We study the abstract problem of rounding fractional bipartite $b$-matchings online. The input to the problem is an unknown fractional bipartite $b$-matching, exposed node-by-node on one side. The objective is to maximize the \emph{rounding ratio} of the output matching $M$, which is the minimum over all fractional $b$-matchings $\mathbf{x}$, and edges $e$, of the ratio $\Pr[e\in M]/x_e$. In offline settings, many dependent rounding schemes achieving a ratio of one and strong negative correlation properties are known (e.g., Gandhi et al., J.ACM'06 and Chekuri et al., FOCS'10), and have found numerous applications. Motivated by online applications, we present \emph{online dependent-rounding schemes} (ODRSes) for $b$-matching. For the special case of uniform matroids (single offline node), we present a simple online algorithm with a rounding ratio of one. Interestingly, we show that our algorithm yields \emph{the same distribution} as its classic offline counterpart, pivotal sampling (Srinivasan, FOCS'01), and so inherits the latter's strong correlation properties. In arbitrary bipartite graphs, an online rounding ratio of one is impossible, and we show that a combination of our uniform matroid ODRS with repeated invocations of \emph{offline} contention resolution schemes (CRSes) yields a rounding ratio of $1-1/e\approx 0.632$. Our main technical contribution is an ODRS breaking this pervasive bound, yielding rounding ratios of $0.646$ and $0.652$ for $b$-matchings and simple matchings, respectively. We obtain these results by grouping nodes and using CRSes for negatively-correlated distributions, together with a new method we call \emph{group discount and individual markup}, analyzed using the theory of negative association. We present a number of applications of our ODRSes to online edge coloring, several stochastic optimization problems, and algorithmic fairness.

In SODA 2025