Fully Dynamic Algorithms for Graph Spanners via Low-Diameter Router Decomposition

A $t$-spanner of an undirected $n$-vertex graph $G$ is a sparse subgraph $H$ of $G$ that preserves all pairwise distances between its vertices to within multiplicative factor $t$, also called the \emph{stretch}. We investigate the problem of maintaining spanners in the fully dynamic setting with an adaptive adversary. Despite a long line of research, this problem is still poorly understood: no algorithm achieving a sublogarithmic stretch, a sublinear in $n$ update time, and a strongly subquadratic in $n$ spanner size is currently known. One of our main results is a deterministic algorithm, that, for any $512 \leq k \leq (\log n)^{1/49}$ and $1/k\leq δ\leq 1/400$, maintains a spanner $H$ of a fully dynamic graph with stretch $poly(k)\cdot 2^{O(1/δ^6)}$ and size $|E(H)|\leq O(n^{1+O(1/k)})$, with worst-case update time $n^{O(δ)}$ and recourse $n^{O(1/k)}$. Our algorithm relies on a new technical tool that we develop, called low-diameter router decomposition. We design a deterministic algorithm that maintains a decomposition of a fully dynamic graph into edge-disjoint clusters with bounded vertex overlap, where each cluster $C$ is a bounded-diameter router, meaning that any reasonable multicommodity demand over the vertices of $C$ can be routed along short paths and with low congestion. A similar graph decomposition notion was introduced by [Haeupler et al., STOC 2022] and strengthened by [Haeupler et al., FOCS 2024]. However, in contrast to these and other prior works, the decomposition that our algorithm maintains is proper, ensuring that the routing paths between the pairs of vertices of each cluster $C$ are contained inside $C$, rather than in the entire graph $G$. We show additional applications of our router decomposition, including dynamic algorithms for fault-tolerant spanners and low-congestion spanners.

💡 Research Summary

The paper tackles the long‑standing open problem of maintaining sparse graph spanners in the fully‑dynamic setting against an adaptive adversary. While static spanners achieve the optimal trade‑off of stretch (2k-1) and size (O(n^{1+1/k})), no dynamic algorithm has simultaneously achieved sub‑logarithmic stretch, sub‑linear (or even (n^{o(1)})) worst‑case update time, and a spanner size strictly below (n^{2-\varepsilon}). The authors present the first deterministic algorithm that meets all these criteria for a wide range of parameters.

Main result. For any integer (k) with (512 \le k \le (\log n)^{1/49}) and any (\delta) satisfying (1/k \le \delta \le 1/400), the algorithm maintains a spanner (H) of a fully‑dynamic undirected graph (G) with:

• Stretch (poly(k)\cdot 2^{O(1/\delta^{6})}) (polynomial in (k) and exponential only in (1/\delta^{6}));
• Edge count (|E(H)| = O!\bigl(n^{1+O(1/k)}\bigr));
• Worst‑case update time (n^{O(\delta)});
• Recourse (the number of edge changes per update) (n^{O(1/k)}).

All guarantees are deterministic, thus holding against an adaptive adversary. The framework also extends to fault‑tolerant spanners, connectivity certificates, and low‑congestion spanners.

Technical core – Low‑Diameter Router Decomposition.
The authors introduce a novel graph‑decomposition tool called a low‑diameter router decomposition. A router is a subgraph that can route any reasonable multicommodity demand among its vertices using short paths and low congestion; it is closely related to length‑constrained expanders. Prior works (Haeupler et al., STOC 2022; FOCS 2024) used router‑like decompositions but allowed routing paths to leave the cluster. In contrast, this paper’s decomposition is proper: every routing path for a pair of vertices inside a cluster (C) remains entirely within (C). This property is crucial for dynamic updates because clusters can be modified independently without affecting the rest of the graph.

The decomposition is built through a sequence of deterministic sub‑tools:

1. Construction of a “nice router.” Starting from the input graph, the authors assign vertex weights and use conductance arguments to obtain a subgraph (W_k) with bounded diameter and strong expansion properties.

2. Pruning of (W_k). Using an Even‑Shiloach tree and vertex‑splitting techniques, they iteratively delete vertices that are not essential for routing while preserving the router guarantees. The resulting subgraph (W_k’) is called a properly pruned router.

3. Embedding routers into the whole graph. Two problems are defined: EmbedOrSeparate (decide whether a router can be embedded into (G) or must be separated) and EmbedOrScatter (embed a router while scattering its edges). Deterministic algorithms solve both, guaranteeing that the embedding respects the properness of clusters and keeps congestion low.

4. Router witness. A combinatorial certificate is constructed to prove that a given subgraph truly functions as a router. This witness also shows resilience to bounded‑degree vertex failures, enabling fault‑tolerant spanner constructions.

5. Decremental low‑diameter clustering. When edges are inserted or deleted, clusters are updated using a ball‑growing procedure and a careful processing step that maintains the low‑diameter and edge‑disjointness properties. The update cost of this clustering is (n^{O(\delta)}).

All components are deterministic and run in worst‑case time polynomial in (n^{\delta}). The clustering ensures that the total number of vertices across all clusters is (O!\bigl(n^{1+O(1/k)}\bigr)), and the majority of edges of (G) belong to some cluster.

Putting the pieces together.
Initially, the algorithm builds a router decomposition of the empty graph, creates a router certificate for each cluster, and stores a dense subgraph (C’\subseteq C) with prescribed density and routing guarantees. Upon each edge insertion or deletion, the algorithm:

• Updates the affected clusters using the decremental low‑diameter clustering routine;
• Re‑prunes and re‑embeds the routers of those clusters;
• Updates the router certificates and the auxiliary subgraphs (C’);
• Adjusts the global spanner (H) by adding or removing edges from the updated (C’) structures.

Because each update touches only (n^{O(1/k)}) clusters and each cluster contains (n^{O(1/k)}) edges, the recourse bound follows. The spanner size bound follows from the fact that each cluster contributes at most (O!\bigl(|V(C)|^{1+O(1/k)}\bigr)) edges, and the sum of (|V(C)|) over all clusters is (O!\bigl(n^{1+O(1/k)}\bigr)).

Applications.

• Fault‑tolerant spanners: Using the router witness, the authors show that each cluster remains a router even after the removal of up to (\Delta) vertices, yielding a spanner that tolerates bounded‑degree failures with only a modest increase in stretch.
• Connectivity certificates: The decomposition provides a compact representation of the graph’s connectivity that can be updated quickly.
• Low‑congestion spanners: Because each router routes demands with low congestion, the overall spanner inherits a congestion guarantee of (O(\log n)). This is valuable for dynamic flow and shortest‑path algorithms that rely on low‑congestion embeddings.

Significance.
The work simultaneously achieves four desiderata that have been elusive in prior literature: (i) sub‑logarithmic stretch, (ii) near‑optimal spanner size, (iii) worst‑case update time (n^{o(1)}) (for constant (\delta)), and (iv) deterministic guarantees against an adaptive adversary. The low‑diameter router decomposition is a versatile tool that may find further use in dynamic graph algorithms beyond spanners, such as dynamic cut sparsifiers, dynamic low‑stretch trees, and dynamic network design problems.

Future directions include tightening the dependence on (\delta) to achieve truly polylogarithmic update time, extending the router decomposition to directed or weighted graphs, and exploring its interaction with other dynamic primitives like dynamic matching or dynamic spectral sparsification.