---
title: Lossless Online Rounding in Bipartite Matching
url: https://www.emergentmind.com/topics/lossless-online-rounding
type: topic
---

# Lossless Online Rounding in Bipartite Matching

Searching arXiv for the cited paper and closely related work on online rounding and online bipartite matching.
Lossless online rounding is an online randomized procedure that converts the decisions of a fractional online algorithm into an actual online integral solution while preserving the prescribed marginals exactly. In the online bipartite matching setting studied by Pollner, Saberi, and Wajc, this means that if a fractional algorithm assigns values \(x_{i,t}\) to edges \((i,t)\) as online vertices arrive, then the rounded online matching \(\mathcal M\) must satisfy \(\Pr[(i,t)\in \mathcal M]=x_{i,t}\) for every edge, not merely match the same total expected value [2106.04863]. The subject is technically subtle because offline integrality of the bipartite matching polytope does not imply that such rounding can be performed online under irrevocability and partial information. The 2021 paper “Lossless Online Rounding for Online Bipartite Matching (Despite its Impossibility)” reframed the area by showing that exact marginal preservation is impossible for arbitrary online fractional matchings, yet achievable for an important structured subclass via additional non-convex constraints and carefully maintained negative dependence [2106.04863].

## 1. Definition and conceptual distinction

In the online bipartite matching model, a bipartite graph \(G=(L,R,E)\) has offline vertices \(L=[n]\) known in advance and online vertices \(t\in R\) arriving one by one. Upon arrival of \(t\), all incident edges \((i,t)\) are revealed and the algorithm must decide immediately and irrevocably whether and how to match \(t\) [2106.04863]. In the vertex-weighted version, each offline vertex \(i\) has weight \(w_i>0\), and the objective is to maximize the total weight of matched offline vertices; in the unweighted case, all \(w_i=1\) [2106.04863].

A fractional online algorithm maintains values \(x_{i,t}\in[0,1]\), chosen online and irrevocably, subject to the online matching LP constraints
\[
\sum_t x_{i,t}\le 1 \quad \forall i\in L,\qquad \sum_i x_{i,t}\le 1 \quad \forall t\in R.
\]
Writing
\[
x_i^{(t)} := \sum_{t'<t} x_{i,t'},
\]
this quantity is the fractional degree of offline node \(i\) before time \(t\), and the fractional objective is \(\sum_{i,t} w_i x_{i,t}\) [2106.04863].

An integral or randomized online algorithm instead outputs an actual matching \(\mathcal M\), possibly using randomness, where each online vertex is either unmatched or matched to at most one currently free offline neighbor [2106.04863]. The strongest possible relation between the two notions is lossless online rounding: an online randomized procedure that, given the fractional decisions as they are produced, outputs a random matching \(\mathcal M\) such that
\[
\Pr[(i,t)\in \mathcal M]=x_{i,t}.
\]
This preserves the fractional solution edge-by-edge in expectation and is therefore strictly stronger than preserving only the total expected objective [2106.04863].

The central conceptual distinction is between offline and online integrality. Offline, the bipartite matching polytope is integral, so every feasible fractional matching can be rounded without loss. But that offline fact says nothing about whether the rounding can be done online, preserving every marginal irrevocably as vertices arrive [2106.04863]. This distinction is one of the defining ideas of the subject.

## 2. Impossibility and the meaning of “despite its impossibility”

The point of departure is the observation, attributed in the paper to Devanur, Jain, and Kleinberg (SODA’13), that lossless online rounding is impossible in general [2106.04863]. The representative counterexample in the paper uses a two-choice fractional algorithm that assigns \(x_{i,t}=\tfrac12\) to every revealed edge. If the first two online vertices have neighborhoods \(\{1,2\}\) and \(\{3,4\}\), then preserving marginals forces each of those arrivals to be matched with probability \(1\). After those two steps, some cross-pair \((i,j)\in\{1,2\}\times\{3,4\}\) is simultaneously matched with probability at least \(1/4\). If a third online vertex arrives adjacent exactly to that pair \(\{i,j\}\), then it can be matched with probability at most \(3/4\), while the fractional solution gives it value \(1\) [2106.04863]. Thus even a feasible two-choice fractional process need not be losslessly roundable online.

This impossibility already applies to natural strong fractional algorithms such as Balance [2106.04863]. The negative result therefore concerns unrestricted feasible online fractional matchings, not only artificial constructions.

The 2021 paper’s main conceptual refinement is that impossibility does not mean lossless online rounding is conceptually hopeless. Any randomized online algorithm trivially induces a fractional online algorithm with the same competitive ratio by setting \(x_{i,t}=\Pr[(i,t)\in\mathcal M]\), and such an induced fractional algorithm is, by definition, losslessly roundable online—namely by running the original randomized algorithm [2106.04863]. This suggests that the obstacle is not the idea of exact marginal preservation itself, but the insufficiency of the basic offline LP constraints in the online setting.

The paper therefore uses the phrase “online integrality gap” informally to denote the gap between what can be achieved by online fractional algorithms and what can be achieved by online integral or randomized algorithms under online information and irrevocability constraints [2106.04863]. A plausible implication is that “despite its impossibility” refers to a shift from trying to round arbitrary feasible fractional solutions to identifying subclasses of fractional processes whose marginals are compatible with an online randomized matching.

## 3. Sound two-choice algorithms and the lossless rounding theorem

The main positive theory is developed for two-choice algorithms, meaning that at each time \(t\), the fractional algorithm assigns positive mass to at most two neighbors. If
\[
P_t:=\{i: x_{i,t}>0\},
\]
then \(|P_t|\le 2\) [2106.04863].

The key definition is soundness. A two-choice fractional online algorithm is sound if for every arrival \(t\),
\[
\sum_{i\in P_t} x_{i,t}\le 1-\prod_{i\in P_t} x_i^{(t)}.
\]
If equality holds for every \(t\), the algorithm is maximal [2106.04863]. When \(|P_t|=2\), say \(P_t=\{1,2\}\), this becomes
\[
x_{1,t}+x_{2,t}\le 1-x_1^{(t)}x_2^{(t)}.
\]

The paper motivates this inequality by imagining that the matched statuses of offline vertices behave independently up to time \(t\). Then \(x_i^{(t)}\) is the probability that \(i\) is already matched, so the probability that both neighbors are unavailable would be \(x_1^{(t)}x_2^{(t)}\), and hence the probability that the new online vertex can be matched at all is at most \(1-x_1^{(t)}x_2^{(t)}\) [2106.04863]. What the paper proves is that for two-choice algorithms, this condition is not merely necessary-looking but sufficient when combined with a rounding scheme that maintains strong negative dependence.

The central theorem states that if \(\mathcal A\) is a sound two-choice online fractional algorithm with output \(\vec x\), then there exists a randomized online algorithm whose output matching \(\mathcal M\) satisfies
\[
\Pr[(i,t)\in \mathcal M]=x_{i,t}\qquad \forall (i,t)\in E.
\]
If \(\mathcal A\) is maximal, this randomized algorithm can be implemented in polynomial time [2106.04863].

This result gives an exact characterization only for the two-choice sound subclass, not for all online fractional matchings. That restriction is substantive rather than cosmetic. Later work on online dependent rounding schemes for arbitrary fractional bipartite matchings studies a different objective—approximate preservation via a rounding ratio—and explicitly notes that exact ratio \(1\) is impossible in general [2301.08680]. The 2021 theorem is therefore best understood as a precise positive island inside a broader impossibility landscape [2106.04863].

## 4. Negative dependence, maximality, and rounding mechanics

The proof is organized around events \(F_{i,t}\), where \(F_{i,t}\) denotes that offline vertex \(i\) is free by time \(t\), and for \(I\subseteq[n]\),
\[
F_{I,t}:=\bigwedge_{i\in I} F_{i,t}.
\]
The rounding maintains the exact edge marginals together with the monotonicity condition
\[
\Pr[F_{i,t}\mid F_{K,t}] \le \Pr[F_{i,t}\mid F_{J,t}]
\qquad \forall J\subseteq K\subseteq [n]\setminus\{i\},
\]
whenever \(\Pr[F_{K,t}]>0\) [2106.04863]. Equivalently, \(f(I):=\Pr[F_{I,t}]\) is log-submodular, yielding the reverse-FKG type inequality
\[
\Pr[F_{A\cap B,t}]\,\Pr[F_{A\cup B,t}] \le \Pr[F_{A,t}]\,\Pr[F_{B,t}].
\]
This implies pairwise negative correlation among free-status indicators, hence also among matched-status indicators [2106.04863].

These correlations are not incidental. Soundness compares \(\sum_{i\in P_t}x_{i,t}\) to a product of marginals, and negative correlation ensures that the true probability that all candidates are unavailable is at most that product [2106.04863]. This is the probabilistic mechanism that makes exact online marginal preservation feasible.

For maximal inputs, the maintained structure is stronger:
\[
\Pr[F_{i,t}] = 1-x_i^{(t)},
\]
and for every \(I\subseteq[n]\),
\[
\Pr[F_{I,t}] \in \left\{0,\ \prod_{i\in I}(1-x_i^{(t)})\right\}.
\]
Thus each set is either “negative” or fully independent [2106.04863]. This dichotomy is a key reason the maximal-case algorithm admits a polynomial-time implementation.

In the maximal case, if arrival \(t\) has two relevant neighbors \(1,2\), with current loads \(x_1=x_1^{(t)}\), \(x_2=x_2^{(t)}\), and increments \(\Delta x_i=x_{i,t}\), then if \(\{1,2\}\) is negative, the algorithm matches to the sole free one \(i\) with probability
\[
\frac{\Delta x_i}{1-x_i}.
\]
If \(\{1,2\}\) is independent and both are free, it matches to \(1\) with probability
\[
\frac{1-x_2-\Delta x_2}{(1-x_1)(1-x_2)},
\]
and to \(2\) with probability
\[
\frac{1-x_1-\Delta x_1}{(1-x_1)(1-x_2)}.
\]
Under maximal soundness,
\[
\Delta x_1+\Delta x_2 = 1-x_1x_2,
\]
so these probabilities sum to \(1\) [2106.04863].

For non-maximal inputs, the paper gives a more general probability-setting program using \(a_i\) for the both-free case and \(b_i\) for the single-free case, constrained by
\[
a_1+a_2\le 1,\qquad a_i\ge 0,\qquad b_i\le 1,
\]
\[
b_i\ge a_i,\qquad b_i\le \frac{a_i}{1-a_{3-i}},
\]
and the marginal equations
\[
a_i p_{12}+b_i(1-x_i-p_{12})=\Delta x_i\qquad (i=1,2),
\]
where \(p_{12}=\Pr[F_{\{1,2\},t}]\) [2106.04863]. The inductive proof uses these constraints to preserve both exact marginals and the reverse-FKG style dependence structure.

## 5. Competitive algorithms and main applications

The lossless-rounding theorem matters algorithmically because the paper also designs competitive sound fractional algorithms. For unweighted online matching it gives a maximal sound two-choice “restricted water level” algorithm: on each arrival, pick the two neighbors of minimum current load \(x_1^{(t)}\le x_2^{(t)}\), and raise them to the common level
\[
x_f=\frac{x_1^{(t)}+x_2^{(t)}+1-x_1^{(t)}x_2^{(t)}}{2},
\]
so
\[
x_{1,t}=x_f-x_1^{(t)},\qquad x_{2,t}=x_f-x_2^{(t)}.
\]
This satisfies
\[
x_{1,t}+x_{2,t}=1-x_1^{(t)}x_2^{(t)},
\]
hence is maximal sound [2106.04863].

A primal-dual analysis yields competitiveness
\[
\alpha_g := \min_{x\in[0,1]} \frac{1-x^2}{1-3g(x)+2g\!\left(x+\frac{1-x^2}{2}\right)}
\]
for any increasing convex bijection \(g:[0,1]\to[0,1]\). Choosing \(g(x)=\frac{a^x-1}{a-1}\) with \(a=1.6\) gives a lower bound of about \(0.531\), and the paper gives an adversarial instance showing the ratio is at most about \(0.536\). After lossless rounding, the resulting randomized online algorithm is \(0.536\)-competitive [2106.04863].

For vertex-weighted bipartite matching, the paper gives a \(2\)-level maximal sound two-choice fractional algorithm with competitive ratio
\[
\frac{11}{21}\approx 0.524.
\]
After rounding, this yields a \(0.524\)-competitive randomized online algorithm, beating the deterministic \(1/2\) barrier [2106.04863].

These guarantees support one of the paper’s principal applications: the amount of randomness or advice needed to beat deterministic online matching. Previous upper bounds were \(O(n)\) or \(O(n\log n)\), while Peña and Borodin proved a lower bound of \((1-o(1))\log\log n\) bits for obtaining \((1/2+\varepsilon)\)-competitiveness. The paper closes this gap by showing that
\[
(1\pm o(1))\log\log n
\]
random bits, or advice bits, are both necessary and sufficient to achieve competitive ratio
\[
\frac12+\Omega(1)
\]
for online vertex-weighted bipartite matching [2106.04863]. The “doubly-exponential improvement” refers to lowering sufficiency from polynomial or linear in \(n\) to doubly logarithmic in \(n\).

The second major application is online correlated selection. Fahrbach et al. introduced \(\gamma\)-semi-OCS, where an element appearing in \(k\) pairs remains unchosen with probability at most
\[
2^{-k}(1-\gamma)^{k-1}.
\]
Gao et al. later showed the optimum is \(\gamma=\tfrac12\). The paper derives this optimum through lossless online rounding by defining a sound fractional process with cumulative load
\[
x_i^{(t)} = 1-2^{-2^{\ell_i}+1}
\]
for an item that has appeared in \(\ell_i\) pairs so far, and increment
\[
x_{i,t}=2^{-2^{\ell_i-1}+1}-2^{-2^{\ell_i}+1}.
\]
The resulting selector satisfies
\[
\Pr[i\text{ never picked}]=2^{-2^k+1},
\]
and since
\[
2^{-2^k+1}=2^{-k}(1-\tfrac12)^{k-1},
\]
this gives a \(\tfrac12\)-semi-OCS, matching the optimum from a different derivational route [2106.04863].

## 6. Scope, limitations, and relation to broader online rounding research

The positive characterization is explicitly limited to two-choice fractional algorithms. A natural extension to multiple-choice algorithms would require
\[
\sum_{i\in I} x_{i,t}\le 1-\prod_{i\in I} x_i^{(t)}
\qquad \forall I\subseteq P_t,
\]
but the paper proves that these stronger constraints are still not sufficient for lossless online rounding: there exists a three-choice fractional algorithm satisfying them that cannot be losslessly rounded online [2106.04863]. The theory is therefore not a full characterization of online roundability.

This limitation sharply distinguishes exact lossless online rounding from later approximate frameworks. “Online Dependent Rounding Schemes for Bipartite Matchings, with Applications” formalizes a rounding ratio \(\rho=\inf_{\mathbf{x},e}\Pr[e\in M]/x_e\) for arbitrary online-revealed fractional bipartite \(b\)-matchings, proving ratios \(0.652\) for simple matchings and \(0.646\) for \(b\)-matchings, while also emphasizing that exact ratio \(1\) is impossible in general [2301.08680]. That line of work studies unrestricted inputs but settles for approximate marginal preservation. By contrast, the 2021 lossless theory attains exact preservation but only under structural constraints [2106.04863].

A different but related perspective appears in the tutorial “Randomized Rounding Approaches to Online Allocation, Sequencing, and Matching,” which surveys sequential randomized rounding methods that often preserve marginals only after a multiplicative scaling, such as \(\Pr[\text{start job }i\text{ at time }t]=y_{it}/2\) in stochastic knapsack or \(\Pr[\text{match }j\text{ to }i]=x_{ji}/2\) in simple online stochastic matching [2407.20419]. This suggests a broader taxonomy: exact lossless online rounding is rare, scaled exactness is more common, and both depend heavily on the structure of the underlying feasibility system.

Other works use the phrase “lossless online rounding” in different models. In online stochastic matching with correlated arrivals, a new rounding scheme in the \(\Indep\) model preserves per-type routing marginals exactly,
\[
\Pr\!\left[\sum_{\ell=1}^{D_j}\mathbf 1\{\pi_j(\ell)=i\}=1\right]=x_{i,j},
\]
but this is a demand-aware single-type construction rather than an unrestricted online bipartite matching theorem [2208.02229]. In online multi-class selection with overlapping classes, “lossless” means preservation of expected total utility and, for proportional fairness, of each class’s expected utility, not edge-by-edge matching marginals [2510.21055]. These differences indicate that the term is model-sensitive.

The 2021 paper’s enduring contribution is therefore methodological as much as algorithmic. It replaces the coarse statement “lossless online rounding is impossible” with a finer one: arbitrary feasible online fractional matchings cannot be losslessly rounded online, but by adding the right non-convex online constraints—soundness for two-choice algorithms—and maintaining carefully designed negative dependence invariants, one obtains a constructive family that is losslessly roundable online and already strong enough to beat deterministic bounds [2106.04863]. A plausible implication is that the main open direction is not whether online lossless rounding exists at all, but how far the class of exactly roundable online fractional processes can be extended beyond two choices.

Source: https://www.emergentmind.com/topics/lossless-online-rounding