Near-Exact Hopsets in Graph Algorithms

• Near-exact hopsets are graph augmentations that add weighted shortcut edges so every vertex pair has a path within (1+ε) of the true distance using a bounded number of hops.
• They achieve favorable trade-offs between sparsity and hop-bound through constructions like Thorup–Zwick emulators and sampling-and-bunch frameworks.
• These structures underpin efficient sequential, parallel, distributed, and streaming algorithms for approximating shortest paths in diverse graph settings.

Near-exact hopsets are graph augmentations that add weighted shortcut edges to a graph so that every pair of vertices admits a path with a bounded number of hops and length within a factor $1+\epsilon$ of the true shortest-path distance. In the standard terminology, a $(\beta,\epsilon)$-hopset enforces a hop-bound $\beta$ and multiplicative stretch $1+\epsilon$, with the regime of small $\epsilon$ constituting the near-exact setting. Introduced by Cohen in the context of parallel computation of approximate shortest paths, hopsets later became a basic primitive in dynamic graph algorithms, distributed computing, and the streaming model (Elkin et al., 2016).

1. Formal definition and variants

For a weighted undirected graph $G=(V,E,\omega)$, let $d_G(u,v)$ denote the shortest-path distance, and for an integer $\beta \ge 1$, let

$d_{G'}^{(\beta)}(u,v)$

denote the length of a shortest $u$-to-$(\beta,\epsilon)$0 path in $(\beta,\epsilon)$1 that uses at most $(\beta,\epsilon)$2 edges. A graph $(\beta,\epsilon)$3 is a $(\beta,\epsilon)$4-hopset for $(\beta,\epsilon)$5 if, for every $(\beta,\epsilon)$6,

$(\beta,\epsilon)$7

Here $(\beta,\epsilon)$8 is the hop-bound and $(\beta,\epsilon)$9 is the multiplicative stretch parameter (Elkin et al., 2016).

The same basic idea extends to directed graphs, but the formulation is less uniform. One directed definition requires exact distance preservation in the augmented graph together with a low-hop approximate path: for every $\beta$0,

$\beta$1

and there exists a $\beta$2 path in $\beta$3 of at most $\beta$4 edges and weight at most $\beta$5 (Bernstein et al., 2022). A broader directed formulation fixes $\beta$6 and asks that every reachable ordered pair $\beta$7 admit an $\beta$8 path in $\beta$9 of at most $1+\epsilon$0 edges and total weight at most $1+\epsilon$1; the case $1+\epsilon$2 is the directed near-exact regime (Hoppenworth et al., 2024).

Two parameters dominate the theory. The first is sparsity, usually expressed as $1+\epsilon$3 or by specialized linear-size bounds. The second is the hop-bound $1+\epsilon$4, whose dependence on $1+\epsilon$5, the sparsity exponent, and in directed settings the aspect ratio, is the central structural quantity throughout the literature.

2. Undirected trade-offs and existential bounds

The undirected theory developed through a sequence of progressively sharper size-hopbound trade-offs. An early milestone gave linear-size hopsets with substantially improved hopbound: setting $1+\epsilon$6 in the Elkin–Neiman construction yields $1+\epsilon$7 and

$1+\epsilon$8

while the general bound is $1+\epsilon$9 with

$\epsilon$0

for a runtime parameter $\epsilon$1 (Elkin et al., 2017).

A subsequent construction exhibited the first sparse hopsets with a constant number of hops. For parameters $\epsilon$2, $\epsilon$3, and $\epsilon$4, one can construct a $\epsilon$5-hopset of expected size

$\epsilon$6

where

$\epsilon$7

and $\epsilon$8 is independent of $\epsilon$9 for fixed $G=(V,E,\omega)$0; the leading $G=(V,E,\omega)$1 in the size can be removed at the expense of increasing the exponent of $G=(V,E,\omega)$2 by $G=(V,E,\omega)$3 (Elkin et al., 2016).

The later survey literature emphasizes a parallel line of results based on scale-free constructions. For every $G=(V,E,\omega)$4-vertex weighted graph, every $G=(V,E,\omega)$5, and integer $G=(V,E,\omega)$6, there exists a $G=(V,E,\omega)$7-hopset of size

$G=(V,E,\omega)$8

with

$G=(V,E,\omega)$9

and a refined analysis removes the leading $d_G(u,v)$0. The same parameters can be achieved by a single, $d_G(u,v)$1-independent construction, so that one hopset simultaneously satisfies the $d_G(u,v)$2 guarantee for all $d_G(u,v)$3 (Elkin et al., 2020).

Construction	Size	Hop-bound
Linear-size Elkin–Neiman specialization (Elkin et al., 2017)	$d_G(u,v)$4	$d_G(u,v)$5
Constant-hop hopsets (Elkin et al., 2016)	$d_G(u,v)$6	$d_G(u,v)$7
Scale-free hopsets (Elkin et al., 2020)	$d_G(u,v)$8	$d_G(u,v)$9
Thorup–Zwick optimal trade-off (Huang et al., 2017)	$\beta \ge 1$0	$\beta \ge 1$1

These parameterizations are not identical: $\beta \ge 1$2 and $\beta \ge 1$3 index different sparsity curves. Nonetheless, they chart the main undirected progression from linear-size quasi-polylogarithmic hopbound, to constant-hop sparse constructions, to sparsity-optimal Thorup–Zwick-based hopsets.

A further structural synthesis appears in the unified framework of Neiman and Shabat. In the size regime $\beta \ge 1$4, their sampling-and-bunch formalism recovers a simultaneous $\beta \ge 1$5-hopset with

$\beta \ge 1$6

and proves a matching lower bound of a different form: any $\beta \ge 1$7-hopset of size $\beta \ge 1$8 must satisfy

$\beta \ge 1$9

(Neiman et al., 2021). For near-exact hopsets, where $d_{G'}^{(\beta)}(u,v)$0, this gives a coarse but general obstruction to simultaneously minimizing multiplicative stretch and hopbound.

3. Thorup–Zwick emulators as universally optimal hopsets

A decisive simplification came from the observation that Thorup–Zwick sublinear-additive emulators are also hopsets. In the construction, one fixes an integer $d_{G'}^{(\beta)}(u,v)$1, forms nested samples

$d_{G'}^{(\beta)}(u,v)$2

and for each $d_{G'}^{(\beta)}(u,v)$3 adds weighted edges from $d_{G'}^{(\beta)}(u,v)$4 to every $d_{G'}^{(\beta)}(u,v)$5, where

$d_{G'}^{(\beta)}(u,v)$6

and each edge $d_{G'}^{(\beta)}(u,v)$7 receives weight $d_{G'}^{(\beta)}(u,v)$8 (Huang et al., 2017).

Following Pettie, the sampling probabilities are set to

$d_{G'}^{(\beta)}(u,v)$9

With this choice,

$u$0

The key analytic step is to convert Thorup–Zwick’s sublinear additive emulator guarantee into multiplicative $u$1-stretch by explicitly bounding the number of hops in $u$2. Defining $u$3 and

$u$4

one obtains $u$5, and choosing $u$6 yields

$u$7

(Huang et al., 2017).

This trade-off is described in the paper as universally optimal. Abboud–Bodwin–Pettie proved that any hopset of size $u$8 must have

$u$9

whereas the Thorup–Zwick modification achieves size $(\beta,\epsilon)$00 with

$(\beta,\epsilon)$01

matching the lower-bound curve up to $(\beta,\epsilon)$02-polynomial factors (Huang et al., 2017).

The same construction also has corollaries outside the hopset setting. For unweighted graphs it gives a sublinear-additive emulator of size $(\beta,\epsilon)$03 and stretch

$(\beta,\epsilon)$04

and, combined with standard “stretch $(\beta,\epsilon)$05” pruning, yields a spanner with stretch $(\beta,\epsilon)$06, additive error $(\beta,\epsilon)$07, and size

$(\beta,\epsilon)$08

where $(\beta,\epsilon)$09 (Huang et al., 2017). This is one of the clearest demonstrations that near-exact hopsets, near-additive emulators, and sparse spanners often occupy the same combinatorial template.

4. Construction paradigms and computational models

The first constant-hop construction is also notable for its algorithmic generality. Its centralized algorithm uses a hierarchy of clusterings, interleaving super-clustering and interconnection over distance scales. For fixed $(\beta,\epsilon)$10, it constructs in

$(\beta,\epsilon)$11

time a $(\beta,\epsilon)$12-hopset of expected size $(\beta,\epsilon)$13, with space $(\beta,\epsilon)$14; parallel, Congested Clique, CONGEST, and streaming implementations with explicit resource bounds are also given (Elkin et al., 2016).

In the EREW PRAM model, this framework yields a construction using

$(\beta,\epsilon)$15

processors and time

$(\beta,\epsilon)$16

after which $(\beta,\epsilon)$17-approximate shortest paths from $(\beta,\epsilon)$18 sources can be computed via $(\beta,\epsilon)$19-limited Bellman–Ford in $(\beta,\epsilon)$20 further time (Elkin et al., 2016). In the Congested Clique, one can build the hopset in

$(\beta,\epsilon)$21

rounds, with a choice of $(\beta,\epsilon)$22 making $(\beta,\epsilon)$23 constant in $(\beta,\epsilon)$24 (Elkin et al., 2016).

Deterministic parallelization required a separate advance. The first efficient deterministic parallel hopset algorithm builds a $(\beta,\epsilon)$25-hopset of size

$(\beta,\epsilon)$26

with

$(\beta,\epsilon)$27

in CREW-PRAM time

$(\beta,\epsilon)$28

and work

$(\beta,\epsilon)$29

yielding the first deterministic polylogarithmic-time PRAM algorithm for $(\beta,\epsilon)$30-SSSP with slightly super-linear work (Michael et al., 2020).

Distributed settings impose different sparsity-round trade-offs. In the Congested Clique, a randomized algorithm constructs, with high probability, a $(\beta,\epsilon)$31-hopset of size

$(\beta,\epsilon)$32

hopbound

$(\beta,\epsilon)$33

and round complexity

$(\beta,\epsilon)$34

providing the first polylogarithmic-round Congested Clique algorithm for sparse hopsets with polylogarithmic hopbound (Nazari, 2019).

Taken together, these results show that the near-exact regime is not only an existential theory of sparse augmentations. It is also a design framework for sequential, PRAM, Congested Clique, CONGEST, streaming, and deterministic shortest-path algorithms.

5. Directed near-exact hopsets

Directed graphs are markedly harder. The first nearly linear-work directed construction produces a hopset with $(\beta,\epsilon)$35 edges and

$(\beta,\epsilon)$36

in sequential time $(\beta,\epsilon)$37, and a parallel version with work $(\beta,\epsilon)$38 and span $(\beta,\epsilon)$39 (Cao et al., 2019). This already established that near-exact directed hopsets can be built efficiently, but not yet with the same linear-size hopbound known for directed shortcut sets.

That gap was closed by Bernstein and Wein. For any $(\beta,\epsilon)$40-vertex directed graph with integer weights in $(\beta,\epsilon)$41, fixed $(\beta,\epsilon)$42, and $(\beta,\epsilon)$43, there exists a $(\beta,\epsilon)$44-hopset $(\beta,\epsilon)$45 of size

$(\beta,\epsilon)$46

In particular, choosing

$(\beta,\epsilon)$47

gives $(\beta,\epsilon)$48 and hop-bound $(\beta,\epsilon)$49. More generally, for every $(\beta,\epsilon)$50, one can set

$(\beta,\epsilon)$51

to obtain a $(\beta,\epsilon)$52-hopset of size

$(\beta,\epsilon)$53

matching, up to polylogarithmic factors, the known shortcut-set trade-off (Bernstein et al., 2022).

The construction combines several ingredients: a weighted transitive closure, a “nice path collection” of vertex-disjoint shortest paths of exactly $(\beta,\epsilon)$54 hops, Raz–Spira forward 2-hop emulators on those paths, hierarchical vertex-path sampling, and a new sixth-hop backward-shortcutting gadget that handles the cyclic obstructions absent in DAG-like settings (Bernstein et al., 2022).

Lower bounds reveal that the directed near-exact regime is highly sensitive to weight structure. For arbitrary weighted directed graphs, every $(\beta,\epsilon)$55-size approximate hopset with any given finite stretch must have

$(\beta,\epsilon)$56

while $(\beta,\epsilon)$57-size exact hopsets in unweighted directed graphs require $(\beta,\epsilon)$58, and $(\beta,\epsilon)$59-size shortcut sets require $(\beta,\epsilon)$60 (Hoppenworth et al., 2024). The $(\beta,\epsilon)$61 bound establishes a separation from the polynomial-aspect-ratio setting, where $(\beta,\epsilon)$62-size approximate hopsets with hopbound $(\beta,\epsilon)$63 are known (Hoppenworth et al., 2024). A common misunderstanding is therefore to treat the $(\beta,\epsilon)$64 linear-size bound as fully general in directed graphs; the lower bound shows that unbounded aspect ratio rules this out.

6. Reductions, unification, and optimization

Near-exact hopsets have long been studied alongside near-additive spanners and emulators. The survey literature highlighted that the results of Elkin–Peleg for spanners and Elkin–Neiman for hopsets are “strikingly similar,” and explicitly posed the problem of a black-box reduction between near-additive emulators and near-exact hopsets (Elkin et al., 2020).

One direction was clarified through a reduction from hopsets to spanners, emulators, and distance preservers. The central notion is an $(\beta,\epsilon)$65-missing $(\beta,\epsilon)$66-spanner: a subgraph $(\beta,\epsilon)$67 such that every pair $(\beta,\epsilon)$68 has an approximate path of length at most $(\beta,\epsilon)$69 omitting at most $(\beta,\epsilon)$70 edges of $(\beta,\epsilon)$71. Given a sequence of hopsets $(\beta,\epsilon)$72 with parameters $(\beta,\epsilon)$73, one can build in polynomial time a subgraph $(\beta,\epsilon)$74 that is an

$(\beta,\epsilon)$75

spanner of size

$(\beta,\epsilon)$76

The same framework yields new weighted near-additive spanners, directed weighted pairwise preservers, and lower bounds such as $(\beta,\epsilon)$77 for linear-size exact weighted hopsets (Kogan et al., 2022).

The reverse direction was established more recently. Any algorithm $(\beta,\epsilon)$78 that constructs a $(\beta,\epsilon)$79-emulator of size $(\beta,\epsilon)$80 for undirected unweighted graphs can be used as a black box to obtain, for any $(\beta,\epsilon)$81, a $(\beta,\epsilon)$82-hopset for an undirected weighted graph with

$(\beta,\epsilon)$83

and size

$(\beta,\epsilon)$84

This answers the reduction question for sparse graphs, while leaving open the removal of the linear dependence on $(\beta,\epsilon)$85 in the size bound (Aeri et al., 8 Jul 2026).

The unification program also reaches the level of algorithmic templates. A single sampling-and-bunch framework can attain all state-of-the-art spanners and hopsets for general graphs by choosing the appropriate input parameters, and in the $(\beta,\epsilon)$86-size regime it resolves an open problem by proving that every $(\beta,\epsilon)$87-hopset must satisfy $(\beta,\epsilon)$88 (Neiman et al., 2021).

A further development shifts attention from worst-case existence to graph-specific optimization. The optimization variant asks, for a given graph instance, for the minimum number of edges in a hopset with prescribed hopbound and stretch. In the near-exact regime, there is a randomized polynomial-time algorithm that, for directed graphs with $(\beta,\epsilon)$89, returns a $(\beta,\epsilon)$90-hopset of size

$(\beta,\epsilon)$91

and in undirected graphs with moderate $(\beta,\epsilon)$92, a randomized polynomial-time algorithm returning a $(\beta,\epsilon)$93-hopset of size

$(\beta,\epsilon)$94

for $(\beta,\epsilon)$95 (Dinitz et al., 10 Feb 2025).

Open questions remain explicit and technically sharp. The survey literature asks whether the classical near-exact hopbound can be reduced to a polynomial in $(\beta,\epsilon)$96 and $(\beta,\epsilon)$97, describes the gap as “wide open,” and lists directed analogues, black-box reductions, universal constructions, and dynamic or streaming maintenance as central directions (Elkin et al., 2020). The emulator-to-hopset reduction leaves open a size bound independent of $(\beta,\epsilon)$98 (Aeri et al., 8 Jul 2026). In directed graphs, the recent lower bounds isolate aspect ratio as a decisive parameter and ask for a finer characterization of how the best achievable $(\beta,\epsilon)$99 depends on weight spread (Hoppenworth et al., 2024).