---
title: Directed Expander Decomposition
url: https://www.emergentmind.com/topics/directed-expander-decomposition
type: topic
---

# Directed Expander Decomposition

Searching arXiv for recent papers on directed expander decomposition and closely related variants.
{"query":"directed expander decomposition arXiv directed expander decomposition near-optimal algorithm capacitated length-constrained", "max_results": 10}
I found several relevant arXiv papers:
- 2403.04542 — “Near-Optimal Algorithm for Directed Expander Decompositions”
- 2507.09729 — “Improved Directed Expander Decompositions”
- 2503.23217 — “Length-Constrained Directed Expander Decomposition and Length-Constrained Vertex-Capacitated Flow Shortcuts”
- 2204.02519 — “Maintaining Expander Decompositions via Sparse Cuts”
- 2005.02369 — “The Expander Hierarchy and its Applications to Dynamic Graph Algorithms”

I’ll use these to ground the article.
Directed expander decomposition is a family of structural and algorithmic frameworks for partitioning a directed graph into induced subgraphs that are themselves directed expanders, while isolating only a controlled set of inter-component edges that violate a global acyclic orientation. In the conductance-based formulation used by recent work, the central difficulty is that directed acyclic graphs cannot contain nontrivial directed expanders, so a decomposition cannot hope to bound all inter-cluster edges; instead, it bounds only those edges that go “backwards” relative to a topological order of the quotient graph. Since 2024, the area has moved from almost-optimal unweighted decremental frameworks to near-optimal static and dynamic algorithms, near-linear capacitated algorithms, and length-constrained variants with explicit multicommodity routing characterizations [2403.04542] [2507.09729] [2503.23217].

## 1. Directed expansion and the decomposition primitive

For a directed graph \(G=(V,E)\), recent work uses the undirected degree
\[
\deg_G(v)=|\{e\in E:e\text{ incident to }v\}|
\]
and the corresponding volume
\[
\vol_G(S)=\sum_{v\in S}\deg_G(v)
\]
for \(S\subseteq V\). If \(E_G(A,B)\) denotes directed edges from \(A\) to \(B\), and \(e_G(A,B)=|E_G(A,B)|\), then a cut \((S,\bar S)\) is \(\phi\)-out-sparse when \(\vol_G(S)\le e(G)\) and \(e_G(S,\bar S)<\phi\cdot \vol_G(S)\). A graph is a \(\phi\)-out-expander if it has no \(\phi\)-out-sparse cut. A directed \(\phi\)-expander is defined symmetrically: both \(G\) and its reversal \(G^{\mathrm{rev}}\) must be \(\phi\)-out-expanders, equivalently every nontrivial \(S\subset V\) with \(\vol_G(S)\le e(G)\) satisfies
\[
\min\{e_G(S,\bar S),e_G(\bar S,S)\}\ge \phi\cdot \vol_G(S).
\]
This is a purely combinatorial notion; no spectral definition is involved [2403.04542].

The basic directed decomposition problem asks for a partition \(\mathcal X\) of \(V\) and a set of removed or special edges \(E^r\subseteq E\) such that every induced subgraph \(G[X]\), \(X\in\mathcal X\), is a directed \(\Omega(\phi)\)-expander, the number of special edges is bounded by \(|E^r|\le \beta\phi m\), and the quotient graph \((G\setminus E^r)/\mathcal X\) is acyclic. The DAG condition is the directed analogue of the small-boundary condition in undirected expander decomposition. It records that inter-component edges are globally well-oriented except for the few edges in \(E^r\). A common misconception is that directed decomposition should simply minimize all inter-cluster edges; the DAG obstruction shows why this is impossible in general [2403.04542].

A related decremental formulation defines a directed expander decomposition of a directed graph \(G\) as a partition \(\mathcal X\) and a set \(E_{\mathrm{Rest}}\subseteq E\) such that each \(G[X]\) is a directed \(\phi\)-expander, \(|E_{\mathrm{Rest}}|\le \beta\phi m\), and the contracted graph \(G/\{X_i\}_i\setminus E_{\mathrm{Rest}}\) is a DAG. In that setting, the decomposition is maintained under edge deletions, vertex splits, and self-loop insertions, and the algorithmic objective is often stronger than a one-shot decomposition: the partition should refine over time, while \(E_{\mathrm{Rest}}\) should only grow [2204.02519].

Capacitated variants replace edge counts by total capacities. For a directed capacitated graph \(G=(V,E,\mathbf c)\), one writes \(\delta_G(S,T)\) for the total capacity from \(S\) to \(T\), and some formulations measure conductance relative to a vertex weighting \(\mathbf d\):
\[
\Phi_{G,\mathbf d}(S)=\frac{\min(\delta_G(S,V\setminus S),\delta_G(V\setminus S,S))}{\min(\mathbf d(S),\mathbf d(V\setminus S))}.
\]
A graph is then a \((\phi,\mathbf d)\)-expander if this quantity is at least \(\phi\) for all \(S\subseteq V\). This weighted formulation is the basis for near-linear capacitated decomposition algorithms [2507.09729].

## 2. Witnesses, cut–matching, and flow certificates

A defining feature of the directed literature is the use of witness graphs. In the augmented framework, a pair \((W,\Pi)\) is a \(\psi\)-witness for \(G\) if \(W\) is itself a \(\psi\)-expander on the same vertex set, \(\Pi\) embeds every edge of \(W\) into a path of \(G\) with bounded congestion, and degree bounds satisfy
\[
\deg_G(v)\le \deg_W(v)\le \frac{\deg_G(v)}{\psi}.
\]
A standard fact quoted in the 2024 near-optimal algorithm is that if \((W,\Pi)\) is a \(\psi\)-witness for \(G\), then \(G\) is a \(\phi\)-expander. Intuitively, a sparse cut in \(G\) would induce a sparse cut in \(W\), contradicting that \(W\) is an expander. This witness formalism is what the dynamic algorithms actually maintain [2403.04542].

The witness viewpoint also appears in decremental sparse-cut frameworks. There, an \((R,\phi,\psi)\)-witness of \((G,r)\) is a directed graph \(W\) on the same vertex set together with an embedding into \(G\), a slack vector \(r\), and a balance vector \(\gamma\), such that \(W\) certifies out-expansion in both \(G\) and \(\overleftarrow{G}\). The key implication is that if such a witness exists for \((G,0)\), then \(G\) is a \(\psi^2\phi\)-expander. Witness degradation under deletions is handled by repairing the embedding or extracting a genuine sparse cut [2204.02519].

Static decomposition algorithms typically begin with a cut–matching game. In the 2024 near-optimal algorithm, the directed cut–matching game either returns a balanced sparse cut or returns a witness \((W,\Pi)\) certifying that \(G\cup \mathcal F\) is a \(\phi\)-expander for a small set of fake edges \(\mathcal F\). The fake edges are then removed by a pruning subroutine, which extracts a large expander core and recursively decomposes the remaining pieces. This “witness plus pruning” paradigm is central to the 2024 framework [2403.04542].

The 2025 improved framework replaces the fake-edge cleanup paradigm by the first implementation and analysis of the non-stop cut-matching game for directed, capacitated graphs. The algorithm outputs disjoint sparse cuts \(C_1,\dots,C_k\), and either the total removed weight is already substantial or the remaining vertex weighting mixes in the graph with congestion \(O(T/\phi)\), where \(T=O(\log n\log(nW))\). This is the directed analogue of the Saranurak–Wang undirected template, but its analysis requires active/deleted vertex clones, directed bisections, and separate row- and column-mixing arguments [2507.09729].

## 3. Algorithmic development and quantitative guarantees

The modern literature can be organized around three milestones: a decremental sparse-cut framework, a near-optimal witness-and-pruning framework, and a capacitated non-stop cut-matching framework.

| Paper | Model | Stated guarantee |
|---|---|---|
| “Maintaining Expander Decompositions via Sparse Cuts” [2204.02519] | Directed dynamic graph under edge deletions, vertex splits, and self-loop insertions | Maintain a \((\gamma,\phi/\gamma)\)-expander decomposition in total time \(\tilde O(m^{1+1/L_{\max}}\cdot \gamma/\phi^2)\), with refinement and monotone \(E_{\mathrm{Rest}}\) |
| “Near-Optimal Algorithm for Directed Expander Decompositions” [2403.04542] | Static and decremental directed graphs | Static time \(O(m\log^4 m/\phi)\); dynamic initialization \(O(m\log^5 m/\phi)\); amortized deletion time \(O(\log^7 m/\phi^2)\) |
| “Improved Directed Expander Decompositions” [2507.09729] | Directed capacitated graphs | Uncapacitated time \(O(m\log^4 n/\phi)\); capacitated time \(O(m/\phi+m\log^7(nW))\) |

In the decremental sparse-cut framework, the guiding meta-algorithm is direct: repeatedly find a \(\phi\)-sparse cut inside some cluster, split the cluster, and add the smaller directed side of the cut to the global removed-edge set. The significance of the 2022 result is not only the \(m^{o(1)}/\phi^2\)-type update bound, but also the structural guarantees: the maintained partition is always a refinement of the previous one, and the cut set is monotonically nondecreasing. The paper emphasizes that this is the first approach guaranteeing a sublinear \(\phi m^{1+o(1)}\) bound on the total number of edges crossing between clusters across the entire update sequence [2204.02519].

The 2024 near-optimal algorithm improves over previous directed algorithms that were optimal only up to subpolynomial factors. For \(\phi\le c/\log^2 m\), it maintains a \((\beta,\phi')\)-directed expander decomposition under edge deletions with
\[
\beta=O(\log^6 m),\qquad
\phi'=\Omega(\phi/\log^8 m),
\]
initialization time
\[
O\!\left(\frac{m\log^5 m}{\phi}\right),
\]
and amortized deletion time
\[
O\!\left(\frac{\log^7 m}{\phi^2}\right).
\]
Its static algorithm computes a \(\bigl(O(\log^6 m),\phi,O(1/\log^2 m)\bigr)\)-expander decomposition in time
\[
\tilde O(m/\phi)=O\!\left(\frac{m\log^4 m}{\phi}\right).
\]
For constant \(\phi\) up to polylogarithmic factors, this is near-linear and near-optimal [2403.04542].

The 2025 improved results match the guarantees of Saranurak and Wang for undirected expander decomposition, now in directed graphs. For uncapacitated graphs, the algorithm computes a partition \(V_1,\ldots,V_k\) and an acyclic subgraph \(D\) such that each induced \(G[V_i]\) is a \(\phi\)-expander, the number of inter-component edges in \(G-D\) is \(O(\phi m\log^3 n)\), and the running time is \(O(m\log^4 n/\phi)\). For capacitated graphs with capacities in \(\{1,\dots,W\}\), it computes a partition and acyclic subgraph with each part a \(\phi\)-expander, total capacity of inter-component edges in \(G-D\) equal to \(O(\phi m\log^8(nW))\), and runtime
\[
O(m/\phi+m\log^7(nW)).
\]
This is stated as the first directed expander decomposition algorithm for capacitated graphs in near-linear time with optimal dependence on \(\phi\) [2507.09729].

## 4. Dynamic maintenance and the push–pull–relabel framework

The decisive technical innovation of the 2024 near-optimal decremental algorithm is the push–pull–relabel framework, a generalization of classical push–relabel from pre-flows to arbitrary pseudo-flows. Classic push–relabel warm-starts successfully in undirected expander pruning because restricting a solved flow to an induced subgraph preserves the pre-flow property. In the directed setting, after cutting off a subset, the restricted object may have negative excess; the resulting object is only a pseudo-flow. This breaks the standard invariants and explains the extra complexity of earlier directed frameworks [2403.04542].

The new framework maintains a state \((f,\ell)\), where \(f\) is a pseudo-flow and \(\ell:V\to\{0,\dots,h\}\) is a label function satisfying a saturation condition on edges that are “too far downhill.” A valid state additionally requires vertices with negative excess to lie at level \(0\) and vertices with positive excess to lie at level \(h\). The data structure supports two updates: `IncreaseSource(δ)`, which increases the source vector by a nonnegative integer vector, and `RemoveVertices(S)`, which restricts the flow problem to the induced subgraph on the remaining vertices. These are exactly the operations needed when deleted or fake edges create new source mass and when recursive pruning cuts off a set of vertices [2403.04542].

Algorithmically, the framework combines two symmetric routines. `PushRelabel` resolves positive excess using push operations from higher to lower labels and relabels by increasing \(\ell(v)\). `PullRelabel` resolves negative excess using a new pull operation that pulls flow back from a higher-level neighbor \(u\) to \(v\) when \(\ell(u)=\ell(v)+1\), with relabels now decreasing \(\ell(v)\). Because labels can move both upward and downward, monotonicity is lost; the analysis instead uses the potential
\[
\Phi(t)=\sum_v \Delta_t^+(v)\ell_t(v)+\sum_v \Delta_t^-(v)(h-\ell_t(v)).
\]
The resulting complexity bound states that over a sequence of source increases and vertex removals, the total time to maintain valid states is
\[
O\!\left(h\cdot\Bigl(\|\Delta\|_1+\sum_t \min\bigl((\tilde V_{t+1},S_t),(S_t,\tilde V_{t+1})\bigr)+\nabla(S_t)\Bigr)\right),
\]
where \(h=O(\log n/\phi)\). This is the mechanism that makes full dynamization possible with only polylogarithmic overhead [2403.04542].

The 2022 decremental sparse-cut framework solves the same high-level maintenance problem differently. It maintains multiple levels of witnesses \(W_{X,l}\), slack vectors \(r_{X,l}\), and balance vectors \(\gamma_X\) for each cluster \(X\), and triggers repair whenever \(\|r_{X,l}\|_1\) exceeds a scale-dependent threshold. Repair invokes a local directed flow computation `PruneOrRepair`, which either returns a directed sparse out-cut or produces a new witness with reduced \(\psi\). The resulting decomposition always refines previous decompositions, and the removed-edge set only grows. This monotonicity is central in applications such as decremental SCC maintenance [2204.02519].

The 2025 capacitated framework returns to a cut-matching-plus-trimming organization, but now with a native directed analysis. Partially matched vertices are handled by splitting each vertex into active and deleted clones; mixing is certified through two potentials, one for rows and one for columns of the flow matrix; and a grafting step either reincorporates difficult vertices into the near-expander or identifies additional sparse cuts. This shows that the natural undirected non-stop cut-matching approach applies even to directed graphs, though only after substantial modifications [2507.09729].

## 5. Capacitated and length-constrained variants

Capacitated directed expander decomposition is not a straightforward corollary of unweighted results. The 2025 improved algorithms regularize the vertex weighting by
\[
\mathbf d(v):=\deg_G(v)+\frac{\widetilde{\deg}_G(v)}{2m}\cdot \deg_G(V),
\]
where \(\widetilde{\deg}_G(v)\) is the unweighted degree. This weighting satisfies \(\mathbf d(v)\ge \deg_G(v)\) and \(\mathbf d(V)=2\deg_G(V)\), and it allows the dynamic push-relabel analysis inside trimming to bound the total work by \(O(m)\) times the height factor. Because \(\mathbf d\) dominates the usual degree measure, a \((\phi,\mathbf d)\)-expander is automatically a \((\phi,\deg_G)\)-expander. This regularization is what permits near-linear capacitated runtimes with optimal \(1/\phi\) dependence [2507.09729].

A distinct generalization is length-constrained directed expander decomposition. Here the graph has edge lengths \(\ell_G:E\to\mathbb N^+\) and capacities \(u_G:E\to\mathbb R_{>0}\), demands are constrained to pairs at directed distance at most \(h\), and “cuts” become moving cuts \(C:E\to\{0,\frac1h,\frac2h,\dots,1\}\) that increase edge lengths rather than delete edges. The size of a moving cut is
\[
|C|=\sum_{e\in E}u_G(e)C(e),
\]
and its effect is to replace \(\ell_G(e)\) by \(\ell_G(e)+hC(e)\). Separation is measured by how much demand becomes farther than a threshold after the length increase [2503.23217].

Given a node-weighting \(A\), the \((h,s)\)-length conductance \(\mathrm{cond}_{(h,s)}(A)\) is the minimum, over \(h\cdot s\)-length moving cuts, of the worst-case sparsity against symmetric \(A\)-respecting \(h\)-length demands. One says that \(A\) is \((h,s)\)-length \(\phi\)-expanding if \(\mathrm{cond}_{(h,s)}(A)\ge \phi\). An \((h,s)\)-length \(\phi\)-expander decomposition for \(A\) with cut slack \(\kappa\) is then an \(h\cdot s\)-length moving cut \(C\) such that \(A\) is \((h,s)\)-length \(\phi\)-expanding in \(G-C\) and
\[
|C|\le \kappa\phi|A|.
\]
This is a genuine extension of the decomposition paradigm: the inter-cluster defect is encoded as added length, not deleted capacity [2503.23217].

The corresponding routing theorem states that if \(A\) is \((h,s)\)-length \(\phi\)-expanding in \(G\), then every \(A\)-respecting symmetric \(h\)-length demand can be routed with congestion \(O(\log N/\phi)\) and length at most \(h\cdot s\). Conversely, if \(A\) is not \((h,s)\)-length \(\phi\)-expanding, then there exists an \(A\)-respecting symmetric \(h\)-length demand that cannot be routed with congestion at most \(1/(2\phi)\) and length at most \(h\cdot s/2\). The decomposition theorem then states that for any directed graph \(G\), node-weighting \(A\), \(h>1\), \(\alpha\ge 1\), \(\phi<1\), and \(s=O(\log n)\), there exists a directed \((h,s)\)-length \(\phi\)-expander decomposition for \(A\) with cut slack
\[
\kappa=O\bigl(n^{O(1/s)}\log n\bigr).
\]
This extends length-constrained expander decomposition beyond the undirected edge-capacitated setting [2503.23217].

## 6. Structural significance, applications, and open directions

Directed expander decomposition has become a standard primitive in modern graph algorithms for flows, shortest paths, and dynamic reachability. The 2024 near-optimal algorithm is already used as a building block in the currently fastest min-cost flow algorithm, and the authors further state that it can speed up and simplify recent breakthroughs in combinatorial graph algorithms toward fast maximum flow algorithms. The same paper explicitly connects directed expander decompositions to min-cost flow and to recent combinatorial maximum-flow and bipartite-matching algorithms, where decomposition localizes augmenting paths, reduces global congestion, and supports dynamic maintenance of residual structures [2403.04542].

In decremental directed graphs, the monotone decomposition of the 2022 sparse-cut framework yields especially clean applications. The paper derives a simple decremental SCC algorithm by combining the monotone set \(E_{\mathrm{Rest}}\) with a theorem of Łącki for maintaining SCCs under monotone edge deletions outside the current SCC partition. The conceptual point is that refinement of clusters and monotonicity of removed edges are not merely cosmetic; they are algorithmically load-bearing properties [2204.02519].

The relation to expander hierarchies is more delicate. The 2020 “expander hierarchy” framework is formulated for undirected graphs and explicitly states that there is no sensible generalization of expander hierarchies to directed graphs in that development. Its boundary-linked decomposition and hierarchy tree are nevertheless relevant context because they explain why dynamic graph algorithms seek stronger cluster interfaces than plain conductance guarantees. A plausible implication is that directed decomposition still lacks a comparably canonical hierarchy formalism, even though directed dynamic maintenance now exists [2005.02369].

Several misconceptions can be stated precisely. Directed expander decomposition is not an undirected decomposition of the underlying graph followed by arbitrary orientation; the decomposition primitive itself is directional, and the DAG condition is essential. Nor is the area confined to ordinary edge deletion: length-constrained decomposition replaces deletion by moving cuts that increase edge lengths. Finally, fake edges are not intrinsic to the directed setting. The 2024 near-optimal algorithm still uses them as an intermediate certificate that must be pruned away, but the 2025 improved algorithms show that a direct non-stop cut-matching approach can avoid them entirely, even in capacitated graphs [2403.04542] [2507.09729].

The open directions stated across the recent literature are concrete. The 2024 near-optimal paper points to reducing the \(\log^8 m\) loss in \(\phi'\), the \(\log^6 m\) factor in \(\beta\), and the \(\log^7 m\) factor in the update time, extending beyond the decremental model to fully dynamic directed graphs, and generalizing more sharply to weighted or capacitated settings. The length-constrained work identifies almost-linear-time algorithms as a target for future work and suggests that directed flow shortcuts remain out of reach despite the new decomposition theorem. The 2025 improved results indicate no fundamental barrier to matching the exact Saranurak–Wang polylogarithmic factors for capacitated directed graphs. Taken together, these results suggest that directed expander decomposition has progressed from a difficult specialized tool to a central, modular primitive, but that the fully dynamic, hierarchy-level, and length-aware theories are still evolving [2403.04542] [2503.23217] [2507.09729].

Source: https://www.emergentmind.com/topics/directed-expander-decomposition