Separated Low-Diameter Decompositions

• Separated low-diameter decompositions are families of partitions that maintain bounded cluster diameters and probabilistic or deterministic controls on the separation of nearby points or edges.
• They are applicable in arbitrary metric spaces as well as undirected and directed graphs, utilizing random shifts, exponential sampling, and deterministic methods to achieve controlled cut probabilities and separation.
• Recent approaches integrate approximate distance techniques, blurry ball growing, and distributed algorithms to enhance performance and ensure rigorous separation in various graph families.

Searching arXiv for relevant papers on separated low-diameter decompositions and related variants. Separated low-diameter decompositions are families of partitions or clusterings that combine a diameter bound with an explicit control on separation. In the metric formulation, a distribution over partitions is required to place every block inside diameter $\Delta$ while ensuring that nearby points are separated with probability at most proportional to $d(x,y)/\Delta$. In graph-theoretic formulations, the same idea appears as a bound on the probability that an edge is cut, or as a deterministic requirement that distinct clusters be non-adjacent or lie at graph-distance at least $D$. More recent directed variants replace symmetry by an ordered partial clustering in which short directed paths from later clusters to earlier clusters are forbidden (Gupta et al., 2013, Rozhoň et al., 2022, Haeupler et al., 29 Sep 2025).

1. Formal notions and parameterizations

The term covers several closely related definitions rather than a single universal formalism. In the metric-space setting, a distribution $\mathcal{D}$ over partitions $\{V_1,\dots,V_q\}$ of an $n$-point metric space $(V,d)$ is a $\Delta$-LDD with separation parameter $\beta$ if every part has diameter at most $\Delta$ and, for every $d(x,y)/\Delta$0,

$d(x,y)/\Delta$1

This is the standard separated low-diameter decomposition (SLDD) definition used for arbitrary metrics (Gupta et al., 2013).

In undirected weighted graphs, the same principle is often stated edgewise. A randomized partition into connected clusters of strong diameter $d(x,y)/\Delta$2 is required to satisfy

$d(x,y)/\Delta$3

for every edge $d(x,y)/\Delta$4 of length $d(x,y)/\Delta$5; here $d(x,y)/\Delta$6 is the quality of the decomposition (Dou et al., 2024). Related work on probabilistic low-diameter decompositions uses the notation $d(x,y)/\Delta$7-decomposition or $d(x,y)/\Delta$8-low-diameter decomposition, with strong diameter at most $d(x,y)/\Delta$9 and per-edge cut-probability at most $D$0 in weighted graphs (Miller et al., 2013, Forster et al., 2021).

A distinct deterministic formulation makes separation literal rather than probabilistic. A $D$1-separated, $D$2-strong-diameter decomposition of a weighted graph $D$3 is a partition into clusters $D$4 such that every $D$5 has diameter at most $D$6 and every pair of distinct clusters satisfies $D$7 (Rozhoň et al., 2022). In distributed work, an even stricter local notion is used: a clustering is separated, or non-adjacent, if for all distinct clusters $D$8 one has $D$9 (Rozhoň et al., 2022).

Directed graphs require a different notion of separation. An ordered clustering $\mathcal{D}$0 together with a mark function $\mathcal{D}$1 is a separated low-diameter decomposition if clustered vertices form strongly connected pieces of diameter at most $\mathcal{D}$2, and if $\mathcal{D}$3 then every $\mathcal{D}$4 and $\mathcal{D}$5 satisfy $\mathcal{D}$6; marked vertices are declared unclustered (Haeupler et al., 29 Sep 2025). This makes separation directional and inherently asymmetric.

The literature also distinguishes weak from strong diameter. Metric decompositions and some probabilistic constructions control weak diameter, whereas distributed and parallel graph algorithms often require strong diameter in the induced subgraph. A related strengthening is the tree-supported decomposition (TSD), in which each cluster is equipped with a supporting tree; the relevant parameters then include the maximum tree diameter and the load $\mathcal{D}$7 of every edge (Becker et al., 2019).

2. Classical randomized constructions in metrics and undirected graphs

A central metric construction is the random-rates algorithm. Given $\mathcal{D}$8, one first computes a $\mathcal{D}$9-net $\{V_1,\dots,V_q\}$0, then samples for each terminal $\{V_1,\dots,V_q\}$1 a truncated exponential $\{V_1,\dots,V_q\}$2 with $\{V_1,\dots,V_q\}$3, sets $\{V_1,\dots,V_q\}$4, and assigns every $\{V_1,\dots,V_q\}$5 by

$\{V_1,\dots,V_q\}$6

The resulting partition is a $\{V_1,\dots,V_q\}$7-LDD with $\{V_1,\dots,V_q\}$8 on arbitrary $\{V_1,\dots,V_q\}$9-point metrics, and $n$0 on metrics of doubling dimension $n$1. The same paper states that $n$2 is tight on arbitrary metrics, while in doubling metrics the bound becomes dimension-dependent (Gupta et al., 2013).

For unweighted undirected graphs, the shifted shortest-path framework provides a graph-native analogue. Each vertex $n$3 samples $n$4, a super-source $n$5 is connected to $n$6 by an edge of length $n$7, and a single parallel BFS assigns every vertex to the center minimizing shifted distance $n$8. The resulting partition is a $n$9-decomposition; under CRCW-PRAM it has expected $(V,d)$0 work and $(V,d)$1 parallel time, with constant success probability that can be boosted by repetition (Miller et al., 2013).

A later improvement replaces true exponentials by a capped geometric distribution. For $(V,d)$2, it sets $(V,d)$3 and

$(V,d)$4

samples delays from $(V,d)$5, and runs a Dijkstra-like propagation from a super-source for exactly $(V,d)$6 synchronous CONGEST rounds. The paper proves a random partition with strong diameter $(V,d)$7 and per-edge cut-probability

$(V,d)$8

with the notable feature that both the diameter and the running time are deterministic rather than merely high-probability bounds (Forster et al., 2021).

These constructions expose a recurring design pattern: random perturbations induce Voronoi-type regions, and the main analytical burden is to show that nearby points, balls, or edges are separated only with probability proportional to their scale.

3. Approximate distances, blurry ball growing, and tree-supported decompositions

A major technical obstacle in large-scale models is that many classical decompositions depend heavily on exact SSSP. The approximate-distance approach of Becker, Emek, and Lenzen addresses this by introducing blurry ball growing. The procedure $(V,d)$9 takes a seed set $\Delta$0, a scale $\Delta$1, and $\Delta$2, and outputs a superset $\Delta$3 such that every $\Delta$4 satisfies $\Delta$5 and every edge $\Delta$6 satisfies

$\Delta$7

The algorithm proceeds in $\Delta$8 rounds, interleaving random radius choices with $\Delta$9-approximate SSSP computations (Becker et al., 2019).

This technique is combined with the exponential-delay idea of Miller et al. (SPAA 13) in the algorithm $\beta$0. In each of $\beta$1 phases it performs exponential shifts, constructs Voronoi-style cells using a $\beta$2-approximate SSSP from a virtual source, shrinks each cell by a second approximate SSSP from its boundary, and then blurs the interior by calling $\beta$3. The main theorem gives a randomized algorithm that, on input $\beta$4 and target radius $\beta$5, produces with high probability a tree-supported decomposition $\beta$6 satisfying

$\beta$7

and, for every edge $\beta$8 of length $\beta$9,

$\Delta$0

The complexity bounds are model-specific: $\Delta$1 rounds in CONGEST, polylog$\Delta$2 depth and $\Delta$3 work in PRAM, and polylog$\Delta$4 passes with $\Delta$5 space in semi-streaming (Becker et al., 2019).

The same framework yields hierarchical TSDs, projected trees, and dominating $\Delta$6-HSTs. The resulting embeddings have expected stretch $\Delta$7, depth $\Delta$8, and, for fixed $\Delta$9, expected $d(x,y)/\Delta$00-stretch $d(x,y)/\Delta$01. The projected-tree construction additionally preserves a load bound $d(x,y)/\Delta$02 on every original edge, which the paper identifies as useful for capacitated problems and for simulating CONGEST algorithms on the embedding tree (Becker et al., 2019).

4. Deterministic and distributed separated clustering

Deterministic separated clustering arose in distributed symmetry breaking as a route to strong-diameter network decompositions. One local process in CONGEST constructs a separated, or non-adjacent, clustering $d(x,y)/\Delta$03 in which clusters have strong diameter $d(x,y)/\Delta$04, at least half of all vertices are clustered, and the computation takes $d(x,y)/\Delta$05 rounds. Iterating this procedure $d(x,y)/\Delta$06 times yields a full network decomposition with $d(x,y)/\Delta$07 colors and strong diameter $d(x,y)/\Delta$08 in $d(x,y)/\Delta$09 CONGEST rounds (Rozhoň et al., 2022).

The mechanism is not based on random shifts. Instead, it maintains rooted forests over the alive vertices and repeatedly recolors trees red or blue according to successive ID bits. Blue subtrees propose attachment to adjacent red trees; red trees either accept the proposals and grow, or decline and delete the proposing subtrees. Three invariants drive the analysis: a ruling invariant, a separation invariant based on shared ID prefixes within connected components, and a deletion invariant ensuring progress. The result is explicitly separated clustering rather than merely low expected cut-frequency (Rozhoň et al., 2022).

A more general deterministic framework for weighted graphs replaces random shifts by bit-splitting over approximate distance profiles. Given a weighted graph and separation parameter $d(x,y)/\Delta$10, the algorithm runs for $d(x,y)/\Delta$11 phases. In phase $d(x,y)/\Delta$12, it splits centers by the $d(x,y)/\Delta$13-st ID bit, invokes an approximate-distance oracle $d(x,y)/\Delta$14, computes a distance difference $d(x,y)/\Delta$15 between the two center classes for each vertex, buckets vertices by $d(x,y)/\Delta$16 with $d(x,y)/\Delta$17, and deletes a smallest bucket. The paper proves a deterministic $d(x,y)/\Delta$18-separated strong-diameter decomposition with strong diameter $d(x,y)/\Delta$19, separation at least $d(x,y)/\Delta$20, and coverage at least $d(x,y)/\Delta$21; it also states a randomized version covering $d(x,y)/\Delta$22 vertices with diameter $d(x,y)/\Delta$23 and separation at least $d(x,y)/\Delta$24 (Rozhoň et al., 2022).

This line of work reframes separation as a constructive barrier between cluster centers rather than a probabilistic event. A plausible implication is that deterministic separation is especially natural when the decomposition is meant to support subsequent local computations inside each cluster.

5. Separator-sensitive improvements and restricted graph families

For graph classes with additional structure, the separation parameter can improve asymptotically. Kamma and Krauthgamer study $d(x,y)/\Delta$25-path-separable graphs, defined recursively through balanced vertex-separators that are unions of at most $d(x,y)/\Delta$26 shortest paths. Their main theorem states that every $d(x,y)/\Delta$27-path-separable $d(x,y)/\Delta$28-vertex graph admits an efficiently samplable $d(x,y)/\Delta$29-decomposition with

$d(x,y)/\Delta$30

The construction has two stages: a deterministic recursive choice of centers by taking $d(x,y)/\Delta$31-nets on separator paths in carefully chosen subgraphs, and a randomized carving phase in which each center samples

$d(x,y)/\Delta$32

A key structural lemma bounds the number of centers that can threaten any fixed vertex by $d(x,y)/\Delta$33 (Kamma et al., 2015).

The stated consequences are class-sensitive. Every treewidth-$d(x,y)/\Delta$34 graph is $d(x,y)/\Delta$35-path-separable, giving $d(x,y)/\Delta$36, while every $d(x,y)/\Delta$37-minor-free graph is $d(x,y)/\Delta$38-path-separable, giving $d(x,y)/\Delta$39. The same source notes, however, that excluded-minor graphs admit $d(x,y)/\Delta$40-padded decompositions by other techniques, so the path-separable bound is not the final word for that class (Kamma et al., 2015).

A recent distributed-and-parallel treatment extends approximate-SSSP techniques to arbitrary graphs and to graphs separable through $d(x,y)/\Delta$41 shortest paths. For arbitrary graphs, an edge of length $d(x,y)/\Delta$42 is cut with probability $d(x,y)/\Delta$43; for $d(x,y)/\Delta$44-path-separable graphs, the bound improves to $d(x,y)/\Delta$45. The framework combines a pseudo-padded decomposition based on random shifts and one approximate SSSP, an inner-core extraction by two SSSP computations, blurry ball growing to obtain a low-diameter clustering that covers at least half the nodes, and recursive repetition to obtain a full partition. In the restricted-family case it further replaces explicit separators by a weak $d(x,y)/\Delta$46-separator built from randomly sampled short $d(x,y)/\Delta$47-approximate paths (Dou et al., 2024).

The model-specific bounds are also explicit: $d(x,y)/\Delta$48 depth and $d(x,y)/\Delta$49 work in PRAM, $d(x,y)/\Delta$50 time in HYBRID, and $d(x,y)/\Delta$51 for arbitrary graphs versus $d(x,y)/\Delta$52 for $d(x,y)/\Delta$53-path-separable graphs in CONGEST. The paper presents these as distributed and parallel implementations of sequential divide-and-conquer algorithms in which exact shortest paths are replaced by approximate shortest paths (Dou et al., 2024).

6. Directed separated decompositions

The directed case changes the problem qualitatively because undirected arguments based on symmetric distances no longer apply. A recent definition of separated low-diameter decomposition for directed graphs uses an ordered partial clustering $d(x,y)/\Delta$54 with marks $d(x,y)/\Delta$55. Each cluster must be strongly connected with

$d(x,y)/\Delta$56

and if $d(x,y)/\Delta$57 then every $d(x,y)/\Delta$58 and $d(x,y)/\Delta$59 must satisfy $d(x,y)/\Delta$60. Vertices with $d(x,y)/\Delta$61 are left unclustered (Haeupler et al., 29 Sep 2025).

Two near-linear-time sampling routines are given. Algorithm A uses truncated-exponential in-balls and out-balls over $d(x,y)/\Delta$62 scales, marking all vertices within distance at most $d(x,y)/\Delta$63 of each boundary layer. Algorithm B uses random-radius rounds, heavy/light classifications, and alternating out-ball and in-ball cuts, again marking the layer $d(x,y)/\Delta$64. The separation guarantee is produced by the marking step: vertices near a boundary are forbidden from later clustering, so any short directed path that would violate cluster order must cross marked territory (Haeupler et al., 29 Sep 2025).

The principal theorem states that for any $d(x,y)/\Delta$65, $d(x,y)/\Delta$66, and $d(x,y)/\Delta$67, there is a distribution over ordered partial clusterings and marks such that

$d(x,y)/\Delta$68

clusters have diameter at most $d(x,y)/\Delta$69, the ordering is $d(x,y)/\Delta$70-separated, every edge $d(x,y)/\Delta$71 satisfies

$d(x,y)/\Delta$72

and sampling takes $d(x,y)/\Delta$73 time. The same work emphasizes that setting $d(x,y)/\Delta$74 already yields a nontrivial separated decomposition with sub-logarithmic diameter (Haeupler et al., 29 Sep 2025).

The paper also introduces guarantees absent from prior directed LDDs: exponential-tail bounds for paths and edge-sets, and independence results for far-away edges. This suggests that “separation” in the directed setting is not merely a translation of the undirected condition, but a more structured interface between decomposition, causality, and boundary management.