Distance Preservers in Graphs

• Distance preservers are sparse subgraphs that exactly preserve shortest-path distances between designated demand pairs in a graph.
• Recent research leverages techniques such as consistent tiebreaking and novel peeling methods to establish nearly optimal upper and lower bounds.
• They are pivotal in applications like graph sparsification, metric compression, and algorithm design, with extensions to fault-tolerant and weighted/directed settings.

A distance preserver is a sparse subgraph of a given graph that exactly preserves the shortest-path distance for a prescribed set of pairs, termed demand pairs. The study of distance preservers is central to extremal combinatorics, graph sparsification, metric compression, and distance labeling. Recent work has yielded sharp, nearly optimal upper and lower bounds on their size, powerful reductions between directed and undirected instances, and insights into both their algorithmic construction and computational complexity.

1. Formal Definition and Foundational Properties

Let $G=(V,E)$ be a (possibly directed, possibly weighted) graph, and let $P \subseteq V \times V$ be a set of $p$ demand pairs. A subgraph $H \subseteq G$ is a distance preserver for $(G, P)$ if: $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$ For undirected graphs, $(s, t)$ may be treated as unordered. Two important special cases are pairwise preservers (arbitrary $P$) and subsetwise (all pairs in $S \subseteq V$). The extremal quantity of interest is, for each $n$ and $P \subseteq V \times V$0, the smallest possible number of edges in a preserver for some $P \subseteq V \times V$1-node graph and $P \subseteq V \times V$2 demand pairs.

Trivial lower bounds are immediate: a preserver must retain every edge forming a unique shortest path for its associated pair. In an unweighted clique with $P \subseteq V \times V$3 demands, $P \subseteq V \times V$4 edges are needed. In a path of length $P \subseteq V \times V$5 and $P \subseteq V \times V$6, all $P \subseteq V \times V$7 edges are required (Bodwin et al., 2015).

2. Combinatorial Upper and Lower Bounds

Historically, distance preserver construction relied on properties of consistent tiebreaking for shortest paths. Here, a tiebreaking scheme $P \subseteq V \times V$8 is consistent if, for any four vertices $P \subseteq V \times V$9, the intersection $p$0 is either empty or a contiguous subpath in both (Bodwin et al., 2015).

Coppersmith–Elkin (2006) established that, under any consistent scheme, the union of $p$1 respective demand-paths yields a preserver with

$p$2

edges in an $p$3-node undirected unweighted graph (Bodwin et al., 2015). This was shown tight for consistent schemes via a finite field grid-line construction.

Recent advances decisively surpassed these bounds. Bodwin and Vassilevska Williams introduced a fundamentally new peeling technique, demonstrating that

$p$4

edges always suffice in an undirected unweighted $p$5-vertex graph, for any set of $p$6 demand pairs (Bodwin et al., 2015). The proof leverages a density-based iterative rerouting via low-diameter “chokepoint” sets and branching event arguments. The resulting curve interpolates and improves on all prior regimes. Currently, the best known upper bounds for key settings are summarized below.

Graph type	Upper Bound	Reference
Undirected, unweighted	$p$7	(Bodwin et al., 2015)
Directed, weighted	$p$8	(Bodwin, 2016)
Exact, unweighted directed	$p$9	(Hoppenworth et al., 2024)
Undirected, unweighted, linear-in-$H \subseteq G$0 regime	$H \subseteq G$1 (see below)	(Bodwin, 2016)

Lower bounds are known from classical and recent combinatorial constructions:

• At least $H \subseteq G$2 edges are required for certain parameter regimes in both weighted and directed settings (Bodwin, 2016, Hoppenworth et al., 2024).
• For subsetwise preservers, there are parameter regions where $H \subseteq G$3 edges are necessary to preserve all pairwise distances within a set of $H \subseteq G$4 nodes, even in unweighted undirected graphs (Bodwin, 2016).

A major refinement, due to Bodwin (Bodwin, 2016), established that in undirected unweighted graphs with $H \subseteq G$5, there are preservers of size $H \subseteq G$6, where $H \subseteq G$7 is the Ruzsa–Szemerédi function—the maximal $H \subseteq G$8 such that any graph of $H \subseteq G$9 edges can be partitioned into $(G, P)$0 induced matchings. The best known is $(G, P)$1.

3. Directed, Weighted, and Approximate Preservers

The situation in directed and/or weighted graphs is more intricate:

• In directed weighted graphs, $(G, P)$2 suffices, and $(G, P)$3 can be achieved if $(G, P)$4 (Bodwin, 2016).
• For unweighted directed graphs, recent work shows an upper bound of $(G, P)$5 (Hoppenworth et al., 2024), improving prior $(G, P)$6 from consistent tiebreaking.
• Critically, a directed-to-undirected reduction has been established: improvements in the undirected preserver regime immediately translate (via an explicit exponent transformation) to better directed preservers (Hoppenworth et al., 2024).

For $(G, P)$7–approximate preservers (allowing for a finite stretch), lower bounds reveal that approximate preservers can require asymptotically as many edges as exact ones in certain regimes; for all finite $(G, P)$8, $(G, P)$9 edges may still be required (Hoppenworth et al., 2024). For near-exact $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$0–preservers, the best bounds in weighted directed graphs are

$$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$1

as a consequence of reductions from hopset constructions (Kogan et al., 2022).

4. Algorithmic Construction and Complexity

Several explicit algorithms have been developed for constructing (sub)optimal distance preservers:

• The “hitting set” framework and slack stitching arguments yield polylog-free ($$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$2-factor) constructions, modulo randomness, for numerous related preserves and additive spanners (Bodwin, 2020).
• The parameterized complexity of the minimum distance preserver problem has recently been mapped (Simonov et al., 22 Mar 2026). Both the subsetwise and pairwise variants are NP-hard and W[1]-hard parameterized by numbers of terminals even on grid subgraphs; however, they become FPT on full grids or graphs of bounded treewidth.
• The minimum-weight preserver and cost-sharing variants are highly inapproximable. For weighted graphs, the cost-sharing pairwise preserver (CSPDP) cannot be approximated within a factor $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$3 unless Label-Cover$$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$4 improves, and the best available algorithm achieves $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$5-approximation in time $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$6 (Abdolmaleki et al., 2020).

5. Structural and Fault-Tolerant Generalizations

Several extensions of the preserver paradigm have been studied:

• Distance-preserving trees: Characterize when a single spanning tree preserves distances from two vertices; necessary and sufficient conditions are given in terms of unique shortest paths and a partitioning dictated by unique nearest points (Araki et al., 2014). The general $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$7-vertex case remains open.
• Fault-tolerant preservers: For an integer $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$8, an $$\forall (s, t) \in P :\quad \operatorname{dist}_H(s,t) = \operatorname{dist}_G(s,t)$$9-fault-tolerant preserver ensures distances are preserved even after up to $(s, t)$0 edge (or vertex) failures. For $(s, t)$1, subquadratic-size preservers exist, but for $(s, t)$2 and weighted (or directed) graphs, $(s, t)$3 edges may be necessary (Bodwin et al., 2017).
• Distributed preserver construction: In the CONGEST model, distributed algorithms yield 2-fault-tolerant preservers with $(s, t)$4 edges and $(s, t)$5 rounds for $(s, t)$6 sources (Parter, 2020).

6. Connections, Generalizations, and Open Problems

Distance preservers are deeply connected to:

• Additive spanner constructions: The best-known preservers power additive spanners with $(s, t)$7 edges and $(s, t)$8 stretch (Bodwin et al., 2015).
• Hopsets: General reductions from hopset constructions have enabled improved near-exact preserver bounds in both directed and undirected settings (Kogan et al., 2022). However, reductions are currently one-way (hopsets $(s, t)$9 preservers).
• Induced-matchings and the Ruzsa–Szemerédi problem: In specific parameter regimes, the size of the best preserver is governed by $P$0, which is a central open problem in extremal combinatorics related to induced-matching decompositions (Bodwin, 2016).

Several open problems remain:

• The conjectured optimal bound in undirected unweighted graphs is $P$1, but the precise dependence on $P$2 for small $P$3 remains unresolved (Bodwin et al., 2015).
• The subsetwise preserver regime for intermediate subset size still has a gap between upper and lower bounds (Bodwin, 2016).
• Uniform improvements in undirected preservers, particularly in the challenging $P$4 regime, would automatically transfer via current reductions to the directed setting (Hoppenworth et al., 2024).
• Generalizations to multiroot distance-preserving trees and to negative-weight or directed graphs invite further investigation (Araki et al., 2014).

In summary, distance preservers constitute a central object at the interface of extremal graph theory, algorithmic metric compression, and graph sparsification. The landscape of optimal size bounds is nearly but not completely mapped, especially beyond the simplest undirected, unweighted, pairwise regime. Recent advances have revealed deep combinatorial phenomena, clarified computational hardness, and created bridges between previously disparate sparsification techniques. The field remains open to both combinatorial advances and algorithmic innovation.