One-Sided Matching Market with Advice

• The paper demonstrates that auxiliary advice compresses global structure, enabling optimal allocation despite online, informational, or computational constraints.
• Methodologies include geometric partitioning, information-theoretic bounds, and equilibrium certificates to precisely quantify and leverage advice complexity (e.g., n-1 bits).
• Implications span online matching, bandit learning, and welfare maximization, showing how limited side information can interpolate between local decisions and global optimality.

Searching arXiv for the cited papers to ground the article in the latest records. One-sided matching market with advice denotes a family of allocation and learning models in which one side of a market is passive or fixed—items, servers, firms, or arms—while the other side—agents, requests, or players—is matched subject to online, informational, or computational constraints, and auxiliary information is supplied to improve performance. Across the recent literature, “advice” takes several technically distinct forms: an oracle’s binary tape in online matching on the line (Csaba et al., 2024), known preferences on one side of a two-sided stable matching problem (Pagare et al., 17 Sep 2025), equilibrium prices and allocations in the Hylland–Zeckhauser scheme (Vazirani et al., 2020), and a bounded number of value queries per agent in welfare-maximizing one-sided matching (Amanatidis et al., 2020). Despite their differences, these models share a common theme: limited side information can sharply reduce uncertainty, alter lower bounds, and interpolate between purely local online decisions and globally coordinated allocations.

1. Conceptual scope and market interpretations

In the online geometric model studied in "On the Advice Complexity of Online Matching on the Line" (Csaba et al., 2024), the market consists of a fixed set of servers on the real line and a sequence of requests arriving online. Each request must be matched irrevocably to an unmatched server, and the objective is to minimize the total $L_1$ distance. The paper explicitly connects this to a one-sided matching market interpretation: servers are fixed items or facilities with no preferences, requests are sequentially arriving agents, and each agent’s objective is solely distance minimization.

In the pure-exploration stable matching model of "Optimal Algorithms for Bandit Learning in Matching Markets" (Pagare et al., 17 Sep 2025), the underlying environment is two-sided, but the one-sided-advice regime is defined by asymmetry of information: arm-side preferences are known to the platform, whereas player-side preferences must be learned from stochastic rewards. Here advice is not a price or a query answer but a full ranking $\succ_{a_k}$ for each arm $a_k$, which constrains the blocking pairs that matter for identifying the unique stable matching.

In the welfare-maximization framework of "A Few Queries Go a Long Way: Information-Distortion Tradeoffs in Matching" (Amanatidis et al., 2020), one-sided matching is the standard setting with $n$ agents and $n$ items, where agents’ ordinal preferences are induced by hidden cardinal valuations. Advice is operationalized as elicited cardinal information through value queries, and the performance criterion is distortion, namely the ratio between optimal welfare and achieved welfare.

In the Hylland–Zeckhauser line of work formalized in "Computational Complexity of the Hylland-Zeckhauser Scheme for One-Sided Matching Markets" (Vazirani et al., 2020), the market is one-sided and fractional: agents buy probability shares of items subject to equal budgets and a size constraint. Advice is most naturally interpreted as a communicated equilibrium object—typically a price vector $p$ or a certificate $(p,x)$—that lets agents optimize locally while the market clears globally.

These formulations do not define a single universal model of advice. Rather, they instantiate a broader research program in which side information modifies the feasible decision set, the adversary’s alternate set, or the algorithm’s online ambiguity.

2. Formal models of advice

The online matching-on-the-line model fixes servers

$S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$

and processes requests $r_1,\ldots,r_n$ online. If $\pi$ denotes the assignment permutation, the matching cost is

$\succ_{a_k}$0

and competitiveness is measured by $\succ_{a_k}$1 (Csaba et al., 2024). Advice is provided in the tape model: an oracle writes an infinite binary tape, and the online algorithm may read bits as needed. The advice complexity is the total number of bits read.

The one-sided learning model in bandit matching markets fixes players

$\succ_{a_k}$2

with stochastic rewards

$\succ_{a_k}$3

where preferences of player $\succ_{a_k}$4 are induced by the means $\succ_{a_k}$5 (Pagare et al., 17 Sep 2025). A matching is stable if there is no blocking pair

$\succ_{a_k}$6

In the one-sided-advice setting, the arm-side rankings $\succ_{a_k}$7 are known to the platform, whereas player-side preferences are unknown and learned through sampling.

In the distortion framework, each agent $\succ_{a_k}$8 has a valuation function $\succ_{a_k}$9 over items, inducing an ordinal ranking, and a mechanism is allowed at most $a_k$0 value queries per agent (Amanatidis et al., 2020). The objective is to output a perfect matching maximizing social welfare

$a_k$1

Advice here is the collection of elicited values, and the worst-case loss is measured by distortion.

In the Hylland–Zeckhauser scheme, each agent chooses a fractional bundle $a_k$2 solving

$a_k$3

with equal budget $a_k$4 and item prices $a_k$5 (Vazirani et al., 2020). Advice takes the form of equilibrium prices and allocations that satisfy the KKT and market-clearing conditions, so that local optimization and global feasibility coincide.

A plausible implication is that “advice” in one-sided matching research is best understood functionally rather than syntactically: it is any exogenous information that compresses the latent global structure needed for efficient or stable allocation.

3. Online one-sided matching on the line: exact optimality and trade-offs

The central exact result for online matching on the line is that there exists a $a_k$6-competitive deterministic online algorithm with advice complexity $a_k$7, and no $a_k$8-competitive deterministic online algorithm can read fewer than $a_k$9 bits (Csaba et al., 2024). Thus the exact advice complexity is tight.

The upper bound is realized by the algorithm LR. Let $n$0 be an optimal permutation, and partition requests into

$n$1

When a request arrives and unmatched servers exist on both sides, one advice bit specifies whether the request belongs to $n$2 or $n$3. LR then matches to the greatest unmatched server less than the request if the bit is $n$4, and to the least unmatched server greater than the request if the bit is $n$5. If all unmatched servers lie on one side, or an exact server is present, no advice is needed. The last request never needs advice, so the total is at most $n$6 bits (Csaba et al., 2024).

The correctness argument relies on two structural propositions for optimal matchings on the line. The first imposes order constraints that forbid crossing patterns that would increase cost. The second gives a switching symmetry: if two requests lie on the same side of both matched servers, then swapping the matched servers preserves optimality. Induction together with these propositions shows that LR reproduces an optimal matching (Csaba et al., 2024).

The matching lower bound is information-theoretic. The paper constructs a family $n$7 of size $n$8 with servers $n$9 and recursively designed request sequences. Distinct instances force different optimal decisions at a common prefix, which implies the required advice words must be pairwise prefix-free. By the Kraft inequality,

$n$0

so a prefix-free family of size $n$1 requires maximum word length at least $n$2 (Csaba et al., 2024).

The same work establishes a tunable advice–competitiveness trade-off. For each integer $n$3 with $n$4, there exists a $n$5-competitive online algorithm with advice complexity

$n$6

where $n$7 is the span of the servers and $n$8 is the competitive ratio of the best online algorithm without advice on inputs of size $n$9 (Csaba et al., 2024). The algorithm, DIVIDE$p$0, partitions servers into $p$1 contiguous blocks, uses advice to encode all cross-block matches in an optimal solution, serves those optimally with LR, and delegates the remaining intra-block subproblems to any no-advice subroutine $p$2.

The trade-off interpolates continuously. With $p$3, the advice is $p$4 and the algorithm reduces to running $p$5 globally, achieving $p$6. With $p$7, the structure collapses to exact optimality, consistent with the $p$8-bit characterization (Csaba et al., 2024). For real coordinates, a rescaling and rounding argument preserves the asymptotic guarantees up to negligible additive error.

4. One-sided learning with known arm preferences

In the bandit learning formulation, one-sided advice means that arm-side preferences are known. This known preference structure sharply reduces the alternate set that a lower-bound or testing procedure must consider (Pagare et al., 17 Sep 2025).

For a candidate stable matching $p$9, define

$(p,x)$0

the set of arms that would prefer player $(p,x)$1 over their current match, and let

$(p,x)$2

The potential blocking arms for player $(p,x)$3 are

$(p,x)$4

Any alternate instance that makes the true stable matching unstable must induce, for some $(p,x)$5 and some $(p,x)$6, the misordering

$(p,x)$7

This characterization is the key way advice enters both the lower bound and the algorithm (Pagare et al., 17 Sep 2025).

The information-theoretic lower bound is expressed through the convex program LO1:

$(p,x)$8

From this, the paper derives the asymptotic lower bound

$(p,x)$9

where

$S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$0

and the maximizing proportion $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$1 is unique (Pagare et al., 17 Sep 2025).

The algorithmic response is ATT1, a deterministic Top-Two-style method. At each iteration it estimates means, computes a current stable matching via an Arm-proposing Gale–Shapley subroutine $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$2, constructs the empirical sets $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$3 using arm-side advice, and samples according to empirical indices and anchor functions motivated by the KKT conditions of LO1 (Pagare et al., 17 Sep 2025).

ATT1 stops when the estimated stable matching is unique and a generalized-likelihood-ratio index exceeds

$S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$4

It is $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$5-correct and asymptotically optimal:

$S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$6

The paper further develops a system of fluid ODEs that characterizes the idealized trajectory ATT1 tracks (Pagare et al., 17 Sep 2025).

Relative to the two-sided learning setting, the one-sided-advice regime eliminates arm-side divergence terms from the lower bound and confines attention to edges in $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$7. This suggests that known preferences on one side act as a structural sparsifier of the hypothesis space.

5. Query-based advice and distortion in one-sided matching

In the welfare-maximization literature, advice appears as a limited budget of value queries per agent. The starting point is stark: with unit-sum valuations, the best deterministic ordinal-only mechanism has distortion $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$8 (Amanatidis et al., 2020). The main result is that a small number of carefully structured queries greatly improves this guarantee.

The core mechanism is $S=\{s_1,\ldots,s_n\},\qquad s_1\le \cdots \le s_n,$9-TSF. For each agent $r_1,\ldots,r_n$0, it queries the top item and value $r_1,\ldots,r_n$1, defines thresholds

$r_1,\ldots,r_n$2

and uses value queries plus binary search to partition the ranked list into $r_1,\ldots,r_n$3 blocks $r_1,\ldots,r_n$4. It then defines simulated values $r_1,\ldots,r_n$5 for items in block $r_1,\ldots,r_n$6 and $r_1,\ldots,r_n$7 otherwise, computes a maximum-weight matching under $r_1,\ldots,r_n$8, and outputs it (Amanatidis et al., 2020).

The guarantee is explicit: $r_1,\ldots,r_n$9-TSF makes

$\pi$0

queries per agent and achieves distortion at most

$\pi$1

By choosing $\pi$2, one gets $\pi$3-type distortion with $\pi$4 queries per agent. By choosing $\pi$5, one obtains constant distortion with $\pi$6 queries per agent (Amanatidis et al., 2020).

The lower bounds are correspondingly strong. For unrestricted valuations, any deterministic mechanism using $\pi$7 value queries per agent has distortion

$\pi$8

For unit-sum valuations, any deterministic mechanism using

$\pi$9

queries per agent has distortion

$\succ_{a_k}$00

A corollary is that constant distortion requires nearly logarithmic queries per agent (Amanatidis et al., 2020).

The paper also identifies a structured class of instances, the $\succ_{a_k}$01-well-structured instances, on which a simpler mechanism $\succ_{a_k}$02-FMM makes exactly $\succ_{a_k}$03 queries per agent and achieves distortion $\succ_{a_k}$04 (Amanatidis et al., 2020). For unit-sum valuations, the adaptive two-query mechanism FirstPositionAdaptive achieves

$\succ_{a_k}$05

distortion, yielding the first deterministic two-query sublinear-distortion guarantee in that regime (Amanatidis et al., 2020).

A notable feature of this line is that advice is endogenous to the mechanism. The designer selects what to query, and the welfare guarantee is a function of the amount of cardinal information elicited.

6. Price-based advice and the Hylland–Zeckhauser scheme

The Hylland–Zeckhauser scheme is not an online or learning model, but it provides a distinct and influential interpretation of advice in one-sided matching markets: prices and allocations can be communicated as equilibrium guidance (Vazirani et al., 2020).

Each agent optimizes a linear program with a size constraint and a budget constraint, and equilibrium is characterized by KKT conditions involving dual variables $\succ_{a_k}$06 and $\succ_{a_k}$07:

$\succ_{a_k}$08

Complementary slackness implies that if $\succ_{a_k}$09 then $\succ_{a_k}$10, and if $\succ_{a_k}$11 then the budget is tight. The resulting “bang-per-buck with offset” relation is

$\succ_{a_k}$12

for all consumed items $\succ_{a_k}$13 (Vazirani et al., 2020).

From a market-design perspective, a price vector $\succ_{a_k}$14 is actionable advice: given $\succ_{a_k}$15, each agent can compute an optimal bundle, and the designer can verify whether a proposed pair $\succ_{a_k}$16 is an equilibrium certificate by checking feasibility, budget conditions, complementarity, and support equalities (Vazirani et al., 2020). The paper emphasizes that such certificates can be checked efficiently to a prescribed precision, even though exact equilibria may be irrational.

The computational results sharply delineate when exact price advice is tractable. For $\succ_{a_k}$17 utilities, extended to bivalued utilities via an agent-wise affine transformation,

$\succ_{a_k}$18

the paper gives a combinatorial, strongly polynomial-time algorithm (Vazirani et al., 2020). In contrast, it constructs a $\succ_{a_k}$19 instance with only irrational equilibria and disconnected equilibrium sets, implying that exact equilibrium computation is not in PPAD and does not admit a convex programming formulation (Vazirani et al., 2020). The general problem is placed in FIXP via a polynomial-size algebraic fixed-point map $\succ_{a_k}$20 over a compact domain.

This body of results makes advice computationally meaningful. In simple utility domains, exact equilibrium prices can be generated and disseminated. In richer domains, advice must typically be approximate, algebraic, or verified numerically rather than represented as an exact rational object.

7. Comparative themes, limitations, and open directions

The four notions of advice differ in syntax, but they exhibit parallel structural roles. In online matching on the line, advice resolves left-versus-right ambiguity at precisely those arrivals where local geometry does not determine the optimal move (Csaba et al., 2024). In one-sided bandit learning, advice prunes the alternate set to potential blocking arms and thereby determines the critical KL comparisons (Pagare et al., 17 Sep 2025). In query-based welfare maximization, advice creates blockwise lower bounds on values that support optimization over a simulated instance (Amanatidis et al., 2020). In Hylland–Zeckhauser, advice provides equilibrium prices and allocations that certify optimal local behavior and aggregate clearing (Vazirani et al., 2020).

Several limitations recur. The online line results rely on the non-crossing structure of the line; extending them to trees or higher-dimensional metrics is identified as an open question (Csaba et al., 2024). The one-sided learning guarantees assume a unique stable matching, independent and stationary SPEF rewards, and asymptotics as $\succ_{a_k}$21 (Pagare et al., 17 Sep 2025). The distortion results are deterministic and do not analyze incentives or fairness (Amanatidis et al., 2020). The Hylland–Zeckhauser framework faces exact computational barriers arising from irrational equilibria, disconnected solution sets, and possible FIXP-hardness, which remains open (Vazirani et al., 2020).

A common misconception is that advice merely accelerates computation. The cited works show a stronger phenomenon: advice can change the attainable performance frontier itself. Exact optimality in online matching requires exactly $\succ_{a_k}$22 bits (Csaba et al., 2024); asymptotically optimal pure exploration becomes possible when one side’s preferences are known (Pagare et al., 17 Sep 2025); constant-distortion welfare guarantees require nearly logarithmic value queries per agent (Amanatidis et al., 2020); and equilibrium price guidance may exist but be computationally non-rational in general (Vazirani et al., 2020).

Taken together, these results define one-sided matching market with advice not as a single mechanism class but as a research area centered on how partial global information can be encoded, elicited, or certified so as to overcome the intrinsic myopia of online, ordinal, stochastic, or decentralized matching procedures.