Edge Arrival Model in Graph Algorithms

• Edge Arrival Model is defined as a sequential framework where edges, not vertices, are the primary units, enabling detailed online optimization and probabilistic analysis.
• It underpins various applications from adversarial graph streams to KIID stochastic matching, demonstrating rigorous space complexity and competitive ratio results.
• The model also drives advances in graph diffusion and generation, using reverse edge-removal techniques to faithfully reconstruct and simulate sparse network behavior.

In graph algorithms, online optimization, and statistical network modeling, an edge arrival model is a sequential formalism in which edges are the primary revealed, sampled, or generated objects. In the adversarial streaming formulation, the input is an edge stream $S=\langle e_1,e_2,\dots,e_m\rangle$ over a pre-determined vertex set $V$, with edges arriving one by one in arbitrary order (Cormode et al., 2018). In online stochastic matching, the arrivals are edge types sampled from a known distribution on a type-graph, including the KIID setting in which arrivals are known, independently and identically distributed (Feng et al., 15 Jun 2026). In statistical network modeling, the corresponding symmetry principle is often edge exchangeability, defined by invariance under permutations of edge labels rather than vertex labels (Crane et al., 2016). More recently, graph diffusion models have used “edge arrival” in a constructive sense, by generating graphs through reverse edge removal on sparse active-node subsets (Wu et al., 2023).

1. Canonical meanings and formal variants

The literature does not use a single universal formal definition. Instead, several closely related meanings recur, distinguished by what is random, what is revealed online, and whether the objective is optimization, inference, or generation.

Usage	Arriving object	Representative formalization
Edge stream	individual graph edges	$S=\langle e_1,\dots,e_m\rangle$ over fixed $V$
Online edge matching	feasible matching opportunities	immediate accept/reject, possibly with preemption
KIID edge arrivals	IID edge types from a known type-graph	$\Pr[e_k=e]=m_e/m$
Edge exchangeable network	interaction events indexed by edges	$\mathcal Y^\sigma \overset d= \mathcal Y$
Reverse edge-arrival generation	edges reconstructed over denoising steps	reverse edge-removal with $p=0$

In the graph-stream setting, the edge-arrival model is explicitly defined as a sequence of edges arriving one by one in arbitrary order, with the full edge set given by $E=\{e_i:i\in[m]\}$ (Cormode et al., 2018). In online weighted matching, the formulation is similarly sequential but decision-theoretic: when an edge $e=(u,v)$ arrives, its realized weight is revealed and the algorithm must make an immediate and irrevocable decision, subject to maintaining a matching (Ezra et al., 2020). In KIID stochastic matching, all vertices are known in advance, the type-graph need not be bipartite, and the online objects are edges sampled independently from a common edge-type distribution (Feng et al., 15 Jun 2026).

A different but closely connected formulation arises in edge-exchangeable probability models. There, the primitive object is an interaction process, and exchangeability is imposed on edge labels or interaction indices rather than on vertex labels. A random edge-labeled network $\mathcal Y\in\mathcal E_S$ is edge exchangeable if $V$0 for every permutation $V$1 (Crane et al., 2016). This makes the order of arrival statistically irrelevant, while still permitting a sequential interpretation via mixtures of i.i.d. edge-events.

2. Adversarial edge streams and streaming complexity

In streaming complexity, edge arrival is most often the adversarial insertion-only model. The strongest known separations from vertex-arrival appear precisely here. For independent sets, a one-pass edge-arrival stream is substantially harder than either explicit or implicit vertex-arrival. The sharp negative result is that every randomized constant error one-pass streaming algorithm in the edge arrival model that computes a maximal independent set requires $V$2 space; the paper interprets this as saying that one must essentially store all edges (Cormode et al., 2018). The same work defines approximate maximality by

$V$3

and proves that even this relaxation remains hard: every randomized constant error one-pass streaming algorithm that computes a $V$4-maximal independent set requires space $V$5, while every such algorithm computing a $V$6-maximal independent set requires space $V$7 for every $V$8 (Cormode et al., 2018).

For maximum matching, the edge-arrival model is likewise provably harder than classical one-sided vertex arrival in the semi-streaming regime. Any single-pass streaming algorithm that finds a $V$9-approximate matching in an $S=\langle e_1,\dots,e_m\rangle$0-vertex bipartite graph, for any constant $S=\langle e_1,\dots,e_m\rangle$1, with probability at least $S=\langle e_1,\dots,e_m\rangle$2, must use $S=\langle e_1,\dots,e_m\rangle$3 bits of space (Kapralov, 2021). A plausible implication is that the edge-arrival model should be treated as a distinct computational regime rather than as a minor perturbation of vertex-arrival assumptions.

Arrival order also changes what is achievable in semi-streaming edge colouring. In the W-streaming model, uniformly random edge order permits a one-pass algorithm using $S=\langle e_1,\dots,e_m\rangle$4 space and $S=\langle e_1,\dots,e_m\rangle$5 colours with high probability when $S=\langle e_1,\dots,e_m\rangle$6, yielding a $S=\langle e_1,\dots,e_m\rangle$7 colouring (Charikar et al., 2020). Under adversarial edge arrival, the best guarantee in the same work is $S=\langle e_1,\dots,e_m\rangle$8 colours in $S=\langle e_1,\dots,e_m\rangle$9 space (Charikar et al., 2020). The contrast isolates the technical role of order randomness: random chunks behave like hypergeometric samples, while adversarial order destroys that concentration structure.

3. Online matching under edge arrival

In online matching, edge arrival is a particularly stringent model because feasibility depends on the state of both endpoints at the moment the edge appears. This two-sided contention is the main difference from one-sided vertex-arrival stochastic matching and is repeatedly identified as the source of stronger hardness (Feng et al., 15 Jun 2026).

For adversarial-order stochastic weighted matching, prophet-inequality methods currently give the strongest positive guarantees. In general graphs with independently drawn edge weights and oblivious adversarial edge order, there is a $V$0-competitive prophet inequality for matching under edge arrivals (Ezra et al., 2020). The analysis is based on an online contention resolution scheme on the fractional matching polytope,

$V$1

together with a refined lower bound on the probability that both endpoints of an arriving edge remain unmatched (Ezra et al., 2020).

For stochastic edge arrival with known distributions, the KIID model formalizes the arrival process directly at the edge level. The input is a known type-graph $V$2, each edge type $V$3 has integral expected arrival count $V$4, the total number of rounds is $V$5, and each round samples an edge type independently with probability $V$6 (Feng et al., 15 Jun 2026). In this setting, Greedy is at most $V$7-competitive, Suggested Matching achieves $V$8 under integral arrival rates, and the two-stage Boosted Suggested Matching algorithm is $V$9-competitive, improving to $\Pr[e_k=e]=m_e/m$0 when the type-graph admits a perfect matching (Feng et al., 15 Jun 2026). The same paper proves that no online algorithm can be better than $\Pr[e_k=e]=m_e/m$1-competitive (Feng et al., 15 Jun 2026).

The role of free disposal is especially clear on acyclic graphs. With free disposal, unweighted online matching on Growing Trees has a tight competitive ratio $\Pr[e_k=e]=m_e/m$2, Forests admit $\Pr[e_k=e]=m_e/m$3, and weighted online matching on Growing Trees admits a simple ordinal $\Pr[e_k=e]=m_e/m$4-competitive algorithm that is optimal among ordinal algorithms (Jiang et al., 2024). The paper emphasizes that without free disposal, weighted edge-arrival matching has no bounded competitive ratio even on a star, whereas in the unweighted setting these positive ratios give the first explicit separations showing extra power from free disposal on Growing Trees and Forests (Jiang et al., 2024).

4. Edge arrival as a probabilistic and statistical network principle

In statistical network theory, edge arrival is often encoded through edge exchangeability. The central motivation is that many datasets are built from interaction events rather than from sampling a fixed vertex set and observing all pairwise adjacencies. This leads to the principle that edges, not vertices, are the statistical units (Crane et al., 2016). Formally, every edge exchangeable network can be represented as a mixture of i.i.d. edge-events; in the blip-free case,

$\Pr[e_k=e]=m_e/m$5

where conditional on $\Pr[e_k=e]=m_e/m$6, the interactions $\Pr[e_k=e]=m_e/m$7 are i.i.d. (Crane et al., 2016). This directly induces a sequential edge-arrival interpretation and supports sparse and power-law network behavior unavailable under ordinary countable vertex exchangeability.

The same perspective has been extended to extremely sparse graph sequences. An edge-exchangeable graph sequence $\Pr[e_k=e]=m_e/m$8 can be generated from latent node weights $\Pr[e_k=e]=m_e/m$9 by sampling Bernoulli edge-multiplicity increments

$\mathcal Y^\sigma \overset d= \mathcal Y$0

independently across time and pairs, conditional on the latent CRM (Kilian, 20 Jun 2026). Under the rapidly varying Lévy-tail condition

$\mathcal Y^\sigma \overset d= \mathcal Y$1

the paper proves

$\mathcal Y^\sigma \overset d= \mathcal Y$2

yielding an “extremely sparse” regime in which the number of observed edge types is near-linear in the number of observed vertices up to a slowly varying factor (Kilian, 20 Jun 2026). This shows that edge-arrival constructions are not limited to online optimization; they also form a nonparametric probabilistic framework for sparse network asymptotics.

A related generalization replaces dyadic edges by directed hyperedge events. The Hyperedge Event Model treats timestamped events with one sender and one or more receivers, or vice versa, and couples a dynamic ERGM-style receiver model with a survival model for event timing (Kim et al., 2018). The observed event is generated by a sender-specific race,

$\mathcal Y^\sigma \overset d= \mathcal Y$3

so the next arrival is the earliest proposed hyperedge event rather than a dyadic edge (Kim et al., 2018). This suggests that edge-arrival models sit on a continuum ranging from individual dyads to structured event streams with joint recipient selection and continuous-time timing.

5. Reverse edge-arrival in graph diffusion models

A newer usage appears in graph generation, where edge arrival is not an external input model but a latent construction path. EDGE reformulates large-graph generation as an edge-arrival, or equivalently reverse edge-removal, process (Wu et al., 2023). Starting from adjacency matrix $\mathcal Y^\sigma \overset d= \mathcal Y$4, the forward process deletes edges with $\mathcal Y^\sigma \overset d= \mathcal Y$5,

$\mathcal Y^\sigma \overset d= \mathcal Y$6

with per-edge retention probability $\mathcal Y^\sigma \overset d= \mathcal Y$7, so that $\mathcal Y^\sigma \overset d= \mathcal Y$8 is the empty graph and reverse denoising reconstructs $\mathcal Y^\sigma \overset d= \mathcal Y$9 from $p=0$0 (Wu et al., 2023). The model introduces binary active-node variables $p=0$1, because only nodes incident to edge changes at time $p=0$2 need to be scored; this is the main scaling mechanism on large sparse graphs (Wu et al., 2023).

EDGE++ keeps this framework but modifies both schedule design and sampling. It introduces an active-node schedule $p=0$3, derives the expected number of active nodes $p=0$4, and solves for a degree-specific retention schedule $p=0$5 by minimizing a schedule-matching objective under the constraint $p=0$6 (Wu et al., 2023). It also adds volume-preserved reweighting of node and edge distributions during sampling so that active-node counts and edge counts remain aligned with the intended reverse trajectory (Wu et al., 2023). Empirically, on Polblogs and PPI with $p=0$7, the paper reports that EDGE++ improves or matches graph-statistics fidelity relative to EDGE, winning on 7 of 8 metrics, while reducing training memory from 22.4 GB to 15.4 GB on Polblogs and from 57.6 GB to 34.1 GB on PPI (Wu et al., 2023).

This usage shifts the meaning of edge arrival from “online revelation of an unknown graph” to “latent sequential reconstruction of a graph.” A plausible implication is that edge-arrival thinking now functions as an algorithmic design principle as well as a data-access model.

6. Scope, misconceptions, and distinct usages

A persistent misconception is that edge arrival is merely vertex arrival with the neighborhood information dispersed over time. The literature does not support that equivalence. In vertex-arrival streams, the greedy algorithm is a one-pass $p=0$8-space maximal independent set algorithm for both implicit and explicit vertex streams, whereas in one-pass edge arrival the exact maximal independent set problem requires $p=0$9 space (Cormode et al., 2018). For bipartite matching, one-sided vertex arrival admits the classical $E=\{e_i:i\in[m]\}$0 phenomenon, while single-pass edge arrival already incurs super-semistreaming lower bounds beyond $E=\{e_i:i\in[m]\}$1 (Kapralov, 2021).

Another source of confusion is terminological drift outside graph theory. In edge computing service-hosting work, arrivals may denote user requests driving online hosting decisions, but that paper explicitly “does not propose a new probabilistic ‘edge arrival model’”; its analytical models are an arbitrary online sequence $E=\{e_i:i\in[m]\}$2 and an i.i.d. Bernoulli process $E=\{e_i:i\in[m]\}$3 (Narayana et al., 2021). In mobile-edge computing scheduling, arrivals are mobile devices entering a cell at the beginning of each frame with probability $E=\{e_i:i\in[m]\}$4, together with task and location marks (Huang et al., 2019). In heliospheric forecasting, “arrival” may refer to the leading edge of an ICME shock or compression wave, operationally detected in WSA–ENLIL+Cone by a threshold crossing of the derivative of simulated dynamic pressure (Wold et al., 2018). These are distinct notions of “edge” and “arrival,” and they should not be conflated with graph-edge arrival.

Taken together, the cited literature shows that edge arrival is best understood as a family of sequential edge-centric formalisms rather than a single model class. In adversarial graph streams it is a hard online information regime; in stochastic matching it is a distribution-aware arrival process over edge types; in edge-exchangeable probability it is a sampling principle with de Finetti-type structure; and in diffusion-based graph generation it is a constructive reverse-time parameterization of sparse graph synthesis (Cormode et al., 2018, Feng et al., 15 Jun 2026, Crane et al., 2016, Wu et al., 2023).