Node Sampling using Random Centrifugal Walks

Sampling a network with a given probability distribution has been identified as a useful operation. In this paper we propose distributed algorithms for sampling networks, so that nodes are selected by a special node, called the \emph{source}, with a given probability distribution. All these algorithms are based on a new class of random walks, that we call Random Centrifugal Walks (RCW). A RCW is a random walk that starts at the source and always moves away from it. Firstly, an algorithm to sample any connected network using RCW is proposed. The algorithm assumes that each node has a weight, so that the sampling process must select a node with a probability proportional to its weight. This algorithm requires a preprocessing phase before the sampling of nodes. In particular, a minimum diameter spanning tree (MDST) is created in the network, and then nodes’ weights are efficiently aggregated using the tree. The good news are that the preprocessing is done only once, regardless of the number of sources and the number of samples taken from the network. After that, every sample is done with a RCW whose length is bounded by the network diameter. Secondly, RCW algorithms that do not require preprocessing are proposed for grids and networks with regular concentric connectivity, for the case when the probability of selecting a node is a function of its distance to the source. The key features of the RCW algorithms (unlike previous Markovian approaches) are that (1) they do not need to warm-up (stabilize), (2) the sampling always finishes in a number of hops bounded by the network diameter, and (3) it selects a node with the exact probability distribution.

💡 Research Summary

The paper introduces a novel class of random walks called Random Centrifugal Walks (RCW) and shows how they can be used to implement exact‑distribution sampling services in distributed networks. An RCW starts at a designated source node and is constrained to move strictly away from the source at each step; consequently, the walk can never exceed the network’s diameter in length. The key idea is to pre‑compute, for each node, a “stay probability” (the probability that the walk stops at that node) and a set of “hop probabilities” to farther neighbours. When a walk reaches a node, it stops with the stay probability; otherwise it hops to a farther neighbour according to the hop probabilities. By carefully choosing these probabilities, the overall probability that a node is selected equals a prescribed target distribution.

Two main scenarios are addressed.

1. Arbitrary weighted graphs.
   Each node x carries a positive weight w(x) and must be selected with probability proportional to w(x). The authors first construct a Minimum‑Diameter Spanning Tree (MDST) of the network. The MDST’s depth provides an upper bound on the walk length (≤ D, the network diameter). A distributed weight‑aggregation phase runs on the tree: each node i computes, for every incident tree edge (i, x), the total weight of the subtree reachable via that edge, denoted T_i(x). These aggregates are stored locally and later used to compute the stay and hop probabilities. Specifically, the stay probability q(i) and hop probabilities h(i, y) are derived so that the product of the visit probability v(i) (the probability the walk ever reaches i) and q(i) equals w(i)/∑w, guaranteeing exact sampling. After the one‑time preprocessing (MDST construction O(n) and weight aggregation O(n)), any node can act as a source and an arbitrary number of independent samples can be drawn. Each sample requires at most D hops, no warm‑up period, and incurs only O(1) communication per hop.

2. Distance‑based distributions on structured topologies.
   Here the desired probability depends only on the hop distance k from the source: all nodes at distance k share the same selection probability p_k (e.g., uniform, Kleinberg’s harmonic, or any custom distance‑based function). For such cases the authors avoid any preprocessing. They treat the network as a set of concentric rings R_k (k = 0…R). Two concrete families are examined: (a) a 2‑D Manhattan grid, where the source is at (0,0) and rings correspond to Manhattan distance; (b) “concentric‑rings networks” where each node in ring k has the same number of neighbours in rings k‑1 (γ_k) and k+1 (δ_k). In both families the RCW is defined so that every node in the same ring has identical stay probability q_k and identical hop probabilities to the next outer ring. The authors prove that if v_k (the visit probability of any node in ring k) and q_k satisfy p_k = v_k·q_k, then the walk yields the exact distance‑based distribution. Because each hop moves one ring outward, the walk terminates after at most R hops.

   For non‑uniform concentric‑rings networks (γ_k, δ_k vary across nodes), the paper proposes building an overlay network that adds virtual edges to enforce uniform connectivity. Simulations show that the overlay construction succeeds in a large fraction of random instances, after which the same RCW algorithm for uniform rings can be applied.

Key contributions and insights

• No warm‑up required. Unlike Metropolis‑Hastings or other Markov‑chain samplers, RCW’s termination probability is designed to match the target distribution at every step, so the first sample is already exact.
• Bounded latency. The walk length is deterministically bounded by the network diameter (general graphs) or by the number of rings (structured graphs), guaranteeing predictable sampling latency.
• One‑time preprocessing. For arbitrary weighted graphs, the MDST and weight‑aggregation are performed once; all subsequent samples reuse the same data, making the approach scalable to many sources and many samples.
• Applicability to distance‑based sampling without preprocessing. The grid and concentric‑rings algorithms enable exact distance‑based sampling with only local state (the per‑ring probabilities), which is attractive for applications such as landmark‑free positioning, epidemic broadcasting, or constructing small‑world overlays.
• Complexity. Preprocessing costs O(n) messages and time; each sampling walk costs O(D) or O(R) hops and O(1) messages per hop. This is dramatically lower than the thousands of hops reported for random‑walk‑based samplers to reach stationarity.
• Robustness. The MDST can be recomputed or maintained under link failures using known distributed algorithms; the overlay construction can be re‑run when topology changes.

Potential impact
The RCW framework provides a practical, low‑overhead method for exact network sampling, which is a building block for many distributed algorithms: random‑peer selection, load balancing, network monitoring, and topology‑aware service placement. By eliminating the warm‑up phase and guaranteeing bounded sampling time, RCW makes it feasible to deploy sampling as a real‑time service in large‑scale peer‑to‑peer systems, sensor networks, and social‑graph analytics platforms. The paper’s blend of rigorous probability analysis, algorithmic design, and simulation validation makes it a solid contribution to the field of distributed graph algorithms.