A fast parallel algorithm for minimum-cost small integral flows

We present a new approach to the minimum-cost integral flow problem for small values of the flow. It reduces the problem to the tests of simple multi-variate polynomials over a finite field of characteristic two for non-identity with zero. In effect, we show that a minimum-cost flow of value k in a network with n vertices, a sink and a source, integral edge capacities and positive integral edge costs polynomially bounded in n can be found by a randomized PRAM, with errors of exponentially small probability in n, running in O(k\log (kn)+\log^2 (kn)) time and using 2^{k}(kn)^{O(1)} processors. Thus, in particular, for the minimum-cost flow of value O(\log n), we obtain an RNC^2 algorithm.

💡 Research Summary

The paper tackles the classic minimum‑cost integral flow problem (MCIF) under the regime where the required flow value k is relatively small compared to the size of the network. While polynomial‑time sequential algorithms for MCIF are well‑known, efficient parallel (RNC) algorithms have remained elusive, especially when one seeks both polylogarithmic time and a polynomial number of processors. The authors introduce a novel reduction that translates the existence of a minimum‑cost flow of value k into the non‑zeroness test of a carefully constructed multivariate polynomial over a finite field of characteristic two.

The reduction proceeds by assigning two variables to each edge: one indicating whether the edge participates in the flow and another encoding the amount of flow on that edge. Flow‑conservation, capacity, and cost constraints are expressed as algebraic equations, and the overall flow configuration corresponds to the product of these equations. To isolate the optimal flow, the authors apply a variant of the Isolation Lemma: random binary weight vectors are added to the variables so that, with high probability, the minimum‑cost flow becomes the unique monomial of highest total degree in the polynomial. Consequently, checking whether the polynomial evaluates to a non‑zero value is equivalent to confirming the existence of the optimal flow.

Because the field has characteristic two, addition corresponds to XOR and multiplication can be performed with simple bit‑wise operations, making polynomial evaluation highly amenable to parallelization. On a PRAM, each processor independently simulates a random assignment of the variables and computes a partial contribution to the polynomial. A parallel reduction then aggregates these contributions in O(log (kn)) steps. The total running time of the algorithm is O(k log (kn) + log² (kn)). The number of processors required is 2^k·(kn)^{O(1)}; thus, when k = O(log n) the processor count remains polynomial, yielding an RNC² algorithm for the minimum‑cost flow of logarithmic value.

Error probability is driven by the size of the field. By choosing the field extension degree m = Θ(log n), the probability that the random weighting fails to isolate the optimal flow drops to 2^{‑Ω(log n)}, i.e., exponentially small in n. Repeating the procedure further reduces the error to negligible levels.

Compared with prior work, which either relied on sequential augmenting‑path methods (O(k·m·n) time) or parallel schemes demanding Ω(n³) processors, the presented method confines the exponential dependence on k solely to the processor count, while achieving polylogarithmic time. This makes the approach particularly attractive for networks where the desired flow is modest (e.g., O(log n) or constant).

The authors discuss several extensions: handling larger k by employing processor‑compression techniques, adapting the method to networks with negative or real‑valued costs via appropriate transformations, and implementing the algorithm on real distributed systems to assess practical performance. In summary, the paper offers a fresh algebraic perspective on MCIF, leverages randomization and characteristic‑two arithmetic to enable highly parallel computation, and establishes the first RNC² algorithm for minimum‑cost flows of logarithmic magnitude, thereby advancing both the theory of parallel algorithms and the practical toolkit for network optimization.