---
title: Directed Low-Diameter Decompositions
url: https://www.emergentmind.com/topics/directed-low-diameter-decompositions
type: topic
---

# Directed Low-Diameter Decompositions

Searching arXiv for recent papers on directed low-diameter decompositions and closely related work.
Found and cross-checking the core recent papers on directed low-diameter decompositions.
Directed low-diameter decompositions (directed LDDs) are randomized decompositions of an edge-weighted directed graph \(G=(V,E,w)\) with a diameter parameter \(\Delta\), obtained by deleting a random subset of edges so that the remaining graph breaks into strongly connected components of bounded diameter while each edge is cut with probability proportional to \(w(e)/\Delta\). In the 2025–2026 literature, the subject developed from the first \(O(\log^2 n)\)-loss directed constructions into near-optimal \(O(\log n\log\log n)\)-loss schemes, faster near-linear-time algorithms, stronger guarantees for paths and separated clusterings, and substantially better bounds on structured digraphs such as bounded-pathwidth graphs [2502.05687][2505.10244][2509.24565][2606.31560].

## 1. Formal definitions and principal variants

A standard formalization fixes a directed, edge-weighted graph \(G=(V,E,\ell)\) and a parameter \(D\ge 1\), and defines a directed LDD with loss \(L\) as a distribution over edge sets \(S\subseteq E\) such that for any two vertices \(u,v\) in the same strongly connected component of \(G\setminus S\), one has \(d_G(u,v)\le D\) and \(d_G(v,u)\le D\), while every edge \(e\in E\) satisfies
\[
\Pr(e\in S)\le \frac{\ell(e)}{D}\cdot L.
\]
Recent algorithmic papers present a closely related formulation in which every strongly connected component in the remaining graph has weak diameter at most \(\Delta\), and each edge \((u,v)\) is deleted with probability at most \(\frac{\ell(m)}{\Delta}w(u,v)\) [2502.05687][2505.10244][2510.22721].

The weak-diameter requirement is central. One cited distinction is that weak diameter is measured in the original graph or via the underlying undirected metric, whereas a strong-diameter requirement would force short paths inside the remaining graph itself. The weak version is the focus of the near-optimal results; the same line of work notes that strong-diameter LDDs are harder, with \(O(\log^3 n)\) loss known, and that matching the weak-diameter bounds is unlikely by current methods [2502.05687].

A second principal notion is the directed quasipartition. Here the requirement is stronger: whenever one vertex can still reach another after the edge removal, the two must be close in the original directed metric. Every quasipartition yields an LDD, but the converse does not necessarily hold. This distinction is especially relevant in structured digraphs and in flow-cut applications [2606.31560].

A common misconception is to regard directed LDDs as a routine lift of undirected ball carving. The directed setting is qualitatively different: in directed graphs, out-balls or in-balls need not be strongly connected components, and the decomposition must explicitly manage directionality, SCC structure, and the interaction between forward and backward metrics [2510.22721].

## 2. From \(O(\log^2 n)\) to near-optimal \(O(\log n\log\log n)\)

The modern theory starts from the extension of LDDs to directed graphs in the negative-length single-source shortest path breakthrough cited in later work. That construction cut each edge with probability at most \(O\!\left(\frac{1}{D}\log^2 n\right)\). The 2025 near-optimal result improved this to
\[
O\!\left(\frac{1}{D}\log n\log\log n\right),
\]
which is only a \(\log\log n\) factor from the lower bound \(\Omega\!\left(\frac{1}{D}\log n\right)\), stated to hold even for undirected graphs [2502.05687].

Subsequent work preserved the same loss factor while reducing the runtime and simplifying the construction. One paper gave a simpler and faster algorithm matching the \(O(\log n\log\log n)\) loss and improving the running time to \(O((m+n\log\log n)\log^2 n)\) [2505.10244]. A later paper improved this further to expected
\[
O((m+n\log\log n)\log n\log\log n)
\]
while still achieving \(\ell(n)=O(\log n\log\log n)\) and succeeding with high probability [2510.22721].

| Paper | Main guarantee | Runtime / setting |
|---|---|---|
| [2502.05687] | Loss \(O(\log n\log\log n)\) | Deterministic \(\widetilde O(m\cdot \mathrm{poly}(D))\); randomized expected \(\widetilde O(m)\) |
| [2505.10244] | Same loss with simpler algorithm | \(O((m+n\log\log n)\log^2 n)\) |
| [2510.22721] | Same loss with faster expected algorithm | \(O((m+n\log\log n)\log n\log\log n)\) |
| [2606.31560] | \((O(\mathsf{pw}),\Delta)\)-LDD | Structured digraphs of pathwidth \(\mathsf{pw}\) |

The significance of the loss factor is algorithmic rather than merely formal. In the cited literature, the loss directly governs the approximation or recursion overhead in applications such as cut sparsifiers, network design, multicut, and negative-weight shortest paths; therefore the improvement from \(O(\log^2 n)\) to \(O(\log n\log\log n)\) propagates into downstream bounds [2505.10244][2510.22721].

## 3. Structural foundations: lopsided expanders, MWU, and quasipartitions

The near-optimal existence theory introduced a conceptual link between directed LDDs and expander decompositions. The key insight is that approximate cost minimization for LDDs can be reduced to constructing a suitable expander decomposition, but not the standard form. The paper defines a lopsided expansion parameter
\[
\psi(U)=\frac{c(U,\overline U)}{\minvol(U)\cdot \log \frac{\vol(V)}{\minvol(U)}},
\]
so that the denominator incorporates a logarithmic dependence on cut imbalance. This produces a lopsided expander decomposition in which every SCC has no \(\psi\)-lopsided sparse cut [2502.05687].

This lopsided notion improves the diameter-expansion tradeoff. Standard expanders yield diameter \(O(\phi^{-1}\log n)\), whereas the cited directed lopsided expander lemma gives diameter
\[
O(\psi^{-1}\log\log n+\log n).
\]
That improvement is precisely what allows a decomposition tuned to the \(\log n\log\log n\) regime rather than the earlier \(\log^2 n\) regime [2502.05687].

The second conceptual ingredient is a multiplicative-weights-update reduction. The paper proves that if, for every cost function \(c\), one can delete edges of total cost at most \(c(E)\cdot \frac{L}{D}\) so that all SCCs have diameter at most \(D\), then an LDD with per-edge cut probability \(O(L/D)\) exists. This converts a global cost-minimization statement into a per-edge probabilistic decomposition statement and is one of the main reasons the 2025 result is both existential and constructive [2502.05687].

Quasipartitions sit adjacent to this picture. They are stronger than LDDs and are particularly useful in embedding and flow-cut arguments. The structured-digraph results of 2026 explicitly separate the two notions: quasipartitions remain the stronger object, but LDDs can sometimes achieve substantially better loss in restricted graph classes [2606.31560].

## 4. Algorithmic schemes: CKR sampling, heavy vertices, and size-aware recursion

The fastest general constructions are recursive partitioning algorithms based on random in-balls and out-balls. At a high level, they repeatedly sample vertices, choose random radii, carve balls in random order, recurse on induced subgraphs, and remove cut edges so that the final SCCs have bounded weak diameter [2505.10244][2510.22721].

A notable simplification is the replacement of geometric ball-growing by the Calinescu-Karloff-Rabani method with uniformly random radii and randomized orderings. The later papers emphasize that this avoids delicate probabilistic tuning tied to memorylessness while retaining the \(O(\log n\log\log n)\) loss factor. In these schemes, vertices are sampled in a degree-sensitive way, balls are processed in random order, and the analysis bounds the number of times an edge can be threatened through a harmonic-number argument over nearby samples [2505.10244][2510.22721].

Heavy vertices are a technical obstacle because very large in-balls or out-balls can disrupt both recursion depth and cut-probability accounting. One paper introduces a modular reduction that lets the main recursion assume all vertices are out-light; if heavy vertices exist, the algorithm either extracts a subgraph of diameter at most \(\Delta\) directly or splits the instance so that the remaining vertices are light [2505.10244]. A later refinement makes heavy-vertex detection sample-efficient: instead of paying a full \(\Theta(\log n)\)-factor overhead from repeated Dijkstra computations, it samples only \(O(\log\log n)\) vertices, allows occasional misclassification, and restarts only with low enough probability to preserve the expected near-linear time bound [2510.22721].

The runtime improvements are not purely analytic. The implementations use shortest-path pruning, careful ordering and skipping of SSSP work, and Thorup’s integer priority queue, yielding \(O(m+n\log\log n)\)-time Dijkstra phases on integral or poly-bounded weights. The recursion depth is controlled by ensuring rapid decrease in edge count, and one analysis shows that each vertex is only considered \(O(\log n\log\log n)\) times in expectation [2505.10244][2510.22721].

A further refinement is Seymour’s trick, or size-aware penalties: larger pieces are cut with a smaller penalty, while smaller pieces can tolerate larger penalties because they recur fewer times. This balances the cumulative loss across recursion levels and is used to keep the total edge-cut probability at \(O\!\left(\frac{\log n\log\log n}{\Delta}\right)w(e)\) [2510.22721].

## 5. Stronger guarantees: separation, exponential non-cut bounds, and small \(D\)

The next stage of the theory strengthened the probabilistic guarantees rather than merely the loss bound. The 2025 paper on stronger directed LDDs introduced the first separated low-diameter decompositions in directed graphs, the first guarantees that remain meaningful for \(D=\Omega(\log\log n)\), and new independence statements for far-apart edge sets [2509.24565].

For a directed graph \(G\) and diameter \(D\), the paper gives ordered \(D\)-diameter clusterings with
\[
\Pr(\text{edge }e\text{ is not cut}) \ge \exp\!\left(-\frac{d_e}{D}\cdot (\log n\log\log n)\right),
\]
and analogous bounds for paths and arbitrary edge sets \(\Gamma\) of total length \(d_\Gamma\). This is stronger than an additive per-edge cut bound because it remains informative when \(d_e\) is a constant fraction of \(D\), whereas additive bounds become vacuous once \(d_e\) exceeds roughly \(D/\log n\) [2509.24565].

The same work formalizes independence at distance. If two edge sets or balls are \(\Omega(D)\) apart in the undirected sense, then the events of not being cut behave as if independent; moreover, the paper states that these events are never worse than independent and may in some cases be positively correlated. Such statements were previously unavailable in directed LDDs [2509.24565].

Separation is obtained for partial clusterings. For parameters \(G,D,d\), there is a distribution over partial clusterings of ordered, strong diameter at most \(D\) such that
\[
\Pr(\text{vertex }v\text{ is clustered}) \ge \exp\!\left(-\frac{d}{D}\cdot (\log n\log\log n)\right),
\]
and if cluster \(C\) is ordered after \(C'\), then every \(u\in C\) and \(v\in C'\) satisfy \(\operatorname{dist}_G(u,v)>d\). The implementation uses truncated exponential radii, eliminating restarts, and a vertex-marking rule that leaves some vertices unclustered near ball boundaries to enforce separation [2509.24565].

An important limitation is explicit: preserving entire neighborhoods, as in some undirected formulations, is impossible in the directed setting. The paper presents the separation property for partial clusterings as the best achievable generalization in this direction [2509.24565].

## 6. Applications, structured digraphs, and current limits

Directed LDDs function as a black-box primitive in several algorithmic settings. The cited work lists cut sparsifiers, network design, multicut, directed flow problems, approximation algorithms, and parallel, distributed, and dynamic graph algorithms as natural application domains [2505.10244][2510.22721].

The most explicit 2025 application is to negative-weight single-source shortest paths on integer-weighted digraphs. Using the faster directed LDD, one paper obtains an algorithm running in
\[
O((m+n\log\log n)\log(nW)\log n\log\log n),
\]
where \(W\) is the maximum absolute value of a negative edge. The decomposition is applied to \(G_{\ge 0}\), the graph obtained by setting negative edges to zero, and then integrated into a scaling framework that localizes the problem to small-diameter components and bounds the number of inter-component crossings on negative paths [2510.22721].

In structured digraphs, the theory departs sharply from the general \(O(\log n\log\log n)\) regime. The 2026 paper on structured digraphs proves that every directed graph of pathwidth \(\mathsf{pw}\) admits an \((O(\mathsf{pw}),\Delta)\)-LDD, improving the previous best-known \((2^{O(\mathsf{pw}^2)},\Delta)\)-LDD that had been obtained only implicitly via quasipartitions. Its main technical device is “scissors carving,” which alternates out-balls from one endpoint of a critical path and in-balls from the other, ensuring that each recursive step either shortens the relevant subpath, reduces pathwidth, or already meets the diameter target [2606.31560].

The same structured-graph line also improves the integrality gap of the Directed Non-Bipartite Sparsest-Cut LP relaxation on \(n\)-vertex graphs of treewidth \(\mathsf{tw}\) to \(O(\mathsf{tw}\log n)\), from the earlier \(O(\mathsf{tw}\log^2 n)\). That result is obtained through a refined analysis of quasipartitions rather than ordinary LDDs, illustrating again that LDDs and quasipartitions are related but non-equivalent tools [2606.31560].

Several limits remain visible in the current literature. Weak diameter is much better understood than strong diameter; meaningful guarantees for small \(D\) only recently reached \(D=\Omega(\log\log n)\); and in dynamic settings the strongest fully dynamic LDD results in the cited corpus are still for undirected multigraphs. The dynamic random-shift hierarchy work notes that decremental ES-tree methods can be adapted to directed weighted graphs with positive integer weights, which suggests a route toward directed dynamic LDDs, but the fully dynamic guarantees themselves are not established there for directed graphs [1804.04928].

Source: https://www.emergentmind.com/topics/directed-low-diameter-decompositions