Random Linear Streaming Codes (RLSCs)

• RLSCs are streaming-oriented erasure-control schemes that form random linear combinations of source packets under causal delay constraints to improve delivery reliability.
• They employ diverse designs—including generation-based, sliding-window, and multi-hop approaches—using full-rank decoding criteria to balance throughput and energy consumption.
• Empirical and analytical studies reveal trade-offs between coding overhead and decoding complexity, guiding optimal configurations for varying channel conditions.

Random Linear Streaming Codes (RLSCs) are streaming-oriented erasure-control schemes in which transmitted packets are random linear combinations of source data formed under causal delay constraints. In the literature represented here, the term spans several closely related models: generation-based application-layer streaming in which packets are partitioned into generations and served in round-robin order, sliding-window non-systematic and systematic encoders over symbol or packet erasure channels, large-field pipelined schemes for multi-hop relay networks, and sparse broadcast formulations linked to sparse random linear network coding. Across these settings, recovery is determined by full-rank or row-space conditions of accumulated random coding matrices, while performance is measured through delivery packet count, net throughput, energy consumption, or long-run slot error probability (Li et al., 2012, Huang et al., 2 Sep 2025, Huang et al., 17 Dec 2025, Brown et al., 2017).

1. Scope and canonical constructions

In the generation-based formulation, a file is chopped into $N$ fixed-size blocks (“packets”) of $B$ bits each and partitioned into $n$ disjoint subsets $G_1,\dots,G_n$ of equal size $g$, so $N=n\cdot g$. Each subset is a generation, and coding mixes packets only within a single generation. To form the $t$-th coded packet from generation $G_j$, the encoder draws a coefficient vector $c=[c_1,\dots,c_g]\in GF(q)^g$ with entries chosen i.i.d. uniformly in $GF(q)$, and computes the coded payload $B$0. The packet header carries enough information to reconstruct $B$1, either as an explicit header of $B$2 field elements or as a short PRNG seed with a deterministic generator. Generations are then served in cyclic order according to $B$3, so one coded packet per generation is emitted in each round, without feedback until the end (Li et al., 2012).

A later single-hop streaming formulation replaces generations with a sliding memory of length $B$4. In the non-systematic case, at each timeslot $B$5, $B$6 new source symbols $B$7 arrive and are stored with the previous $B$8 slots. The encoder multiplies the $B$9 symbols $n$0 by a random generator matrix $n$1, with all entries chosen i.i.d. uniformly over nonzero field elements, and transmits $n$2. The systematic variant keeps the same memory length $n$3, but partitions each transmitted packet into a systematic block $n$4 and a parity block consisting of random linear combinations of the $n$5 most recent source symbols (Huang et al., 2 Sep 2025).

In the multi-hop relay formulation, node $n$6 at time $n$7 stores all correctly received packets up to time $n$8 in memory $n$9, and transmits one packet of length $G_1,\dots,G_n$0 over $G_1,\dots,G_n$1 according to $G_1,\dots,G_n$2, where $G_1,\dots,G_n$3 is an $G_1,\dots,G_n$4 matrix whose non-zero entries are drawn uniformly at random from $G_1,\dots,G_n$5. The analysis assumes $G_1,\dots,G_n$6 and the GMDS property, namely that any full-matching submatrix of $G_1,\dots,G_n$7 is invertible (Huang et al., 17 Dec 2025).

These constructions share the same underlying principle—random linear mixing of recently available source data—but differ in how “recently available” is defined: by generation, by sliding window, or by per-node memory. This suggests that “RLSC” denotes a family of causally encoded random-linear streaming schemes rather than a single fixed encoder architecture.

2. Decoding criteria and computational structure

For generation-based round-robin transmission, the receiver maintains, for each generation $G_1,\dots,G_n$8, a buffer of received $G_1,\dots,G_n$9 pairs. Let $g$0 denote the collected coding vectors and $g$1 the corresponding payloads. Once $g$2, the decoder attempts Gaussian elimination on $g$3, $g$4. If $g$5, it recovers the original $g$6 blocks by solving $g$7, and after recovery it drops any further packets of that generation from the buffer. Storing up to $g$8 coding vectors requires $g$9 words of $N=n\cdot g$0 memory plus $N=n\cdot g$1 bits of payload buffer, while Gaussian elimination costs $N=n\cdot g$2 arithmetic operations and $N=n\cdot g$3 in the worst case $N=n\cdot g$4. Thus per-generation decoding has cubic field-operation complexity and $N=n\cdot g$5 storage (Li et al., 2012).

In sliding-window single-hop RLSCs, the receiver accumulates per-slot observation matrices $N=n\cdot g$6, obtained by restricting $N=n\cdot g$7 to the rows corresponding to received symbols. These are stacked into cumulative matrices $N=n\cdot g$8. A source symbol $N=n\cdot g$9 is decodable by time $t$0 if the corresponding unit vector lies in the row-space of $t$1. Under the large-$t$2 almost-MDS assumption, decoding succeeds if and only if enough linear equations arrive within the decoding deadline (Huang et al., 2 Sep 2025).

In multi-hop relay networks, if $t$3 denotes $t$4 with erased rows removed, then the received equations at the destination satisfy

$t$5

where $t$6 and $t$7 is the overall receiver matrix obtained from the products of the $t$8's and $t$9's across the $G_j$0 hops. A symbol $G_j$1 is $G_j$2-decodable at time $G_j$3 if the unit vector $G_j$4 lies in the row-space of $G_j$5. Equivalently, the destination must have collected at least $G_j$6 independent equations involving the first $G_j$7 source blocks by time $G_j$8 (Huang et al., 17 Dec 2025).

Across all three formulations, decoding is therefore an algebraic rank test on accumulated random linear equations. The principal implementation distinction is not the decoding rule itself, but the state variable that determines which source symbols can still participate in future equations.

3. Delivery count, throughput, and energy in generation-based RLSCs

For generation-based pure random linear coding over $G_j$9, let $c=[c_1,\dots,c_g]\in GF(q)^g$0 be the memoryless packet-erasure probability. After $c=[c_1,\dots,c_g]\in GF(q)^g$1 transmissions from one generation, the probability of successful decoding is

$c=[c_1,\dots,c_g]\in GF(q)^g$2

where the product is the probability that a random $c=[c_1,\dots,c_g]\in GF(q)^g$3 matrix over $c=[c_1,\dots,c_g]\in GF(q)^g$4 has rank $c=[c_1,\dots,c_g]\in GF(q)^g$5. When $c=[c_1,\dots,c_g]\in GF(q)^g$6 is large, this is approximated by the corresponding binomial tail. If $c=[c_1,\dots,c_g]\in GF(q)^g$7 is the number of generations and $c=[c_1,\dots,c_g]\in GF(q)^g$8 is the probability that all $c=[c_1,\dots,c_g]\in GF(q)^g$9 generations are decoded by time $GF(q)$0, the expected delivery packet count satisfies

$GF(q)$1

The net throughput is

$GF(q)$2

and, with physical-layer rate $GF(q)$3, the expected delivery time per generation is $GF(q)$4. On a battery-powered terminal, if energy is dominated by radio-on time, then $GF(q)$5 (Li et al., 2012).

The same framework permits direct comparison among pure RL, systematic RL, and MDS coding. Systematic RL sends the $GF(q)$6 original packets first and then random combinations; its per-generation success probability is

$GF(q)$7

MDS codes, such as $GF(q)$8 Reed–Solomon codes, guarantee recovery as soon as any $GF(q)$9 out of the $B$00 codewords are received, and their round-robin success probability can be written in the closed-form double-sum given as Eq. (5) in the paper. The resulting comparison is not purely asymptotic: the model tracks the interaction between erasures, generation size, round-robin service, and coding overhead, and thereby translates algebraic recovery behavior into throughput and energy terms (Li et al., 2012).

A recurring conclusion in this line of work is that coding overhead, delivery packet count, and energy are tightly coupled. In this model, improved rank accumulation reduces expected communication time, and reduced communication time lowers energy consumption when the radio dominates the power budget.

4. Information debt and exact error analysis in stochastic channels

For non-systematic RLSCs in a Gilbert–Elliott symbol erasure channel, the channel state $B$01 evolves as a hidden two-state Markov chain with transition matrix

$B$02

and stationary distribution $B$03. If $B$04 symbols are received at slot $B$05, the analysis is organized around the information debt $B$06, which counts the missing equations required before the decoder can clear all past unknowns. With $B$07,

$B$08

Zero-hitting times $B$09 and ceiling-hitting times $B$10 partition the process into renewal cycles. A renewal-reward argument yields the long-run slot-error probability

$B$11

where $B$12 counts undecodable slots in cycles with no ceiling hit, and $B$13 account for partial cycles that do hit $B$14. The paper then derives closed forms for the cycle-length distribution, the stationary start distribution at zero hits, and the expectation terms using debt-transition matrices $B$15, block matrices built from them, and sums of matrix powers (Huang et al., 2 Sep 2025).

The same paper analyzes systematic RLSCs in the packet erasure channel, where each packet is either received perfectly with probability $B$16 or completely erased with probability $B$17. Here the immediate availability of the systematic symbols forces a modified debt process with a time-varying ceiling $B$18 and reset time $B$19. The resulting error characterization is more subtle than in the non-systematic case. In each renewal cycle, only erasure slots can fail; if no $B$20 occurs, every $B$21 with $B$22 is an error, while if some $B$23 occurs, the last such $B$24 determines the failing interval $B$25. The paper also gives a counter-intuitive example: with $B$26, $B$27, and $B$28, an 8-slot erasure pattern exists in which NRLSC decodes all symbols but SRLSC fails on 3 of them, due to de-correlation of parity with earlier symbols once some systematic slots arrive. For $B$29 and coding rate $B$30, an exact expression is derived for $B$31 in terms of an integral involving a Toeplitz transition matrix and a Catalan-number expansion for the expected number of error slots (Huang et al., 2 Sep 2025).

This body of analysis replaces worst-case bounded-erasure reasoning with exact stochastic characterization under i.i.d. and Markov channels. It also shows that systematic transmission changes more than latency: it changes the combinatorics of which erasure patterns are recoverable by deadline.

5. Multi-hop relay networks and band-structured Markov analysis

Large-field RLSCs in multi-hop relay networks are analyzed through a notion of detained symbols and information debt generalized from point-to-point networks. If $B$32 denotes the total number of source symbols represented in the memory of node $B$33 at time $B$34, the detained symbols between nodes $B$35 and $B$36 are

$B$37

At the destination, $B$38 is the number of equations collected, and the destination information debt is $B$39. For the two-hop case, the detained-symbol process obeys coupled recurrences: $B$40

$B$41

$B$42

$B$43

$B$44

The times $B$45 at which $B$46 define rounds; in each round, one decodes $B$47 symbols from $B$48 equations (Huang et al., 17 Dec 2025).

The stochastic analysis models $B$49 as an infinite Markov chain, then truncates the state variables at large caps $B$50 and constructs a joint transition matrix with a “two-level band” structure. The basic building block is the matrix $B$51, whose deliver band resets the state to zero with probability $B$52, while its erasure band advances by $B$53 with probability $B$54. By nesting these blocks, the analysis forms $B$55 for $B$56, then $B$57 for $B$58, and finally embeds the $B$59 transitions into four sparse matrices $B$60, $B$61, $B$62, and $B$63, corresponding respectively to $B$64, $B$65 nonzero, nonzero $B$66 nonzero, and nonzero $B$67. The stationary distribution $B$68 at renewal instants then yields the inter-hit probabilities

$B$69

$B$70

From these matrices, the long-run slot error rate is expressed as a ratio of steady-state expectations over one round, and the denominator admits a closed-form information-balance identity with $B$71. The same nesting procedure extends verbatim to an arbitrary number of hops by enlarging the band-structured transition matrices (Huang et al., 17 Dec 2025).

This framework is significant because it converts a causal multi-hop streaming problem into a renewal-theoretic Markov analysis driven by detained-symbol random walks. The key conceptual move is that loss propagation across hops is represented not by explicit pathwise decoding states, but by the amount of source information still “in flight” between consecutive nodes.

6. Sparse broadcast variants, empirical results, and design limits

In point-to-multipoint broadcast, a related sparse formulation encodes a $B$72-packet source message $B$73, $B$74, into rateless coded packets

$B$75

with coding coefficients distributed as

$B$76

If $B$77 out of $B$78 transmissions are received, the receiver forms a $B$79 matrix $B$80, and decoding succeeds if $B$81 has full rank $B$82. The delivery probability is

$B$83

Because exact full-rank probabilities for sparse random matrices are intractable, the paper uses zero-row conditioning together with Stein–Chen Poisson approximation for minimal linear-dependence events. Writing $B$84, it obtains the approximation

$B$85

with $B$86 and $B$87 given by an inclusion–exclusion recursion. Monte Carlo results for $B$88, $B$89, and $B$90 show maximal absolute error below $B$91 for binary fields and mean-squared error below $B$92; for $B$93, the gap never exceeds $B$94, with MSE under $B$95 (Brown et al., 2017).

Experimental and simulation results in the other RLSC settings expose the central implementation trade-offs. In Nokia N8 smartphone experiments at $B$96, low-rate MDS codes such as $B$97 are best for small $B$98; systematic RL overtakes for intermediate $B$99; and pure RL has the lowest overhead in theory at large $n$00, but on a 1 GHz ARM11 phone its cubic decoding cost causes dropped packets once $n$01. Net throughput and energy track inversely, because maximizing throughput reduces time-on-air and hence energy. In the multi-hop setting, simulations with two-hop parameters $n$02, $n$03, $n$04, $n$05, Monte Carlo horizon $n$06, and state caps $n$07 verify the matrix analysis: the relative error falls below $n$08 for $n$09 and $n$10, and exceeds $n$11 accuracy if $n$12. In the typical regime $n$13, large-field RLSCs achieve up to an order-of-magnitude lower error rates than adversarial DF schemes because the relay forwards “everything it has seen” in a pipelined manner rather than waiting for full decoding. In the stochastic single-hop setting, simulations show $n$14 in slot-error rate for most parameters when $n$15, but the counter-example above demonstrates that systematic transmission is not uniformly dominant (Li et al., 2012, Huang et al., 17 Dec 2025, Huang et al., 2 Sep 2025).

Several design boundaries recur across these studies. Increasing generation size $n$16 or enlarging the effective memory window reduces erasure overhead, but raises decoding complexity and memory requirements. Larger fields reduce linear-dependence failures, but the large-$n$17 analyses explicitly omit rare finite-field penalties; one paper notes that future work can bound the $n$18 term. Sparsity lowers decoding complexity from $n$19 toward $n$20 or better, yet increases the probability of degenerate coefficient patterns and therefore the transmission overhead required for a target delivery probability. The open directions stated in the literature include ACK/NACK hybrids for systematic streaming, extensions to multi-state hidden Markov channels, adaptive selection of the systematic rate $n$21, and explicit low-complexity constructions such as sparse generator matrices with analyses comparable to the current random-ensemble results.