Search for the end of a path in the d-dimensional grid and in other graphs

We consider the worst-case query complexity of some variants of certain \cl{PPAD}-complete search problems. Suppose we are given a graph $G$ and a vertex $s \in V(G)$. We denote the directed graph obtained from $G$ by directing all edges in both directions by $G’$. $D$ is a directed subgraph of $G’$ which is unknown to us, except that it consists of vertex-disjoint directed paths and cycles and one of the paths originates in $s$. Our goal is to find an endvertex of a path by using as few queries as possible. A query specifies a vertex $v\in V(G)$, and the answer is the set of the edges of $D$ incident to $v$, together with their directions. We also show lower bounds for the special case when $D$ consists of a single path. Our proofs use the theory of graph separators. Finally, we consider the case when the graph $G$ is a grid graph. In this case, using the connection with separators, we give asymptotically tight bounds as a function of the size of the grid, if the dimension of the grid is considered as fixed. In order to do this, we prove a separator theorem about grid graphs, which is interesting on its own right.

💡 Research Summary

The paper studies the worst‑case query complexity of a class of PPAD‑complete search problems that can be phrased as “find an endpoint of a hidden directed path”. The setting is as follows: we are given an undirected graph G and a distinguished start vertex s. From G we form G′ by replacing each undirected edge with two opposite directed arcs. An unknown subgraph D of G′ is promised to consist of vertex‑disjoint directed paths and cycles, and exactly one of those paths originates at s. A query consists of picking a vertex v∈V(G) and receiving the set of D‑edges incident to v together with their directions. The goal is to locate a terminal vertex of the path that starts at s while using as few queries as possible.

General graphs – upper bound.
The authors first apply classic separator theory. Any n‑vertex graph can be separated by a vertex set S of size O(√n) into two subgraphs each containing at most 2n/3 vertices. The algorithm queries every vertex of S. For each queried vertex we learn whether the hidden path passes through it and, if so, which direction it continues. This information tells us in which of the two subgraphs the remaining part of the path lies. Recursively applying the same procedure yields a total of O(√n) queries (the logarithmic factor can be eliminated by a careful balanced‑separator strategy). Thus the endpoint can always be found with Θ(√n) queries.

General graphs – lower bound.
To prove optimality, the paper constructs a hard distribution over D. In a dense graph such as a complete bipartite graph, the hidden path can be placed uniformly at random among all possible s‑originating paths. Each query reveals at most two bits of information (presence/absence of an incident edge and its direction). Since the number of possible D’s is exponential in n, any deterministic or randomized algorithm needs Ω(√n) queries to reduce the uncertainty enough to pinpoint the endpoint. This matches the upper bound.

d‑dimensional grid graphs.
The main technical contribution concerns the d‑dimensional grid G_d(L) with side length L (so N = L^d vertices). The authors prove a new “grid separator theorem”: fixing one coordinate and cutting along a hyperplane yields a separator of size O(L^{d‑1}) = O(N^{1‑1/d}) that splits the grid into two roughly equal parts, each of which is again a d‑dimensional grid. Using the same recursive query strategy as for general graphs, the algorithm now needs only Θ(N^{1‑1/d}) queries. This is substantially better than the Θ(√N) bound that would follow from the generic separator theorem, and it scales naturally with the dimension.

Grid lower bound.
For the lower bound, the authors consider a random “snake” path that visits every vertex of the grid exactly once (or a collection of disjoint long paths). They show that any single query hits the separator with probability O(N^{-1/d}), so on average an algorithm must ask at least N^{1‑1/d} queries before it can expect to intersect the hidden path. A more formal information‑theoretic argument yields an Ω(N^{1‑1/d}) lower bound, which together with the upper bound establishes Θ(N^{1‑1/d}) as the optimal query complexity for fixed dimension d.

Single‑path case.
Even when D consists of a single path (the most restrictive scenario), the same lower bound holds. The authors introduce a “symmetry‑flip” technique: after each query they can imagine the adversary reflecting the remaining unknown portion of the path, preserving the distribution of possible endpoints. Consequently the algorithm cannot gain more than O(N^{-1/d}) information per query, forcing Ω(N^{1‑1/d}) queries.

Practical implications.
The model captures situations where a system can probe only local information (e.g., a sensor can read its incident links) and must locate a hidden route or failure line. In sensor networks laid out on a grid, the Θ(N^{1‑1/d}) bound translates directly into energy or time budgets for fault detection. Moreover, because the underlying problem is a PPAD‑complete search task, the results shed light on the intrinsic difficulty of finding equilibria or fixed points when only local queries are allowed.

Conclusion.
The paper delivers a tight characterization of query complexity for endpoint‑finding in both arbitrary graphs and fixed‑dimension grids. The optimal complexities are Θ(√n) for general n‑vertex graphs and Θ(N^{1‑1/d}) for d‑dimensional grids. The new separator theorem for grids is of independent interest in graph theory. These findings bridge complexity theory, combinatorial geometry, and practical distributed‑system design, offering both theoretical insight and concrete guidance for algorithms that must locate hidden paths with minimal probing.