Stable Matching–Iterative Balancing (SM-IB)

• Stable Matching–Iterative Balancing (SM-IB) is a framework that iteratively reduces instability by updating matchings through controlled truncations of the Gale–Shapley process.
• It employs bounded-degree bipartite graph models to achieve ε-stable matchings and (2+ε)-approximations for maximum-weight matching in a distributed setting.
• The approach extends to parallel blocking-pair correction and iterative deletion in constrained markets, highlighting its versatility across diverse matching scenarios.

Stable Matching–Iterative Balancing (SM-IB) is an interpretive label for stable-matching procedures that repeatedly update a matching, or the feasible representation of a market, so that instability is progressively reduced. In the supplied literature, its clearest realization is truncated Gale–Shapley on bounded-degree bipartite graphs, where the propose–accept process is stopped after finitely many synchronous rounds and still yields an $\epsilon$-stable matching whose instability is controlled relative to matching size (0812.4893). Related iterative mechanisms appear in parallel blocking-pair correction, in iterated deletion of unattractive alternatives for constrained many-to-one markets, and in decentralized blocking-pair dynamics whose behavior also clarifies the limits of local stabilization (Wynn et al., 2024, Gutin et al., 2022, Rudov, 2024).

1. Core definitions and conceptual scope

The common substrate of SM-IB-style methods is the blocking-pair view of stability. In the bicoloured-graph formulation, the market is a bipartite graph $G=(R\cup B,E)$, where each node has a linear preference order over its neighbors. A matching $M\subseteq E$ is stable if it has no unstable edges. An edge $\{u,v\}\in E\setminus M$ is unstable relative to $M$ when $u$ is unmatched or prefers $v$ over its current match in $M$, and $v$ is unmatched or prefers $u$ over its current match in $G=(R\cup B,E)$0 (0812.4893). The paper then defines

$G=(R\cup B,E)$1

In the one-to-one matching model, the same logic is expressed in terms of firms and workers. A pair $G=(R\cup B,E)$2 is a blocking pair for matching $G=(R\cup B,E)$3 if they are not matched to each other under $G=(R\cup B,E)$4, but each strictly prefers the other to their current partner: $G=(R\cup B,E)$5 A matching is stable if it has no blocking pairs (Rudov, 2024). In the many-to-one setting with worker set $G=(R\cup B,E)$6, firm set $G=(R\cup B,E)$7, firm quotas $G=(R\cup B,E)$8, and strict preference lists, stability is the usual Gale–Shapley blocking notion adapted to the possibility that a firm at quota may drop a worse current assignee (Gutin et al., 2022).

The supplied papers also use two different notions of near stability. In the bounded-degree distributed setting, “almost stable” means instability is small relative to matching size via the $G=(R\cup B,E)$9-stable condition (0812.4893). In the decentralized-dynamics setting, an “almost stable” matching is one that is just one blocking pair away from a stable matching (Rudov, 2024). This difference matters, because one notion is quantitative and ratio-based, while the other is purely local in the state graph of matchings.

2. Truncated Gale–Shapley as the canonical SM-IB mechanism

The central SM-IB construction in the supplied material is a distributed version of Gale–Shapley that is run for only a fixed number of rounds. Each round consists of two turns: a blue turn and a red turn. Blue nodes accept the most preferred incoming proposal, possibly breaking with an old partner, and red nodes propose to their best remaining candidate if unmatched. The state after round $M\subseteq E$0 is denoted by a matching $M\subseteq E$1. The analysis tracks three monotonic features: blue nodes only move to more preferred matches, red nodes only lose candidates from their remaining preference lists, and the set of lost edges grows monotonically (0812.4893).

The paper formalizes the resulting tradeoff through lost-edge and potential arguments. Let $M\subseteq E$2 be the set of lost edges by the end of round $M\subseteq E$3, and let $M\subseteq E$4 be a potential over red nodes measuring how much better the next possible proposal could still be. The key inequality is

$M\subseteq E$5

For the unweighted case $M\subseteq E$6, one has $M\subseteq E$7, and the number of unstable edges $M\subseteq E$8 satisfies

$M\subseteq E$9

Choosing $\{u,v\}\in E\setminus M$0 yields

$\{u,v\}\in E\setminus M$1

once

$\{u,v\}\in E\setminus M$2

This is the paper’s linear-growth phenomenon: the ratio of matched individuals to blocking pairs grows linearly with the number of propose–accept rounds. A faithful summary given in the supplied material is that each extra round removes some candidate edges, the total remaining potential for instability decreases, and after $\{u,v\}\in E\setminus M$3 rounds the unstable-edge-to-matching ratio is $\{u,v\}\in E\setminus M$4 (0812.4893).

The formal algorithmic consequence is Theorem 1: $\{u,v\}\in E\setminus M$5 This is obtained by running the truncated Gale–Shapley procedure for

$\{u,v\}\in E\setminus M$6

rounds, each round taking two synchronous communication steps. Because $\{u,v\}\in E\setminus M$7 is assumed constant, the running time depends only on $\{u,v\}\in E\setminus M$8 and $\{u,v\}\in E\setminus M$9, not on graph size or diameter. The output is therefore local: in $M$0 synchronous steps, information propagates only to distance $M$1, so the decision associated with an edge depends only on a bounded-radius neighborhood (0812.4893).

3. Weighted objectives, ties, and constant-time estimation

The same truncated machinery extends beyond unweighted almost stability. In edge-weighted bicoloured graphs, preferences are sorted by edge weights, and the paper proves

$M$2

The core inequality is

$M$3

where $M$4 is an optimum matching. Combined with the potential bound $M$5, this gives

$M$6

Thus the same iterative balancing process is not only almost stable in the blocking-pair sense, but also sufficiently structured to support comparison with maximum-weight matching (0812.4893).

The paper further states that the algorithm still works when ties are allowed in preference lists, with an analogous proof. This indicates that the truncation and locality phenomenon is not restricted to strict rankings. Exact strict rankings are therefore not required for the local algorithm or for the approximation guarantees stated in the supplied material (0812.4893).

A further consequence is a centralized randomized constant-time approximation scheme for estimating the size of a stable matching. The theorem is stated as follows: $M$7

$M$8

queries and outputs an estimate $M$9 such that, with probability at least $u$0,

$u$1

where $u$2 is the size of a stable matching in $u$3. The estimate targets the size of an exact stable matching, not merely an $u$4-stable matching. Operationally, the algorithm samples nodes uniformly at random to estimate the number of red nodes and the fraction of red nodes matched after $u$5 rounds of the truncated algorithm, then transfers the guarantee from $u$6 to $u$7. To determine whether a sampled red node is matched after $u$8 rounds, it suffices to inspect preferences inside radius $u$9, whose size is bounded by

$v$0

This locality bound is what makes the oracle algorithm constant-time in the query sense (0812.4893).

4. Parallel blocking-pair correction: PII-RMD

A different SM-IB-style mechanism appears in the parallel stable-matching literature built on the Parallel Iterative Improvement (PII) framework. The setting is the standard stable marriage problem with $v$1 men, $v$2 women, and a preference matrix $v$3, where each pair $v$4 has a left value $v$5, the rank of woman $v$6 in man $v$7’s list, and a right value $v$8, the rank of man $v$9 in woman $M$0’s list. A pair $M$1 is a blocking pair for a matching $M$2 if

$M$3

where $M$4 and $M$5 are the current matched partners of $M$6 and $M$7 (Wynn et al., 2024).

PII starts from a random matching $M$8, finds all blocking pairs, chooses NM1-generating pairs row-wise and NM1 pairs column-wise, replaces conflicting current matches, and fills open rows and columns. With $M$9 processors, each iteration takes $v$0 time, and total runtime was empirically around $v$1 when convergence occurred within $v$2-many iterations. The key weakness reported in the supplied material is that the original PII algorithm converged only about $v$3 of the time in earlier experiments and could cycle indefinitely in the remaining cases (Wynn et al., 2024).

The paper’s augmented method, PII-RMD, introduces two selection rules and a preprocessing step:

Component	Criterion	Stated role
Right-Minimum Selection	Accept $v$4 as potential NM1-generating pair iff $v$5	Enforces right-side improvement
Dynamic Selection	Accept $v$6 iff $v$7	Makes each accepted row update a new left-side minimum over time
Quick Initialization	Each man proposes to his best remaining woman; accepted woman is removed from all other men’s lists	Produces a faster starting matching

Right-Minimum Selection is interpreted in the paper as a monotonicity constraint more favorable to women, “similar in manner to the original Gale–Shapley algorithm.” The paper proves NM1-cycle freeness and full cycle freeness under this rule, using strict decrease of right values along replacement chains. Dynamic Selection maintains a minimum pointer per row and uses a wait time $v$8, with the re-entry condition

$v$9

where $u$0 is the current iteration and $u$1 is the last iteration at which the left-minimum processor was compared. The authors report that increasing $u$2 each time the left-minimum processor is selected works best, and they set $u$3 when a new left-minimum processor is chosen. Quick Initialization is claimed to run in $u$4 parallel time with $u$5 processors, faster than Smart Initialization at $u$6 (Wynn et al., 2024).

The empirical findings are explicit. Right-Minimum Selection alone improves convergence at larger $u$7 from about $u$8 to over $u$9, with no visible decline as $G=(R\cup B,E)$00 grows. PII-RMD reaches $G=(R\cup B,E)$01 convergence within $G=(R\cup B,E)$02 iterations for any initialization method, whereas PII-SC with Smart Initialization needs about $G=(R\cup B,E)$03 iterations for the same level. Over $G=(R\cup B,E)$04 million trials, PII-RMD achieved $G=(R\cup B,E)$05 convergence within $G=(R\cup B,E)$06 iterations, and an additional scalability experiment for $G=(R\cup B,E)$07, totaling $G=(R\cup B,E)$08 trials, also converged. The claimed asymptotic behavior is $G=(R\cup B,E)$09 average runtime with $G=(R\cup B,E)$10 processors, but the authors explicitly state that this remains an empirical guarantee rather than a proven worst-case bound, and that it is unclear whether PII-RMD fully converges theoretically (Wynn et al., 2024).

5. Iterated pruning under assignment constraints

In constrained matching markets, SM-IB-style reasoning appears as iterative pruning rather than repeated modification of a single matching. The setting is a many-to-one two-sided market with workers $G=(R\cup B,E)$11, firms $G=(R\cup B,E)$12, quotas $G=(R\cup B,E)$13, strict preference lists, and possibly incomplete lists. The paper studies assignment constraints requiring certain pairs to be included, forbidding others, and asks whether such constraints are compatible with stability. The participant-level constraints are reduced to a pairwise formulation: $G=(R\cup B,E)$14

$G=(R\cup B,E)$15

Here $G=(R\cup B,E)$16 are required pairs and $G=(R\cup B,E)$17 are forbidden pairs (Gutin et al., 2022).

The algorithmic core is the iterated deletion of unattractive alternatives (IDUA). After converting the many-to-one market into a one-to-one market by splitting each firm of quota $G=(R\cup B,E)$18 into $G=(R\cup B,E)$19 identical copies, the paper represents the market as a directed graph $G=(R\cup B,E)$20 whose vertices are acceptable pairs and whose arcs encode preference comparisons: $G=(R\cup B,E)$21

$G=(R\cup B,E)$22

$G=(R\cup B,E)$23

A matching is an independent set in this digraph, and a stable matching is a kernel: independent and such that every vertex outside the matching has an out-neighbor in the matching (Gutin et al., 2022).

IDUA repeatedly applies the reduction rule $G=(R\cup B,E)$24. If no arc in $G=(R\cup B,E)$25 leaves $G=(R\cup B,E)$26, delete all $G=(R\cup B,E)$27 such that $G=(R\cup B,E)$28. If no arc in $G=(R\cup B,E)$29 leaves $G=(R\cup B,E)$30, delete all $G=(R\cup B,E)$31 such that $G=(R\cup B,E)$32. Writing

$G=(R\cup B,E)$33

the process stops at the normal form

$G=(R\cup B,E)$34

Lemma 1 states that $G=(R\cup B,E)$35 and $G=(R\cup B,E)$36 contain exactly the same stable matchings. This gives the iterative-pruning interpretation: alternatives that cannot survive in any stable matching are deleted, and each deletion can force additional deletions (Gutin et al., 2022).

The normal form yields extremal stable matchings

$G=(R\cup B,E)$37

identified as the worker-optimal and firm-optimal stable matchings in the normal form. The recursive constrained algorithm then enforces required pairs, iteratively deletes forbidden vertices $G=(R\cup B,E)$38 satisfying

$G=(R\cup B,E)$39

reruns IDUA, and branches on any $G=(R\cup B,E)$40 when the two extremal matchings differ. Its main guarantee is: $G=(R\cup B,E)$41 where $G=(R\cup B,E)$42 and $G=(R\cup B,E)$43 is the number of feasible stable matchings output; the time between two consecutive output stable matchings is $G=(R\cup B,E)$44 (Gutin et al., 2022).

This suggests an SM-IB interpretation in which balancing is performed over the feasible set rather than over a single evolving matching. The paper itself presents the method as normal-form reduction plus branch-and-bound enumeration, but the iterative deletion phase is structurally close to an iterative balancing or fixed-point computation.

6. Fragility, path dependence, and limits of iterative stabilization

The most important qualification to any broad SM-IB narrative comes from decentralized blocking-pair dynamics. In the one-to-one market $G=(R\cup B,E)$45, the process studied in the supplied material starts from any matching $G=(R\cup B,E)$46, repeatedly selects a blocking pair or best blocking pair at random, satisfies that pair, makes the previous partners unmatched, and leaves all other matches unchanged. The classical theorem of Roth and Vande Vate is recalled in the form: for any unstable matching, there exists a finite sequence of blocking pairs that leads to a stable matching (Rudov, 2024).

The paper argues that this convergence guarantee is weak from a fragility perspective. Its first theorem characterizes when “anything goes.” A subset $G=(R\cup B,E)$47 of equal size is a fragment if, in the induced submarket, there is a stable matching $G=(R\cup B,E)$48 such that every agent inside the fragment prefers their partner in $G=(R\cup B,E)$49 to every agent outside. The theorem states that the following are equivalent: from any unstable matching $G=(R\cup B,E)$50 and any stable matching $G=(R\cup B,E)$51, there exists a finite sequence of (best) blocking pairs from $G=(R\cup B,E)$52 to $G=(R\cup B,E)$53; the same holds already for every almost stable matching; and there are no non-trivial fragments. The corollary is that, in the absence of non-trivial fragments, random decentralized dynamics can reach any stable matching with positive probability (Rudov, 2024).

The second theorem addresses time to stabilization. For any sequence of markets with a unique stable matching, and for any $G=(R\cup B,E)$54, after a small augmentation there exist $G=(R\cup B,E)$55-unstable initial matchings such that, under the random decentralized dynamics,

$G=(R\cup B,E)$56

The proof uses a biased random walk on

$G=(R\cup B,E)$57

where $G=(R\cup B,E)$58 is the unique stable matching. In the critical region close to stability, there are $G=(R\cup B,E)$59 destabilizing blocking pairs and only $G=(R\cup B,E)$60 stabilizing ones, with the destabilizing side dominating once $G=(R\cup B,E)$61 is small. The paper therefore shows not merely slow convergence, but a regime in which many participants remain mismatched for extended periods (Rudov, 2024).

There is also a positive case: if the market has a nested structure of trivial fragments, such as a sequence of top-top pairs that can be peeled off one by one, expected stabilization time is on the order of $G=(R\cup B,E)$62 under uniform random blocking-pair selection. Even here, however, the supplied material stresses that small perturbations or augmentations can destroy the easy dynamics and restore exponential slowdown (Rudov, 2024).

Taken together, these results delimit the scope of SM-IB as an interpretive framework. Truncated Gale–Shapley yields a deterministic local algorithm with explicit quantitative control over residual instability; PII-RMD provides a parallel empirical anti-cycling strategy; and IDUA gives an iterative pruning method for constrained enumeration. But decentralized satisfaction of blocking pairs is not, in general, a robust or rapidly stabilizing balancing process. A plausible implication is that “iterative balancing” is best treated not as a single theorem about stable matching, but as a family of update principles whose guarantees depend sharply on the model of interaction, the notion of near stability, and the structural properties of the market.