Congestion-Preserving DAG Projections

• The paper introduces a randomized algorithm that constructs DAG projections from directed graphs, preserving maximum flow within a n^(o(1)) congestion factor.
• It employs expander decomposition and recursive layering to ensure that network cuts and flows are approximately maintained between the original graph and its DAG projection.
• Its efficient, nearly-linear and parallelizable construction paves the way for practical applications in max-flow, cut, and distance problem approximations in large-scale networks.

Congestion-preserving DAG projections are algorithmic constructions that transform arbitrary directed graphs into directed acyclic graphs (DAGs) in a manner that approximately preserves network flow capacities between vertex subsets. These projections facilitate the transfer of combinatorial and parallel algorithmic advances—traditionally applicable only on DAGs—to the broader class of general directed graphs. Central to their utility are efficient projection algorithms, bounded blow-up in edge count and vertex width, and quantifiable guarantees on flow and cut preservation. The paradigm provides the first non-trivial reductions for approximating maximum flow in general digraphs via simpler, more structured DAG instances, with substantial algorithmic implications for flow, cut, and distance problems.

1. Formal Definition and Projection Structure

Let $G=(V,E,U_G)$ be a directed graph with positive capacities $U_G(e)$. A congestion-preserving DAG projection consists of a triplet: a DAG $D=(V',E',U_D)$, a projection map $\pi: V' \to V\cup\{\bot\}$, and associated projection algorithms. Vertices mapping to $V$ are copies, while those mapping to $\bot$ are Steiner (dummy) vertices. The projection maintains that $D$ is acyclic, and that any edge between copies of real vertices corresponds directly to an edge in $G$, exactly replicating its capacity.

Define the width as $w = \max_{v\in V} |\pi^{-1}(v)|$. The crucial property is the $(\kappa, \delta)$-congestion-preserving guarantee: for parameters $U_G(e)$0, there exist efficient algorithms mapping flows and cuts between $U_G(e)$1 and $U_G(e)$2 such that for feasible flows and corresponding cuts: $U_G(e)$3 For the leading constructions, $U_G(e)$4 and $U_G(e)$5 (Haeupler et al., 6 Apr 2026).

2. Main Theorems and Existential Guarantees

The foundational result establishes that, for any directed graph $U_G(e)$6 with $U_G(e)$7 vertices and $U_G(e)$8 edges, and integer capacities up to $U_G(e)$9, there exists a randomized algorithm with $D=(V',E',U_D)$0 running time and $D=(V',E',U_D)$1 parallel depth that outputs a DAG projection $D=(V',E',U_D)$2 satisfying:

• $D=(V',E',U_D)$3
• $D=(V',E',U_D)$4
• $D=(V',E',U_D)$5

This ensures that all-pairs or subset-pairs maximum flows in $D=(V',E',U_D)$6 can be approximated within an $D=(V',E',U_D)$7 factor via computations on $D=(V',E',U_D)$8 (Haeupler et al., 6 Apr 2026).

The construction leverages structural decompositions and offers strong existential and algorithmic guarantees. The only prior related results for distance preservation suffered $D=(V',E',U_D)$9 loss [Assadi, Hoppenworth, Wein, STOC'25; Filtser, SODA'26], and no comparable max-flow-preserving transformation was previously established.

3. Construction via Expander Hierarchy

The design of the projection is grounded in expander decompositions. Core subroutines include:

• The directed cut–matching game, producing near-perfect bipartite matchings for balanced demand vectors.
• Nonstop expander decomposition, which recursively removes sparse cuts until only expander components remain.

Composing these over $\pi: V' \to V\cup\{\bot\}$0 phases yields an expander hierarchy:

• Edge sets $\pi: V' \to V\cup\{\bot\}$1 partition $\pi: V' \to V\cup\{\bot\}$2.
• Ordered clusterings $\pi: V' \to V\cup\{\bot\}$3 with each $\pi: V' \to V\cup\{\bot\}$4 being $\pi: V' \to V\cup\{\bot\}$5-constrained and $\pi: V' \to V\cup\{\bot\}$6-expanding ($\pi: V' \to V\cup\{\bot\}$7).
• Each phase creates layers in the DAG, modeling the recursive structure and inter-cluster connectivity.

The resulting DAG is created by (i) concatenating projections for child clusters, (ii) replicating the union to obtain multiple copies, (iii) introducing Steiner vertices joining replicas with prescribed capacities, and (iv) scaling capacities to control congestion. This recursive gluing maintains acyclicity and prepares the structure for efficient routing with bounded congestion.

4. Congestion Analysis and Guarantees

Projections preserve flow and cut properties up to a factor governed by the hierarchy structure and expander properties. Any flow in $\pi: V' \to V\cup\{\bot\}$8 can be simulated in $\pi: V' \to V\cup\{\bot\}$9 with congestion increase by at most $V$0, arising from the following sources:

• Width blow-up by at most $V$1 due to the number of copies.
• At each expander hierarchy layer, routing through dummy vertices (Steiner hops) demands an additional congestion blow-up of $V$2.
• The total number of layers is $V$3, compounding the overall congestion to within $V$4.

These congestion bounds are symmetric for projection of flows from $V$5 to $V$6 and vice versa, with corresponding guarantees for projected cuts.

5. Algorithmic Parallelization and Runtime

All construction and projection phases—expander decomposition, cut–matching, reachability, and flow decomposition—admit algorithms with $V$7 work and $V$8 parallel depth on CRCW PRAM models. As a result, the entire pipeline from input graph to flow-preserving DAG projection and associated mappings can be performed efficiently in almost-linear time and subpolynomial parallel depth (Haeupler et al., 6 Apr 2026).

This enables scalable use in large-scale processing and unlocks practical pathways for imposing acyclicity on otherwise highly cyclic networks without losing essential flow structure.

6. Applications and Reductions

Congestion-preserving DAG projections permit direct reductions of classical flow problems in general digraphs to more tractable settings. Notably, for any $V$9-approximate max-flow/min-cut oracle on DAGs, the maximum flow computation on general $\bot$0 reduces as follows:

1. Compute a DAG projection $\bot$1 in $\bot$2 time.
2. Solve the max-flow problem in $\bot$3 between projected vertex sets with approximation $\bot$4.
3. Project the solution back to $\bot$5, incurring an extra factor $\bot$6.

Consequently, exact max-flow in $\bot$7 reduces to an $\bot$8-approximate computation on DAGs in near-linear time (Haeupler et al., 6 Apr 2026). Further, these projections improve state-of-the-art for distance preservers, hop-set constructions, and approximate cut computations in directed graphs ([Hoppenworth, Xu, Xu, SODA'25], [Cheung, Lau, Leung, SICOMP'13], [Kogan, Parter, SODA'22], [Bernstein, Wein, SODA'23], [Bernstein, Blikstad, Saranurak, Tu, FOCS'24], [Bernstein, Blikstad, Li, Saranurak, Tu, FOCS'25]).

7. Limitations, Implications, and Future Directions

The subpolynomial congestion factor $\bot$9 is attributable to (a) the number of hierarchy layers ($D$0), (b) expansion slack ($D$1), and (c) replica count ($D$2). Any algorithmic advance in reducing these contributions—such as sharper expander decompositions or improved hierarchical partitioning—would tighten the congestion bound toward polylogarithmic regimes.

A plausible implication is that combining improved DAG-flow oracles with these projections could yield further consequences for max-flow and cut problems in digraphs. The method also reduces open problems in exact single-source shortest paths and maximum flow on directed graphs to their comparable problems on DAGs or undirected graphs, highlighting a pathway for transferring structural simplicity to broader algorithmic domains (Haeupler et al., 6 Apr 2026).

---

For full proofs, algorithmic specifics, and references, consult (Haeupler et al., 6 Apr 2026).