP video streaming, the protected unit is explicitly the frame. Video frames are each wrapped into separate, contiguous RTP packets using a custom RTP header extension; each RTP packet contains only data from a single video frame; and additional RTP header information indicates frame type, frame boundaries, and GOP distance. Protection is applied at the frame level: all packets for a frame are protected or not as a unit. The design allows the protection module to operate using only RTP headers, eliminating deep inspection of video bitstreams and enabling real-time operation with very modest overhead (Díaz et al., 2024).

QUIC-based systems expose a packet-level analogue. QUIC-FEC introduces a generic FEC frame carrying repair symbols and scheme-specific identifiers, supports XOR, Reed-Solomon, and Convolutional RLC, and protects only packets carrying user data after the handshake. Because FEC-recovered packets would otherwise mask congestion losses, the design adds a Recovered frame so that the sender can remove the packet from the retransmission queue while still triggering a congestion window decrease as if the packet was lost (Michel et al., 2019). A related design for QUIC also evaluates multipath scheduling; in single-path scenarios, RLC delivers more data than the two other schemes with short loss bursts, whereas Reed-Solomon outperforms RLC when the bursts are longer (Michel et al., 2018).

In LoRa-based federated learning, FEC is inserted after update sparsification, quantization, compression, and fragmentation. The simulator assumes an MDS code, and the duty-cycle penalty is explicit: if $k = \left\lceil U/b \right\rceil$4 is airtime per frame, the total airtime required to transmit the update is $k = \left\lceil U/b \right\rceil$5, and the next round must wait at least

$k = \left\lceil U/b \right\rceil$6

to remain duty-cycle compliant (Singh et al., 14 Aug 2025).

Wireless standard proposals relocate the coding layer altogether. One 2026 study inserts a Random Linear Network Coding layer between PDCP and RLC and transmits generations of $k = \left\lceil U/b \right\rceil$7 packets with extra coded packets, replacing repeated HARQ and ARQ rounds by forward erasure correction with block acknowledgment only when additional coded transmissions are needed (Vasudevan et al., 4 Jan 2026).

5. Performance metrics and evaluation practice

Because frame-erasure correction operates over finite codewords or finite protected units, frame-level metrics are often more informative than raw bit error counts. A survey of optical communication systems with FEC emphasizes frame-based operation, treats FER and post-FEC BER as the central outcome measures, and argues that thresholds such as pre-FEC BER, GMI, NGMI, and MI are only predictors whose reliability depends on code universality. The recommended practical method is to include actual encoding and decoding in experiments: collect transmitted and received symbols, regenerate bit streams and LLRs, and feed them into the real FEC decoder so that post-FEC BER and FER are measured directly rather than inferred (Schmalen, 2022).

Application-specific studies use more operational metrics. For VA-ULP, the reported metrics are packet loss recovery rate by frame type and average PSNR at redundancy rates from $k = \left\lceil U/b \right\rceil$8 to $k = \left\lceil U/b \right\rceil$9, and the paper concludes that all approaches deliver similar overall data recovery but that VA-ULP improves quality by protecting the most impactful frames (Díaz et al., 2024). QUIC studies use download completion time and DCT ratio, while adaptive and coding-aware transport studies report goodput, file delivery delay, mean and maximum in-order delivery delay, and timely delivery ratio (Michel et al., 2019, Michel et al., 2022, Cohen et al., 2019, Chen et al., 19 Jun 2025).

In systems where the communication substrate affects a learning loop, convergence itself becomes a reliability metric. The LoRa FL simulator reports accuracy per round, completion time, and device airtime. The study states that, without FEC, performance is extremely poor; with FEC, for any given spreading factor, performance is drastically improved; and at round $b$0 with SF$b$1, completion time with $b$2 is about $b$3 higher than with $b$4, rather than a naive doubling, because updates become more compressible as training progresses (Singh et al., 14 Aug 2025).

6. Limitations, alternatives, and emerging directions

A recurring misconception is that more FEC is always better. QUIC evaluations show the opposite regime as well: for long data transfers or when the loss rate and delay are small, the FEC overhead negatively impacts the download completion time, whereas high packet loss rates and long delays or smaller files are the operating points where FEC drastically reduces completion time by avoiding costly retransmission timeouts (Michel et al., 2019). Video-streaming studies make the same point at the application layer: static, coarse-grained configurations create unnecessary redundancy during stable periods and insufficient protection during bursty losses, especially when shallow buffers and oversized blocks increase stall risk (Sidhu et al., 10 Feb 2026).

Video conferencing exposes a second limitation: irrecoverable frame loss may trigger long dependency chains. One study states that, if too little redundancy is provided, the receiver cannot decode frames, leading to extended freezes and quality degradation, while too much redundancy wastes bandwidth (Li et al., 2023). The same paper proposes a generative alternative, Reparo, which does not use redundancy or retransmission and empirically outperforms state-of-the-art FEC-based video conferencing solutions in PSNR, SSIM, LPIPS, and the occurrence of video freezes; it fails to render only $b$5, $b$6, and $b$7 of frames under low, medium, and high loss rates, respectively, compared with $b$8, $b$9, and $n = k/r$0 for VP9+Tambur (Li et al., 2023). This does not invalidate FEC, but it shows that frame-erasure recovery can also be pursued by reconstruction rather than correction.

A plausible implication is that future forward frame-erasure correction will be increasingly cross-layer, adaptive, and modular. LoRa FL results explicitly state that frame-level FEC is indispensable and that spreading factor and FEC rate should be chosen jointly (Singh et al., 14 Aug 2025). Wireless standardization work argues for greater modularity in protocol stack design so that coding-enabled slices can coexist with HARQ/ARQ-based slices and novel reliability mechanisms can be integrated above existing layers (Vasudevan et al., 4 Jan 2026). Across these works, the dominant direction is not the replacement of FEC, but the replacement of static FEC by mechanisms that adapt redundancy, block geometry, symbolization, or repair scheduling to the protected application unit and the current network state.