Optimal Metric Distortion for Learning-Augmented Matching on the Line

Abstract: We revisit the problem of matching on the line with ordinal preferences. In the classic setting, there are $n$ agents and $n$ items in a shared unknown line metric, and the goal is to find a low-cost perfect matching using only the agents' rankings of the items by distance. A mechanism has distortion $α$ if it always outputs a matching whose cost is within a factor of $α$ of the optimum, in every consistent line metric. In the learning-augmented setting, the mechanism is also supplied with a prediction that conveys additional information about the instance. The quality of this prediction is unknown, and the goal is to optimize the mechanism's distortion when the prediction is accurate (consistency), while preserving worst-case guarantees when the prediction is arbitrarily inaccurate (robustness). We propose a mechanism that takes a matching as its prediction and guarantees $1$-consistency and $3$-robustness. By recovering an optimal matching when the prediction is perfectly accurate while retaining the optimal prediction-free distortion guarantee when it is arbitrarily inaccurate, we resolve an open question of Filos-Ratsikas et al. (IJCAI, 2025).

• The paper presents a deterministic mechanism that achieves a distortion of min{η,3} by recovering the optimal prediction (consistency 1) and ensuring robustness (3) with adversarial advice.
• The mechanism refines the OrderMatch approach using swap-based local improvements and canonical structure theorems to secure a global 3-approximation for various top-k cost objectives.
• The paper proves that only full outcome predictions can reduce distortion, establishing that partial or one-sided advice offers no extra benefit in matching on the line.

Optimal Metric Distortion for Learning-Augmented Matching on the Line: An Expert Analysis

Problem Overview and Motivation

This work revisits one-sided bipartite matching under the ordinal preferences setting, where agents must be matched to items based on rank orderings—rather than direct cardinal costs—induced by an unknown underlying line metric. The challenge is to design mechanisms that, given only ordinal information (preference rankings), can still select matchings whose total or top-$k$ cost is within a guaranteed factor (the “distortion”) of the matching that would have been optimal with full metric information. The line metric is especially relevant for applications in computational social choice, allocation, and school choice, as it models alignment along a continuous dimension while exhibiting enough structure to enable detailed analysis.

The classical results show that in general metric spaces, ordinal mechanisms suffer high worst-case distortion, with $\Omega(n)$ lower bounds and no deterministic mechanism breaking $O(n)$ [Anari et al., 2023; Haqi & Gharan, 2026]. On the line, however, constant distortion is achievable: recent work established that a deterministic mechanism can guarantee distortion $3$ for total cost, and that this is tight [Filos-Ratsikas et al., 2025].

There is growing interest in the “learning-augmented” paradigm, which considers mechanisms that supplement their traditional input (here: ordinal rankings) with a machine-learned or domain-predicted “advice”—while demanding strong guarantees across the spectrum of advice quality, i.e., the best of both worlds in terms of consistency (performance if the advice is perfect) and robustness (performance if the advice is adversarial). In the matching-on-the-line setting, prior work posed the open question of whether such advice can drive the distortion below $3$, without sacrificing robustness [Filos-Ratsikas et al., 2025]. This paper resolves that question, delivering strong structural, algorithmic, and optimality results.

Technical Contributions

1. Stronger Characterization of Optimal Prediction-Free Distortion

The authors generalize and strengthen the OrderMatch mechanism for prediction-free matching on the line. The original OrderMatch recovers a maximal partial ordering of items using agents’ top and bottom preferences, and then greedily matches “sortable” items and agents. The refined mechanism described here demonstrates that a significantly wider range of matchings—essentially, any matching that assigns ambiguous “outlier” items arbitrarily but matches the rest assortatively—achieves distortion at most $3$. The constructive proof employs a sequence of swap-based local improvements, combined with canonical structure theorems, to confirm a global $3$-approximation on any top-$k$ cost objective. This optimal bound is robust under a wide family of fair cost objectives (monotone symmetric norms).

2. Best-of-Both-Worlds Learning-Augmented Matching Mechanism

The main algorithmic result is a deterministic, learning-augmented mechanism that receives, along with the ordinal input, a predicted perfect matching $\widehat{M}$. The mechanism:

• Assigns ambiguous “outlier” items to the agents designated by the prediction $\widehat{M}$
• Refines all residual agent orderings (when multiple agents have the same top item) according to $\Omega(n)$0 (breaking ties according to predicted assignments)
• Matches the sorted agents to sortable items in line order

This design ensures that (i) when the prediction is optimal, the mechanism recovers it exactly (consistency $\Omega(n)$1), and (ii) for arbitrary predictions, the mechanism is at worst as bad as the best prediction-free ordinal mechanism (robustness $\Omega(n)$2). Importantly, it achieves $\Omega(n)$3-distortion whenever the prediction itself has distortion $\Omega(n)$4.

3. Matching Lower/Upper Bounds and Impossibility for Alternative Prediction Models

The mechanism’s guarantees are proven optimal: no deterministic oracle-augmented mechanism can obtain distortion strictly below $\Omega(n)$5 in this setting—even if the advice is full knowledge of either agent or item locations (but not the entire ground-truth metric and assignment), as shown by refined indistinguishability/projection arguments. This clarifies that matching-on-the-line is “advice-sensitive” only through full outcome predictions; partial or side-specific predictions confer no extra power for distortion.

Key Theoretical Results

Guarantee Type	Bound Attained	Tightness
Consistency (prediction optimal)	$\Omega(n)$6	Tight (Exact Recovery)
Robustness (adversarial prediction)	$\Omega(n)$7	Tight (Lower Bound)
Distortion vs. advice accuracy ($\Omega(n)$8)	$\Omega(n)$9	Tight with respect to underlying prediction

The approach generalizes to any top-$O(n)$0 cost (including fairness objectives), not just total cost.

Methodological Innovation

This paper leverages the geometric structure of the line to recover maximal partial orders, and then sharply characterizes the impact of outliers (ambiguously-ordered items or agents) on the feasible region. The key technical ingredient is a swap-based local-global argument which, using careful decomposition into internal/external assignments, shows that arbitrary choices for outlier allocation are benign with respect to global distortion. The learning-augmented extension reframes ambiguous allocation and agent ordering as degrees of freedom to be fixed by advice, which is algorithmically injected via tie-breaking and reserved assignments.

Practical and Theoretical Implications

Practical

• In real-world allocation settings (school choice, kidney exchanges, assignment of scarce resources with spatial or alignment constraints), even modest prediction of optimal matchings—via machine learning or historical data—enables mechanisms to interpolate gracefully between optimality (if advice is precise) and best-known robustness (when advice is misleading).
• Advice in the form of outcomes (predicted matchings) is algorithmically more valuable than predictions about only one market side or local agent/item positions.

Theoretical

• The “best of both worlds” result confirms that, on the line, learning-augmented matching can always delegate the core ambiguity to the advice, but is fundamentally limited by the structure imposed by ambiguous outliers.
• The results suggest a general principle: for certain structured metric domains (e.g., lines, trees), the performance of ordinal mechanisms is governed more by the set of ambiguous assignments than by finer details of preference structure. Thus, learning-augmented algorithm design can focus on exploiting and covering this ambiguity optimally.
• The proof techniques and mechanism design pattern suggest possible generalizations to other structured metrics (trees, cycles), and inspire investigation of similar “advice sensitivity” in multi-agent aggregation and resource allocation problems.

Future Research Directions

• The primary open problem in general metrics remains: closing the $O(n)$1 gap for prediction-free matching distortion, which underpins the possibility of robust, advice-driven improvements.
• Can analogous “viable family” approaches—identifying robust sets of low-distortion solutions among which predictions select—yield improved learning-augmented guarantees in higher-dimensional, less structured metric spaces?
• Is it possible to design randomized or multi-advice mechanisms that can, in expectation or in a Bayesian sense, beat the limitations proven here for deterministic learning-augmented approaches in the worst-case?

Conclusion

This paper rigorously establishes that learning-augmented matching on the line admits a deterministic mechanism that is simultaneously optimally consistent and optimally robust—attaining distortion $O(n)$2 when the prediction is perfect, and $O(n)$3 otherwise—across all top-$O(n)$4 and symmetric norm objectives. The mechanism’s structure, and the sharp upper and lower bounds, clarify that on the line the only effective “advice” is a prediction of the optimal assignment itself. The analytic techniques point toward broader blueprint for learning-augmented mechanism design in settings where partial orders and ambiguous regions can be explicitly characterized. These insights lay strong groundwork for further exploration of advice-robust ordinal mechanisms in richer metric and combinatorial domains.

---

References:

• "Optimal Metric Distortion for Learning-Augmented Matching on the Line" (2607.09038)
• "Optimal Metric Distortion for Matching on the Line" [Filos-Ratsikas et al., IJCAI 2025]
• "Distortion in metric matching with ordinal preferences" [Anari et al., EC 2023]
• "On Thin Perfect Matchings up to Polylogarithmic Factors" [Haqi & Gharan, (Haqi et al., 31 May 2026)]
• "Learning-Augmented Metric Distortion via (p,q)-Veto Core" [Berger et al., EC 2024]