Accomplice Manipulation of the Deferred Acceptance Algorithm
Transcript of Accomplice Manipulation of the Deferred Acceptance Algorithm
Accomplice Manipulation of the Deferred AcceptanceAlgorithm
Hadi HosseiniPennsylvania State University
University Park PA, [email protected]
Fatima UmarRochester Institute of Technology
Rochester NY, [email protected]
Rohit VaishTIFR
Mumbai, [email protected]
ABSTRACTThe deferred acceptance algorithm is an elegant solution to the stablematching problem that guarantees optimality and truthfulness forone side of the market. Despite these desirable guarantees, it issusceptible to strategic misreporting of preferences by the agentson the other side. We study a novel model of strategic behaviorunder the deferred acceptance algorithm: manipulation through anaccomplice. Here, an agent on the proposed-to side (say, a woman)partners with an agent on the proposing sideβan accompliceβtomanipulate on her behalf (possibly at the expense of worseninghis match). We show that the optimal manipulation strategy foran accomplice comprises of promoting exactly one woman in histrue list (i.e., an inconspicuous manipulation). This structural resultimmediately gives a polynomial-time algorithm for computing anoptimal accomplice manipulation. We also study the conditionsunder which the manipulated matching is stable with respect to thetrue preferences. Our experimental results show that accomplicemanipulation outperforms self manipulation both in terms of thefrequency of occurrence as well as the quality of matched partners.
ACM Reference Format:Hadi Hosseini, Fatima Umar, and Rohit Vaish. 2021. Accomplice Manipula-tion of the Deferred Acceptance Algorithm. In Appears at the 3rd Games,Agents, and Incentives Workshop (GAIW 2021). Held as part of the Work-shops at the 20th International Conference on Autonomous Agents andMultiagent Systems., London, UK, May 2021, IFAAMAS, 15 pages.
1 INTRODUCTIONThe deferred acceptance (DA) algorithm [9] is a crowning achieve-ment of the theory of two-sided matching [18], and forms the back-bone of a wide array of real-world matching markets such as entry-level labor markets [21, 24] and school choice [1, 2]. Under thisalgorithm, one side of the market (colloquially, the men) makesproposals to the other side (the women) subject to either immediaterejection or tentative acceptance. A key property of the DA algorithmis stability which says that no pair of unmatched agents should prefereach other over their assigned partners. This property has playeda significant role in the long-term success of several real-worldmatching markets [22, 23].
Permission to make digital or hard copies of part or all of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. Copyrights for third-party components of this work must be honored.For all other uses, contact the owner/author(s).Appears at the 3rd Games, Agents, and Incentives Workshop (GAIW 2021). Held aspart of the Workshops at the 20th International Conference on Autonomous Agents andMultiagent Systems., Aziz, Ceppi, Dickerson, Hosseini, Lev, Mattei, McElfresh, Zick(chairs), May 2021, London, UK. Β© 2021 Copyright held by the owner/author(s).
The attractive stability guarantee of the DA algorithm, however,comes at the cost of incentives, as any stable matching procedure isknown to be vulnerable to strategic misreporting of preferences [20].The special proposal-rejection structure of the DA algorithm makestruth-telling a dominant strategy for the proposing side, i.e., themen [8, 20], implying that any strategic behavior must occur on theproposed-to side, i.e., the women. This model of strategic behaviorby a womanβwhich we call self manipulationβhas been the subjectof extensive study in economics and computer science [6β8, 11, 16,17, 26, 27].
Our interest in this work is in studying a different model of strate-gic behavior under the DA algorithm called manipulation through anaccomplice [4]. Under this model, a woman reports her preferencestruthfully, but asks an agent on the proposing side (a.k.a. an accom-plice) to manipulate the outcome on her behalf, possibly worseninghis match in the process.
Such a strategic alliance can naturally arise in the assignment ofstudents to schools under a school-proposing setup, where a βwell-connectedβ student could have a school administrator manipulateon his/her behalf, possibly at a small loss to the school. Similarly,in a student-proposing setting, schools can strategize by makingthemselves appear less attractive to students from low-income back-grounds (say, by increasing rent of dormitories or requiring studentsto purchase expensive uniforms), thus forcing a change in the stu-dentsβ preferences [12]. Yet another justification for accomplicemanipulation comes from thinking about strategic behaviour as acontrol problem, wherein a woman can bribe a man to lie on herbehalf. Bribery has been extensively studied in computational socialchoice in the context of voting, and our work can be seen as investi-gating this phenomenon in the two-sided matching framework.
At first glance, manipulation through an accomplice might notseem any more powerful than self manipulation, as the latter providesdirect control over the preferences of the manipulator. Interestingly,there exist instances where this intuition turns out to be wrong.
Example 1.1 (Accomplice vs. self). Consider the following pref-erence profile where the DA outcome is underlined. The notationβπ1 : π€3 π€2 π€1 π€4β denotes that for manπ1, the first choice womanis π€3, the second choice is π€2, and so on.
π1: π€β3 π€2 π€1 π€4 π1: π4 πβ
3 π1 π2
π2: π€1 π€β4 π€2 π€3 π€2: πβ
4 π3 π2 π1
π3: π€2 π€4 π€β1 π€3 π€3: π3 πβ
1 π2 π4
π4: π€β2 π€1 π€3 π€4 π€4: πβ
2 π1 π3 π4
Suppose π€1 seeks to improve her match via manipulation. Theoptimal self manipulation strategy forπ€1 is truth-telling, asπ2 is theonly man who proposes to her under the DA algorithm. On the other
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
Figure 1: Comparing no-regret accomplice manipulation andself manipulation against truthful reporting (top) and againsteach other (bottom).
hand, π€1 can improve her outcome by asking π1 to misreport onher behalf. Indeed, if π1 misreports by declaring β»β²
π1B π€1 β» π€3 β»π€2 β» π€4, then π€1βs match improves from π2 to π3 (the new DAmatching is marked by β). Notice that the accompliceπ1 preserveshis initial match, meaning he does not incur any βregretβ. β‘
The above example highlights that accomplice manipulationcould, in principle, have an advantage over self manipulation. How-ever, it is not a priori clear how frequent such an advantage might bein a typical matching scenario. To investigate the latter question, wetake a quick experimental detour.
Accomplice manipulation is a viable strategic behavior. We sim-ulate a two-sided matching scenario for an increasingly larger setof agents (specifically, π β {3, . . . , 40}, where π is the number ofmen/women) and for each setting, generate 1000 preference profilesuniformly at random. For each profile, we compute the optimal selfmanipulation under the DA algorithm for a fixed woman [26], aswell as the optimal accomplice manipulation by any man (we allowany man to be chosen as an accomplice as long as he is not worseoff, i.e., a no-regret accomplice manipulation). Figure 1 illustratesthe fraction of instances where accomplice and self manipulationare strictly more beneficial than truthful reporting, as well as howthey compare against each other.
Example 1.1 and Figure 1 suggest that the incentive for manipu-lation through an accomplice is not only present but actually moreprevalent than self manipulation. Additionally, as we discuss laterin our experimental results, women are expected to receive bettermatches when manipulating through an accomplice (Figure 2). Thesepromising observations call for a systematic study of the structuraland computational aspects of the accomplice manipulation problem,which is the focus of our work.
Our contributions. We consider two models of strategic behaviorβno-regret manipulation (wherein the accompliceβs own match doesnβtworsen upon misreporting) and with-regret manipulation (where theaccomplice could get a worse match)βand make the followingcontributions:
β’ No-regret manipulation: Our main theoretical result (Theorem 4.3)is that any optimal no-regret accomplice manipulation can be sim-ulated by promoting exactly one woman in the true preferencelist of the accomplice; in other words, the manipulation is in-conspicuous [27]. This structural finding immediately gives apolynomial-time algorithm for computing an optimal manipula-tion (Corollary 2). We also show that the inconspicuous no-regretstrategy results in a matching that is stable with respect to the truepreferences (Corollary 3).
β’ With-regret manipulation: For the more permissible strategyspace that allows the accomplice to incur regret, the optimal ma-nipulation strategy once again turns out be inconspicuous (Theo-rem 5.2). However, in contrast to the no-regret case, the inconspic-uous with-regret strategy is no longer guaranteed to be stability-preserving (Example 5.1). Nevertheless, any blocking pair can beshown to necessarily involve the accomplice (Proposition 2). Thisproperty justifies the use of an accomplice who can be encouragedto tolerate some regret to benefit a woman.
β’ Experiments: On the experimental front, we work with prefer-ences generated uniformly at random, and find that accomplicemanipulation outperforms self manipulation with respect to thefrequency of occurrence, the quality of matched partners, and thefraction of women who can improve their matches (Section 6).
2 RELATED WORKThe impossibility result of Roth [20] on the conflict between strat-egyproofness and stability has led to extensive follow-up research.Much of the earlier work in this direction focused on truncationmanipulation [5, 10, 14, 25], where the misreported preference listis required to be a prefix of the true list. In the context of accom-plice manipulation, however, the truncation manipulation problembecomes trivial. Indeed, for any fixed accomplice, it is easy to seethat a truncation strategy is never better than truthful reporting as theset of proposals under the former is always a subset of those underthe latter.
The literature on self manipulation via permutation is more recentand has focused on computational aspects. Teo et al. [26] provided apolynomial-time algorithm for computing the optimal permutationmanipulation by a woman under the men-proposing DA algorithm.Vaish and Garg [27] showed that an optimal permutation manipula-tion is, without loss of generality, inconspicuous. They also studiedconditions under which the manipulated outcome is stable with re-spect to the true preferences. Deng et al. [7] studied permutationmanipulation by a coalition of women.
Huang [13] has studied (weakly) Pareto improving permutationmanipulation by a coalition of men, revisiting the result of Dubinsand Freedman [8] on the impossibility of manipulations that arestrictly improving for every member of the coalition.
The accomplice manipulation model was proposed by Bendlinand Hosseini [4], who noted that manipulation through an accom-plice can be strictly more preferable for the woman than optimal selfmanipulation. However, they left the structural and computationalquestions open.
Balinski and SΓΆnmez [3] studied a closely related problem inschool choice wherein the students have an incentive to performworse on tests, or make themselves appear less preferable to collegeswhen the college-optimal algorithm is used. Hatfield et al. [12]similarly showed that in a student-optimal mechanism, schools havethe incentive to deliberately make themselves look less attractive toβundesirableβ students. For example, a private school that is legallyrequired to cap its tuition fee for low-income students could makeitself less attractive by increasing the rent in dormitories or requiringthe students to purchase expensive uniforms.
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
3 PRELIMINARIES3.1 Stable Matching Problem
Problem setup. An instance of the stable marriage problem [9] isspecified by the tuple β¨π,π , β»β©, where π is a set of π men,π is aset of π women, and β» is a preference profile which consists of thepreference lists of all agents. The preference list of any manπ β π ,denoted by β»π , is a strict total order over all women inπ (for anyπ€ βπ , the list β»π€ is defined analogously).
We use the shorthand π€1 βͺ°π π€2 to denote βeither π€1 β»π π€2or π€1 =π π€2β (the latter relation denotes that man π is indifferentbetween π€1 and π€2), and write β»βπ to denote the preference listsof all men and women except manπ; thus, β»= {β»βπ, β»π}.
Stable matchings. A matching is a function ` : π βͺπ β π βͺπsuch that ` (π) β π for all π β π , ` (π€) β π for all π€ β π , and` (π) = π€ if and only if ` (π€) =π. A matching ` admits a blockingpair with respect to the preference profile β» if there is a man-womanpair (π,π€) who prefer each other over their assigned partners under`, i.e., π€ β»π ` (π) and π β»π€ ` (π€). A stable matching is one thatdoes not admit any blocking pair. We will write πβ» to denote theset of all matchings that are stable with respect to β». In addition,for any pair of matchings `, ` β², we will write ` βͺ°π ` β² to denote` (π) βͺ°π ` β²(π) for allπ β π (and ` βͺ°π ` β² for the women).
Deferred acceptance algorithm. Given a preference profile β», theDeferred Acceptance (DA) algorithm of Gale and Shapley [9] pro-ceeds in rounds. In each round, the algorithm consists of a proposalphase, where each man who is currently unmatched proposes tohis favorite woman from among those who have not rejected himyet, followed by a rejection phase where each woman tentativelyaccepts her favorite proposal and rejects the rest. The algorithmterminates when no further proposals can be made. Gale and Shap-ley [9] showed that given any profile β» as input, the DA algorithmalways returns a stable matching as output; we denote this matchingby DA(β»). They also observed that this matching is men-optimal,i.e., it assigns each man his favorite stable partner among all stablematchings in πβ». McVitie and Wilson [19] subsequently showed thatthis matching is also women-pessimal.
PROPOSITION 1 ([9, 19]). Let β» be a preference profile and let` B DA(β»). Then, ` β πβ». Furthermore, for any ` β² β πβ», ` βͺ°π ` β²
and ` β² βͺ°π `.
Accomplice manipulation. Under this model of strategic behav-ior, a woman π€ , instead of misreporting herself, has a man π pro-vide a manipulated preference list, say β»β²
π , in order to improve hermatch. Given a preference profile β», we say that π€ can manipu-late through accomplice π if ` β²(π€) β»π€ ` (π€), where ` B DA(β»),β»β²B {β»βπ, β»β²
π}, and ` β² B DA(β»β²). We will often refer to (π,π€)as the manipulating pair (not to be confused with a blocking pair).
Throughout this paper, any manipulation will be assumed tobe optimal unless stated otherwise. That is, there exists no otherlist β»β²β²
π for the accomplice π such that ` β²β²(π€) β»π€ ` β²(π€), whereβ»β²β²B {β»βπ, β»β²β²
π}, and ` β²β² B DA(β»β²β²). Note that we assume thatthe manipulator has full information about the preferences of otheragents. Extending our results to incomplete or uncertain informationsettings is an interesting direction for future research.
No-regret and with-regret manipulation. We say that the accom-plice π incurs regret if his match worsens upon misreporting, i.e.,` (π) β»π ` β²(π). It is known that the DA algorithm is strategyprooffor the proposing side [8], which means that no man can improvehis match by unilaterally misreporting his preferences. Therefore,for any man π β π and for any misreport β»β²
π , we have that` (π) βͺ°π ` β²(π). Thus, equivalently, we say that man π incursregret if ` (π) β ` β²(π).
We will consider two models of accomplice manipulation in thispaper: no-regret manipulation wherein only those misreports β»β²
π areallowed under which ` (π) = ` β²(π), and with-regret manipulation,where the accomplice is allowed (but not required) to incur regret.Thus, any no-regret strategy is also a with-regret strategy. Recall thatthe misreport in Example 1.1 was a no-regret manipulation.
Stability relaxations. For any preference profile β» and a fixedmanπ β π , we say that a matching ` isπ-stable [4] with respect toβ» if any blocking pair (if one exists) in ` involves the manπ. That is,for any pair (πβ²,π€ β²) that blocks ` under β», we haveπβ² =π. Clearly,a stable matching is also π-stable. Under accomplice manipulation,it can be shown that any matching ` β² that is stable with respectto the manipulated profile (in particular, when ` β² = DA(β»β²)) is π-stable with respect to the true profile β» (Proposition 2). We note thatProposition 2 strengthens a result of Bendlin and Hosseini [4] whoproved a similar statement only for a DA matching. The proof ofthis result, along with all other omitted proofs, can be found in theappendix.
PROPOSITION 2. Let β» denote the true preference profile. Forany manπ, let β»β²B {β»βπ, β»β²
π}, and let ` β² β πβ»β² be any matchingthat is stable with respect to β»β². Then, ` β² is π-stable with respect toβ».
3.2 Structural ObservationsPush up/push down operations. Note that given a profile β», the
preference list of any manπ can be written as β»π= (β»πΏπ, ` (π), β»π
π),where ` = DA(β») and β»πΏ
π (respectively, β»π π) is the set of women
that π prefers to (respectively, finds less preferable than) ` (π).Interestingly, the DA outcome does not change even if each manπ arbitrarily permutes the parts of his list above and below hisDA-partner ` (π). This result, due to Huang [13], is recalled below.
PROPOSITION 3 ([13]). Let β» be a preference profile and let ` BDA(β»). For any man π β π , let β»β²
πB (ππΏ (β»πΏπ), ` (π), ππ (β»π
π)),where ππΏ and ππ are arbitrary permutations of β»πΏ
π and β»π π , respec-
tively. Let β»β²B {β»βπ, β»β²π}, and let ` β² B DA(β»β²). Then, ` β² = `.
Proposition 3 considerably simplifies the structure of accomplicemanipulations that we need to consider. Indeed, we can assumethat any manipulated list β»β²
π is such that the relative ordering ofagents in the parts above and below ` β²(π) is the same as underthe true list β»π , where ` β² B DA(β»β²) and β»β²B {β»βπ, β»β²
π} are thepost-manipulation DA outcome and preference profile, respectively.
This observation implies that, without loss of generality, any ma-nipulated list β»β²
π can be obtained from the true list β»π by onlypush up and push down operations, wherein a set of women ispushed up above the true match ` (π), and another disjoint set is
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
pushed below ` (π). Importantly, no permutation or shuffling op-eration is required as part of the manipulation. Formally, startingwith the true list β»π= (β»πΏ
π, ` (π), β»π π), we say that man π per-
forms a push up operation for a set π β π if the new list isβ»π βπ B (β»πΏ
π βͺπ, ` (π), β»π π \π ). Likewise, a push down operation
of a set π βπ results in β»π βπ B (β»πΏ
π \π, ` (π), β»π π βͺπ ).
For manipulation via push down operations only, Huang [13] hasshown that the resulting matching is weakly improving for all men.Together, with the fact that the DA algorithm is strategyproof for theproposing side (in our case the men) [8], we get that the DA partnerof the accomplice remains unchanged after a push down operation.
PROPOSITION 4 ([8, 13]). Let β» be the true preference profileand let ` B DA(β»). For any subset of women π βπ and any fixedaccomplice π β π , let β»β²B {β»βπ, β»π β
π } and ` β² B DA(β»β²). Then,` β² βͺ°π ` and ` β²(π) = ` (π).
The effect of push down operations for the proposed-to side is theexact opposite, as the resulting matching makes all women weaklyworse off.
Lemma 1. Let β» be the true preference profile and let ` B DA(β»).For any subset of women π β π , let β»β²B {β»βπ, β»π β
π } and ` β² BDA(β»β²). Then, ` βͺ°π ` β².
Lemma 1 shows that in order to improve the partner of thewoman π€ , the use of push up operations (by the accomplice) isnecessary. However, it is not obvious upfront whether push up alonesuffices; indeed, it is possible that the optimal strategy involves somecombination of push up and push down operations. To this end, ourtheoretical results will show that, somewhat surprisingly, pushingup at most one woman achieves the desired optimal manipulation(Theorems 4.3 and 5.2). This strategy is known in the literature asinconspicuous manipulation, which we define next.
Inconspicuous manipulation. Given a profile β» of true prefer-ences and any fixed accompliceπ, the manipulated list β»β²
π is said tobe an inconspicuous manipulation if the list β»β²
π can be derived fromthe true preference list β»π by promoting exactly one woman andmaking no other changes. The notion of inconspicuous manipula-tion has been previously studied in the context of self manipulation(whereπ€ misreports herself), where it was shown that an optimal selfmanipulation is, without loss of generality, inconspicuous [7, 27].
4 NO-REGRET ACCOMPLICEMANIPULATION
Let us start by observing that the DA matching after an arbitrary, i.e.,not necessarily push up, no-regret accomplice manipulation may notbe stable with respect to the true preferences.
Example 4.1. Consider the following preference profile wherethe DA outcome is underlined.π1 : π€β
2 π€β 1 π€3 π€4 π€5 π€1 : π
β 1 π3 πβ
2 π4 π5
π2 : π€β1 π€
β 2 π€3 π€4 π€5 π€2 : π
β 2 πβ
1 π3 π4 π5
π3 : π€1 π€β,β 3 π€4 π€2 π€5 π€3 : π
β,β 3 π1 π2 π4 π5
π4 : π€4 π€β,β 5 π€1 π€2 π€3 π4 : π
β,β 5 π3 π1 π2 π4
π5 : π€5 π€β,β 4 π€1 π€2 π€3 π€5 : π
β,β 4 π1 π2 π3 π5
Suppose the manipulating pair is (π3,π€4). The DA matches afterthe accompliceπ3 submits the manipulated list β»β²
π3B π€4 β» π€3 β»
π€1 β» π€2 β» π€5 are marked by β. The manipulation results in π€4being matched with her top choice π5 (i.e., β»β²
π3 is an optimal ma-nipulation), an improvement over her true matchπ4. Althoughπ3does not incur regret, the manipulated matching admits a blockingpair (π3,π€1) with respect to the true preferences. β‘
Notice that if insteadπ3 were to submit β»β²β²π3B π€4 β» π€1 β» π€3 β»
π€2 β» π€5 as his preference list in Example 4.1, then the resulting DAmatching (indicated by β ) would be stable with respect to the truepreferences while still allowing π€4 to match with π5 (i.e., β»β²β²
π3 isalso optimal). The manipulated list β»β²β²
π3 is derived from the true listβ»π3 through a no-regret push up operation. Our first main result ofthis section (Theorem 4.2) shows that this is not a coincidence: Theset of all stable matchings with respect to a profile after a no-regretpush up operation is always contained within the stable set of thetrue preference profile.
THEOREM 4.2 (NO-REGRET PUSH UP IS STABILITY PRESERV-ING). Let β» be a preference profile, and let ` B DA(β»). For anysubset of women π βπ and any manπ, let β»β²B {β»βπ, β»π β
π }, and` β² B DA(β»β²). Ifπ does not incur regret, then πβ»β² β πβ».
A primary consequence of Theorem 4.2 is that the DA matchingafter a no-regret accomplice manipulation is weakly preferred overthe true DA outcome by all women, while the opposite is true for themen.
Corollary 1. Let β» be a preference profile and let ` B DA(β»). Forany man π, let β»β²B {β»βπ, β»π β
π } and ` β² B DA(β»β²). If π does notincur regret, then ` β² βͺ°π ` and ` βͺ°π ` β².
PROOF. Since π does not incur regret, it follows from Theo-rem 4.2 that ` β² β πβ». Then, from Proposition 1, we have that` β² βͺ°π ` and ` βͺ°π ` β². β‘
As observed in Section 3.2, any manipulation by the accomplicecan be, without loss of generality, assumed to comprise only of pushup and push down operations. We will now show that combiningthese operations is not necessary. That is, any manipulation that isachieved by a combination of push up and push down operationscan be weakly improved by a push up operation alone (Lemma 2).We note that this result does not require the no-regret assumption,and applies to the with-regret setting as well.
Lemma 2. Let (π,π€) be a manipulating pair and let β» be a pref-erence profile. For any subsets of women π β π and π β π , letβ»β²B {β»βπ, β»π β
π } denote the preference profile after pushing up theset π , and β»β²β²B {β»βπ, β»π β,π β
π } denote the profile after pushing upπ and pushing down π in the true preference list β»π of man π. Let` B DA(β»), ` β² B DA(β»β²), and ` β²β² B DA(β»β²β²). If ` β²β²(π€) β»π€ ` (π€),then ` β²(π€) βͺ°π€ ` β²β²(π€).
Having narrowed down the strategy space to push up operationsalone, we will now turn our attention to inconspicuous manipulations(recall that such a manipulation involves promoting exactly onewoman in the accompliceβs true preference list to a higher position).We will show that any match for the manipulating woman π€ thatcan be obtained by pushing up a set of women can also be achievedby promoting exactly one woman in that set (Lemma 3). In otherwords, any no-regret push up operation is, without loss of generality,
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
inconspicuous. We note that although Lemma 3 assumes no regretfor the accomplice, the corresponding implication actually holdseven in the with-regret setting (see Lemma 4).
Lemma 3. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up without incurring regret.Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
We will now use the foregoing observations to prove our mainresult (Theorem 4.3).
THEOREM 4.3. If there is an optimal no-regret accomplice ma-nipulation, then there is an optimal inconspicuous no-regret accom-plice manipulation.
PROOF. From Proposition 3 (and subsequent remarks), we knowthat any accomplice manipulation can be simulated via push upand push down operations. Lemma 2 shows that any beneficialmanipulation that is achieved by some combination of pushing upa set π β π and pushing down π β π can be weakly improvedby only pushing up π . Finally, from Lemma 3, we know that anymatch for the manipulating woman π€ that is achieved by pushingup π βπ is also achieved by pushing up exactly one woman in π ,thus establishing the desired inconspicuousness property. β‘
Theorem 4.3 has some interesting computational and structuralimplications. First, the inconspicuousness property implies a straight-forward polynomial-time algorithm for computing an optimal no-regret accomplice manipulation (Corollary 2). Second, the DAmatch-ing resulting from an inconspicuous no-regret manipulation is stablewith respect to the true preferences (Corollary 3). Together, theseresults reconcile the seemingly conflicting interests of the manipula-tor (who wants to compute optimal manipulation efficiently) and thecentral planner (who wants the resulting matching to be stable withrespect to the true preferences).
Corollary 2. An optimal no-regret accomplice manipulation strat-egy can be computed in O(π3) time.
Corollary 3. The DA outcome from an inconspicuous no-regret ac-complice manipulation is stable with respect to the true preferences.
In summary, recall from Example 4.1 that an arbitrary optimalno-regret strategy may not be stability-preserving. Nevertheless, anyoptimal no-regret strategy admits an equivalent inconspicuous strat-egy (Theorem 4.3) which indeed preserves stability (Corollary 3).
5 WITH-REGRET ACCOMPLICEMANIPULATION
No-regret manipulations come at no cost for the accomplice and thusare a viable strategic behavior (as shown in Figure 1). Yet, a morepermissive strategy space may allow for the accomplice to incursome regret. Such with-regret manipulations may be justifiable inpractice: An accompliceβs idiosyncratic preference may be tolerantto a small loss in exchange of gain for the partnering woman, or awoman may persuade a man to withstand some regret by providingside-payments.
We will start by illustrating that a with-regret accomplice manip-ulation can be strictly more beneficial compared to its no-regret andself manipulation counterparts.
Example 5.1 (With-regret vs. no-regret). Consider the followingpreference profile where the DA outcome is underlined.
π1 : π€β4 π€
β 1 π€2 π€5 π€3 π1 : π
β 1 πβ
2 π3 π4 π5
π2 : π€2 π€4 π€β1 π€
β 5 π€3 π€2 : π
β,β 3 π5 π1 π2 π4
π3 : π€1 π€β,β 2 π€4 π€3 π€5 π€3 : π2 πβ
5 π1 πβ 4 π3
π4 : π€1 π€β 3 π€β
5 π€2 π€4 π€4 : π4 π3 πβ1 π
β 5 π2
π5 : π€1 π€β 4 π€β
3 π€5 π€2 π€5 : πβ4 π
β 2 π5 π1 π3
Suppose the manipulating pair is (π1,π€1). The DA matchingafter π1 submits the optimal no-regret1 manipulated list β»β²
π1Bπ€2 β» π€4 β» π€1 β» π€5 β» π€3 and the optimal with-regret manipulatedlist β»β²β²
π1B π€1 β» π€4 β» π€2 β» π€5 β» π€3 are marked by β and β ,respectively. Both manipulation strategies improve π€1βs matchingcompared to truthful reporting, butπ€1 strictly prefers the with-regretoutcome. β‘
Example 5.1 highlights two key differences between optimalno-regret and with-regret manipulations. First, the matching afterthe inconspicuous with-regret manipulation (marked by β ) admitsa blocking pair (π1,π€4) with respect to the true profile β». Thisis in contrast to the no-regret case which is stability preserving(Theorem 4.2). Second, in contrast to Corollary 1, an optimal with-regret manipulation is not guaranteed to weakly improve or worsenthe matching for all agents on one side; indeed the women π€3 andπ€5 are strictly worse off while π€1 is strictly better off. Similarly, themanπ1 is strictly worse off whileπ4 andπ5 strictly improve.
The primary distinction between no-regret and with-regret ma-nipulation lies in the push up operations. If pushing up a set ofwomen does not cause regret for the accomplice, then pushing upany subset thereof does not either. By contrast, if by pushing up aset of women the accomplice incurs regret, then there exists exactlyone woman in that set who causes the same level of regret whenpushed up individually. As previously mentioned, with-regret pushup operations do not uniformly affect all men and all women (incontrast to Corollary 1). Moreover, the set of attained matchingsafter a with-regret manipulation are no longer stable with respectto true preferences (in contrast to Theorem 4.2), which makes theanalysis challenging.
Despite these structural differences, we are able to prove an ana-logue of Lemma 3 for with-regret push up operations (Lemma 4).Our proof of this result relies on the fact that all proposals that occurwhen the accomplice pushes up a set of women are contained in theunion of sets of proposals that occur when pushing up individualwomen in that set. This is relatively easy to prove for the no-regretcase, since the DA matchings after these push up operations are allstable with respect to true preferences (Theorem 4.2). Although wecannot rely on the same stability result for the with-regret case, wecircumvent the issue by reasoning about the sets of proposals ingreater detail. The full proof of Lemma 4, along with an extensivediscussion, is deferred to the appendix.1To see why β»β²
π1 is an optimal no-regret manipulation, note that the woman-optimalstable matching (with respect to β») matches π€1 withπ2. From Theorem 4.3 and Corol-lary 3, an optimal no-regret manipulation is, without loss of generality, stability preserv-ing, and from Proposition 1,π2 is the best stable partner for π€2.
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
Figure 2: Comparing no-regret accomplice and self manipula-tion in terms of the improvement in the rank of the matchedpartner of π€ . The solid bars, whiskers, and dots denote the in-terquartile range, range excluding outliers, and outliers, respec-tively.
Lemma 4. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up (while incurring regret).Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
Subsequently, an optimal with-regret manipulation is, withoutloss of generality, inconspicuous. The proof is similar to that of theno-regret case (Theorem 4.3) with the only difference being the useof Lemma 4 in place of Lemma 3.
THEOREM 5.2. If there is an optimal with-regret accomplicemanipulation, then there is an optimal inconspicuous with-regretaccomplice manipulation.
Theorem 5.2 immediately implies a polynomial-time algorithmfor computing an optimal with-regret accomplice manipulation.Moreover, the DA outcome from any inconspicuous with-regret ac-complice manipulation is π-stable with respect to the true prefer-ences (Proposition 2).
Corollary 4. An optimal with-regret accomplice manipulation strat-egy can be computed in O(π3) time.
6 EXPERIMENTAL RESULTSIn addition to the experiments described in Section 1, we performeda series of simulations to analyze the performance of accomplicemanipulation. As for the previous experimental setup, we generated1000 profiles uniformly at random for each value of π β {3, . . . , 40}(where π is the number of men/women) and allowed any man to bechosen as an accomplice for each experiment unless stated otherwise.
Comparing the Quality of Partners. We first compare the qualityof partners that a fixed strategic woman π€ is matched with throughno-regret accomplice and self manipulation. Figure 2 illustratesthe distributions of improvement (in terms of rank difference) outof only those instances where π€ is strictly better off through thetwo strategies individually. In other words, the self (respectively,accomplice) manipulation boxplots only reflect the data for when self(respectively, accomplice) manipulation is successful. It is evidentthat, in expectation, π€ is matched with better partners through no-regret accomplice manipulation.
The Fraction of Women Who Improve. We additionally comparethe fraction of women who are able to improve through no-regret ac-complice and self manipulation individually. Teo et al. [26] reportedthat 5.06% of women were able to improve using self manipula-tion when π = 8. In our experiment, this value is similarly 4.18%.However, 9.99% of women are able to improve through no-regretaccomplice manipulation. As illustrated in Figure 3, the fractionof women who benefit from no-regret accomplice manipulation isconsistently more than double that of self manipulation.
Figure 3: Comparing no-regret accomplice and self manipula-tion in terms of the fraction of women who benefit.
So far, our experiments have taken the optimistic approach ofallowing the strategic woman to pick any man of her choice as theaccomplice. However, we show that a with-regret strategy outper-forms self manipulation even when a single accomplice is randomlychosen in advance, and no-regret accomplice manipulation is, onaverage, better than self manipulation when there is a fixed pool offour or more men to choose from. The complete discussion of thisexperiment, along with other experimental results, is deferred to theappendix.
7 CONCLUDING REMARKSWe showed that accomplice manipulation is a viable strategic behav-ior that only requires inconspicuous misreporting of preferences andis frequently more beneficial than the classical self-manipulationstrategy. A natural avenue for future research is to investigate asetting with multiple accomplices working together to manipulatethe outcome for the strategic woman. Additionally, one can thinkof a broader strategy space in which both the accomplice and themanipulating woman are able to misreport their preference listssimultaneously. Analyzing the benefits of such coalitional manipula-tion strategiesβwith or without regretβon one or both sides, andstudying their structural and algorithmic properties are intriguingdirections for future work.
ACKNOWLEDGMENTSHH acknowledges support from NSF grant #1850076. RV acknowl-edges support from ONR#N00014-171-2621 while he was affiliatedwith Rensselaer Polytechnic Institute, and is currently supported byproject no. RTI4001 of the Department of Atomic Energy, Govern-ment of India. Part of this work was done while RV was supportedby the Prof. R Narasimhan postdoctoral award.
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
REFERENCES[1] Atila Abdulkadiroglu, Parag A Pathak, and Alvin E Roth. 2005. The New York
City High School Match. American Economic Review 95, 2 (2005), 364β367.[2] Atila Abdulkadiroglu, Parag A Pathak, Alvin E Roth, and Tayfun SΓΆnmez. 2005.
The Boston Public School Match. American Economic Review 95, 2 (2005),368β371.
[3] Michel Balinski and Tayfun SΓΆnmez. 1999. A Tale of Two Mechanisms: StudentPlacement. Journal of Economic Theory 84, 1 (1999), 73β94.
[4] Theodora Bendlin and Hadi Hosseini. 2019. Partners in Crime: Manipulating theDeferred Acceptance Algorithm through an Accomplice. In Proceedings of theAAAI Conference on Artificial Intelligence, Vol. 33. 9917β9918.
[5] Peter Coles and Ran Shorrer. 2014. Optimal Truncation in Matching Markets.Games and Economic Behavior 87 (2014), 591β615.
[6] Gabrielle Demange, David Gale, and Marilda Sotomayor. 1987. A Further Noteon the Stable Matching Problem. Discrete Applied Mathematics 16, 3 (1987),217β222.
[7] Yuan Deng, Weiran Shen, and Pingzhong Tang. 2018. Coalitional PermutationManipulations in the Gale-Shapley Algorithm. In Proceedings of the 17th Interna-tional Conference on Autonomous Agents and MultiAgent Systems. 928β936.
[8] Lester E Dubins and David A Freedman. 1981. Machiavelli and the Gale-ShapleyAlgorithm. The American Mathematical Monthly 88, 7 (1981), 485β494.
[9] David Gale and Lloyd S Shapley. 1962. College Admissions and the Stability ofMarriage. The American Mathematical Monthly 69, 1 (1962), 9β15.
[10] David Gale and Marilda Sotomayor. 1985. Ms. Machiavelli and the Stable Match-ing Problem. The American Mathematical Monthly 92, 4 (1985), 261β268.
[11] David Gale and Marilda Sotomayor. 1985. Some Remarks on the Stable MatchingProblem. Discrete Applied Mathematics 11, 3 (1985), 223β232.
[12] John William Hatfield, Fuhito Kojima, and Yusuke Narita. 2016. ImprovingSchools through School Choice: A Market Design Approach. Journal of EconomicTheory 166 (2016), 186β211.
[13] Chien-Chung Huang. 2006. Cheating by Men in the Gale-Shapley Stable MatchingAlgorithm. In European Symposium on Algorithms. Springer, 418β431.
[14] Paula Jaramillo, Γagatay KayΔ±, and Flip Klijn. 2014. On the Exhaustiveness ofTruncation and Dropping Strategies in Many-To-Many Matching Markets. SocialChoice and Welfare 42, 4 (2014), 793β811.
[15] Donald Knuth. 1997. Stable Marriage and its Relation to Other CombinatorialProblems: An Introduction to the Mathematical Analysis of Algorithms. Vol. 10.American Mathematical Society.
[16] Hirotatsu Kobayashi and Tomomi Matsui. 2009. Successful Manipulation inStable Marriage Model with Complete Preference Lists. IEICE Transactions onInformation and Systems 92, 2 (2009), 116β119.
[17] Hirotatsu Kobayashi and Tomomi Matsui. 2010. Cheating Strategies for the Gale-Shapley Algorithm with Complete Preference Lists. Algorithmica 58, 1 (2010),151β169.
[18] David Manlove. 2013. Algorithmics of Matching under Preferences. Vol. 2. WorldScientific.
[19] David G McVitie and Leslie B Wilson. 1971. The Stable Marriage Problem.Commun. ACM 14, 7 (1971), 486β490.
[20] Alvin E Roth. 1982. The Economics of Matching: Stability and Incentives.Mathematics of Operations Research 7, 4 (1982), 617β628.
[21] Alvin E Roth. 1984. The Evolution of the Labor Market for Medical Interns andResidents: A Case Study in Game Theory. Journal of Political Economy 92, 6(1984), 991β1016.
[22] Alvin E Roth. 1991. A Natural Experiment in the Organization of Entry-LevelLabor Markets: Regional Markets for New Physicians and Surgeons in the UnitedKingdom. The American economic review (1991), 415β440.
[23] Alvin E Roth. 2002. The Economist as Engineer: Game Theory, Experimentation,and Computation as Tools for Design Economics. Econometrica 70, 4 (2002),1341β1378.
[24] Alvin E Roth and Elliott Peranson. 1999. The Redesign of the Matching Market forAmerican Physicians: Some Engineering Aspects of Economic Design. AmericanEconomic Review 89, 4 (1999), 748β780.
[25] Alvin E Roth and Uriel G Rothblum. 1999. Truncation Strategies in MatchingMarketsβIn Search of Advice for Participants. Econometrica 67, 1 (1999),21β43.
[26] Chung-Piaw Teo, Jay Sethuraman, and Wee-Peng Tan. 2001. Gale-Shapley StableMarriage Problem Revisited: Strategic Issues and Applications. ManagementScience 47, 9 (2001), 1252β1267.
[27] Rohit Vaish and Dinesh Garg. 2017. Manipulating Gale-Shapley Algorithm:Preserving Stability and Remaining Inconspicuous. In Proceedings of the 26thInternational Joint Conference on Artificial Intelligence. 437β443.
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
Appendix
8 OMITTED MATERIAL FROM SECTIONS 3.1& 3.2
8.1 Additional PreliminariesLattice of stable matchings. Given any preference profile β» and
any pair of stable matchings `, ` β² β πβ», let us define their meet`β§ B ` β§ ` β² as follows: For every manπ β π ,
`β§ (π) ={` β²(π) if ` (π) β»π ` β²(π)` (π) otherwise,
and for every women π€ βπ ,
`β§ (π€) ={` (π€) if ` (π€) β»π€ ` β²(π€)` β²(π€) otherwise.
Similarly, the join `β¨ B ` β¨ ` β² is defined as follows: For every manπ β π ,
`β¨ (π) ={` (π) if ` (π) β»π ` β²(π)` β²(π) otherwise,
and for every women π€ βπ ,
`β¨ (π€) ={` β²(π€) if ` (π€) β»π€ ` β²(π€)` (π€) otherwise.
A well-known result, attributed to John Conway [15], establishesthat the set of stable matchings is closed under meet and join opera-tions.
PROPOSITION 5 ([15]). Let β» be a preference profile and let`, ` β² β πβ». Then, `β§, `β¨ β πβ».
8.2 Self Manipulation vs. No-Regret AccompliceManipulation
The self manipulation and no-regret accomplice manipulation strat-egy spaces are not contained in one another. Example 1.1 shows aninstance where no-regret accomplice manipulation is better than selfmanipulation. In Example 8.1, we provide an instance where selfmanipulation is better than no-regret accomplice manipulation.
Example 8.1. Consider the following preference profile wherethe DA outcome is underlined.
π1: π€2 π€3 π€β1 π€4 π1: πβ
1 π2 π3 π4π2: π€3 π€β
2 π€4 π€1 π€2: πβ2 π3 π4 π1
π3: π€1 π€β3 π€4 π€2 π€3: πβ
3 π1 π4 π2π4: π€1 π€β
4 π€2 π€3 π€4: π3 π1 πβ4 π2
Suppose π€1 seeks to improve her match via manipulation. Theoptimal self manipulation strategy for π€1 is β»β²
π€1=π4 β»π3 β»π1 β»π2, which allows her to match with her top choiceπ1 (the new DAmatching is marked by β). Regardless of the choice of accomplice,the optimal no-regret accomplice manipulation strategy, on the otherhand, is truth-telling. β‘
8.3 Proof of Proposition 2PROPOSITION 2. Let β» denote the true preference profile. For
any manπ, let β»β²B {β»βπ, β»β²π}, and let ` β² β πβ»β² be any matching
that is stable with respect to β»β². Then, ` β² is π-stable with respect toβ».
PROOF. Suppose, for contradiction, that ` β² is notπ-stable withrespect to β». Then, there must exist a man-woman pair (πβ²,π€ β²) thatblocks ` β² with respect to β» such that πβ² β π, i.e., π€ β² β»πβ² ` β²(πβ²)andπβ² β»π€β² ` β²(π€ β²). Sinceπ is the only agent whose preferences dif-fer between β» and β»β², we have that β»πβ² =β»β²
πβ² and β»π€β² =β»β²π€β² . Thus,
π€ β² β»β²πβ² `
β²(πβ²) andπβ² β»β²π€β² `
β²(π€ β²), implying that the pair (πβ²,π€ β²)blocks ` β² with respect to β»β², which contradicts the assumption that` β² β πβ»β² . Thus, ` β² must beπ-stable with respect to the true profileβ». β‘
8.4 Proof of Lemma 1Lemma 1. Let β» be the true preference profile and let ` B DA(β»).For any subset of women π β π , let β»β²B {β»βπ, β»π β
π } and ` β² BDA(β»β²). Then, ` βͺ°π ` β².
PROOF. Suppose, for contradiction, that there exists a womanπ€ β² such that πβ² β»π€β² ` (π€ β²), where πβ² B ` β²(π€ β²). We infer thatπ€ β² βπβ² ` (πβ²), otherwise the stability of ` with respect to β» iscompromised. Since πβ² β ` (π€ β²), we have that π€ β² β ` (πβ²), andtherefore ` (πβ²) β»πβ² π€ β². However, Proposition 4 guarantees ` β² βͺ°π`, thus posing a contradiction. β‘
9 OMITTED MATERIAL FROM SECTION 49.1 Proof of Theorem 4.2
THEOREM 4.2 (NO-REGRET PUSH UP IS STABILITY PRESERV-ING). Let β» be a preference profile, and let ` B DA(β»). For anysubset of women π βπ and any manπ, let β»β²B {β»βπ, β»π β
π }, and` β² B DA(β»β²). Ifπ does not incur regret, then πβ»β² β πβ».
PROOF. Suppose, for contradiction, that there exists a matchingπ β πβ»β² \ πβ». Then, there must be a pair (πβ²,π€ β²) that blocks π
with respect to β». It follows from Proposition 2 thatπβ² =π. Thus,π€ β² β»π π (π) andπ β»π€β² π (π€ β²).
From Proposition 1, we have that ` β²(π) βͺ°β²π π (π). Since π does
not incur regret, we have ` (π) = ` β²(π), and thus, ` (π) βͺ°β²π π (π).
All women below ` (π) in β»β²π are also below ` (π) in β»π by the
push up assumption. Since ` (π) βͺ°β²π π (π), this implies that all
women below π (π) in β»β²π are also below π (π) in β»π . Thus, if
π (π) β»β²π π€ β², then π (π) β»π π€ β², which contradicts the blocking
pair condition above. Therefore, we must have that π€ β² β»β²π π (π)
(note that π€ β² β π (π) by the blocking pair condition).Furthermore, since β»β²
π€β² = β»π€β² , the blocking pair condition alsoimplies thatπ β»β²
π€β² π (π€ β²). Combined with the conditionπ€ β² β»β²π π (π),
this contradicts the assumption that π β πβ»β² . Thus, πβ»β² β πβ». β‘
9.2 Proof of Lemma 2Lemma 2. Let (π,π€) be a manipulating pair and let β» be a pref-erence profile. For any subsets of women π β π and π β π , letβ»β²B {β»βπ, β»π β
π } denote the preference profile after pushing up theset π , and β»β²β²B {β»βπ, β»π β,π β
π } denote the profile after pushing up
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
π and pushing down π in the true preference list β»π of man π. Let` B DA(β»), ` β² B DA(β»β²), and ` β²β² B DA(β»β²β²). If ` β²β²(π€) β»π€ ` (π€),then ` β²(π€) βͺ°π€ ` β²β²(π€).
PROOF. Our proof will use case analysis based on whether or not` β²(π) = ` (π).
Case I (when ` β²(π) = ` (π)): The list β»β²β²π can be considered as
being derived from β»β²π via a push down operation on the set π (see
Figure 4). From Lemma 1, we know that a push down operation isweakly worse for all women; thus, in particular, we get ` β²(π€) βͺ°π€
` β²β²(π€), as desired. Note that the relative ordering of π and π above` β²(π) in the list β»β²
π is not important in light of Proposition 3.
β»π : Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β· π Β· Β· Β·
β»β²π : Β· Β· Β· π Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β·
β»β²β²π : Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β· π Β· Β· Β·
Figure 4: An illustration of man πβs preference lists under theprofiles β», β»β², and β»β²β² in the proof of Lemma 2.
Case II (when ` β²(π) β ` (π)): Suppose ` β²(π) β π . Then, the listβ»β²β²π can be considered as being derived from β»β²
π via a permutationof the women below ` β²(π) (see Figure 4). By Proposition 3, thisimplies that ` β² = ` β²β², and in particular ` β²(π€) = ` β²β²(π€), as desired.
Therefore, for the remainder of the proof, let us assume that` β²(π) β π . Since β»β²
π is derived from β»π via a push up operationon the set π , and since ` β²(π) β ` (π) by assumption, we have that` (π) β»π ` β²(π). By Proposition 3, we can assume, without lossof generality, that ` β²(π) is positioned immediately below ` (π) inthe list β»π . By construction, the same property also holds for thelists β»β²
π and β»β²β²π . Thus, β»β²β²
π can be considered as being obtainedfrom β»β²
π via a push down operation on the set π (note that thisoperation is defined with respect to ` β²(π)). By Lemma 1, we have` β²(π€) βͺ°π€ ` β²β²(π€), as desired. β‘
9.3 Proof of Lemma 3Lemma 3. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up without incurring regret.Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
Preliminaries for the proof of Lemma 3: Let β» be a true pref-erence profile. Given an accomplice π and a set of women π =
{π€π,π€π ,π€π , . . . } such that ` (π) β»π π€π₯ for all π€π₯ β π , we defineβ»ππ as the list derived from β»π where π pushes up π€π , β»ππ as thelist derived from β»π where π pushes up π€π , etc., and β»ππ as the listwhereπ pushes up all women in π ; the corresponding profiles areβ»π, β»π , β»π , . . . , β»π . Additionally, let `π B DA(β»π), `π B DA(β»π), `π B DA(β»π ), . . . , `π B DA(β»π ). Note that the placement ofwomen being pushed up above ` (π) in β»ππ, β»ππ, β»ππ, . . . , β»ππ doesnot affect `π, `π , `π , . . . , `π by Proposition 3.
Example 9.1. Suppose π = {π€π,π€π }. Figure 5 illustrates thepossible configurations ofπβs preference lists under the profiles β»,β»π , β»π , and β»π .
β»π : Β· Β· Β· ` (π) Β· Β· Β· π€π Β· Β· Β· π€π Β· Β· Β·
β»ππ : Β· Β· Β· π€π Β· Β· Β· ` (π) Β· Β· Β· π€π Β· Β· Β·
β»ππ : Β· Β· Β· π€π Β· Β· Β· ` (π) Β· Β· Β· π€π Β· Β· Β·
β»ππ : Β· Β· Β· π€π Β· Β· Β· π€π Β· Β· Β· ` (π) Β· Β· Β·
Figure 5: An illustration of man πβs preference lists under theprofiles β», β»π , β»π , and β»π when π = {π€π,π€π }.
It can be shown that π does not incur regret under any of theprofiles β»π, β»π , and so on (Lemma 5).
Lemma 5. Let π = {π€π,π€π ,π€π , . . . } be an arbitrary finite set ofwomen that the accompliceπ can push up without incurring regret.Then, for every π€π₯ β π ,π does not incur regret under the matching`π₯ B DA(β»π₯ ), where β»π₯B {β»βπ, β»π€π₯ β
π }.
PROOF. We will prove the lemma for the fixed profile β»π (anidentical argument works for other profiles).
Suppose, for contradiction, thatπ incurs regret in the profile β»π .That is, ` (π) β»π `π (π) where `π B DA(β»π). From Lemma 7 (seeSection 10.1), we get that `π (π) = π€π . Further, using Proposition 3,we can assume, without loss of generality, that π€π is positionedimmediately below ` (π) in the true list β»π , and immediately aboveit in the manipulated lists β»ππ , as well as β»ππ . This implies that thetransition from β»ππ to β»ππ is a with-regret push up operation involv-ing the promotion of π \ {π€π} (since, according to the list β»ππ , thenew partner ` (π) is strictly worse than π€π). Again, from Lemma 7,it follows that `π (π) β π \ {π€π}. By the no-regret assumption forthe set π , we know that `π (π) = ` (π). This, however, is a contra-diction since all women in π \ {π€π} are placed below ` (π) in thelist β»π , and hence must be different from it. β‘
Given any profile β», let πβ» denote the set of all proposals thatoccur in the execution of the DA algorithm on β». Formally, forany man ππ β π and woman π€ π β π , the ordered pair (ππ ,π€ π )belongs to the set πβ» if ππ proposes to π€ π during the execution ofDA algorithm on the profile β».
Lemma 6. Let π = {π€π,π€π ,π€π , . . . } be an arbitrary finite set ofwomen that the accompliceπ can push up without incurring regret.Then, any proposal that occurs under β»π also occurs under at leastone of the profiles β»π, β»π , β»π , . . . .
PROOF. Suppose, for contradiction, that (π1,π€1) is the first pro-posal to occur during the DA execution on β»π such that it is not anelement of πβ»π βͺ πβ»π βͺ πβ»π βͺ . . . . Note that the proposals made bythe accompliceπ in πβ»π are only to the women above and including` (π) in β»π as well as the women in {π€π,π€π ,π€π , . . . }, all of whomhe proposes to in πβ»π βͺ πβ»π βͺ πβ»π βͺ . . . . Thus,π1 β π, implyingthatπ1 is a truthful agent.
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
Since men propose in decreasing order of their preference, π1must have been rejected by a woman, say π€2, whom he ranks im-mediately above π€1 in β»π1 , before proposing to π€1 under β»π .Further, π1 must propose to π€2 under at least one of the profilesβ»π, β»π , β»π , . . . , and must not be rejected by her (otherwise, he willpropose to π€1). Thus, π1 must be matched with π€2 under at leastone of the matchings `π, `π , `π , . . . . Without loss of generality, letus assume that `π (π1) = π€2.
Since π€2 is not matched with π1 under `π , she must have re-ceived a more preferable proposal from some man, say π2; thus,π2 β»π€2 π1. Due to our assumption that (π1,π€1) is the first pro-posal during the DA execution on β»π that does not occur in πβ»π βͺπβ»π βͺ πβ»π βͺ . . . , π2 must have proposed to π€2 under at least oneof the profiles β»π, β»π , β»π , . . . . Without loss of generality, he pro-poses under β»π . Since women match with their best proposers,`π (π€2) βͺ°π€2 π2. This, combined withπ2 β»π€2 π1 = `π (π€2), we get`π (π€2) β»π€2 `π (π€2).
We infer thatπ1 does not propose toπ€2 under β»π , since otherwiseπ€2 (eventually) rejects π1 causing him to propose to π€1, whichwould contradict our assumption that (π1,π€1) β πβ»π βͺπβ»π βͺπβ»π βͺ. . . . Thus, `π (π1) β»π1 π€2 = `π (π1).
Notice that the profiles β»π and β»π are obtained from the truepreference profile β» by no-regret push up operations. Therefore,from Lemma 5, we get that the matchings `π and `π are stable withrespect to the true preference profile (i.e., `π, `π β πβ»).
Now consider the join `β¨ B `π β¨ `π of the two matchingswith respect to the true preference profile β», wherein each man isassociated with his more preferred partner between `π and `π , andeach woman is associated with her less preferred partner (refer toSection 8.1 for the relevant definitions). Thus,π1 is associated with`π (π1) β π€2 and π€2 is associated with `π (π€1) =π1. The resultingassignment `β¨ is not a valid matching, which contradicts the stablelattice result (Proposition 5). β‘
We are now ready to prove Lemma 3.
Lemma 3. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up without incurring regret.Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
PROOF. Letππ B `π (π€), where `π B DA(β»π ). From Lemma 5,we know that the accompliceπ does not incur regret under any of thematchings `π, `π , `π , . . . , and therefore `π (π) = `π (π) = `π (π) =Β· Β· Β· = ` (π). Ifππ =π, then the lemma follows trivially sinceπ ismatched with his `π -partner, namely π€ , under each of the match-ings `π, `π , `π , . . . , and therefore the `π -partner of π€ can also beachieved under any of the profiles β»π, β»π , β»π , . . . . Thus, for the re-mainder of the proof, we will assume thatππ β π; in other words,ππ is a truthful agent.
Suppose, for contradiction, thatππ is not matched toπ€ under anyof the profiles β»π, β»π , β»π , . . . . Starting from the profile β»π , we canobtain the profile β»π via a no-regret push up operation of the set π \{π€π} in the list β»ππ of the accomplice. Therefore, from Corollary 1,we have that `π (ππ ) βͺ°ππ
`π (ππ ) = π€ , where `π B DA(β»π).Since `π (ππ ) β π€ by the contradiction assumption, we furtherobtain that `π (ππ ) β»ππ
π€ . By a similar argument, we get that
the women `π (ππ ), `π (ππ ), . . . are also placed above π€ in the listβ»ππ
.Since ππ is matched with π€ under `π , he must propose to π€
during the execution of DA algorithm on β»π , i.e., (ππ ,π€) β πβ»π .From Lemma 6, we have that (ππ ,π€) β πβ»π βͺ πβ»π βͺ πβ»π . . .. Sinceππ is a truthful agent, his preference list remains unchanged, andtherefore he ranks the women `π (π), `π (ππ ), `π (ππ ), . . . strictlyabove π€ under each of the profiles β»π, β»π , β»π , . . . . This, however,poses a contradiction since men propose in decreasing order of theirpreference. β‘
9.4 Proof of Corollary 2Corollary 2. An optimal no-regret accomplice manipulation strat-egy can be computed in O(π3) time.
PROOF. (sketch) The algorithm simply promotes each womanthat is below ` (π) in the accompliceβs true preference list to someposition above ` (π) and checks the DA outcome. The total numberof such checks is O(π), and for each check, running the DA algorithmtakes O(π2) time. β‘
9.5 Proof of Corollary 3Corollary 3. The DA outcome from an inconspicuous no-regret ac-complice manipulation is stable with respect to the true preferences.
PROOF. (sketch) An inconspicuous manipulation is a specialcase of a push up operation, which was shown in Theorem 4.2 to bestability preserving. β‘
10 OMITTED MATERIAL FROM SECTION 510.1 Proof of Theorem 5.2
THEOREM 5.2. If there is an optimal with-regret accomplicemanipulation, then there is an optimal inconspicuous with-regretaccomplice manipulation.
Recall from Lemma 3 that the match for the manipulating womanπ€ obtained by a no-regret push up operation of a set π βπ by theaccomplice can also be achieved by pushing up exactly one womanin π . The following result (Lemma 4) establishes the with-regretanalogue of this result.
Lemma 4. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up (while incurring regret).Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
Preliminaries for the proof of Lemma 4: The relevant notation issimilar to that used in the proof of Lemma 3, which we recall belowfor the sake of completeness.
Let β» be a true preference profile. Given an accomplice π anda set of women π = {π€π,π€π ,π€π , . . . } such that ` (π) β»π π€π₯ forall π€π₯ β π , we define β»ππ as the list derived from β»π where π
pushes up π€π , β»ππ as the list derived from β»π where π pushes upπ€π , etc., and β»ππ as the list where π pushes up all women in π ;the corresponding profiles are β»π, β»π , β»π , . . . , β»π . Additionally, let`π B DA(β»π), `π B DA(β»π ), `π B DA(β»π ), . . . , `π B DA(β»π ).
We will now show that under a with-regret push up operation,the DA algorithm matches the accomplice to one of the women he
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
pushes up (Lemma 7). In particular, if the accomplice promotes onlyone woman and incurs regret, then he must be matched to her in theresulting matching.
Lemma 7. Let β» be a true preference profile and ` B DA(β»). Forany fixed man π and any subset π βπ of women, let β»β²B {β»βπ, β»π β
π } denote the preference profile after pushing up the women inπ in β»π and let ` β² B DA(β»β²). If π incurs regret (i.e., if ` (π) β»π` β²(π)), then ` β²(π) β π .
PROOF. Suppose, for contradiction, that ` β²(π) β π . Because ofProposition 3, without loss of generality, we have that the set πis placed immediately below ` (π) in the true list β»π . By strate-gyproofness of the DA algorithm,π cannot be matched under ` β² toany woman who, according to his true list, is preferred over ` (π),i.e., ` β²(π) β {π§ β π : π§ β»π ` (π)}. By the contradiction assump-tion, π cannot be matched with any woman in π , and because ofthe with-regret assumption,π cannot be matched with ` (π) either.Therefore, the woman π¦ B ` β²(π) must be such that ` (π) β»β²
π π¦,which, by the push up assumption, implies that ` (π) β»π π¦.
Once again, due to Proposition 3. we can assume, without lossof generality, that π¦ is placed immediately below the set π in thetrue list β»π . This, in turn, implies that π¦ is immediately below ` (π)in the manipulated list β»π β
π . Therefore, starting with the list β»π βπ ,
one can obtain the true list β»π simply by permuting the agents thatare above π¦. Proposition 3 would then imply that the partner ofπ does not change in the process, i.e., ` β²(π) = ` (π), which is acontradiction. Hence, we must have ` β²(π) β π . β‘
Lemma 7 implies that the accomplice π matches with somewoman in π under the matching `π ; say `π (π) = π€π . We willnow show that π is matched with π€π under the matching `π as well(Lemma 8). Notice that in light of Proposition 3, we can assume,without loss of generality, that in the lists β»ππ and β»ππ , the womanπ€π
is placed immediately above ` (π). That is, π€π is the least-preferredwoman in π according to the list β»ππ .
Lemma 8. Let π€π β π be the woman who the accomplice π
matches with under `π . Then, π also matches with π€π under `π ,where `π B DA(β»π) and β»πB {β»βπ, β»π€πβ
π }.
PROOF. Starting with the profile β»π , we can obtain β»π by push-ing down all women in π \ {π€π} (recall from the aforementioned ob-servation that π€π is the least-preferred woman in π according to thelist β»ππ). Then, from Proposition 4, we get that `π (π) = `π (π). β‘
From Lemma 8, it follows that β»π is a with-regret profile. Bycontrast, Lemma 9 will show that the rest of the profiles β»π , β»π . . .
do not cause regret for the accomplice.
Lemma 9. Let π€π β π be the woman who the accomplice π
matches with under `π . Then, for any π€π§ β π \ {π€π}, π doesnot incur regret under the profile β»π§B {β»βπ, β»π€π§ β
π }.
PROOF. Let οΏ½οΏ½ B π \ {π€π}, and let β»οΏ½οΏ½B {β»βπ, β»π βπ } be the
profile whereπ pushes up all women in οΏ½οΏ½ starting from the true listβ»π .
We claim that π must match with the woman ` (π) under thematching `οΏ½οΏ½ B DA(β»οΏ½οΏ½ ). Indeed, if that is not the case, then pro-moting οΏ½οΏ½ is a with-regret push up operation. Then, from Lemma 7,
the man π must be matched with some woman in οΏ½οΏ½ under `οΏ½οΏ½ . This,however, would contradict the strategyproofness of DA algorithm,as π is able to strictly improve in going from the βtrueβ list β»ππ(where he is matched with π€π , who is the least-preferred woman inπ according to the list β»ππ) to the βmanipulatedβ list β»οΏ½οΏ½ (where hispartner is some woman in οΏ½οΏ½ ).
Thus,π must be matched with the woman ` (π) under `οΏ½οΏ½ , imply-ing that promoting οΏ½οΏ½ is a no-regret push up operation. Lemma 5 nowimplies that for everyπ€π§ β οΏ½οΏ½ , β»π§B {β»βπ, β»π€π§ β
π } is also a no-regretprofile, as desired. β‘
It can also be shown that if the accomplice π pushes up allwomen in π \ {π€π} simulataneously, then he does not incur regret(Lemma 10).
Lemma 10. Let β»π be the profile obtained by pushing up π Bπ \ {π€π} in the accomplice πβs true preference list. Then, π doesnot incur regret under β»π (i.e,π matches with ` (π) under β»π ).
PROOF. Suppose, for contradiction, that β»π is a with-regret pro-file. Then, from Lemmas 7 and 8, we have that for some π€π§ β π ,the profile β»π§B {β»βπ, β»π€π§ β
π } is also with-regret. This, however,contradicts the implication of Lemma 9 that β»π§ is no-regret for everyπ€π§ β π . β‘
Recall that given any profile β», πβ» denotes the set of all proposalsthat occur in the execution of the DA algorithm on β». Formally, forany man ππ β π and woman π€ π β π , the ordered pair (ππ ,π€ π )belongs to the set πβ» if ππ proposes to π€ π during the execution ofDA algorithm on the profile β». Our next result (Lemma 11) showsthat the set of proposals that occur under a true preference profile β»is contained in the set of proposals that occur under a profile β»β² thatis obtained via a no-regret push up operation on β».
Lemma 11. Let β» be a preference profile and ` B DA(β»). For anyfixed man π, let β»β²= {β»βπ, β»π β
π } and ` β² = DA(β»β²) such that πdoes not incur regret (i.e., ` β²(π) = ` (π)). Then, πβ» β πβ»β² .
PROOF. Since π does not incur regret under β»β², we have that` βͺ°π ` β² (Corollary 1). Under the DA algorithm, men propose indecreasing order of their preference. Therefore, any proposal madeby a truthful man under β» is also made under β»β². Furthermore, thepush up assumption implies that the accompliceπ proposes to thewomen in β»πΏ
π (i.e., the woman strictly preferred by π over ` (π)according to his true list β»π) under β»π as well. β‘
Let πβ»\β»β² B πβ» \πβ»β² denote the set of proposals that occur underthe profile β» but not under β»β². Our next result (Lemma 12) shows thatthe set πβ»π§\β»π , where β»π§ is a no-regret profile obtained by pushingup some woman π€π§ β π \ {π€π} (as established in Lemma 9), iscontained in the set πβ».
Lemma 12. For any woman π€π§ β π \ {π€π}, let β»π§ be the no-regretprofile obtained by pushing up π€π§ (as discussed in Lemma 9). Then,πβ»π§\β»π β πβ».
PROOF. We start by showing that any proposal in πβ»π§\β»π mustoccur after π proposes to ` (π) during the DA execution on β»π§ .Suppose, for contradiction, that this is not true. Then, let (π1,π€1) bethe first proposal during the DA execution on β»π§ such that it does not
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
belong to πβ»π and occurs before (π, ` (π)). Note that the proposalsmade by the accomplice π before ` (π) under β»π§ are only to thewomen above and including ` (π) in β»π as well as π€π§ , all of whomhe also proposes to under β»π§ . Thus,π1 β π, implying thatπ1 is atruthful agent.
Since men propose in decreasing order of their preference andit is assumed that (π1,π€1) β πβ»π , π1 must have been rejected byπ€2 B `π (π1) under β»π§ before proposing to π€1. Then, under β»π§ ,π€2 must have received a proposal from some man, say π2, suchthat π2 β»π€2 π1. We assumed (π1,π€1) to be the first proposalduring the DA execution on β»π§ to not belong to πβ»π . Since (π2,π€2)occurs before (π1,π€1) under β»π§ , we must have that (π2,π€2) alsooccurs under β»π . We already established that π2 β»π€2 π1 and knowthat women are truthful. Since women match with their favoriteproposers, this implies that π€2 does not match with π1 under β»π .However, this contradicts the statement that π€2 = `π (π1).
We proceed by showing that any proposal that occurs after πproposes to ` (π) during the DA execution on β»π§ must also occurin πβ». Suppose, for contradiction, that this is not true. Then, let(π1,π€1) be the last proposal during the DA execution on β»π§ suchthat it does not belong to πβ» and occurs after (π, ` (π)). We inferthat π€1 accepts and matches withπ1 under β»π§ , since otherwiseπ1must make additional proposals, contradicting our assumption that(π1,π€1) is the last proposal to not occur in πβ». Note that π doesnot propose to anyone after ` (π) by the no-regret assumption of β»π§ .Thus,π1 β π, implying thatπ1 is a truthful agent.
Let π2 be the man π€1 is temporarily engaged to before she re-ceives a proposal fromπ1 under β»π . We assumed thatπ€2 = `π (π1)which means that π€1 rejects π2 in favor of π1 under β»π§ , causingπ2 to propose to the next woman in β»π§π2 , say π€2. Since πβ» β πβ»π§
(Lemma 11), we know that (` (π€1),π€1) β πβ»π§ . We also definedπ2 to be π€1βs favorite proposer under β»π§ before matching withπ1,which implies that π2 βͺ°π€1 ` (π€1). If π€1 β»π2 ` (π2), then the pair(π2,π€1) blocks ` with respect to β». Thus, we have that ` (π2) βͺ°π2π€1. Note that the accompliceπ does not make any proposals after(π1,π€1) under β»π§ , since (π1,π€1) occurs after (π, ` (π)) and β»π§ isa no-regret profile. We know that π2 proposes to π€2 after (π1,π€1),implying that π2 β π. Therefore, ` (π2) βͺ°π2 π€1 β»π2 π€2, and π2does not propose to π€2 under β». However, this contradicts the as-sumption that (π1,π€1) is the last proposal during the DA executionon β»π to not belong to πβ».
Thus, we have shown that (1) any proposal that belongs to πβ»π§\β»π
occurs after (π, ` (π)) during the DA execution on β»π§ , and (2) anyproposal that occurs after (π, ` (π)) during the DA execution on β»π§belongs to πβ». These two statements imply that πβ»π§\β»π β πβ». β‘
Our next result (Lemma 13) shows that the set πβ»π \β»οΏ½οΏ½ is con-tained in the set πβ»π .
Lemma 13. Let β»π be the profile obtained by pushing up π Bπ \{π€π} in the accompliceπβs true preference list. Then, πβ»π \β»οΏ½οΏ½ βπβ»π .
PROOF. For this proof, let β»π be the βtrueβ preference profile.Remember, from Lemma 10, thatπ matches with ` (π) under β»π .Thus, β»π is a with-regret profile with respect to β»π , and is obtainedby pushing up π B π βͺ {` (π)} in β»ππ . On the other hand, π
matches with π€π under β»π and β»π (Lemma 8). Thus, β»π is a no-regret profile with respect to β»π , and is obtained by pushing upπ = π \ {` (π)} = π \ {π€π}.
For any womanπ€π§ β π , let β»ππ§ be the profile obtained by pushingup π€π§ in β»ππ . It is easy to see that β»ππ§ is a no-regret profile withrespect to β»π . Indeed, suppose β»ππ§ is with-regret. Then, π mustmatch with π€π§ under β»ππ§ (Lemma 7). In light of Proposition 3 wecan assume, without loss of generality, thatπ ranks the set π strictlyabove π€π in the list β»ππ . Given profile β»π , π could then manipulateby submitting β»ππ§π in order to match with π€π§ . However, this wouldcontradict strategyproofness of the DA algorithm since π€π§ β»π π€π .Therefore, β»ππ§ is no-regret with respect to β»π .
Given this observation, from Lemma 12, we get that for anywoman π€π§ β π = π \ {` (π)}, πβ»ππ§ \ πβ»οΏ½οΏ½ β πβ»π . Sinceπ matcheswith π€π under β»ππ§ and β»π , we have that β»ππ§ is no-regret withrespect to β»π . Thus, we invoke Lemma 6 to claim that any proposalthat occurs under β»π is contained in πβ»ππ§1 βͺ πβ»ππ§2 βͺ Β· Β· Β· βͺ πβ»ππ§π ,whereπ B {π€π§1 ,π€π§2 , . . . ,π€π§π }. Since {πβ»ππ§1 βͺπβ»ππ§2 βͺΒ· Β· Β·βͺπβ»ππ§π }\πβ»οΏ½οΏ½ β πβ»π , we get that πβ»π \β»οΏ½οΏ½ β πβ»π . β‘
Recall from Lemma 6 that, in the no-regret setting, any proposalthat occurs during the DA execution on the profile β»π must also occurduring the DA execution on at least one of the profiles β»π, β»π , β»π , . . . .Our next result (Lemma 14) establishes the with-regret analogue ofthis result.
Lemma 14. Let π = {π€π,π€π ,π€π , . . . } be an arbitrary finite set ofwomen such that the accomplice π incurs regret after pushing up π .Then, any proposal that occurs under β»π also occurs under at leastone of the profiles β»π, β»π , β»π , . . . .
PROOF. Let β»π be the profile afterπ pushes up π B π \ {π€π}in his true preference list. Then, from Lemma 13, we know thatπβ»π \ πβ»οΏ½οΏ½ β πβ»π . Additionally, from Lemma 10, we know thatβ»π is a no-regret profile. It then follows from Lemma 6 that πβ»οΏ½οΏ½ βπβ»π βͺ πβ»π βͺ . . . .
Any proposal that occurs under β»π is contained in either πβ»π \πβ»οΏ½οΏ½ or πβ»οΏ½οΏ½ . By combining the aforementioned observations, weget that any such proposal is contained in πβ»π βͺ πβ»π βͺ πβ»π βͺ . . . , asdesired. β‘
We are now ready to prove Lemma 4.
Lemma 4. Let (π,π€) be a manipulating pair, and let π βπ be anarbitrary set of women that π can push up (while incurring regret).Then, the match for π€ that is obtained by pushing up all women inπ can also be obtained by pushing up exactly one woman in π .
PROOF. Let ππ B `π (π€), where `π B DA(β»π ). We assumethat the accomplice π matches with a woman π€π β π (Lemma 7)under β»π . From Lemma 8, we know that he also matches with π€π
under β»π . If ππ = π, then the lemma follows trivially since π ismatched with his `π -partner, namely π€ , under the matching `π , andtherefore the `π -partner of π€ can also be achieved through profileβ»π . Thus, for the remainder of the proof, we assume thatππ β π;in other words,ππ is a truthful agent.
Suppose, for contradiction, thatππ is not matched toπ€ under anyof the profiles β»π, β»π , β»π , . . . . Starting from the with-regret profileβ»π (as established in Lemma 8), we can obtain the profile β»π via a
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
no-regret push up operation of the set π \ {π€π} in the list β»ππ of theaccomplice. Therefore, from Corollary 1, we have that `π (ππ ) βͺ°ππ
`π (ππ ) = π€ , where `π B DA(β»π). Since `π (ππ ) β π€ by thecontradiction assumption, we further obtain that `π (ππ ) β»ππ
π€ .Now, consider the no-regret profiles β»π , β»π , β»π , . . . (as estab-
lished in Lemma 9). Suppose `π (ππ ) β»ππ`π (ππ ), where `π B
DA(β»π ). Then, since men propose in decreasing order of their pref-erence, (ππ , `
π (ππ )) β πβ»π\β»π . Lemma 12 consequently impliesthat (ππ , `
π (ππ )) β πβ», and thus `π (ππ ) βͺ°ππ` (ππ ). Since
` (ππ ) βͺ°ππ`π (ππ ) by Corollary 1, we infer that `π (ππ ) =
` (ππ ). This, combined with `π (ππ ) β»ππ`π (ππ ), gets us π€ =
`π (ππ ) β»ππ` (ππ ). From the accomplice manipulation assump-
tion, we also know that ππ = `π (π€) β»π€ ` (π€). However, this im-plies that the pair (ππ ,π€) blocks ` with respect to β», posing a con-tradiction. Thus, we have shown that π€ = `π (ππ ) βππ
`π (ππ ).Since `π (ππ ) β π€ by the initial contradiction assumption, we fur-ther obtain that `π (ππ ) β»ππ
π€ . By a similar argument, we get thatthe women `π (ππ ), `π (ππ ), . . . are also placed above π€ in the listβ»ππ
.Sinceππ is matched withπ€ under `π , he must propose toπ€ dur-
ing the execution of DA algorithm on β»π , i.e., (ππ ,π€) β πβ»π . FromLemma 14, we have that (ππ ,π€) β πβ»π βͺ πβ»π βͺ πβ»π . . .. Sinceππ
is a truthful agent, his preference list remains unchanged, and there-fore he ranks the women `π (π), `π (ππ ), `π (ππ ), . . . strictly aboveπ€ under each of the profiles β»π, β»π , β»π , . . . . This, however, poses acontradiction since men propose in decreasing order of their prefer-ence. β‘
We are now ready to prove Theorem 5.2.
THEOREM 5.2. If there is an optimal with-regret accomplicemanipulation, then there is an optimal inconspicuous with-regretaccomplice manipulation.
PROOF. From Proposition 3 (and subsequent remarks), we knowthat any accomplice manipulation can be simulated via push upand push down operations. Lemma 2 shows that any beneficialmanipulation that is achieved by some combination of pushing up aset π βπ and pushing down π βπ can be weakly improved byonly pushing up π βπ . Finally, from Lemma 4, we know that anymatch for the manipulating woman π€ that is achieved by pushingup π βπ is also achieved by pushing up exactly one woman in π ,thus establishing the desired inconspicuousness property. β‘
10.2 Proof of Corollary 4Corollary 4. An optimal with-regret accomplice manipulation strat-egy can be computed in O(π3) time.
PROOF. (sketch) The algorithm simply promotes each womanthat is below ` (π) in the accompliceβs true preference list to someposition above ` (π) and checks the DA outcome. The total numberof such checks is O(π), and for each check, running the DA algorithmtakes O(π2) time. β‘
11 OMITTED MATERIAL FROM SECTION 6Recall from our previous experiments that we generated 1000 pro-files uniformly at random for each value of π β 3, ..., 40 (where
π is the number of men/women) and allowed any man to be cho-sen as an accomplice. We follow the same setup for all subsequentexperiments unless stated otherwise.
11.1 Fraction of Women Who Improve (Contβd)We revisit the experiment in which we compared the fraction ofwomen who are able to improve through no-regret accomplice andself manipulation. In Table 1, we catalog the number of womenwho benefit from both strategies when π = 20. Not only are theremore instances where at least one woman improves through no-regret accomplice manipulation, but there are also more instanceswhere larger numbers of women improve. For example, there are noinstances where more than ten women improve through self manipu-lation. This is a stark contrast to no-regret accomplice manipulationthrough which sixteen women are able to improve in one of theinstances. Interestingly, there are no instances where exactly onewoman improves through no-regret accomplice manipulation. Thisis due to the sequence of proposals that occur after a no-regret pushup operation. We formalize this observation in Proposition 6.
PROPOSITION 6. Let β» be a preference profile and ` B DA(β»).For any man π, let β»β²B {β»βπ, β»π β
π } and ` β² B DA(β»β²). If πdoes not incur regret and ` β² β `, then there exist at least twodistinct women π€ β²,π€ β²β² β π such that ` β²(π€ β²) β»π€β² ` (π€ β²) and` β²(π€ β²β²) β»π€β²β² ` (π€ β²β²), and at least two distinct men πβ²,πβ²β² β π
such that ` (πβ²) β»πβ² ` β²(πβ²) and ` (πβ²β²) β»πβ²β² ` β²(πβ²β²).Before proving Proposition 6, we show that if a manπ performs
a no-regret push up operation such that all the women he pushesup prefer him less than their original DA partners, then the push upoperation is weak (i.e., the manipulated matching is the same as theoriginal matching).
Lemma 15. Let β» be a preference profile and ` B DA(β»). For anyman π, let β»β²= {β»βπ, β»π β
π } and ` β² = DA(β»β²) such thatπ does notincur regret. If ` (π€ β²) β»π€β² π for all π€ β² β π , then ` β² = `.
PROOF. Sinceπ does not incur regret after a push up operation,` β² must be stable with respect to the true preferences, i.e., ` β² β πβ»(Theorem 4.2). Additionally, Corollary 1 implies that ` β² is weaklybetter for all women (i.e., ` β² βͺ°π `) and weakly worse for all men(i.e., ` βͺ°π ` β²).
Suppose, for contradiction, that ` β² β `. Let π B {π§ β π :` β²(π§) β»π§ ` (π§)} denote the set of women with a strictly morepreferable partner under ` β². Thus, by assumption, π β β . Letπ B {π¦ β π : ` β²(π¦) β π } denote the set of men whose ` β²-partnersare in the set π . Note that any man not in π has the same partnerunder ` and ` β²; in particular, π β π by the no-regret assumption.Also note that each man in π strictly prefers its partner under ` thanunder ` β², i.e., ` β»π ` β² (this is an easy consequence of the stabilityof ` β² with respect to the true profile β») .
Consider the execution of the DA algorithm on the profile β»β².Since the men propose in decreasing order of their preference, eachman in π must be rejected by his `-partner during the algorithm.Letπ1 β π denote the man who is the earliest to be rejected by his`-partner (i.e., the woman he is matched to under the matching `),say π€1 B ` (π1). Then, π€1 must have at hand a more preferableproposal, sayπ2 (i.e.,π2 β»π€1 π1), that she does not receive underthe execution of DA(β»).
GAIWβ21, May 2021, London, UK Hadi Hosseini, Fatima Umar, and Rohit Vaish
No. of women who benefit 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20No. of instances (accomplice) 307 0 101 102 88 79 88 59 49 54 28 21 11 8 3 1 1 0 0 0 0No. of instances (self) 411 151 178 128 68 33 20 7 1 2 1 0 0 0 0 0 0 0 0 0 0
Table 1: The number of instances (out of 1000) where a varying numbers of women benefit through no-regret accomplice manipulationand self manipulation when π = 20.
Sinceπ2 is not rejected by his `-partner yet, we have thatπ€1 β»β²π2
` (π2). Now, if π2 β π, then we get π€1 β»π2 ` (π2), which contra-dicts the stability of ` with respect to β» as (π2,π€2) would constitutea blocking pair. Otherwise, if π2 =π, then it must be that π€1 β π
(since the women in π are the only ones whoπ proposes to underβ»β² but not under β»). Then, from the above condition, we will haveπ = π2 β»π€β² π1 = ` (π€ β²) for some π€ β² β π , which contradicts thecondition given in the lemma. Hence, ` = ` β², as desired. β‘
We are now ready to prove Proposition 6.
PROOF OF PROPOSITION 6. Let π β π be the set of womenthat π pushes up. Since it is assumed that ` β² β `, the contrapos-itive of Lemma 15 implies there exists a woman π€ β² β π suchthat ` (π€ β²) βπ€β² π. By the push up assumption, ` (π) β π€ β²; other-wise π would not have been able to push up π€ β². This implies thatπ β ` (π€ β²) and thus, π β»π€β² ` (π€ β²). Since men propose in decreas-ing order, π proposes to π€ β² under DA(β»β²). By no-regret assumption` (π) = ` β²(π), we know that π€ β² rejects π at some point to bematched with another manπβ² such thatπβ² β»π€β² ` (π€ β²). This impliesthat πβ² did not propose to π€ β² under DA(β»), and thus,πβ² is matchedto a more preferred woman under β», i.e., ` (πβ²) β»πβ² ` β²(πβ²).
Now let οΏ½οΏ½ B ` (π€ β²) be π€ β²βs partner under β». By Corollary 1, itmust be the case that ` (οΏ½οΏ½) β»οΏ½οΏ½ ` β²(οΏ½οΏ½). Let π€ β²β² B ` β²(οΏ½οΏ½). Followingthe same reasoning as above, since ` (π) = ` β²(π) andπ β ` (π€ β²β²),we haveπ β»π€β²β² ` (π€ β²β²). The order of proposals under β»β² indicatesthat π€ β²β² rejects πβs proposal to be matched to a more preferred manπβ²β² under DA(β»β²), which consequently implies that ` β²(π€ β²β²) β»π€β²β²
` (π€ β²β²) while ` (πβ²β²) β»πβ²β² ` (πβ²β²). Therefore, π€ β² and π€ β²β²βs partnersare strictly improved whereas πβ² and πβ²β²βs partners are strictlyworsened off under DA(β»β²). β‘
In the same experiment, we additionally computed the fraction ofinstances in which it is possible for at least one woman to improvethrough no-regret accomplice and self manipulation. The results inFigure 6 once again suggest that no-regret accomplice strategies aremore prevalent than self manipulation.
Figure 6: Comparing fractions of instances that are manipula-ble by at least one woman through no-regret accomplice andself manipulation.
Figure 7: Comparing no-regret accomplice, with-regret accom-plice (with variable-sized pools of potential accomplices forboth), and self manipulation against truthful reporting whenπ = 40.
11.2 How Much Flexibility is Really Needed inChoosing the Accomplice?
So far, our experiments have taken the optimistic approach of al-lowing the strategic woman to pick any man of her choice as theaccomplice. To examine the exact extent of flexibility that this as-sumption requires, we conduct an experiment where the accompliceis chosen from a fixed pool of π men for some π β€ π (for example,when π = 5, we pick the best accomplice from a fixed set of fivemen). For π = 40, we ran the no-regret accomplice, with-regretaccomplice, and self manipulation strategies on 1000 profiles forπ β {1, . . . , 40}. The results are presented in Figure 7.
One would expect with-regret accomplice manipulation to out-perform self manipulation when there is variability in accomplices.Indeed, the strategic woman could simply ask her top choice manto place her at the top of his list. However, we observe that, in ex-pectation, with-regret accomplice manipulation outperforms selfmanipulation for every pool size π (thus, a with-regret strategy out-performs self manipulation even when a single accomplice is ran-domly chosen in advance). Furthermore, the comparatively limitedno-regret accomplice manipulation is also, on average, better thanself manipulation when there are at least four men to choose from.These observations suggest that the superior performance of accom-plice manipulation can be achieved even with a modest amount offlexibility in the choice of the accomplice.
11.3 Regret vs. ImprovementWe examine the tradeoff between regret (of the accomplice) andimprovement (of the strategic woman) under the with-regret manip-ulation model. Rather than allowing any man to be chosen as anaccomplice, we ran the with-regret manipulation strategy on a fixedwoman π€ and recorded the outcomes after individually using eachman as an accomplice. In other words, if there were multiple optimalstrategies for π€ , we chose the one that caused the accomplice to in-cur the least amount of regret. Figure 8 illustrates the distribution of
Accomplice Manipulation of the Deferred Acceptance Algorithm GAIWβ21, May 2021, London, UK
Figure 8: Comparing distributions of improvement for thestrategic woman π€ and regret for the accomplices. The solidbars, whiskers, and dots denote the interquartile range, rangeexcluding outliers, and outliers, respectively.
improvement achieved for π€ and regret incurred by the accomplicesin terms of the difference in the ranks of their matched partnersbefore and after manipulation. Interestingly, the expected regret forthe accomplice is greater than the expected improvement for themanipulating woman.
12 SUBOPTIMAL ACCOMPLICEMANIPULATION
We have already shown that any optimal accomplice manipulationstrategy admits an equivalent inconspicuous strategy (Theorems 4.3and 5.2). Theorem 12.1 strengthens this result by showing that anybeneficial (i.e., optimal or suboptimal) accomplice manipulation ad-mits an equivalent inconspicuous strategy for both the no-regret andwith-regret settings. In order to prove this, we start by introducingsome new results (Lemmas 16 and 17).
Lemma 16. Let (π,π€) be a manipulating pair. Let π be a set ofwomen whoπ can push up and ` (π€) be π€βs match afterπ pushesup all women in π . Letπ€ β² β π be the single womanπ needs to pushup to get π€ matched with ` (π€) (as guaranteed by Lemmas 3 and 4).Then,π can push up any subset of women in π that containsπ€ β² (i.e.,any subset π β π such that π€ β² β π) to get π€ matched with ` (π€).
PROOF. Let β»πB {β»βπ, β»π βπ }, β»πB {β»βπ, β»πβπ }, and β»π€β²
B
{β»βπ, β»π€β²βπ }. Without loss of generality, β»π is derived from β»π if
π pushes down the set π \ π . Similarly, β»π€β²is derived from β»π
if π pushes down the set π \ {π€ β²}. From Lemma 1, we get that`π βͺ°π `π and `π βͺ°π `π€
β², where `π B DA(β»π ), `π B DA(β»π ),
and `π€β²B DA(β»π€β²). Since it is assumed that `π (π€) = `π€
β² (π€), itmust be the case that `π (π€) = `π (π€). Thus, π can push up anysubset to get π€ matched with ` (π€). β‘
The next result (Lemma 17) shows that a strictly beneficial ac-complice manipulation that uses a combination of push up and pushdown operations can be achieved through push up operations alone.This strengthens Lemma 2 which showed that a combination ofpush up and push down operations is weakly worse than push upoperations alone.
Lemma 17. Let (π,π€) be a manipulating pair, and let β» be apreference profile. For any subsets of women π βπ and π βπ , letβ»β²B {β»βπ, β»π β
π } denote the preference profile after pushing up theset π and β»β²β²B {β»βπ, β»π β,π β
π } denote the profile after pushing up
π and pushing down π in the true preference list β»π of man π. Let` B DA(β»), ` β² B DA(β»β²), and ` β²β² B DA(β»β²β²). If ` β²β²(π€) β»π€ ` (π€),then ` β²β²(π€) = ` β²(π€).
PROOF. Suppose, for contradiction, that ` β²β²(π€) β ` β²(π€). This,combined with ` β²(π€) βͺ°π€ ` β²β²(π€) (Lemma 2), gets us ` β²(π€) β»π€
` β²β²(π€). Additionally, we have assumed that ` β²β²(π€) β»π€ ` (π€), whichgets us ` β²(π€) β»π€ ` (π€).
β»π : Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β· π Β· Β· Β·
β»β²π : Β· Β· Β· π Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β·
β»β²β²π : Β· Β· Β· π Β· Β· Β· ` (π) Β· Β· Β· π Β· Β· Β·
β»βπ : Β· Β· Β· ` (π) Β· Β· Β· π Β· Β· Β· π Β· Β· Β·
Figure 9: An illustration of man πβs preference lists under theprofiles β», β»β, β»β², and β»β²β² in the proof of Lemma 17.
Now, let π B π βͺ π . Consider a preference profile β»β derivedfrom the true profile β» by pushing down all women in π below` (π) in the accompliceπβs true preference list (see Figure 9). FromProposition 4, we have that `β (π) = ` (π), where `β B DA(β»β).This implies that a push up/down operation with respect to `β (π)is equivalent to the same operation with respect to ` (π). Therefore,starting with β»β
π , if π pushes up all women in π , then we obtainthe profile β»β². Similarly, starting with β»β
π , if π instead pushes upall women in π (respectively, π ), then we obtain the profile β»β²β²
(respectively, β»).Since profile β»β² is derived from β»β via a push up operation of the
set π , Lemma 3 implies that the same match for π€ , namely ` β²(π€),can be achieved by promoting just one woman, say π€ β² β π , in thepreference list β»β²
π . Since π and π are disjoint sets, we have thateither π€ β² β π or π€ β² β π . If π€ β² β π , then from Lemma 16, wehave that ` β²β²(π€) = ` β²(π€), contradicting our original assumption.On the other hand, if π€ β² β π , then again from Lemma 16 we get` (π€) = ` β²(π€), contradicting the condition ` β²(π€) β»π€ ` (π€) that weshowed above. β‘
THEOREM 12.1. Any beneficial accomplice manipulation is,without loss of generality, inconspicuous.
PROOF. From Proposition 3 (and the subsequent remarks), weknow that any accomplice manipulation can be simulated via pushup and push down operations. Lemma 17 shows that any beneficial(i.e., optimal or suboptimal) manipulation that is achieved by somecombination of pushing up a set π βπ and pushing down π βπ
can also be achieved by only pushing up π β π . Finally, fromLemma 3, we know that any match for the manipulating woman π€
that is achieved by pushing up π βπ is also achieved by pushingup exactly one woman in π , thus establishing the desired inconspic-uousness property. β‘