Limited-Correlation Rounding Overview

• Limited-correlation rounding is a set of techniques that convert fractional solutions into discrete outputs by preserving key expectations and enforcing global constraints.
• The methodology employs swap rounding, pipage rounding, and conditioning approaches to balance marginal preservation with controlled dependence in various applications.
• These methods improve practical outcomes such as variance reduction in estimators, competitive ratios in online allocation, and approximation guarantees in scheduling and matching.

Searching arXiv for papers on limited-correlation rounding and closely related dependent rounding schemes. Limited-correlation rounding is a class of rounding methods for converting fractional solutions, pseudo-distributions, or marginal specifications into discrete objects while preserving selected expectations and hard feasibility constraints, but controlling dependence only where the analysis requires it. In current usage, the term spans several distinct constructions: swap-rounding for fixed-budget treatment assignment, per-unit online rounding for reusable resources, group-based iterative rounding for unrelated-machine scheduling, and conditioning-based correlation rounding in Sherali–Adams relaxations for Ising models (Yamin et al., 15 Jun 2025, Nekouyan et al., 25 Jul 2025, Li, 2024, Jain et al., 2018). This suggests a common design principle: independence is relaxed only on those coordinates, groups, or resources where feasibility, variance, or approximation quality depends on the induced correlations.

1. Formal scope and basic objectives

A recurring starting point is a fractional vector whose marginals are already meaningful, but whose coordinates cannot be sampled independently without violating a global constraint. In budget-constrained experimental design, one is given

$p_i\in[0,1],\qquad \sum_{i=1}^n p_i = B,$

and seeks binary assignments

$x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$

with $\Pr[x_i=1]=p_i$ for all $i$ (Yamin et al., 15 Jun 2025). Independent rounding preserves marginals but generally fails to enforce $\sum_i x_i=B$ exactly.

An analogous objective appears in online allocation. In $k$Rental-Variable, each request $n$ has a fractional target $x_n\in[0,1]$, and the rounded decision $z_n\in\{0,1\}$ should satisfy $E[z_n]=x_n$ while never allocating more than $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$0 reusable units at any time (Nekouyan et al., 25 Jul 2025). Here the obstacle is temporal overlap: some dependencies are indispensable because overlapping requests competing for the same physical unit cannot both be accepted.

In Sherali–Adams-based variational methods for Ising models, the rounding object is not an assignment vector but a pseudo-distribution. Limited-correlation rounding first conditions on a small set $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$1 so that the remaining covariances become small on average, and then replaces the conditional law by a matching product distribution (Jain et al., 2018). In this setting, “rounding” means replacing a high-order relaxation by a tractable distributional surrogate while controlling the loss induced by residual correlation.

Across these settings, the design criteria recur with minor variations: exact or approximate marginal preservation, feasibility with respect to packing or budget constraints, and a controlled dependence structure that is strong enough to enforce constraints but weak enough to support clean variance or approximation analyses.

2. Core mechanisms and foundational constructions

A canonical mechanism is swap-rounding in the budgeted binary setting. Starting from a fractional vector $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$2, the algorithm repeatedly selects a fractional pair $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$3 and redistributes mass between the two coordinates, leaving all others unchanged. If $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$4, the pair is merged below $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$5; if $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$6, the pair is split above $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$7. At each step, at least one coordinate becomes integral, so the procedure terminates after at most $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$8 iterations and returns $x_i \in\{0,1\},\qquad \sum_{i=1}^n x_i = B,$9 with $\Pr[x_i=1]=p_i$0 (Yamin et al., 15 Jun 2025). The method guarantees exact feasibility, marginal preservation, and non-positive pairwise covariance for swapped pairs: $\Pr[x_i=1]=p_i$1

A broader foundational framework is pipage rounding on matroid base polytopes. Given a fractional point $\Pr[x_i=1]=p_i$2 in the base polytope, pipage rounding repeatedly moves along swap directions $\Pr[x_i=1]=p_i$3 until an extreme point is reached. In the randomized version, the final integral vector $\Pr[x_i=1]=p_i$4 satisfies $\Pr[x_i=1]=p_i$5, and the output is negatively cylinder-dependent; in particular, $\Pr[x_i=1]=p_i$6 for distinct coordinates (Harvey et al., 2013). The paper “Pipage Rounding, Pessimistic Estimators and Matrix Concentration” develops the technique of concavity of pessimistic estimators, showing that concentration phenomena unavailable from negative correlation alone can still be obtained under pipage or swap-style dependent rounding (Harvey et al., 2013).

These two constructions exemplify complementary viewpoints. Swap-rounding is a direct coordinate-wise procedure tailored to exact cardinality constraints. Pipage rounding is polyhedral and potential-based, and its importance lies partly in showing that the usefulness of dependent rounding extends beyond pairwise covariance control to matrix concentration, submodular concentration, and SDP rounding.

3. Selective dependence and relaxed correlation conditions

A defining feature of limited-correlation rounding is that it need not impose universal pairwise non-positive correlation. In the unrelated-machine weighted completion-time problem, the rounding framework explicitly relaxes the classical requirement that every pair of assignment events on the same machine be non-positively correlated. Instead, for each machine $\Pr[x_i=1]=p_i$7, jobs are partitioned into size-classes, one marked group is created in each class, and the only required covariance constraint is that every unmarked edge $\Pr[x_i=1]=p_i$8 have non-positive covariance with the entire marked group: $\Pr[x_i=1]=p_i$9 The algorithm then exploits much stronger intra-group exclusion: at most one job from each group is scheduled on a machine (Li, 2024).

The online $i$0Rental-Variable scheme uses a different relaxation pattern. Each physical unit is treated independently, and controlled correlations are introduced only among requests that compete for the same unit. If requests $i$1 are both mapped to unit $i$2 and their intervals overlap, then

$i$3

and more generally any pairwise-overlapping set of requests assigned to the same unit cannot all succeed simultaneously (Nekouyan et al., 25 Jul 2025). Different units use independent randomness, but same-unit decisions are coupled through the threshold $i$4.

In bipartite matching, the recent Brownian-motion-style dependent rounding framework strengthens local negative dependence rather than weakening it. For two edges $i$5 sharing a left-node, the output satisfies

$i$6

while also preserving the right-node matching constraints (Harris, 5 Jun 2026). The paper additionally guarantees a stable-set negative-correlation property. Taken together, these constructions show that “limited correlation” can mean weaker global conditions, stronger local exclusion, or both, depending on which interactions drive the objective.

4. Representative guarantees across problem domains

The literature uses limited-correlation rounding to optimize markedly different quantities: estimator variance, competitive ratio, approximation ratio, and variational error. The guarantees below illustrate the range of formal outcomes.

Setting	Correlation/feasibility feature	Stated guarantee
Budget-constrained experimental design (Yamin et al., 15 Jun 2025)	$i$7, $i$8, $i$9 for swapped pairs	unbiased IPW and general linear estimators; $\sum_i x_i=B$0; $\sum_i x_i=B$1 time or $\sum_i x_i=B$2 with a data structure
$\sum_i x_i=B$3Rental-Variable (Nekouyan et al., 25 Jul 2025)	same-unit overlapping requests cannot both succeed; different units use independent randomness	no lossless online rounding in the variable-duration setting; with $\sum_i x_i=B$4, obtain a $\sum_i x_i=B$5-competitive algorithm
Unrelated-machine weighted completion time (Li, 2024)	non-positive covariance only between an unmarked edge and a marked group; at most one job per group	randomized $\sum_i x_i=B$6-approximation
Fractional bipartite matchings (Harris, 5 Jun 2026)	strong negative correlation for edges sharing a left-node	$\sum_i x_i=B$7
Correlation clustering (Cohen-Addad et al., 2023)	pairwise correlated-rounding error confined to an admissible set after preclustering	$\sum_i x_i=B$8-approximation
Ising models and $\sum_i x_i=B$9-MRFs (Jain et al., 2018)	after conditioning on $k$0, average conditional covariance-squared is at most $k$1 with $k$2	$k$3

These results show that limited-correlation rounding is not tied to a single probabilistic invariant. In some settings the key quantity is covariance; in others it is a joint-probability exclusion, a stable-set inequality, or an entropy–energy comparison after conditioning. The unifying feature is that the dependence is structured and analyzable rather than fully general.

5. Analytical roles in estimation, online algorithms, and approximation

In experimental design, the principal role of negative dependence is variance reduction. For the inverse-propensity weighted estimator

$k$4

the variance decomposition

$k$5

shows directly how negative covariance terms reduce estimator variance under standard positivity and non-negativity assumptions on outcomes (Yamin et al., 15 Jun 2025). The same paper establishes unbiasedness via a martingale argument on partial-rounding estimators $k$6.

In online resource allocation, limited-correlation rounding is embedded inside a price-based fractional algorithm. The fractional targets are generated by Duration-Oblivious Pricing, while the rounding layer turns them into feasible integral allocations using one uniform seed per request and per-unit load bookkeeping. The competitive-ratio proof then proceeds through an online primal-dual or LP-free certificate argument in which the negative-correlation constraint is what preserves dual feasibility (Nekouyan et al., 25 Jul 2025).

In unrelated-machine scheduling, the role of limited correlation is more delicate. The expected weighted completion time can be written in terms of pairwise products $k$7. The marked-group framework is designed so that pairwise non-positive correlation among unmarked edges, together with job-versus-group non-positive covariance, is enough to recover the desired cross-term bound. The algorithm’s stronger intra-group exclusion produces the subtractive term

$k$8

which is the additional gain responsible for improving the approximation factor to $k$9 (Li, 2024).

In matching-based rounding, strong local negative correlation directly improves overload control on shared resources. The fractional-matching paper states that such bounds are useful in rounding-plus-local-list-scheduling methods for minimizing weighted completion time on unrelated machines, because every reduction in the constant $n$0 in

$n$1

translates into a strictly better approximation ratio (Harris, 5 Jun 2026).

In variational inference for Ising models, the analytical role is different again. Correlation rounding is used to find a small conditioning set $n$2 that suppresses average post-conditioning covariance, allowing the conditional law to be replaced by a matching product distribution with controlled loss in the quadratic term. The resulting balance between entropy loss from conditioning and covariance loss from product replacement yields the $n$3 bound (Jain et al., 2018).

Correlation clustering provides a further variant. There, the correlated rounding of Sherali–Adams marginals incurs additive $n$4 error in pairwise probabilities. Preclustering reduces the uncertain part of the instance to $n$5 admissible edges, so this additive error can be absorbed, enabling a mixture of set-based and pivot-based rounding to achieve the $n$6-approximation (Cohen-Addad et al., 2023).

6. Limitations, misconceptions, and related directions

A common misconception is that limited-correlation rounding is synonymous with exact marginal preservation under all constraints. The variable-duration $n$7-rental setting provides a sharp counterexample: no online rounding scheme can be lossless, meaning $n$8 for all $n$9 for every fractional solution, while always satisfying the capacity constraints exactly (Nekouyan et al., 25 Jul 2025). In this sense, limited correlation is sometimes a way of quantifying unavoidable rounding loss rather than eliminating it.

A second misconception is that pairwise non-positive covariance is always the relevant or sufficient notion. The unrelated-machine scheduling paper explicitly departs from universal pairwise non-positive correlation, replacing it with job-versus-group constraints (Li, 2024). Conversely, the matrix-concentration results for pipage and swap rounding show that ordinary negative correlation has limitations: matrix Chernoff bounds do not simply follow from negative dependence, and additional structure in the form of concave pessimistic estimators is needed (Harvey et al., 2013).

A third limitation concerns the power of conditioning-based correlation rounding. The Ising-model analysis proves that the classical $x_n\in[0,1]$0 average covariance bound is sharp and refutes a conjectured $x_n\in[0,1]$1 improvement, using the SK spin glass and Gap-ETH-based tightness arguments (Jain et al., 2018). This places a formal barrier on how far limited-correlation conditioning can be pushed in Sherali–Adams relaxations.

In correlation clustering, the issue is not impossibility but scale sensitivity: additive pairwise rounding error is harmless only after preclustering reduces the uncertain region to the admissible set (Cohen-Addad et al., 2023). Without such preprocessing, the $x_n\in[0,1]$2 accumulation from pairwise errors can dominate the LP value.

A related, but distinct, direction studies correlated rounding as a coupling across different fractional inputs rather than as dependence control within one rounded output. On the hypersimplex $x_n\in[0,1]$3, a maximum-entropy $x_n\in[0,1]$4-subset scheme with common random ordering and common uniform thresholds satisfies

$x_n\in[0,1]$5

and the dummy-coordinate extension gives stretch at most $x_n\in[0,1]$6 for the at-most-$x_n\in[0,1]$7 polytope (Anari et al., 31 May 2026). This is not framed as limited-correlation rounding in the same sense as the scheduling or experimental-design papers, but it illustrates a neighboring theme: carefully engineered dependence can simultaneously preserve marginals and control a secondary stability metric.

Taken together, these works position limited-correlation rounding as a methodological middle ground between independent rounding and full-blown global dependence design. Its central insight is not merely that correlation can help, but that only certain correlations matter, and that isolating those interactions often yields the strongest guarantees.