Fully-Defective Asynchronous Networks

• Fully-defective asynchronous networks are systems that operate without global synchronization and assume severe communication errors, leading to content-oblivious computations based solely on pulse timing.
• Key resilience thresholds, such as 2-edge-connectivity and specific Byzantine bounds, are critical for ensuring correct computations and leader election in these defect-laden settings.
• Techniques like content-oblivious simulation, pulse-based DFS strategies, and modular event-driven dynamics offer practical strategies to overcome the challenges of extreme asynchrony.

Fully-defective asynchronous networks are networked systems studied under the joint absence of global synchronization and the presence of severe communication or dynamical defects. In the strongest explicit formulation, every link may suffer an unlimited number of alteration errors, so that message payloads are unusable and computation becomes content-oblivious, based only on pulse arrivals (Censor-Hillel et al., 2022). Closely related literatures broaden the picture to systems with arbitrary asynchronous schedules, directed or incomplete topologies, bounded or unbounded staleness, and event-driven constraints that can stop, lock, or restart individual components (Bick et al., 2015, Baburin et al., 9 Feb 2025). Taken together, these works define a technical landscape in which correctness is sought without barriers, shared rounds, or trustworthy message contents.

1. Definitions and model families

The explicit distributed-systems definition comes from the fully-defective channel model of Censor-Hillel, Cohen, Gelles, and Sela. The network is an undirected graph $G=(V,E)$; nodes are deterministic; communication is asynchronous; message delays are arbitrary but finite; channels are not assumed FIFO; every sent message eventually causes some delivered message; and channels may alter messages arbitrarily often, while not deleting or injecting them (Censor-Hillel et al., 2022). Because arbitrary alteration destroys all payload semantics, the model is equivalent to content-oblivious computation, where nodes communicate solely via pulses (Chang et al., 11 Jul 2025).

A broader dynamical-systems formalism models an asynchronous network as a quadruple

$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$

with node set $\mathcal N=\{N_0,N_1,\dots,N_k\}$, phase space $M=\prod_{i=1}^k M_i$, a finite family of connection structures $A$, admissible vector fields $\mathcal F=\{f^\alpha\}_{\alpha\in A}$, and an event map $\mathcal E:M\to A$. The active vector field is

$F(X)=f^{\mathcal E(X)}(X).$

In this framework, nodes can evolve independently, be constrained, stop, and later restart, and the interaction pattern can depend on time, state, and stochastic effects (Bick et al., 2015).

In asynchronous cellular and Boolean-network settings, the defect is pushed into scheduling. For asynchronous cellular automata, an update schedule is a map $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$, and the only global requirement is fairness: every cell must be updated infinitely often (Baburin et al., 9 Feb 2025). For Boolean networks under fully asynchronous dynamics, exactly one component is updated at each step, and all one-step transitions are encoded in the asynchronous graph $\Gamma(f)$ (Melliti et al., 2016).

Research line	Formal object	Defect/asynchrony emphasis
Content-oblivious distributed computation	$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$0 with pulse-only communication	Unlimited alteration errors on all links
Event-driven asynchronous dynamics	$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$1	State-dependent topology, stopping, locking
Asynchronous cellular or Boolean systems	$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$2, or $\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$3	Arbitrary fair schedules or one-component updates

This diversity of formalizations suggests that “fully-defective asynchronous network” is best treated as a family resemblance term: the common core is independent local progress under extreme restrictions on coordination.

2. Connectivity thresholds, resilience bounds, and impossibility

In the fully-defective content-oblivious model, the central structural threshold is 2-edge-connectivity. Censor-Hillel, Cohen, Gelles, and Sela prove that if the network is not 2-edge-connected, then no non-trivial computation in the fully-defective setting is possible; if it is 2-edge-connected, any algorithm for a noiseless asynchronous setting can be simulated over the fully-defective network (Censor-Hillel et al., 2022). The impossibility is first shown in the two-party case: over a single fully-defective channel, no deterministic protocol that gives an irrevocable output can compute a non-constant function.

The 2-edge-connected threshold persists in the leader-election refinement of Frei, Gelles, Ghazy, and Nolin. They give an asynchronous content-oblivious leader-election algorithm that quiescently terminates in any 2-edge-connected network with message complexity

$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$4

where $\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$5 is the number of edges, $\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$6 is a known upper bound on the number of nodes, and $\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$7 is the smallest identifier (Chang et al., 11 Jul 2025). Combined with the earlier simulation result, this removes the need for a predesignated leader in that model.

A different, but equally sharp, threshold appears in asynchronous Byzantine agreement over incomplete graphs. In the point-to-point, asynchronous, randomized model with up to $\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$8 Byzantine nodes, exact agreement is solvable iff

$\mathfrak N=(\mathcal N,A,\mathcal F,\mathcal E),$9

where $\mathcal N=\{N_0,N_1,\dots,N_k\}$0 is vertex connectivity (Wang et al., 2020). This transfers Dolev’s synchronous graph condition to the asynchronous randomized setting without weakening it.

In asynchronous shared memory with Byzantine faults and transient failures, the relevant resilience bound is not graph-theoretic but population-theoretic: the randomized self-stabilizing clock-function construction assumes $\mathcal N=\{N_0,N_1,\dots,N_k\}$1, under a full-information dynamic adversary and arbitrary initial corruption (Hoch et al., 2010). These model-specific thresholds show that “defectiveness” does not have a single universal boundary; the admissible defect budget depends on whether the obstacle is topological separation, Byzantine behavior, or self-stabilizing temporal incoherence.

3. Content-oblivious computation, simulation, and leader election

The constructive side of the fully-defective model is based on content-oblivious simulation. The 2022 simulation algorithm has a preprocessing phase that constructs a Robbins cycle $\mathcal N=\{N_0,N_1,\dots,N_k\}$2 and an online phase that simulates any noiseless asynchronous protocol over that cycle. The preprocessing cost is $\mathcal N=\{N_0,N_1,\dots,N_k\}$3, and simulating a message $\mathcal N=\{N_0,N_1,\dots,N_k\}$4 requires

$\mathcal N=\{N_0,N_1,\dots,N_k\}$5

with the communication protocol remaining content-oblivious because nodes must ignore the content of received messages (Censor-Hillel et al., 2022). The key structural fact is Robbins’ theorem: a graph is 2-edge-connected iff it admits a strongly connected orientation.

Communication over the Robbins cycle is driven by pulses interpreted only through direction and phase. The simulation uses token passing and pulse patterns rather than payload-bearing messages. This converts total content corruption into a model where only the existence, direction, and sequencing of communication events matter.

Leader election in the same setting is more delicate because quiescent termination is required before simulation can begin. Frei, Gelles, Ghazy, and Nolin solve this with a two-phase pulse protocol. In the counting phase, each node maintains counters $\mathcal N=\{N_0,N_1,\dots,N_k\}$6, sent-pulse counters $\mathcal N=\{N_0,N_1,\dots,N_k\}$7, and received-pulse counters $\mathcal N=\{N_0,N_1,\dots,N_k\}$8; counting is synchronized by the invariant

$\mathcal N=\{N_0,N_1,\dots,N_k\}$9

and each node keeps sending one pulse per port whenever this minimum increases (Chang et al., 11 Jul 2025). A node exits the counting loop either by reaching $M=\prod_{i=1}^k M_i$0, which happens only for the minimum-ID node, or by detecting the anomaly

$M=\prod_{i=1}^k M_i$1

which can only be caused by DFS notification pulses from a node that has already left the counting phase.

The second phase is a pulse-based DFS notification. Starting from the unique leader $M=\prod_{i=1}^k M_i$2, the algorithm forces each directed edge to carry exactly $M=\prod_{i=1}^k M_i$3 pulses in each DFS direction, which lets downstream nodes infer $M=\prod_{i=1}^k M_i$4 and terminate without ambiguity (Chang et al., 11 Jul 2025). The leader is the last node to terminate. This is significant because the original conjecture that a preselected leader was necessary for any non-trivial content-oblivious task is thereby fully refuted in the 2-edge-connected setting.

4. Event-driven dynamics, causality, and modular decomposition

In the continuous-state theory of asynchronous networks, defectiveness is expressed less through corrupted channels than through constrained motion, switching interaction structure, and the possibility of deadlock. Nodes have phase spaces $M=\prod_{i=1}^k M_i$5, the event map $M=\prod_{i=1}^k M_i$6 partitions $M=\prod_{i=1}^k M_i$7 into event sets, and the resulting dynamics is piecewise smooth. This framework explicitly accommodates nodes that are stopped or partially stopped by constraints, including constraints modeled by foliations of submanifolds (Bick et al., 2015).

When initialization and termination sets factor as

$M=\prod_{i=1}^k M_i$8

the network becomes a functional asynchronous network. For $M=\prod_{i=1}^k M_i$9, the domain

$A$0

collects initial states for which each node reaches its termination set, and the transition function is

$A$1

The theory then isolates dynamical deadlocks, topological deadlocks, partial deadlocks, and livelocks as invariant sets in $A$2 that trap trajectories and prevent the network function from being achieved (Bick et al., 2017).

For regular functional asynchronous networks of simple type, the main structural result is the Modularization of Dynamics Theorem. Such networks can be uniquely represented as feedforward networks connecting primitive dynamical modules, with the generalized transition function factoring as

$A$3

and each $A$4 itself decomposing as a product of primitive-event transition functions (Bick et al., 2017). This is the cleanest formal expression of non-defective asynchronous behavior in the event-driven setting: global function emerges from a partially ordered composition of local events rather than from a globally synchronized clock.

A conceptually adjacent conservative-time formulation appears in fully asynchronous neural simulation. There, each neuron has its own local time line, communication is point-to-point and sporadic, and there is no global synchronizing clock. A postsynaptic neuron $A$5 maintains the latest known time $A$6 for each presynaptic neuron $A$7, and advances only under the safety condition

$A$8

where $A$9 is the synaptic delay (Magalhães et al., 2019). The resulting execution is described as “exhaustive yet not speculative”: every causally admissible interval is integrated, but no rollback is needed because a neuron never advances beyond the earliest possible arrival of an unseen event. This is a concrete causal-consistency mechanism for asynchronous networks with heterogeneous local clocks.

5. Discrete-state complexity, pathological trajectories, and universality

Fully asynchronous Boolean dynamics shows that severe pathologies can coexist with structurally benign update rules. For a Boolean network $\mathcal F=\{f^\alpha\}_{\alpha\in A}$0 with no negative loops, there exists a monotone Boolean network

$\mathcal F=\{f^\alpha\}_{\alpha\in A}$1

such that $\mathcal F=\{f^\alpha\}_{\alpha\in A}$2 is a fixed point of $\mathcal F=\{f^\alpha\}_{\alpha\in A}$3 iff $\mathcal F=\{f^\alpha\}_{\alpha\in A}$4 is a fixed point of $\mathcal F=\{f^\alpha\}_{\alpha\in A}$5, and a path of length $\mathcal F=\{f^\alpha\}_{\alpha\in A}$6 in $\mathcal F=\{f^\alpha\}_{\alpha\in A}$7 corresponds exactly to a path of length $\mathcal F=\{f^\alpha\}_{\alpha\in A}$8 in $\mathcal F=\{f^\alpha\}_{\alpha\in A}$9 (Melliti et al., 2016). The same paper proves that for every even $\mathcal E:M\to A$0, there exists a monotone network, an initial configuration $\mathcal E:M\to A$1, and a fixed point $\mathcal E:M\to A$2 such that

$\mathcal E:M\to A$3

At the same time, if $\mathcal E:M\to A$4 is monotone, then from every initial configuration $\mathcal E:M\to A$5 there exists some fixed point reachable by a fully asynchronous strategy with at most $\mathcal E:M\to A$6 updates (Melliti et al., 2016). The tension between existentially short trajectories and universally long trajectories is one of the discrete-state signatures of defectiveness.

Asynchronous cellular automata push the scheduling defect further by allowing an adversarial update schedule $\mathcal E:M\to A$7, constrained only by fairness. The paper introducing flip automata networks (FAN) shows that FANs have invariant histories, so the logical state-change history is independent of the asynchronous schedule (Baburin et al., 9 Feb 2025). This permits schedule-robust simulation of synchronous automata: a synchronous cellular automaton with $\mathcal E:M\to A$8 states can be simulated by an asynchronous automaton with $\mathcal E:M\to A$9 states in general, and by one with $F(X)=f^{\mathcal E(X)}(X).$0 states in one dimension, improving previously established quadratic bounds.

The same work sharpens the universality frontier. All one-way asynchronous automata are shown to be computationally equivalent to finite-state machines and therefore non-universal, while universal constructions exist as a 6-state first-neighbor automaton in one dimension and a 3-state von Neumann automaton in two dimensions (Baburin et al., 9 Feb 2025). This suggests that fully-defective scheduling does not by itself preclude universality; what matters is whether the local interaction geometry still supports schedule-invariant causal propagation.

6. Optimization, learning, and numerical computation without barriers

In distributed optimization, “fully asynchronous” denotes a milder defect model than total content corruption, but the common feature is the absence of waiting. The asynchronous push-pull gradient algorithm (APPG) works over a strongly connected directed graph, allows no global clock, uses possibly stale information from neighbors, and assumes only bounded activation intervals and bounded transmission delays. Under Lipschitz-continuous gradients and a Polyak–Łojasiewicz condition on the global objective, every node converges to the same optimal solution at a linear rate $F(X)=f^{\mathcal E(X)}(X).$1, where $F(X)=f^{\mathcal E(X)}(X).$2 increases by one no matter which node updates (Zhang et al., 2019). The proof uses two augmented networks with virtual $F(X)=f^{\mathcal E(X)}(X).$3-type and $F(X)=f^{\mathcal E(X)}(X).$4-type nodes so that delayed push and pull processes become a synchronous time-varying linear system.

The same architectural idea appears in distributed reinforcement learning. For policy evaluation over a directed peer-to-peer network, APP-SAG combines push-pull communication with stochastic average gradients and a push-pull augmented graph analysis. Under strong connectivity, bounded asynchrony, and standard MSPBE regularity assumptions, each node’s primal-dual variable converges exactly at a linear rate $F(X)=f^{\mathcal E(X)}(X).$5, and the numerical experiments show linear speedup with respect to the number of nodes together with robustness to straggler nodes (Sha et al., 2020).

Byzantine-robust asynchronous learning adds another defect dimension. Zeno++ removes synchronization barriers and allows anonymous workers, arbitrarily stale worker updates, and an unbounded number of Byzantine workers. The server evaluates a candidate gradient $F(X)=f^{\mathcal E(X)}(X).$6 using a descent score against a validation gradient $F(X)=f^{\mathcal E(X)}(X).$7,

$F(X)=f^{\mathcal E(X)}(X).$8

after normalizing $F(X)=f^{\mathcal E(X)}(X).$9 so that $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$0 (Xie et al., 2019). Under smoothness assumptions, a bounded delay on the validation gradient, and a clean validation set, convergence is proved for non-convex problems despite Byzantine workers.

Numerical simulation of continuous-time networks shows a closely related barrier-free paradigm. In fully asynchronous fully implicit neural simulation, global synchronization barriers are removed, communication becomes point-to-point and event-driven, and time stepping is driven by local causal constraints rather than the shortest network delay. On 64 Cray XE6 compute nodes, the method demonstrates a reduced number of interpolation steps, higher numerical accuracy, and lower time to solution compared with state-of-the-art synchronous methods (Magalhães et al., 2019). In all of these cases, the technical problem is not merely parallelization; it is preserving exactness or convergence when local clocks, message ages, and workloads diverge.

7. Open problems and conceptual frontier

Several open directions remain explicit in the literature. In content-oblivious computation, the main open question is uniformity: the current general 2-edge-connected leader-election algorithm assumes a known upper bound $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$1 on the network size, and it is left open whether content-oblivious leader election is possible in general 2-edge-connected networks without such prior knowledge (Chang et al., 11 Jul 2025). The same work leaves a substantial gap between current upper bounds and known lower bounds for message complexity.

In asynchronous cellular automata, an open question is whether the duality between ACAs and FANs can be generalized to all commutative ACAs (Baburin et al., 9 Feb 2025). A positive answer would unify schedule-invariant asynchronous computation under a single structural principle. In Boolean networks, open problems include extending the monotone embedding to networks with negative loops, reducing the $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$2-component construction to $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$3 components, and characterizing which interaction graphs force large asynchronous diameter (Melliti et al., 2016).

In incomplete-network Byzantine agreement, the authors explicitly point to Byzantine-tolerant synchronizers for incomplete graphs and to the use of more modern asynchronous BA protocols atop the purifying layer as natural next steps (Wang et al., 2020). In the event-driven dynamical-systems line, the frontier is less algorithmic than structural: regular functional asynchronous networks admit feedforward modularization, whereas deadlocks, hidden deadlocks, and locking phenomena mark the transition to defective behavior (Bick et al., 2017).

Across these models, a common synthesis emerges. Fully-defective asynchronous networks are not defined by a single algebraic axiom, but by a regime in which synchrony, trustworthy payloads, or globally ordered rounds are unavailable. The literature then asks which additional structures restore computability: 2-edge-connectivity for content-oblivious communication, $\zeta:\mathbb N\to\mathcal P(\mathbb Z^d)$4 for asynchronous agreement in incomplete Byzantine graphs, commutativity or invariant histories for schedule-robust automata, regularity and feedforward factorization for event-driven dynamics, and augmented-graph contraction or validation-based descent tests for optimization and learning. This suggests that defectiveness is best understood not as a binary label, but as a spectrum of missing coordination mechanisms against which different forms of causal structure, redundancy, and modularity are deployed.