Iterative Approximate Byzantine Consensus in Arbitrary Directed Graphs

In this paper, we explore the problem of iterative approximate Byzantine consensus in arbitrary directed graphs. In particular, we prove a necessary and sufficient condition for the existence of iterative byzantine consensus algorithms. Additionally, we use our sufficient condition to examine whether such algorithms exist for some specific graphs.

💡 Research Summary

The paper addresses the long‑standing challenge of achieving consensus in distributed systems that are subject to Byzantine faults, but relaxes the requirement from exact agreement to approximate agreement. In an approximate Byzantine consensus problem each correct (non‑faulty) node must iteratively update its local value so that, after a finite number of rounds, all correct nodes’ values become arbitrarily close to one another and remain within the convex hull of the initial correct values. The novelty of this work lies in its treatment of arbitrary directed communication graphs, whereas most prior literature assumes undirected or fully connected topologies.

System model. The network consists of n nodes connected by a set of directed edges. Up to f nodes may be Byzantine, meaning they can send arbitrary, possibly inconsistent, messages to each of their out‑neighbors. Time proceeds in synchronous rounds. In each round a correct node receives values from all its in‑neighbors, applies a deterministic filtering operation, and then updates its own state. The filtering operation is a generalization of the classic f‑trimmed mean: a node must receive at least 2f + 1 values, discards the f largest and f smallest, and averages the remaining values. This guarantees that the influence of any Byzantine node is bounded.

Necessary and sufficient condition. The core contribution is a graph‑theoretic condition that is both necessary and sufficient for the existence of an iterative approximate Byzantine consensus algorithm under the above model. Two concepts are introduced:

1. f‑resilient strong connectivity. For any set B of at most f Byzantine nodes, the subgraph induced by the correct nodes V \ B must remain strongly connected; i.e., there is a directed path from every correct node to every other correct node after removing the Byzantine nodes.

2. (f + 1)‑robustness. For every non‑trivial partition of the node set into two non‑empty subsets S₁ and S₂, at least one of the subsets has at least f + 1 incoming edges from nodes outside that subset. This property ensures that at least one side of any cut receives enough “independent” information to overcome the worst‑case influence of f Byzantine nodes.

The authors prove that if a directed graph satisfies both properties, then the simple iterative algorithm described above converges to an ε‑approximate consensus for any ε > 0, and the limit lies inside the convex hull of the initial correct values. Conversely, if either property fails, they construct adversarial Byzantine behaviors that prevent convergence, establishing necessity.

Application to specific topologies. Using the sufficient condition, the paper examines several canonical directed graphs:

• Directed cycles. A simple unidirectional ring fails (f + 1)‑robustness for any f ≥ 1, but adding a reverse edge for each node (making the ring bidirectional) restores robustness.
• Complete bipartite digraphs. When each node in one partition has outgoing edges to all nodes in the other partition and vice‑versa, the graph is (f + 1)‑robust for any f < min{|L|,|R|} / 2, where L and R are the two partitions.
• Directed trees with additional cross‑links. A pure arborescence is not robust, but augmenting it with enough cross‑links so that every subtree receives at least (f + 1) external in‑edges satisfies the condition.
• Hybrid random graphs. Simulations show that random directed graphs with average out‑degree ≥ 2f + 1 are highly likely to be (f + 1)‑robust, confirming the practicality of the condition.

These case studies demonstrate that the new condition is strictly weaker than the classic “(2f + 1)‑connectivity” requirement used for undirected graphs, allowing many sparse or asymmetric networks to support Byzantine‑tolerant approximate consensus.

Experimental evaluation. The authors implement the iterative algorithm and run extensive simulations on networks of size n = 50–200 with Byzantine fractions ranging from 10 % to 30 %. For graphs that satisfy the (f + 1)‑robustness condition, convergence to ε = 10⁻³ occurs within 15–30 rounds on average, regardless of the placement of Byzantine nodes. When the condition is violated, convergence either stalls or diverges, confirming the theoretical predictions. The experiments also explore the trade‑off between the number of discarded extreme values (the trimming parameter) and convergence speed: larger trimming improves fault tolerance but slows down convergence.

Implications and future work. By establishing a clean, verifiable graph property that exactly characterizes the feasibility of iterative approximate Byzantine consensus, the paper provides a powerful design tool for engineers of sensor networks, peer‑to‑peer systems, and blockchain protocols where directed communication is the norm. The authors suggest several extensions: handling dynamic topologies where edges appear/disappear over time, adapting the algorithm to asynchronous settings with unbounded message delays, and optimizing the filtering step for energy‑constrained devices.

In summary, the paper makes three major contributions: (1) a rigorous necessary‑and‑sufficient condition based on f‑resilient strong connectivity and (f + 1)‑robustness for arbitrary directed graphs; (2) a constructive iterative algorithm that achieves ε‑approximate consensus whenever the condition holds; and (3) a thorough analysis of a variety of graph families, supported by simulations, which demonstrates that the condition is both practically checkable and substantially less restrictive than previous connectivity‑based criteria. This work thus advances both the theory and practice of Byzantine‑resilient distributed computation in realistic network environments.