Sparse Length-Constrained Flows

• The paper establishes that sparse length-constrained flows are multi-commodity flows with strictly bounded routing paths, using expander decompositions to ensure efficient flow support.
• It introduces a refined union-of-cuts lemma and decomposition techniques that reduce sparsity loss and maintain structural guarantees in both edge- and vertex-capacitated settings.
• Algorithmic realizations achieve near-linear work and low parallel depth, balancing trade-offs between path-length slack, congestion, and approximation guarantees for multi-commodity flows.

Sparse length-constrained flows are multi-commodity flows in graphs for which the routing paths are strictly bounded in length (typically parameterized by a pair $(h, s)$) and the total path support is restricted (sparse). This paradigm, deeply intertwined with length-constrained expander decompositions, underpins both contemporary algorithm design for min-cost flows and the emergence of efficient parallel and dynamic graph primitives. The foundational results characterize the structural conditions under which such flows exist, the efficient construction of supporting decompositions, and the trade-offs between cut-size, path-length slack, and congestion. The development of sparse length-constrained flows applies to undirected and directed graphs and accommodates both edge- and vertex-capacitated settings (Haeupler et al., 29 Mar 2025, Bodwin et al., 11 Oct 2025, Haeupler et al., 2024).

1. Foundations: Length-Constrained Expansion and Sparsity

A central construct is the moving cut $C$, a function $C:E\to [0,1]$ (or $C:E\cup V \to [0,1]$ for vertex capacities), representing a fractional increase to edge (or edge and vertex) lengths. For a length bound $h$, a node-weighting $A:V\to \mathbb{R}_{\ge0}$, and a demand $D:V\times V\to\mathbb{R}_{\ge0}$ (A-respecting and $h$-length), the cut’s sparsity is quantified as

$$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$

where $|C|$ is the capacity-weighted sum of cut weights, and $C$0 is the maximum $C$1-length, $C$2-respecting demand $C$3 that is $C$4-separated by $C$5 ($C$6 for each $C$7 with $C$8). $C$9 is $C:E\to [0,1]$0-length $C:E\to [0,1]$1-sparse if this ratio is at most $C:E\to [0,1]$2 (Bodwin et al., 11 Oct 2025).

$C:E\to [0,1]$3 is an $C:E\to [0,1]$4-length $C:E\to [0,1]$5-expander (w.r.t.\ $C:E\to [0,1]$6) if every $C:E\to [0,1]$7-length moving cut has sparsity at least $C:E\to [0,1]$8. This condition characterizes the graph’s intrinsic ability to support bounded-length flows without bottlenecks.

2. Length-Constrained Expander Decomposition

The principal decomposition theorem asserts that for any undirected (or directed) graph $C:E\to [0,1]$9 with $C:E\cup V \to [0,1]$0 nodes and $C:E\cup V \to [0,1]$1 edges (or arcs), and for parameters $C:E\cup V \to [0,1]$2, $C:E\cup V \to [0,1]$3, and $C:E\cup V \to [0,1]$4, there exists an $C:E\cup V \to [0,1]$5-length $C:E\cup V \to [0,1]$6-expander decomposition $C:E\cup V \to [0,1]$7 of size at most $C:E\cup V \to [0,1]$8 (Bodwin et al., 11 Oct 2025). This improves upon previous logarithmic and polylogarithmic overheads for the same operation (Haeupler et al., 2024), reducing the sparsity loss in the union of sparse cuts from $C:E\cup V \to [0,1]$9 to $h$0.

The decomposition provides a strategy: repeatedly remove the sparsest $h$1-length cut until the remaining graph is an expander for the chosen parameters. The cumulative cut $h$2 maintains the desired sparsity and slack, ensuring the decomposition is both succinct and efficient for flow routing.

For undirected vertex-capacitated graphs, as well as directed edge-capacitated graphs, analogous decomposition bounds apply, often with an adjustable slack parameter $h$3, yielding cut-size overhead $h$4 for any $h$5 (Haeupler et al., 29 Mar 2025).

3. Union-of-Cuts Lemma and Its Structural Consequences

A crucial structural result is the union-of-cuts lemma: if $h$6 are sequentially found, $h$7-length $h$8-sparse cuts (each after applying the previous), then their sum (after scaling) forms a $h$9-length $A:V\to \mathbb{R}_{\ge0}$0-sparse cut. The proof utilizes the demand-matching graph technique, establishes arboricity bounds for $A:V\to \mathbb{R}_{\ge0}$1-parallel-greedy graphs, and performs dispersed demand decomposition using forest covers. The resulting sparsity bound underpins all existential guarantees for decomposition size and informs practical algorithms (Bodwin et al., 11 Oct 2025, Haeupler et al., 2024).

Parameter	Prior Bound	Improved Bound
Sparsity loss	$A:V\to \mathbb{R}_{\ge0}$2	$A:V\to \mathbb{R}_{\ge0}$3 (Bodwin et al., 11 Oct 2025)

This improvement is not merely technical; it yields substantially tighter existential and algorithmic bounds for all downstream algorithmic applications.

4. Routing and Max-Flow/Min-Cut Equivalence for Sparse Flows

The flow-cut duality carries over to the length-constrained regime: a graph is $A:V\to \mathbb{R}_{\ge0}$4-length $A:V\to \mathbb{R}_{\ge0}$5-expanding (w.r.t.\ $A:V\to \mathbb{R}_{\ge0}$6) if and only if every $A:V\to \mathbb{R}_{\ge0}$7-length $A:V\to \mathbb{R}_{\ge0}$8-respecting demand $A:V\to \mathbb{R}_{\ge0}$9 can be routed in paths of length at most $D:V\times V\to\mathbb{R}_{\ge0}$0 with congestion at most $D:V\times V\to\mathbb{R}_{\ge0}$1. The converse also holds with a matching lower bound: if such an expander property fails, there exists a demand for which any $D:V\times V\to\mathbb{R}_{\ge0}$2-length routing must incur congestion at least $D:V\times V\to\mathbb{R}_{\ge0}$3 (Haeupler et al., 29 Mar 2025, Bodwin et al., 11 Oct 2025).

This correspondence enables length-constrained expander decompositions to directly imply existence and structure of sparse length-constrained flows. Since the expander property guarantees robust separation for all demand patterns, the flows produced can be efficiently decomposed into $D:V\times V\to\mathbb{R}_{\ge0}$4 length-bounded paths, allowing for sparsity in both path support and congestion profiles (Haeupler et al., 2024).

5. Construction and Hierarchies of Length-Constrained Flow Shortcuts

Sparse flow shortcut constructions leverage a recursive application of length-constrained expander decompositions to build small sets of auxiliary edges and vertices (Steiner nodes). Given a graph $D:V\times V\to\mathbb{R}_{\ge0}$5 and parameters $D:V\times V\to\mathbb{R}_{\ge0}$6, the construction proceeds in $D:V\times V\to\mathbb{R}_{\ge0}$7 levels. At each level and for each relevant bucket $D:V\times V\to\mathbb{R}_{\ge0}$8, a corresponding $D:V\times V\to\mathbb{R}_{\ge0}$9-length $h$0-expander cut is computed. The residual clusters are augmented by star graphs centered at Steiner roots, facilitating shortcut routes for path segments that would otherwise be long. The main guarantees are:

• The shortcut edge set has $h$1 total size.
• Any $h$2-length flow in $h$3 is routable in $h$4 with path length $h$5 and step count $h$6.
• Congestion after shortcutting and reverting is bounded by $h$7 (Haeupler et al., 29 Mar 2025).

The hierarchical path decomposition ensures that only the end segments of long paths contribute to recursive rounds, and that all middle segments are efficiently redirected via shortcut stars, maintaining both sparsity and path-length constraints.

6. Algorithmic Realizations and Tradeoffs

For undirected graphs with polynomially bounded integer lengths and capacities, near-optimal sparse length-constrained flows can be constructed in near-linear work $h$8 and low parallel depth $h$9 (Haeupler et al., 2024). The design leverages multiplicative-weights updaters, blaming flows, and lightest-path blockers to keep support size $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$0 and congestion $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$1. Dynamization and extension to weighted demands rely on robust linkage and rounding techniques derived from the union-of-cuts structure.

A trade-off emerges:

• Smaller decomposition size (better sparseness) increases length-slack $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$2 (longer path allowance) and vice versa.
• For any fixed $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$3: path length scales as $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$4, decomposition size grows as $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$5.

Known corollaries include:

• $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$6-approximate multi-commodity min-cost flow in work $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$7 and depth $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$8 for constant $$\mathrm{spars}_{(h,s)}(C, A) = \frac{|C|}{A_{(h,s)}(C)}$$9.
• $|C|$0-approximate distance oracles with $|C|$1 update and $|C|$2 query time in fully dynamic graphs.
• Parallel algorithms for $|C|$3-approximate $|C|$4-commodity flows in $|C|$5 time (Bodwin et al., 11 Oct 2025, Haeupler et al., 2024).

7. Extensions and Generalizations

The theory has been extended from undirected edge-capacitated graphs to:

• Directed, edge-capacitated graphs: all key results (existence, decomposition, shortcutting) generalize directly (Haeupler et al., 29 Mar 2025).
• Undirected, vertex-capacitated graphs: sparsity bounds and shortcut constructions are parallel to those for edge capacities.
• Practical implementations assume integer lengths and capacities; extensions to real weights require technical adaptations.

The length-constrained expander decomposition and sparse length-constrained flows constitute the current foundation for sublinear- and parallel-time algorithms in graph flow, distances, and clustering. Open directions remain in optimizing the constants for cut slack, length slack, and in developing simultaneously constant-factor optimal decomposition and routing length (Bodwin et al., 11 Oct 2025, Haeupler et al., 2024).