IRSA Protocol: Efficient Graph-Based Random Access

• IRSA is an advanced random access protocol that uses irregular packet repetition and graph-based strategies to improve collision resolution and approach channel capacity in grant-free systems.
• It leverages density evolution and optimized degree distributions to balance singleton detection and interference cancellation, substantially enhancing throughput over classical methods.
• Extensions integrate physical layer enhancements, energy harvesting, and online learning to address practical challenges such as AoI, multiuser interference, and dynamic network conditions.

Irregular Repetition Slotted ALOHA (IRSA) is a class of uncoordinated random access protocols that generalize the original Slotted ALOHA paradigm by incorporating graph-based repetition strategies and iterative successive interference cancellation to approach the channel capacity of grant-free multiuser communications. IRSA strategically distributes packet replicas among time slots according to carefully designed, irregular degree distributions, enabling efficient collision resolution and substantially higher throughput compared to classical random access approaches.

1. Protocol Fundamentals and Graph-Theoretic Structure

IRSA operates in discrete frames, each partitioned into a fixed number of slots. Active users wishing to transmit in a frame generate one or more replicas of their packet—the exact number for each user is a random variable $d$ drawn according to a prescribed node-perspective degree distribution, encapsulated by the generating polynomial

$$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$

where $\Lambda_d$ is the probability of sending $d$ replicas and $d_{\max}$ is the maximum allowed repetition degree (Toni et al., 2018, Hansen et al., 2024). Each user places its $d$ replicas in distinct, randomly chosen slots within the frame.

The network operation can be represented by a bipartite graph (cf. [Liva '11]): one set of nodes corresponds to users (variable nodes), the other to time slots (check nodes). Edges exist between a user node and a slot node if a replica from the user is transmitted in that slot. Collisions—multiple packets in a slot—are mapped as higher-degree check nodes. The iterative reception process involves a two-sided message-passing (belief-propagation/AND-OR tree) akin to decoding LDPC codes on the erasure channel.

The main performance parameters are:

• Offered load: $$G = \frac{\text{number of users}}{\text{number of slots per frame}}$$
• Throughput: $T(G) = G (1-p_\infty)$, where $p_\infty$ is the asymptotic fraction of unresolved users after interference cancellation.

2. Density Evolution and Iterative SIC Decoding

Packet recovery leverages successive interference cancellation (SIC). Initial decoding identifies singleton slots (degree one check nodes); upon successful decoding of a single user's packet in such a slot, all replicas of that packet are subtracted (“peeled”) from their respective slots, potentially unmasking new singletons (Hansen et al., 2024, Ngo et al., 2024, Amat et al., 2018). This graph peeling procedure continues iteratively until no further singletons remain.

Performance and design rely on asymptotic density evolution analysis. For a large system, the evolution of the probability $p_i$ that an edge remains unresolved after $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$0 SIC iterations is governed by a recursion:

$$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$1

with fixed point

$$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$2

Here, $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$3. The throughput curve exhibits a distinctive waterfall effect: there exists a threshold $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$4 such that, for $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$5, almost all packets are recovered ($$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$6), while for $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$7, the PLR sharply increases (Nisioti et al., 2018, Dovelos et al., 2016).

For multi-packet reception (MPR)-capable receivers, the evolution is generalized (Fernández-Veiga et al., 2023, Paolini et al., 2022):

$$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$8

where $$\Lambda(x) = \sum_{d=1}^{d_{\max}} \Lambda_d \, x^d$$9 and $\Lambda_d$0 is the edge-perspective user-degree distribution.

3. Optimization of Degree Distributions and Prioritization

Designing the repetition degree distribution $\Lambda_d$1 is central to maximizing IRSA performance. High values of $\Lambda_d$2 (more replicas per user) facilitate early singleton formation and graph connectivity but increase slot occupancy (contention), which can be counterproductive at high load. Optimal distributions are found via density-evolution and potentially further optimized using EXIT chart techniques (Hansen et al., 2024, Tralli et al., 2023, Munari, 2020).

For heterogeneous traffic (priority classes), per-class degree distributions $\Lambda_d$3 can be optimized to maximize weighted system utility, subject to global stability constraints derived from AND-OR tree recursions. Heuristic "safe-region" algorithms enforce the necessary stability for prioritization, aligning collision-resilience with class-specific performance (Toni et al., 2015).

Game-theoretic approaches interpret each user as a selfish agent choosing repetition degree probabilistically to maximize its own utility (success reward minus repetition cost). The unique Nash equilibrium for small-user cases coincides with throughput-maximizing IRSA distributions (e.g., explicit closed forms for the two-user case). For large systems, best-response and evolutionary-learning approaches yield degree distributions achieving up to 34% throughput gains over classical framed ALOHA (Hansen et al., 2024).

4. Stochastic and Learning-Based Protocol Optimization

Traditional (offline) IRSA optimization presumes stationary conditions and static network knowledge. However, practical deployments face unknown and time-varying user populations, traffic, and propagation environments. Model-free, online learning frameworks adapt IRSA parameters in real time to changing conditions.

The online multi-armed bandit (MAB) reformulation considers each candidate (degree distribution $\Lambda_d$4, per-frame packet count $\Lambda_d$5) as an arm, with observed utility reward $\Lambda_d$6 per decision epoch. Bayesian-UCB learning leverages informative priors from asymptotic density evolution, ensuring rapid adaptation with logarithmic regret relative to the optimal strategy (Toni et al., 2018).

Decentralized environments (e.g., distributed sensor networks) motivate reinforcement learning (RL) approaches where nodes make replication decisions solely based on local observations (buffer state). Q-learning in a decentralized partially observable Markov decision process (Dec-POMDP) enables scalable, model-free adaptation, with convergence acceleration via virtual experience: batch Q-iteration over all histories that are equivalent under collision dynamics. This framework can double throughput under heavy loads and smooth the waterfall cliff otherwise observed in standard IRSA (Nisioti et al., 2018).

5. Extensions: Practical and Advanced Variants

IRSA's performance and formulation have been extended in multiple directions.

• Physical Layer Enhancements: Incorporation of intra-slot SIC and capture effect decoding (e.g., power-domain NOMA or satellite-visible diversity) enables the resolution of multiple packets per slot and operation above classical IRSA thresholds, driving the sustained throughput to $\Lambda_d$7 even for overloaded regimes ($\Lambda_d$8) (Interdonato et al., 2016, Recayte et al., 1 Jan 2026).
• Energy-Harvesting Devices and Freshness: For IoT deployments with energy-harvesting sources, IRSA protocol variants account for dropped transmission attempts due to energy shortage. The IDENTIFY receiver method reconstructs dropped-replica patterns to guarantee minimum age-of-information (AoI) and nearly matches unlimited-energy system freshness. Joint degree-distribution/battery-state adaptation further optimizes AoI (Ngo et al., 2024).
• Age of Information: The impact of IRSA frame size and repetition profile on AoI has been rigorously characterized via Markov renewal arguments yielding closed-form AoI and violation probability distributions. IRSA provides up to 50% reduction of average AoI compared to classical slotted ALOHA; age-threshold access with feedback and barring (AT-IRSA) cuts AoI by another 50% and outperforms alternative threshold ALOHA schemes (Munari, 2020, Asgari et al., 2022).
• Finite-Length Analysis: For moderate frame sizes, matrix occupancy combinatorics and finite-length scaling approximations yield precise error probability and PLR estimates, facilitating practical code/design trade-offs (Amat et al., 2018, Dovelos et al., 2016).
• Channel Impairments and Estimation: Realistic uplink deployments with fading, noise, non-orthogonal pilots, and multi-cell interference require revisiting IRSA performance. Pilot contamination, imperfect channel estimation, and cross-cell interference can reduce throughput by up to 70% relative to the ideal single-cell scenario, necessitating longer pilots, more antennas, or reduced repetition degrees for robust operation (Srivatsa et al., 2022, Srivatsa et al., 2021).
• Information-Theoretic Regime: Embedding IRSA in the context of the binary adder channel (BAC) or Gaussian MAC, using random coding and multi-packet reception techniques, establishes rigorous achievability and converse bounds for sum-rate, generalizing the utility of IRSA to unsourced, grant-free random access with tight performance guarantees (Ngo et al., 2023, Paolini et al., 2022, Tralli et al., 2023).

6. Performance Benchmarks, Trade-Offs, and Design Guidelines

In the collision channel model with ideal SIC, optimized IRSA achieves asymptotic throughput thresholds up to $\Lambda_d$9 packets/slot, versus $d$0 for classical slotted ALOHA (Interdonato et al., 2016, Rivas et al., 2023). Properly tailored degree distributions—mixing low-degree (2–3) and occasional high-degree (e.g., 8) users—push these thresholds higher, balance singleton abundance and graph connectivity, and suppress stopping sets (error floor). For MPR, high thresholds scale with MPR order ($d$1), and degree optimization must consider the slot error floor and system efficiency.

Substantial improvements over classical random access have been empirically validated in macro simulations and hardware testbeds for IoT and grant-free mMTC. However, actual system performance is tightly constrained by channel estimation, multiuser detection, pilot contamination, and finite-memory/delay effects, which can substantially degrade the theoretical thresholds if not addressed (Srivatsa et al., 2022, Srivatsa et al., 2021, Fernández-Veiga et al., 2023).

Energy efficiency, application-specific AoI requirements, and fairness (e.g., via user prioritization and game-theoretic access) introduce additional multi-objective design axes. Emerging work emphasizes online, data-driven adaptation (Bayes-UCB, RL, game-theoretic best-reply learning), robust degree-profile optimization, and physical-layer collaboration (capture/SIC, multi-satellite reception, coded slot-Aloha) as necessary pillars for sustaining high throughput in future massive access networks.

7. Research Directions and Broader Impact

Ongoing IRSA research addresses several prominent topics:

• Extension to coded random access on fading and noisy channels using information-theoretic (random coding with BAC/Gaussian MAC) or practical (LDPC-, BCH-based slot coding) approaches (Ngo et al., 2023, Tralli et al., 2023, Paolini et al., 2022)
• Adaptive schemes integrating energy harvesting, age-fairness, and non-stationary load (Ngo et al., 2024, Asgari et al., 2022)
• Ultra-reliable and low-latency access for critical IoT and mission-critical services
• Cross-layer designs combining MAC graph-based coding and PHY multiuser/capture detection
• Realistic multi-cell, multi-antenna, and non-orthogonal pilot settings for dense mMTC

IRSA and its derivatives, supported by a range of analytical, simulation, and learning-based studies, are established as cornerstones for high-throughput, grant-free access in the emerging landscape of IoT, 5G/6G, and beyond (Toni et al., 2018, Hansen et al., 2024, Ngo et al., 2024, Nisioti et al., 2018).