Sequential Randomized Rounding

• Sequential Rounding is a technique for converting LP fractional solutions into binary decisions while preserving path-wise feasibility in online stochastic settings.
• It leverages state tracking, negative correlation, and correlation gap analysis to dynamically adapt to random events and maintain hard constraints.
• Its robust framework underpins constant-factor approximation algorithms for applications such as online contention resolution, stochastic probing, knapsack, and matching.

Sequential Rounding, specifically Sequential Randomized Rounding (SRR), is a methodology for designing approximation algorithms for a wide class of online and stochastic discrete optimization problems, where actions are taken one at a time in the presence of stochastic uncertainty and subject to hard, path-wise constraints. Unlike classical offline randomized rounding, which approximates combinatorial optimization problems from fractional solutions to linear programming (LP) relaxations, SRR must interleave rounded decisions with the evolving state of an online process and adapt to uncontrollable stochastic events, ensuring that feasibility is always maintained on every execution path. This paradigm provides the foundation for constant-factor approximation algorithms in online contention resolution, stochastic probing, knapsack, and matching scenarios, and leverages properties such as state tracking, negative correlation, and correlation gap analysis for performance guarantees (Ma, 2024).

1. Formal Definition and Conceptual Overview

Sequential Randomized Rounding considers online stochastic optimization problems specified by: a set of actions or decision variables $i \in I$; a sequence of random events (e.g., arrivals, acceptances), with known distributions; hard constraints (capacity, matching, or budget) that must hold on all sample paths; and an objective typically expressed as the expected sum of rewards for chosen actions. The standard approach first formulates an LP relaxation: $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$ where $x_i$ are marginals for action $i$.

SRR proceeds by constructing an online policy $\Pi$ that, given the fractional vector $x$, converts it into binary decisions $X_i \in \{0,1\}$ in a manner that satisfies:

• path-wise feasibility of all hard constraints,
• marginal preservation: $\mathbb{E}[X_i] \geq c x_i$ for all $i$ (for constant $c > 0$).

The challenge arises from the fact that the LP constraints are satisfied only in expectation, while SRR requires every realized policy trajectory to obey hard constraints. Decisions are taken as actions become “eligible”, potentially interspersed with random events that modify feasible choices for future steps.

2. Algorithmic Framework and Implementation Paradigm

All SRR algorithms follow a unified template:

1. Solve the LP relaxation to obtain $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$0.
2. Initialize state-tracking variables representing resource availability (e.g., remaining capacity, matched resources).
3. For each event $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$1 as it becomes eligible:
   • Compute an action probability $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$2, where $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$3 encodes the current state, such that the marginal preservation and feasibility conditions hold.
   • Draw a random bit $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$4.
   • If the state $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$5 allows and $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$6, commit to action $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$7 and update the state.

The essential design task is to calibrate $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$8 so that for each $$\max \sum_{i \in I} r_i x_i \quad \text{subject to} \quad A x \leq b, \; 0 \leq x_i \leq 1,$$9, the probability the action is taken is at least $x_i$0, and hard constraints are maintained path-wise. The process interleaves stochastic outcomes and decision-making, and must update tracked variables in real time. The following pseudocode skeleton encapsulates the SRR pattern:

$x$3

3. Theoretical Guarantees and Approximation Bounds

A key theoretical result for SRR, termed the “Master Guarantee”, states: if for every LP-feasible $x_i$1 one can devise an online policy $x_i$2 that (i) executes each action $x_i$3 with probability at least $x_i$4 and (ii) always satisfies the hard constraints, then $x_i$5 achieves a $x_i$6-approximation of the LP optimum. The proof follows from linearity of expectation: $x_i$7 where $x_i$8 is the LP relaxation's value and an upper bound on any (potentially adaptive) policy (Ma, 2024).

4. Representative Applications and Examples

SRR techniques have demonstrated constant-factor approximation in several canonical online stochastic optimization problems. The following table summarizes key applications and their guarantees:

Problem Class	SRR Policy Guarantee	Approximation Factor
Online Contention Resolution (k-unit)	At most $x_i$9 accepts, per-agent accept $i$0	$i$1, e.g. $i$2, $i$3
Stochastic Probing/Sequential Hiring	Up to $i$4 offers, at most $i$5 accepts, negative correlation	$i$6 (for $i$7), $i$8
Stochastic Knapsack	Sequential processing, capacity $i$9, state-tracked sampling	$\Pi$0
Online Stochastic Matching (heterogeneous)	Each resource assigned $\Pi$1 once, per-pair matched $\Pi$2	$\Pi$3 or better

For online contention resolution ($\Pi$4-unit rationing), SRR ensures per-agent acceptance probabilities while strictly respecting cardinality limits. In stochastic probing and hiring, SRR applies negative-correlated rounding to realize adaptive allocation, closing adaptivity gaps. For knapsack, SRR employs state-dependent sampling using LP-derived marginals, and for online matching, SRR provides latent assignments with calibrated marginal probabilities for each agent-resource pair. In several settings, SRR outperforms natural baseline algorithms and achieves approximation ratios matching or exceeding $\Pi$5, depending on problem symmetry and order randomization (Ma, 2024).

5. Underlying Principles and Technical Mechanisms

SRR leverages several algorithmic and probabilistic concepts:

• State Tracking: SRR systematically maintains the probability distribution over key “parallel” states, such as the number of remaining units or resource availabilities ($\Pi$6-variables), which are essential for valid rounding but not for classical dynamic programming.
• Coupling Across Hypothetical States: When actions are contingent on multiple possible system states (e.g., $\Pi$7 for $\Pi$8 remaining units), SRR coordinates rounding to ensure proper marginal preservation across all contingencies.
• Negative Correlation and Correlation Gaps: To compare rounded variable outcomes with the aggregate LP value, SRR exploits negative dependence (ensured by specialized rounding schemes) and tightens bounds using Poisson or binomial limits. Correlation gap tools (e.g., bounding $\Pi$9 in the probing setting) are used to translate concentration to approximation factors.
• Random vs Fixed Order: Introducing randomized action order (e.g., Uniform$x$0 arrival times) can dramatically improve guarantees, as adverse patterns in deterministic ordering are smoothed out and better constant factors are attainable.
• LP Tightening: Adding stringent online-only constraints to the LP (e.g., $x$1 for matching) can improve the SRR approximation quality beyond the baseline $x$2.

6. Policy Design Considerations and Empirical Properties

SRR policies exhibit desirable features such as computational efficiency, parallelizability, and fairness, attributed to their simple randomized structure and provable guarantees. The approach is robust to the curse of dimensionality and often matches or exceeds more complex dynamic programming heuristics. The unified SRR paradigm is thus broadly suitable for a range of online allocation, probing, knapsack, and matching problems.

A plausible implication is that continued refinement of SRR—especially through LP tightening and deeper analysis of the correlation gap—may yield further improvements in approximation bounds across additional classes of online stochastic optimization scenarios.

7. Context, Extensions, and Research Directions

Sequential Randomized Rounding unifies and extends classic randomized rounding techniques to multi-period, stochastic, and dynamically constrained environments. Its evolving theory has prompted research into more refined rounding for complex feasibility structures, adaptivity gaps in probing, and the development of tighter LP relaxations. Extensions incorporating online-only valid inequalities or exponential-size polytopes present promising directions for further constant-factor improvements and deeper structural insight (Ma, 2024). The methodology is pivotal in the design of efficient, near-optimal policies for contemporary online sequential decision-making tasks under uncertainty.