A Model of Optimal Network Structure for Decentralized Nearest Neighbor Search

One of the approaches for the nearest neighbor search problem is to build a network which nodes correspond to the given set of indexed objects. In this case the search of the closest object can be thought as a search of a node in a network. A procedure in a network is called decentralized if it uses only local information about visited nodes and its neighbors. Networks, which structure allows efficient performing the nearest neighbour search by a decentralised search procedure started from any node, are of particular interest especially for pure distributed systems. Several algorithms that construct such networks have been proposed in literature. However, the following questions arise: “Are there network models in which decentralised search can be performed faster?”; “What are the optimal networks for the decentralised search?”; “What are their properties?”. In this paper we partially give answers to these questions. We propose a mathematical programming model for the problem of determining an optimal network structure for decentralized nearest neighbor search. We have found an exact solution for a regular lattice of size 4x4 and heuristic solutions for sizes from 5x5 to 7x7. As a distance function we use L1 , L2 and L_inf metrics. We hope that our results and the proposed model will initiate study of optimal network structures for decentralised nearest neighbour search.

💡 Research Summary

The paper addresses the problem of designing a network topology that enables fast decentralized nearest‑neighbor (NN) search when only local information is available. The authors focus on the Greedy Walk algorithm, which repeatedly moves from the current vertex to the neighbor that is closest to the query according to a chosen distance metric. For Greedy Walk to succeed from any start vertex, the underlying graph must contain at least the Delaunay (Delone) graph of the point set; otherwise the walk can become trapped in a local minimum.

To formalize the design task, the authors introduce a Boolean non‑linear programming model. Binary variables (x_{ij}) indicate whether an edge between vertices (i) and (j) is present. Additional binary indicators (y_{ij}^q) capture whether vertex (j) lies on the Greedy Walk path from start vertex (i) to the target vertex (the vertex closest to query (q)). The objective function minimizes the expected number of distance evaluations performed during the walk, i.e., the average number of distinct vertices whose distance to the query is computed. Both discrete (queries limited to vertices) and continuous (queries anywhere in the domain) formulations are provided.

The model includes several constraints: (1) no self‑loops, (2) the walk must terminate at the true nearest neighbor, (3) path continuity linking (x) and (y) variables, and (4) the greedy selection rule that forces the neighbor with minimal distance to be part of the path. Together these constraints guarantee that any feasible solution defines a graph on which Greedy Walk is guaranteed to reach the correct target from any start node.

Experimental evaluation is performed on two‑dimensional regular lattices. The authors assume a uniform query distribution, so the NN search reduces to locating a specific node. For a 4×4 lattice they solve the model exactly using a branch‑and‑bound algorithm, obtaining optimal average distance‑computation counts (f) of 7.039 for L1, 7.093 for L2, and 7.203 for L∞. For larger lattices (5×5, 6×6, 7×7) they develop a heuristic that iteratively adds or removes edges while respecting the constraints. The heuristic yields near‑optimal structures with f values that grow roughly linearly with lattice size; L1 consistently yields the smallest f, reflecting the compatibility of Manhattan distance with axis‑aligned grids.

Key insights include: (i) the optimal graph is the base lattice plus a carefully chosen set of “shortcut” edges that reduce the number of greedy steps; (ii) the choice of metric strongly influences which shortcuts are beneficial; (iii) the Boolean programming framework is flexible enough to incorporate any metric and can be extended to other objective components (e.g., edge cost, robustness).

The authors acknowledge limitations: the exact approach does not scale beyond 16 vertices due to exponential growth of the search space; the current study is restricted to regular grids, whereas real peer‑to‑peer overlays are irregular and dynamic; and the model addresses exact NN search only, whereas many applications tolerate approximation.

Future work is outlined as follows: develop more scalable meta‑heuristics (genetic algorithms, simulated annealing) for larger graphs; extend the model to non‑Euclidean or high‑dimensional metric spaces; incorporate dynamic operations (node join/leave) and edge maintenance costs; and formulate multi‑objective versions that balance search speed with storage overhead and approximation quality.

In summary, the paper provides a rigorous mathematical formulation for the optimal network structure supporting decentralized Greedy Walk NN search, demonstrates exact and heuristic solutions on small lattices, and opens a pathway for systematic topology design in distributed similarity search systems.