{"id":"64144dc0-f4c9-4e7a-b4be-2a7e4498718e","arxiv_id":"2412.13622","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A school can enforce multiple diversity quotas while maximizing the minimum admission share of every type-combination group, and the paper gives a claimed-unique choice function and flow algorithms for it.","lead":"The paper introduces a fairness constraint, balanced representation, that stops a school from filling diversity quotas with students from only some type combinations, and claims a unique choice function obeys it. It also proposes faster flow-network algorithms, but a load-bearing equivalence in the proof is flawed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 4's floor-based binary search can return a crucial vector that does not achieve balanced representation: for group sizes 3 and 5 with capacity 4, it returns or oscillates toward δ=(1,3), not the optimal (2,2).","rationale":"The reader's verdict is REJECT, and the rejection is warranted, but the most load-bearing flaw is not the one singled out as the weakest assumption. The proof of Theorem 1 does conflate linear cost minimization with lexicographic signature maximization, and the supplied argument is incomplete; however, the theorem may or may not be salvageable within this specific flow network. The decisive problem is the crucial-vector construction: Definition 13 uses floor(α·|S_u|), and Algorithm 4's binary search finds the largest α for which those floor targets are feasible, not the true max-min ratio. The example above shows a valid instance where the algorithm returns (1,3) instead of the balanced optimum (2,2), or fails to terminate. This directly refutes the paper's claim that Algorithms 3 and 5 satisfy balanced representation, which is one of the four headline properties. It also explains why the proof of Theorem 8 ('weakly larger than δ*_u ensures balanced representation') is incorrect: satisfying floor(α·|S_u|) does not imply that every group's selection ratio is at least α. I agree with the reader's overall rejection but identify a different, more concrete and fatal weakness.","tokens_in":22476,"tokens_out":44273,"duration_ms":389808,"concrete_test":"Run Algorithm 4 on the two-group instance with |S_A|=3, |S_B|=5, q=4, and only general seats (so every size-4 matching is rank-maximal). The balanced-optimal counts are (2,2), giving min selection ratio 2/5. Expected: the binary search either loops forever or returns δ=(1,3), and Algorithm 5 then outputs a matching with min selection ratio 1/3. This check isolates the crucial-vector failure from the rank-maximal/flow equivalence issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the proposed algorithms implement balanced representation, which maximizes the minimum selection ratio among maximal-diversity matchings. Algorithm 4 computes the crucial vector δ* by binary search on α, setting δ*_u = floor(α·|S_u|). This floor construction is not equivalent to the integer counts actually needed. Consider U={A,B} with |S_A|=3, |S_B|=5, capacity q=4, and only general seats, so every 4-student matching is maximal-diversity. The true max-min ratio is 2/5, uniquely achieved by counts (2,2): 2/3 and 2/5 have minimum 0.4; any matching with min ratio above 0.4 would need at least 2 from A and 3 from B, totaling 5 > q. Now run Algorithm 4. For any α < 2/3, floor(α·3)=1 and floor(α·5)=3, giving δ=(1,3), which is feasible (1+3=4). For any α ≥ 2/3, δ=(2,3), which is infeasible because the lower bounds sum to 5. Thus the binary search converges to the discontinuity at 2/3, not to α*=0.4: with exact arithmetic it oscillates forever between (1,3) and (2,3), and if it terminates it returns δ=(1,3). Algorithm 5, initialized with the top 1 student from A and top 3 from B, fills capacity and returns (1,3), whose minimum selection ratio is 1/3 ≈ 0.333 < 0.4. Thus the balanced representation property, one of the four fundamental properties in the abstract, is not achieved. The same flawed crucial vector also propagates into the rank-based Algorithm 3. This is a concrete, internal failure of the paper's central algorithmic claim, independent of the rank-maximal-flow equivalence in Theorem 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how a single school should select students when applicants have multiple attributes and diversity goals are expressed as ranked minimum/maximum quotas under the one-to-one convention. It defines groups as type combinations, introduces balanced representation (maximizing the minimum group selection ratio among rank-maximal matchings), and proposes a greedy choice function (Algorithm 1). The paper claims this function is the unique rule satisfying non-wastefulness, maximal diversity, justified envy-freeness, and balanced representation (Theorem 2), and presents polynomial algorithms based on ranked reservation graphs (Algorithms 2 and 3) and on a new four-layer flow network (Algorithms 4 and 5, Theorems 1 and 6-8). Section 9 shows that the choice function violates substitutability in the generalized deferred acceptance setting.","tokens_in":22777,"tokens_out":14137,"duration_ms":131964,"significance":"The motivation is genuinely relevant: real affirmative-action systems count a student under one category, and intersectional groups can be left out even when aggregate type quotas are met. Balanced representation is a natural addition to the smart-reserves literature, and the flow-network compactness idea is attractive when the number of types is small but the applicant pool is large. The paper is honest about relying on prior algorithms, has no fitted parameters, and the greedy choice function is a plausible route to the four stated axioms. However, the current version contains two load-bearing technical errors: the claimed equivalence between rank-maximal matchings and minimum-cost flows is false for three or more ranks, and the floor-based crucial vector does not characterize balanced representation. The algorithmic claims in Sections 7 and 8 are therefore not established, and the main theorems are either false as stated or only sketched.","major_comments":[{"comment":"The equivalence in Theorem 1 is false for r >= 3 because an edge cost equal to the rank index minimizes the sum of ranks, not the lexicographic signature. For signatures (0,2,0) and (1,0,1) with capacity q=2, the total costs are both 4 when rank costs are 1, 2, 3, while (1,0,1) is lexicographically better. Such signatures are realizable in the network of Section 4.2: take one student with types {t1,t0}, one student with type {t0}, and two students with type {t2}; give one rank-1 seat for t1, two rank-2 seats for t2, and rank-3 general seats for t0. A minimum-cost maximum flow can then return a matching that is not rank-maximal, so the flow-based algorithms may fail maximal diversity. All of Theorems 6, 7, and 8 rely on this equivalence.","section":"Section 4.2, Theorem 1"},{"comment":"The condition for a matching to have minimum selection ratio at least alpha is |M_u| >= ceil(alpha |S_u|), not |M_u| >= floor(alpha |S_u|). The floor construction is strictly weaker, and the binary search in Algorithm 4 computes the wrong threshold. Concretely, let |S_A|=3, |S_B|=5, q=4, with no special type quotas (only general seats). Balanced representation requires counts (2,2), giving max-min ratio 0.4. For every alpha < 2/3, Algorithm 4 sets delta=(1,3), which is feasible; for alpha >= 2/3, it sets delta=(2,3), which is infeasible. The search therefore converges to the discontinuity at 2/3 and returns delta=(1,3). Algorithm 5, initialized with the top 1 student from A and the top 3 from B, returns counts (1,3) with minimum selection ratio 1/3 < 0.4. Thus Algorithm 5 fails to implement balanced representation. This also invalidates Theorem 3 and the balanced-representation claim in Theorem 8. The correct target vector should use ceilings rather than floors.","section":"Section 6.1, Definition 13 and Algorithm 4"},{"comment":"The uniqueness proof is an outline rather than a rigorous proof and depends on the two defective ingredients above. The claim that 'any student from the same group as s1 cannot be selected by lambda' is asserted without proving that the exchange preserves maximal diversity and balanced representation, and the disconnected case ('if doing so were to violate maximal diversity... contradiction') presupposes the flow equivalence of Theorem 1. As a result, the central characterization is not established even independently of the algorithmic bugs.","section":"Section 6, Theorem 2"}],"minor_comments":[{"comment":"The proof says the binary search requires O(sqrt(|S|)) iterations, while the preceding paragraph and Algorithm 4 suggest O(log |S|); the iteration bound and termination condition need to be made precise and consistent.","section":"Section 8.1, Theorem 7"},{"comment":"Algorithm 3 presumes that the input matching has the top delta*_u students matched in each group, but no pseudocode step enforces this; Example 8 describes an update ('We first update M...') that is not part of the algorithm.","section":"Algorithm 3"},{"comment":"In the proof of justified envy-freeness, the sentence 'suppose an unmatched student s in S*' should read s not in S*, since the argument concerns a student who was not initially selected.","section":"Section 8.2, Theorem 8"},{"comment":"There are typographical issues, including 'critical vector' for 'crucial vector', 'non-wasfulness' in Theorem 8, and 'studetns' in Section 8.2; these should be corrected in any revision.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The two major errors are in principle repairable: one could use rank-aware or lexicographic costs in the flow network, replace floors with ceilings in the crucial-vector construction, and rewrite the uniqueness proof. If the authors submit a revised version addressing these points, the paper could become a worthwhile contribution. As submitted, however, the central theorems and algorithms are incorrect, so I cannot recommend acceptance or a standard major-revision path within the current statement of results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper identifies a real gap—intersectional balance under the one-to-one convention—and the flow-network formulation is a genuinely more compact way to view the problem. But two load-bearing results are false as stated, so the four-property choice function and the fast algorithms don't currently deliver what the abstract promises.\n\nCredit where it's due. Balanced representation across type combinations is not in the cited work, and the flow network with type-combination nodes is a nice idea: it decouples the complexity from the number of students. The GDA extension is also honest about the substitutability failure. The paper is clearly written and the motivation is concrete.\n\nThe first soft spot is Theorem 1. Minimum-cost maximum flow with edge costs equal to rank index does not enforce rank-maximality. Two matchings with signatures (0,2,0) and (1,0,1) have the same total cost 4 under costs 1,2,3, but the second is lexicographically better. So a flow solver can return a matching that fails maximal diversity, and every flow-based algorithm in Section 8 inherits the problem.\n\nSecond, Algorithm 4's crucial vector uses floors: δ_u = floor(α·|S_u|). That is not equivalent to the integer counts balanced representation actually requires. The stress-test example is correct: with group sizes 3 and 5, capacity 4, and only general seats, the true balanced outcome is (2,2) with minimum ratio 0.4, but Algorithm 4 converges to δ=(1,3) with minimum ratio 1/3. This is a concrete counterexample to the claimed output, not a rounding quibble.\n\nThird, the uniqueness proof in Theorem 2 is a sketch. The flow-based case analysis doesn't fully establish the claim and silently assumes the floor target is the right integer target. It may be repairable, but as written it's not a proof.\n\nThese aren't cosmetic. Balanced representation is one of the four headline axioms; if the implementation misses it, the paper cannot be accepted as a correctness claim. The conceptual model and the combinatorial questions around it are still worth studying.\n\nWould I engage with it? Yes, as a serious referee. The flaws are localizable and the core idea has substance. I'd send it back for major revision, not desk-reject—but I'd want the flow equivalence and the crucial-vector computation fixed before believing the main theorems.","headline":"Nice, well-motivated intersectional balance criterion, but the rank-maximal/min-cost-flow equivalence and the floor-based crucial vector are both wrong, so the headline uniqueness and implementation claims do not hold as written.","tokens_in":23378,"tokens_out":2616,"would_cite":false,"duration_ms":25043,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B68"],"pacs":[],"model":"deepseek-v4-flash","headline":"A simple greedy choice function is the unique school-choice rule that simultaneously achieves maximal diversity, balanced representation across overlapping student types, non-wastefulness, and justified envy-freeness.","keywords":["balanced representation","school choice","one-to-one convention","diversity quotas","rank-maximal matching","flow network","maximal diversity","justified envy-freeness"],"falsifier":"Build a small instance with three quota ranks whose two candidate selections have signatures $\\langle 0,2,0\\rangle$ and $\\langle 1,0,1\\rangle$—the counts of matched seats at ranks 1, 2, and 3. With ranks priced at 1, 2, and 3, both selections have total cost 4, so a minimum-cost maximum-flow solver can return the $\\langle 0,2,0\\rangle$ selection even though $\\langle 1,0,1\\rangle$ is the rank-maximal one. Running the paper's flow-network algorithm on such an instance and checking whether its output is rank-maximal would settle whether Theorem 1, and with it the maximal-diversity guarantee of the flow-based algorithms, holds.","tokens_in":22154,"feed_emoji":"🎓","tokens_out":16843,"duration_ms":133122,"temperature":0.7,"pith_summary":"The paper tackles a gap in school choice with diversity quotas: when a student belongs to several protected types but is counted under only one of them, existing algorithms can meet every type quota while leaving some type combinations entirely unrepresented—for instance, quotas for aborigines and for girls can both be filled while no aboriginal girl is admitted. To fix this, the authors define balanced representation as the maximization of the smallest selection ratio $|M_u|/|S_u|$ across all type-combination groups $u$, among matchings that already achieve maximal diversity. They propose a greedy choice function—admit a student exactly when they can be added without breaking maximal diversity or balanced representation—and prove it is the unique rule satisfying four axioms: non-wastefulness, maximal diversity, justified envy-freeness, and balanced representation. They also reformulate the problem as a flow network whose size depends on the number of types and type combinations rather than the number of students, giving strongly polynomial implementations for the small-type regime found in real admissions systems. If correct, this gives schools a principled and computationally practical answer: one selection rule, and only one, that balances overlapping diversity goals.","feed_headline":"Four axioms single out one fair school-choice rule","feed_subtitle":"Existing quotas can be met while entire student groups are left out; balanced representation closes that gap.","key_machinery":"Three pieces carry the argument. Balanced representation is defined through the selection ratio $|M_u|/|S_u|$ for each type-combination group $u$, with the requirement to maximize the minimum ratio subject to maximal diversity; a supporting definition of a general selection ratio lets the same algorithms handle proportional, egalitarian, or other targets. The greedy choice function (Algorithm 1) processes students in descending priority order and admits a student exactly when the set of previously admitted higher-priority students together with the candidate can still be extended to a matching of size at most $q$ that achieves maximal diversity and balanced representation—this greedy rule is the object whose uniqueness Theorem 2 establishes. The flow network is a four-layer directed graph whose nodes are group, type, (type, rank) quota, and capacity nodes; all edges have capacities, and only edges from a type to its ranked quota nodes carry cost equal to the rank index. The paper's Theorem 1 claims that a minimum-cost maximum flow in this network is exactly a rank-maximal matching, and this equivalence is what lets flow solvers replace bipartite matching algorithms in the implementation.","core_discovery":"The central claim is a uniqueness theorem for a single school choosing students under the one-to-one convention, where each applicant has several types but fills a seat of exactly one type. The new property, balanced representation, requires the chosen set to be matchable in a way that maximizes the minimum selection ratio $\\min_{u\\in U} |M_u|/|S_u|$ over all type-combination groups, subject to the prior requirement of maximal diversity, meaning the chosen set must extend to a rank-maximal matching in the ranked reservation graph. The proposed choice function, Algorithm 1, scans students in decreasing priority order and keeps a student whenever the previously kept students together with this student can still be matched within the school's capacity while satisfying maximal diversity and balanced representation. The paper's Theorem 2 asserts that this greedy rule is the unique choice function satisfying non-wastefulness, maximal diversity, justified envy-freeness, and balanced representation. To implement it, the paper introduces a four-layer flow network whose nodes are groups, types, ranked quotas, and the capacity node, and asserts (Theorem 1) that minimum-cost maximum flows in this network correspond exactly to rank-maximal matchings, yielding strongly polynomial running times independent of the number of students.","pith_inferences":["The uniqueness theorem is tied to the exact maximin definition of balance: a school that wanted to break ties beyond the maximin level—say by then maximizing the second-smallest ratio, or minimizing variance—would get a different rule, so the characterization does not cover these neighboring fairness goals.","The claimed flow-to-matching equivalence holds only if minimizing total rank cost is the same as lexicographic optimization of the rank counts; replacing the rank-index costs with superincreasing weights (growing so fast in the rank that any rank-1 edge outweighs all lower-priority edges) would make the correspondence exact, at the price of large numeric costs.","The imbalance the paper identifies is a general phenomenon of one-to-one quota systems: vaccine rationing, affirmative-action admissions, and reserved-seat elections can all meet per-category quotas while silently excluding intersectional groups, so balanced representation reads as a template for intersectional fairness beyond school choice.","The substitutability failure leaves the multi-school extension without a stability theorem; the paper does not ask which weaker stability or incentive guarantees the deferred-acceptance outcome still possesses, which is a natural next question."],"forward_implications":["A school computing the greedy choice function resolves every fairness dispute over who gets in: its selection is the only one that satisfies non-wastefulness, maximal diversity, justified envy-freeness, and balanced representation simultaneously.","Because the flow network has one node per group, type, and ranked type quota rather than per student, the running times—$O(m\\log n\\,(m+n\\log n))$ per feasibility check and $O(|S|m\\log n\\,(m+n\\log n))$ overall—become strongly polynomial and essentially independent of the number of students whenever the number of privilege types is small, as in Brazilian, Indian, and Chilean admissions.","Balanced representation is reached in two stages: binary search over target quotas computes the maximum equalized selection ratio, then a second pass admits students in priority order while preserving it; both stages reduce to repeated min-cost max-flow computations.","The algorithms are not tied to the percentage selection ratio: any choice of lower and upper group targets from the general selection ratio definition works with the same feasibility machinery, covering proportional and egalitarian balance goals.","Plugging the choice function into the generalized deferred acceptance algorithm yields a multi-school market, but the function does not satisfy substitutability, so the standard matching-with-contracts stability guarantee is not available for that market."],"supporting_citations":[{"why":"Sets up the one-to-one convention and the \"smart reserves\" model of maximal diversity that the paper's balanced representation builds on, from Indian college admissions.","marker":"[Sönmez and Yenmez, 2022]"},{"why":"Supplies the multi-rank reserves framework and ranked reservation graphs, together with the rank-maximal matching routine that Algorithms 2 and 3 call.","marker":"[Aziz and Sun, 2021a]"},{"why":"Defines rank-maximal matchings and the signature of a matching, the lexicographic comparison that underlies maximal diversity and the greedy choice function.","marker":"[Irving et al., 2006]"},{"why":"Introduces dynamic priority and hard versus soft bounds, the interpretation that turns minimum and maximum quotas into two ranks of quotas.","marker":"[Ehlers et al., 2014]"},{"why":"Provides the generalized deferred acceptance algorithm and the substitutability condition that Theorem 9 shows the new choice function fails.","marker":"[Hatfield and Milgrom, 2005]"},{"why":"Supplies the minimum-cost maximum-flow algorithms, residual-graph arguments, and strong polynomial-time bounds used in the flow-network sections.","marker":"[Ahuja et al., 1993]"},{"why":"Source of the general selection ratio in Definition 9, which lets the same validity-checking machinery cover proportional and egalitarian balance targets.","marker":"[Sun et al., 2021]"}],"fun_headline_variants":["Unique school-choice rule balances all type groups","Flow networks unlock fair school diversity selection","One rule, four axioms, balanced school choice","Balanced representation: a unique fair choice algorithm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything the flow-based algorithms promise rests on the claim that minimizing the total of the quota ranks always produces the matching that is best at the most important quotas first; that equivalence is not guaranteed, since two different student selections can have the same total rank cost even when one of them is strictly better at the high-priority quotas.","fun_headline_variants_meta":{"raw":{"variants":["Unique school-choice rule balances all type groups","Flow networks unlock fair school diversity selection","One rule, four axioms, balanced school choice","Balanced representation: a unique fair choice algorithm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000493,"raw_usage":{"total_tokens":2430,"prompt_tokens":965,"completion_tokens":1465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1409}},"tokens_in":581,"tokens_out":1465,"duration_ms":11351,"temperature":1.0,"reasoning_tokens":1409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:58:31.692713+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a small instance with three quota ranks whose two candidate selections have signatures $\\langle 0,2,0\\rangle$ and $\\langle 1,0,1\\rangle$—the counts of matched seats at ranks 1, 2, and 3. With ranks priced at 1, 2, and 3, both selections have total cost 4, so a minimum-cost maximum-flow solver can return the $\\langle 0,2,0\\rangle$ selection even though $\\langle 1,0,1\\rangle$ is the rank-maximal one. Running the paper's flow-network algorithm on such an instance and checking whether its output is rank-maximal would settle whether Theorem 1, and with it the maximal-diversity guarantee of the flow-based algorithms, holds.","supporting_citations":[],"review_version":1}