Balancing Weights, Directed Sparsification, and Augmenting Paths

Abstract: We present a randomized augmenting paths-based algorithm to compute the maximum flow in a directed, uncapacitated graph in almost $m+nF$ time, matching the algorithm of Karger and Levine for undirected graphs (SICOMP 2015). Combined with an initial $\sqrt n$ rounds of blocking flow to reduce the value of $F$, we obtain a maximum flow algorithm with running time $mn^{1/2+o(1)}$. For combinatorial, augmenting paths-based algorithms, this is the first improvement over Dinic's algorithm for moderately sparse graphs. To obtain our algorithm, we introduce a new technique to re-weight the edges of a strongly connected directed graph so that each cut is approximately balanced: the total weight of edges in one direction is within a constant factor of the total weight in the other direction. We then adapt Karger and Levine's technique of sampling edges from the newly weighted residual graph, ensuring that an augmenting path exists in the sampled graph with high probability. One technical difficulty is that our balancing weights have to be dynamically maintained upon changes to the residual graph. Surprisingly, we can black box the dynamic data structure from the recent interior point method-based flow algorithm of van den Brand et al. (FOCS 2024).

• The paper presents an augmenting-paths-based algorithm achieving nearly m+nF runtime for uncapacitated directed maximum flow.
• It employs a dynamic potential function to balance weights, ensuring every cut in the residual digraph meets a specified balance condition.
• Integration with directed sparsification techniques yields the first combinatorial improvement over Dinic's algorithm in moderately sparse regimes.

Balancing Weights and Directed Maximum Flow via Augmenting Paths

Overview

The paper "Balancing Weights, Directed Sparsification, and Augmenting Paths" (2604.14633) introduces an augmenting-paths-based algorithm for the maximum flow problem in uncapacitated, directed graphs, achieving an almost $m + nF$ time complexity, where $m$ is the number of arcs and $F$ the flow value. This matches the seminal Karger-Levine algorithm for undirected graphs. Furthermore, coupling with an initial $\sqrt{n}$ rounds of blocking flows yields a combinatorial maximum flow algorithm with a runtime of $mn^{1/2 + o(1)}$, constituting the first improvement over Dinic's algorithm in the moderately sparse regime.

Directed Balancing via Dynamic Potentials

A key obstacle in applying Karger-Levine-type random sampling approaches to directed graphs is the inability to guarantee cut balance in the residual digraph. The paper innovates by defining a dynamic weighting scheme parametrized by a potential function $\phi: V \to \mathbb{R}$, assigning arc weights as $w(u,v) = 1/(\max\{\phi(v) - \phi(u), 0\} + 1)$. The central algorithmic loop maintains these potentials such that, in the weighted digraph, every cut satisfies $w(\partial^+ S) \ge \frac{1}{2} w(\partial^- S)$.

The balancing potentials are updated iteratively: for any set $S$ violating the balance property, $\phi(v)$ is incremented for all $m$0. Existence and efficient computability of such a potential function are not trivial and are managed through a reduction to maintaining an approximate minimum ratio cut in an undirected, weighted graph induced by the current potential. The essential data structure for this is black-boxed from recent advances in dynamic interior point-based flow algorithms [van den Brand et al., FOCS 2024], which allows dynamic updates in amortized near-linear time.

Sampling and Sparsification in the Directed Setting

After establishing balanced weights, the combinatorial routine samples edges from the residual graph, treating it as an undirected, weighted structure for sampling purposes (with directionality re-encoded in the weights/potentials). To ensure high-probability existence of an augmenting path in the sparsified graph, the paper develops a dynamic random sparsifier with quality $m$1 (for edge weights bounded in $m$2), employing bucketing strategies for weights, multi-level expander decompositions, and careful analysis of cut preservation under sampling.

The dynamic sparsification scheme operates recursively via expander decompositions, ensuring that for any cut with sufficiently large out-arc weight, the sparsifier contains an out-arc of the cut with high probability. The total size of the sparsifier is $m$3 for parameter $m$4.

Algorithmic Details and Energy Analysis

The principal algorithm, at each step, maintains two central dynamic data structures: the balancing weights (fed by the min-ratio cut structure) and the dynamic sparsifier. Upon each successful balancing, the algorithm computes a sparsifier of the current residual graph and searches for an $m$5-$m$6 path. Upon finding an augmenting path, the corresponding arcs are flipped, and the process repeats.

A rigorous amortized analysis, using an energy function over the arc potentials, demonstrates that each balancing operation reduces a global energy metric by at least $m$7, while each augmenting path increases the energy by a controllable amount. This ensures that the total number of rebalancing operations is $m$8, yielding the central runtime claim.

The method enforces all updates while maintaining strong connectivity of the residual digraph—guaranteed via preprocessing and structural restrictions. When the initial flow is not maximal, a carefully chosen set of parallel arcs from $m$9 to $F$0 ensures the strong connectivity necessary for correctness.

Comparison to Prior Work

The runtime achieved matches the undirected Karger-Levine approach and outperforms all previous purely combinatorial, augmenting-paths algorithms for directed flow in the medium-sparse regime (notably improving on the $F$1 bound from Dinic plus Karzanov/Even-Tarjan analysis in the regime $F$2 for $F$3).

The use of dynamic balancing is an extension of techniques previously limited to nearly Eulerian or perfectly balanced directed graphs [Ene et al. 2016, Cohen et al. 2018, Cen et al. 2021], here broadened to arbitrary strongly connected digraphs via dynamically maintained potentials and a reduction to ratio-cut approximations.

Practical and Theoretical Implications

This algorithm demonstrates that purely combinatorial, augmenting-paths-based methods can approach the efficiency of interior point or continuous optimization-based flow algorithms—even in the more challenging directed setting. By leveraging a dynamically maintained potential-based balancing scheme, the approach opens new pathways to transfer powerful sparsification and sampling techniques from the undirected to the directed domain.

Practically, this could lead to more scalable maximum flow solvers for large, sparse directed networks, particularly where combinatorial structure and augmenting path methods are advantageous (e.g., in dynamic or streaming applications where explicit edge capacities are not present, or structural modifications are frequent).

On the theoretical side, the framework shifts the understanding of "balancedness" as a critical property for directed sampling and sparsification. The reduction of balance maintenance to dynamic approximate ratio cuts and extendability of these paradigms—potentially to matching, directed cut, or general multicommodity flow—suggests promising directions for future work.

A crucial limitation is the reliance on strong connectivity for the balancing to function; extending the paradigms to general digraphs with sources or sinks not in a strongly connected component remains open. The barriers and the open question regarding more natural or continuous variants of the balancing function (or a continuous analogue to push-relabel algorithms) are clearly articulated and offer tangible next steps.

Conclusion

The paper establishes a new approach for directed, uncapacitated maximum flow by introducing dynamic, potential-based edge re-weighting to reintroduce balance sufficient for effective random sparsification. When combined with advances in dynamic cut data structures and efficient sampling, the algorithm attains a nearly optimal augmenting-paths-based complexity for moderate sparsity, strictly improving upon long-standing combinatorial bounds. The developed framework extends concepts from undirected to directed settings and indicates broad potential for future generalizations in combinatorial optimization.