Hop-Sets in Graph Algorithms

• Hop-sets are sets of weighted shortcut edges added to a graph that approximate shortest-path distances within a specified hop limit.
• They are constructed using hierarchical sampling and Thorup–Zwick clustering to optimize tradeoffs between stretch, hopbound, and size.
• Hop-sets enable efficient algorithms in parallel, distributed, dynamic, and streaming models, underpinning routing schemes and metric compression.

A hopset for a graph is a set of weighted shortcut edges such that in the augmented graph, all pairwise (or specified) shortest-path distances can be approximated to within a desired stretch using a path of at most a specified number of hops. Hopsets have become foundational primitives in parallel, distributed, dynamic, and streaming graph algorithms, enabling a diverse array of efficient approximate distance computations, metric compression, routing schemes, and graph sparsification.

1. Definition and Basic Properties

Let $G = (V, E, w)$ be a (weighted or unweighted) graph. A $(\beta, \epsilon)$-hopset, or more generally, an $(\alpha, \beta)$-hopset, is a set $H \subseteq V \times V$, where each edge $(u, v) \in H$ is assigned weight $d_G(u, v)$, such that for every $u, v \in V$:

$$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$

or, in the case of multiplicative $(1+\epsilon)$ stretch,

$$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq (1+\epsilon) d_G(u, v)$$

Here, $(\beta, \epsilon)$0 is the minimum length of a $(\beta, \epsilon)$1–$(\beta, \epsilon)$2 path in $(\beta, \epsilon)$3 using at most $(\beta, \epsilon)$4 edges (hops) (Łącki et al., 2020, Neiman et al., 2021).

Exact hopsets correspond to $(\beta, \epsilon)$5 stretch, preserving all distances exactly but forcing all shortest paths to have hop-count at most $(\beta, \epsilon)$6. Exact hopsets are tightly connected to the classical notion of shortcut sets for reachability and diameter reduction (Bodwin et al., 2023).

2. Classical and Modern Constructions

Hierarchical Sampling and Thorup–Zwick–Type Constructions

The dominant paradigm constructs hopsets via random sampling and Thorup–Zwick–style clustering:

• Build a sequence $(\beta, \epsilon)$7 where each $(\beta, \epsilon)$8 is a random sample of $(\beta, \epsilon)$9.
• For each $(\alpha, \beta)$0, define the bunch $(\alpha, \beta)$1.
• Add weighted shortcut edges $(\alpha, \beta)$2 for $(\alpha, \beta)$3, each weighted by $(\alpha, \beta)$4 (Neiman et al., 2021, Huang et al., 2017).

This “generalized TZ” framework allows recovering all state-of-the-art tradeoffs between stretch $(\alpha, \beta)$5, hopbound $(\alpha, \beta)$6, and size $(\alpha, \beta)$7 by appropriate choice of cluster selection and sampling parameters. For example:

Parameters	Stretch ($(\alpha, \beta)$8)	Hopbound ($(\alpha, \beta)$9)	Size ($H \subseteq V \times V$0)	Reference
$H \subseteq V \times V$1-level, multiplicative	$H \subseteq V \times V$2	2	$H \subseteq V \times V$3	(Neiman et al., 2021)
$H \subseteq V \times V$4-stretch, near-additive	$H \subseteq V \times V$5	$H \subseteq V \times V$6	$H \subseteq V \times V$7	(Huang et al., 2017)
Hybrid $H \subseteq V \times V$8	$H \subseteq V \times V$9	$(u, v) \in H$0	$(u, v) \in H$1	(Neiman et al., 2021)

These bounds are essentially optimal; for any $(u, v) \in H$2-hopset of size $(u, v) \in H$3, one must have $(u, v) \in H$4 (Neiman et al., 2021).

Linear-Size, Polylog-Hopbound Constructions

A major breakthrough demonstrated constructions of linear-size hopsets ($(u, v) \in H$5 edges) with polylogarithmic hopbound:

• Setting $(u, v) \in H$6 in the above TZ framework achieves size $(u, v) \in H$7, hopbound $(u, v) \in H$8, stretch $(u, v) \in H$9 (Elkin et al., 2017).
• This nearly exponentially improved previous linear-hopset constructions (which required $d_G(u, v)$0 hopbound).

3. Algorithmic and Computation Models

Parallel, Distributed, and Streaming

Hopsets are efficiently computable in a variety of algorithms models:

• PRAM: Hopsets of size $d_G(u, v)$1, hopbound $d_G(u, v)$2, in polylog time and near-linear work (Elkin et al., 2017).
• Congested Clique: Algorithms yielding $d_G(u, v)$3-size hopsets with polylogarithmic hopbound in polylogarithmic rounds. Sparse, low-hopbound hopsets are constructed via a combination of limited neighborhood covers, clustering, and inter-cluster linking (Nazari, 2019).
• Dynamic/Decremental: Near-optimal decremental hopset algorithms with polylogarithmic hopbound for weighted graphs under deletions, matching static size-hopbound tradeoffs. This enabled (2k−1)(1+ε)-approximate decremental APSP and MSSP oracles with nearly optimal bounds in both update and query time (Łącki et al., 2020).
• Streaming: Hopsets can be constructed with $d_G(u, v)$4 space in $d_G(u, v)$5 streaming passes (Elkin et al., 2016).

Practical Distance Oracles and Routing Schemes

Hopsets directly support the construction of low-space, low-latency routing algorithms and distance oracles:

• Compact distributed routing schemes with per-node memory $d_G(u, v)$6, table size $d_G(u, v)$7, stretch $d_G(u, v)$8, and near-optimal construction time, via hopset-based preprocessing (Elkin et al., 2017).
• Centralized distance oracles in graphs of low highway or skeleton dimension using 3-hopsets, achieving lower shortcut count than classical 2-hop hub-labeling with similar or better query times (Gupta et al., 2018).

4. Lower Bounds and Sensitivity

Hopbound vs. Size

The central tradeoff is: for $d_G(u, v)$9-vertex graphs, achieving hopset size $u, v \in V$0 with $u, v \in V$1-stretch requires $u, v \in V$2; and for linear-size hopsets, $u, v \in V$3 cannot be improved below $u, v \in V$4 (Elkin et al., 2017, Huang et al., 2017, Kogan et al., 2022). For exact hopsets $u, v \in V$5, recent lower bounds show that any $u, v \in V$6-size hopset requires hopbound $u, v \in V$7 (Bodwin et al., 2023), confirming the classical folklore barrier.

Sensitivity and Robustness

Recent work has introduced vertex/edge sensitivity as a quality measure for hopsets, defined as the maximum number of times a vertex (or edge) is bypassed by any hopset edge. The Greedy-Hopset achieves $u, v \in V$8-hopsets with $u, v \in V$9 sensitivity, improving bounds for differentially private range query algorithms, but it is shown that $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$0 for any such construction (Ashvinkumar et al., 2024). Analogous trade-offs are matched for directed graphs and for $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$1-hopsets.

5. Directed Hopsets, Shortcut Sets, and Nonlinear Tradeoffs

A shortcut set for a directed graph is a collection of edges that preserves reachability by bounding the diameter after augmentation. There is now a tight relationship between linear-size directed hopsets (approximate or reachability) and shortcut sets:

• Kogan and Parter [SODA'22] proved that $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$2-size shortcut sets with hopbound $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$3 exist; subsequently, approximate $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$4 hopsets with $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$5 size and $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$6 hopbound were shown, closing the gap between shortcut sets and hopsets in the directed setting (Bernstein et al., 2022).
• The Greedy-Hopset matches existential size-optimality (up to subpolynomial factors) for exact hopsets compared to known lower bounds, and specialized deterministic algorithms can now construct $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$7-size, $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$8-hopbound shortcut sets/directed hopsets in $$d_G(u, v) \leq d_{G \cup H}^{(\beta)}(u, v) \leq \alpha \cdot d_G(u, v)$$9 time (Bals et al., 25 Nov 2025).

6. Hopsets, Spanners, Emulators, and Metric Compression

Hopsets form a bridge between spanners/emulators and distance preservers:

• Any hopset-based compression can be converted to an “$(1+\epsilon)$0-missing $(1+\epsilon)$1-spanner,” and further reduced to pairwise preservers and near-additive spanners, via a general black-box reduction (Kogan et al., 2022).
• Thorup–Zwick’s additive emulators are also universally optimal hopsets, with $(1+\epsilon)$2 parameters and $(1+\epsilon)$3 size, matching known lower bounds (Huang et al., 2017).
• Lower bounds on hopset parameters thus feed directly into the limitations for sparse spanners and metric emulators.

7. Open Problems and Research Directions

• Closing Gaps: Is it possible to obtain linear-size hopsets with hopbound $(1+\epsilon)$4, or are the lower bounds tight? Can the $(1+\epsilon)$5 lower bound for exact hopsets be improved or circumvented in restricted settings?
• Directed and Hybrid Models: Extending the full suite of hopset techniques to directed or mixed-metric graphs remains challenging (Elkin et al., 2017).
• Dynamic and Streaming Maintenance: Efficient dynamic maintenance of small hopbound, linear-size hopsets remains largely open (Łącki et al., 2020).
• Sensitivity/Robustness: Can the lower sensitivity guarantees be improved for hopsets with better hop/stretch parameters (Ashvinkumar et al., 2024)?
• Metric Compression: Deepening the connection between hopset-based and spanner-based metric compression, and determining more general reverse reductions, is a proposed direction (Kogan et al., 2022).

Hopsets remain a central graph-theoretic and algorithmic primitive with fundamental roles in distance approximation, parallel and distributed computation, graph compression, and modern shortest-path algorithms. The rich space of parameter tradeoffs, sensitivity criteria, and algorithmic models continues to drive theoretical advances and practical applications.