Work

Linear additive loss in bipartite matching with strict budgets

Year: 2026
Type: preprint
Abstract: Unrefereed research candidate; a follow-up to the Evidence Press release “Linear discrepancy for vector-labelled perfect matchings” (DOI 10.5281/zenodo.23932844). In one sentence: when a fractional plan for pairing workers with tasks is rounded to a real assignment without breaking any of m hard budgets, the profit lost need never exceed (7m/3) times the value of the single most valuable pairing, down from order m² previously. The problem. A platform pairs workers with tasks (a matching in a bipartite graph). Each pairing has a profit and uses up amounts of m shared resources, such as subsidy, travel or pairings of a protected type, and each resource has a strict budget that may not be exceeded at all. Finding the best such assignment is NP-hard even for one budget, because it contains knapsack. The standard route solves the linear-programming relaxation, which allows fractional pairings, and then rounds. The question is how much profit rounding must lose, measured in units of W, the largest profit of a single pairing. One pairing slightly over its budget shows a loss of about W is unavoidable, and m such pairings (m-dimensional knapsack) show a loss of about mW is unavoidable. What was known. Grandoni, Ravi, Singh and Zenklusen (2014) gave a loss of order m²W in general graphs. Berger, Bonifaci, Grandoni and Schäfer (2011) gave 2W for one budget. The splitting lemma of the earlier Evidence Press release, applied here, already gives about 9mW for bipartite graphs (Remark 6). Results. (1) Theorem 1: for bipartite graphs a budget-feasible matching of profit at least LP − CmW exists, with C1 = 1 (previously 2) and Cm = (14m−5)/6 ≤ 7m/3 for m ≥ 2; it is found deterministically in time NO(m), and heavy-edge guessing gives a deterministic approximation scheme with running time NO(m/ε) (Corollary 7). Proposition 5 gives the sharper support-dependent bound actually proved. (2) Theorem 8: the same accounting improves the earlier release’s bound for vector-labelled perfect matchings from 17k to max{3k+1, (14k−5)/3} ≤ 14k/3 and makes it computable in polynomial time for fixed k, answering that release’s algorithmic question for piecewise-constant data. (3) Theorems 9–10: for the integer-quota fair-representation problem of Aharoni et al., connected examples have exact gap m(1−1/Δ) at maximum degree Δ. Already for one quota and Δ = 3 the gap is 2/3 > 1/2, so the suggested loss m/2 fails for arbitrary edge sets, and the one-quota constant is exactly 1−1/Δ (supremum 1). How. An optimal fractional plan has few branch points. Its alternating paths are laid end to end and cut by one pattern with at most 2m+1 internal cuts that keeps every budget and an adjusted profit exactly; each cut costs at most half a pairing, and branch points are paid for in advance by adjusting the profit. Deleting pairings never raises a nonnegative cost, so every budget stays strict. Limits. Bipartite graphs only; nonnegative upper budgets only (no lower quotas or signed costs); the constant 7/3 is not known to be optimal (the lower bound is 1 per budget); the algorithm is polynomial only for fixed m and is a theoretical guarantee, not a practical solver. The linear order in m was already implicit in the earlier release; the advance is the constant, the algorithm and the exact examples. The one-budget case relies on a published circular splitting theorem of Grandoni, Ravi, Singh and Zenklusen. Contents. The paper (Markdown source, Pandoc LaTeX and PDF) with Figure 1; three standard-library or numpy/scipy checkers with their outputs: exact primal–dual certificates for the lower-bound examples (all 29,992 matchings of the Δ = 3 gadget enumerated), exhaustive finite checks of the chain and branch bookkeeping, and an end-to-end exact rounding test on 300 random instances for each of m = 1, 2, 3 and k = 1, 2, 3; the bibliography from registry records; a machine-readable index; a research-gate record; a rating-free review summary; a replay receipt and SHA-256 sums. Assurance. Written proofs; the computations test conventions and bookkeeping and are not part of any proof. Internal AI-assisted review only (five reviewer roles, an adversarial review and audit by a second AI model, a clean replay, a re-review): not peer review, independent reproduction or formal verification. Creator: Anonymous. Text and data CC0-1.0; code MIT.
Author: Anonymous
Language: English
Open Access status: green
OpenAlex ID: W7233595381
Cites: 19
Cited by:
Related to: