LUND UNIVERSITY

Department of Economics
NEKP01 Master’s Thesis in Economics
Autumn 2015

The Compromise Algorithm in the Swedish PhD Admissions Problem

Dingquan Miao*

Abstract
What mechanism should be designed to allocate PhD applicants to universities in Sweden? We
introduce the Swedish PhD admissions problem, and it is influenced by the college admissions
problem (Gale and Shapley 1962) and the student placement problem (Balinski and Sönmez 1999). In
order to “solve” this problem, we design a novel mechanism, namely the compromise algorithm. We
propose three theorems from this algorithm, i.e. equivalence theorems. The equivalence theorems
specify the equivalence relations among stability, worse and responsiveness. Additionally, we find a
positive result that the number of fields determines the strategy-proofness of the algorithm; meanwhile,
the student optimal stable matching and the university optimal stable matching can be treated as
special cases of our model when we restrict the number of fields. Generally, the compromise
algorithm generates a stable matching that falls in between the student optimal stable matching and the
university optimal stable matching.

KEYWORDS: Matching, Swedish PhD Admissions Problem, Stability, Compromise Algorithm

Supervisor: Dr. Jens Gudmundsson

Date submitted: January 18, 2016
Date examined: January 26, 2016
Discussant: Josefin Kilman

*
I would like to express my deepest gratitude to my supervisor Jens Gudmundsson, for his endless
patience, constant encouragement and invaluable help. His guidance and support helped me overcome
many problems and finish this thesis. I would like to thank my parents for their great confidence in me
all through these years.
Contents
1 INTRODUCTION 3

2 MODEL AND PROPERTIES 8

2.1 PhD Admissions Problem . . . . . . . . . . . . . . . . . . . . 8
2.2 Matching and Properties of Matching . . . . . . . . . . . . . 9

3 ALGORITHMS 11
3.1 Gale-Shapley Deferred Acceptance Algorithm . . . . . . . . . 11
3.2 Compromise Algorithm . . . . . . . . . . . . . . . . . . . . . . 12

4 WORSE 16

5 RESPONSIVENESS 20

6 EQUIVALENCE 22

7 STRATEGY-PROOF 23
7.1 A General Case . . . . . . . . . . . . . . . . . . . . . . . . . . 23
7.2 A Special Case . . . . . . . . . . . . . . . . . . . . . . . . . . 26

8 RESPECTING IMPROVEMENTS 26

9 CONCLUSION 29

References 32

1
List of Tables
1 PREFERENCES FOR EXAMPLE 1 . . . . . . . . . . . . . . 14
2 PREFERENCES FOR EXAMPLE 2 . . . . . . . . . . . . . . 24
3 MATCHING RESULTS FOR STRATEGY-PROOFNESS . . 25
4 PREFERENCES FOR EXAMPLE 3 . . . . . . . . . . . . . . 27
5 MATCHING RESULTS FOR RESPECTING IMPROVEMENTS 29

List of Figures
1 STABILITY OF COMPROMISE ALGORITHM . . . . . . . 30

2
1 INTRODUCTION
With the development of the science and technology civilisation, the number
of postgraduate applicants increases rapidly. Nevertheless, research insti-
tutions and departments from higher education system are generally con-
strained by limited educational resources, such as research funding, grants,
supervisions, and so on. The opportunity of engaging the PhD program ex-
hibits scarcity property, and the admission of the PhD program is exception-
ally competitive. In order to allocate these scare resources more efficiently,
it is appealing and meaningful to investigate the following problems: who
should study PhD program, who is the right person for the research insti-
tution, what system should be designed to select qualified PhD candidates
and so on. In this paper we study the Swedish PhD admission problem
and design a mechanism for the admission system based on its distinctive
attributes. The Swedish PhD program is a vital component of the national
higher education system, and it has several idiosyncrasies: (1) In Sweden,
each university posts PhD positions on its websites and admits PhD stu-
dent individually. There is no centralised “admission centre” to deal with the
national PhD admissions, and this may cause some potential shortcomings.
For example, there is no unified requirement, and the degree of competence
for the admission is unlikely to be the same for different universities, but
all the admitted PhD candidates are able to access education resources from
different universities, such as the U6 network, MIT and etc.1 If PhD candi-
dates have the right to access the education resources equally among different
academic networks within the country, then a centralised admission system
can fully utilise the academic resources among universities, and it is more
efficient to admit students through a centralised admission centre, e.g. a stu-
1
The U6 network consists of universities at Göteborg, Linköping, Lund, Stockholm,
Umeå, Uppsala and Stockholm School of Economics. And MIT is the abbreviation of
Research School of Management and IT, which is a co-operation between twelve Swedish
universities.

3
dent who is not qualified for one university might be acceptable for another
university. (2)The Swedish PhD programs are considered as public goods
for students; meanwhile, the programs select qualified candidates as employ-
ees for the universities, since PhD programs are mainly research-based in
Sweden. Therefore, we may want to take both students’ and universities’
preferences into account. (3) The Swedish PhD program is very competitive,
because of its reputation of high research standard.2 Meanwhile, Swedish uni-
versities consider PhD candidates as employees and pay them salaries, and
this will cause two effects: (i) It assists PhD candidates to focus more on their
research; (ii) The salary, on the other hand, limits the number of positions
offered at each university in Sweden indirectly.3
In contrast with the analysis of traditional economic markets that price
level determines the equilibrium of the market, many real-world situations are
not consistent with traditional theory (Roth 2007). For example, the school
admission problem (Gale and Shapley 1962), the marriage problem (Gale and
Shapley 1962), the housing allocation problem (Shapley and Scarf 1974), and
the human organs for transplants problem (Roth, Sönmez and Ünver 2004)
and so on. The allocation of these cases is no longer depending on price,
and it should generate an “efficient outcome” under different norms rather
than the traditional equilibrium theory. A question arises for how to allocate
the resources of these markets, and the abstract theory developed in the
early 1960s on the basis of the college admissions problem and the marriage
problem (Gale and Shapley 1962). And the solution structured a well-known
matching algorithm that assigns students to universities, namely the Gale-
2
For example, the number of admissions of the four largest universities in Sweden are:
Lund University admits 4-6 students each year out of a number of students, Stockholm Uni-
versity admits 5-10 students each year out of about 100 applications, Uppsala University
admits 8 students each year out of about 100 applications and University of Gothenburg
admits approximately 6 students every second year out of a number of students.
3
Even studentship is a common funding form, other funding forms may exist to com-
plete the studentships, e.g. the PhD program in Economics at Stockholm University also
allows students to be self-financed.

4
Shapely “deferred acceptance” algorithm. More importantly, this deferred
acceptance algorithm generates optimal stable matchings (Gale and Shapley
1962). A matching assigns one set of agents (e.g. students) to another set
of agents (e.g. universities), and a matching is stable if it is individually
rational and there is no blocking pair at that matching. More specifically, a
matching matches two agents A and B with ↵ and , respectively; if no agent
prefers unmatched option than her current matching, then this matching is
individually rational; if no pair of agents prefers each other (A and , B
and ↵) to their partners at this matching, then there is no blocking pair at
this matching. An optimal stable matching is the stable matching for one
agent that is at least as well off under this matching as the agent would be
in any other stable matchings. And the deferred acceptance algorithm can
be classified into two ways. Specifically, for the college admissions problem:
either students propose to universities, namely the student proposing deferred
acceptance (henceforth SP-DA) algorithm, or universities apply to students,
that is, the university proposing deferred acceptance (henceforth UP-DA)
algorithm. The first case leads to the student optimal stable matching, and
the latter one produces the university optimal stable matching.
However, the empirical work of Gale-Shapley algorithm was not recog-
nized until the 1980s, and it started from the medical markets (Roth 1984).
It has been empirically verified the importance of the stability from two
medical markets: the U.S. and the U.K.. In the U.S., a successful centralized
clearinghouse, namely the National Resident Matching Program (NRMP),
which was proved to produce stable matching. It was hypothesized that
the success of the NRMP was intimately correlated to the stability of the
matching it produced (Roth 1984). To test the “stability hypothesis” of med-
ical markets in the U.S., a study was made in the U.K. in the early 1990s
(Roth 1991). And this empirical work verified the “stability hypothesis”,
i.e. for those regions that employed algorithms generate stable matching
had turned out to be successful, while, for the others had broken down in

5
various ways. Along with medical markets, applications in the education
markets have been extensively studied. Like the medical markets, the theo-
retically design of assignment mechanism in the education markets is closely
related to the real world applications. After the famous Gale-Shapley college
admissions problem, one real-life application of the Turkish student place-
ment problem were introduced, namely the multi-category serial dictatorship
mechanism (Balinski and Sönmez 1999). Balinski and Sönmez showed (1999)
that the multi-category serial dictatorship and the Gale-Shapley college op-
timal stable mechanisms are equivalent. Another inspired paper (Abdulka-
diroğlu and Sönmez, 2003) described mechanisms applied in the U.S. at that
time, for instance, Boston student assignment mechanism and Columbus
student assignment mechanism. Additionally, Abdulkadiroğlu and Sönmez
(2003) explained some serious shortcomings of these two mechanisms, and
they suggested two alternative mechanisms that offer practical solution for
those drawbacks, namely the Gale-Shapley student optimal stable mecha-
nism and the top trading cycles mechanism (Shapley and Scarf 1974). Along
with theoretical studies, popular empirical studies of Gale-Shapley and Top
Trading Cycles mechanisms suggest that the efficiency can be improved by
replacing the Boston mechanism with either Gale-Shapley and Top Trading
Cycles mechanism (Chen and Sönmez 2006). After Abdulkadiroğlu and Sön-
mez published their paper (2003), New York City high schools replaced it
allocation mechanism with the student-proposing deferred acceptance mech-
anism (Abdulkadiroğlu, Pathak and Roth 2005).
In order to deal with idiosyncrasies mentioned in the Swedish PhD ad-
mission problem, a novel mechanism is introduced in this paper, namely the
compromise algorithm. This algorithm has some desirable properties, e.g. it
takes both the students and universities’ preferences into account and gen-
erates stable matchings. This paper illustrates the Swedish PhD admissions
problem in comparison with two well-known problems, specifically, the col-
lege admissions problem (Gale and Shapley 1962) and the Turkish student

6
placement problem (Balinski and Sönmez 1999). Sönmez (2015) explained
in his mini-course video that the “no-tie” assumption in the Turkish student
placement problem is acceptable. The “no-tie” assumption indicates univer-
sities’ preferences on students are strict, alternatively, it assumes students
acquire different grades in each field. Since in the very unlikely case there is
a tie, and the tie becomes relevant only for the border line students, e.g. if a
university has 100 slots, only the tie occurs at 100th applicant matters. And
the Turkish admissions system will assign the younger student with a higher
score if two students have the same score on border line. However, in the
Swedish PhD admissions problem, the PhD position for each university in
Sweden is more limited, and “tie” becomes more sensitive than in the Turkish
student placement problem. In order to mitigate the sensitivity of “tie” in
the Swedish PhD admissions problem, we adopt the lexicographic preference
orders over different fields rather than (strict) preference orders over stu-
dents, and the lexicographic preferences impose a less restricted assumption
on students’ grades, i.e. lexicographic preferences allow students to acquire
same grades in some fields as long as there is no two students achieve same
grades in every fields. Note that, the structure of preference relations of the
Swedish PhD problem is a generalisation of the structure preference relations
of the Turkish student placement problem (Balinski and Sönmez 1999).
We present three main contributions of this paper. First, this paper intro-
duces a new matching mechanism, namely the compromise algorithm. This
algorithm is deigned for the Swedish PhD admissions problem, and it gen-
erates a stable matching. The essence of this algorithm is to start from the
SP-DA algorithm, and then universities whose capacities are filled will col-
lectively reject their least preferred students; the assignments subsequently
continue iterating from SP-DA algorithm until the matching is not stable,
and the algorithm selects the stable matching previously adjacent to that
unstable matching. The compromise algorithm eventually produces a stable
matching that falls in between the student optimal stable matching and the

7
university optimal stable matching. Besides the compromise algorithm, three
important theorems are derived from the compromise algorithm, namely the
equivalence theorems. The equivalence theorems suggest the equivalence re-
lations among stability, worse and responsiveness. The equivalence theorems
offer alternative inspections for the stability of the algorithm, which will be
discussed in later section. Furthermore, one minor finding is obtained from
the model, that is, if we restrict the number of fields to one, the compromise
algorithm generates not only stable but also strategy-proof matching, more-
over, this matching is equivalent to the student optimal student matching
and the university optimal stable matching.
The paper is structured as follows. We introduce the basic structures
and properties of the model in Section 2. And we review the Gale-Shapely
deferred acceptance algorithm and present our compromise algorithm in Sec-
tion 3. In Section 4, 5, and 6, we demonstrate the equivalence theorems with
relevant concepts of worse and responsiveness, and then we show how they re-
late to the compromise algorithm. In Section 7, we examine the properties of
strategy-proofness and discuss two related cases. We investigate whether the
compromise algorithm respecting improvements in Section 8. We conclude
in Section 9.

2 MODEL AND PROPERTIES

2.1 PhD Admissions Problem
There is a finite set of students S = {s1 , . . . , sn }, and a set of universities
U = {u1 , . . . , um } is defined for university colleges and universities in Swe-
den. A binary preference relation Rs on U [ {s} is defined for all s 2 S
such that, for each {u, u0 } ✓ U [ {s}, uRs u0 if student s 2 S consid-
ers u is at least as desirable as u0 . Each student s 2 S has a strict
preference relation Ps on U [ {s} such that, for {u, u0 } ✓ U [ {s}, uPs u0

8
if and only if uRs u0 and u0 ¬Rs u. A vector of student preferences PS =
(Ps1 , . . . , Psn ) where Psi is the strict preference relation for student si 2
S. Both types of preference relations are rational , i.e. completeness and
transitivity. There is a finite set of fields F = {f1 , . . . , fk }, e.g. F =
{microeconomics, macroeconomics, econometrics, . . . }. For each field f 2
F , each student s 2 S has a grade gfs 2 R. Let g s = (gfs )f 2F 2 RF be the list
of s’s grades and G = (g s )s2S 2 RF ⇥S be the matrix consists of all students’
grades in all fields. Each university u 2 U has a (strict) preference Pu over
fields. This, together with the grade matrix G, induces a lexicographic
preference relation Pulex over all students and “no-student option”. The
no-student option is ranked prior to those students who are unqualified for
(lex)
university u. For {i, j} ✓ S [ U , if iPj j, then i is acceptable to j. In par-
ticular, for each u 2 U , there exists a list of thresholds g u = (gfu )f 2F 2 RF
such that, for all s 2 S, sPulex u if and only if there is f 2 F such that gfs > gfu
and, for each f 0 2 F such that f 0 Pu f , gfs 0 = gfu0 . And GT = (g u )u2U 2 RF ⇥U
is the threshold matrix consists of all universities’ thresholds in all fields.
Generally, for each u 2 U and {s, s0 } ✓ S [ {u}, sPulex s0 if and only if there is
0 0
f 2 F such that gfs > gfs and, for each f 0 2 F such that f 0 Pu f , gfs 0 = gfs 0 . It
is assumed that for any pair of students {s, s0 } ✓ S there exists at least one
0
field f 2 F such that gfs 6= gfs . In contrast to the student placement problem
(Balinski and Sönmez 1999), the no tie assumption restricts students to pos-
sess different test scores in every skill category. We allow students to achieve
same grades in some fields as long as there is no two students acquire same
grades in every field. Each university u 2 U is constrained by its limited
PhD position(s), and a vector of university capacities q = (qu1 , . . . , qum )
where qu 2 N+ is the capacity of university u.

2.2 Matching and Properties of Matching

A matching µ : S ! U [ S matches students with universities. With some
abuse of notation, a matching is µ such that, for each s 2 S, µ(s) 2 U [ {s},

9
and for each u 2 U , µ(u) = {s 2 S|µ(s) = u}. Also, |µ(u)|  qu . If µ(s) = s,
then s is not assigned a university at µ. If µ(u) = ;, then u has no student
at µ. The set of matchings is M. A matching µ 2 M is individually
rational if, for each student, the student finds his assigned university at
least as desirable as the no-university option at µ, meanwhile, the university
assets the no-student option as the least preferred option among all students
at µ(u).4
DEFINITION 1 (Individual rationality) A matching µ 2 M is individu-
ally rational at R if, for each s 2 S and u 2 U ,

µ(s)Rs s, and 8s 2 µ(u), sPulex u.

Correspondingly, a matching is wastef ul if, a student prefers u to her

current assigned university, and she acquires sufficient grades in terms of the
lexicographical threshold of u, and she is not admitted by u.
DEFINITION 2 (Wastefulness) For a matching µ 2 M is wasteful, if
there is u 2 U and s 2 S,

uPs µ(s), and |µ(u)| < qu , and sPulex u.

A stable matching can be defined as no individual or any pair of agents can

benefit from blocking or deviating from their current matchings. A student-
university pair (s, u) 2 S ⇥ U is a blocking pair for µ 2 M if, uPs µ(s) and
there is s0 2 µ(u) such that sPulex s0 .
DEFINITION 3 (Stability) For a matching µ 2 M is stable if,

• for each u 2 U and s 2 S are individually rational, and

• there is no blocking pair.

The list of students’ preferences and colleges’ preferences, and the capacities
of colleges define the college admissions problem (Gale and Shapley, 1962),
4
This section is mostly guided by Gudmundsson’s working paper (2015).

10
and it is denoted as (PS , PU , q). The student placement problem (Balinski
and Sönmez, 1999) consists of students’ preferences, students’ test scores, and
the capacities of colleges, which can be expressed as (PS , G, q). The Swedish
PhD admissions problem reveals preferences by the lexicographical order on
different fields of economics, and universities rank PhD applicants based on
their grades in each field of economics. The Swedish PhD admissions problem
can thus be defined as (PS , PUlex , q).

3 ALGORITHMS
In this section, we firstly review the classic Gale-Shapely Deferred Acceptance
Algorithm (Gale and Shapley 1962), and then introduce the compromise
algorithm, since the compromise algorithm builds on the SP-DA algorithm.
Moreover, we demonstrate the compromise algorithm with comprehensive
example.

3.1 Gale-Shapley Deferred Acceptance Algorithm

In order to illustrate our algorithm, we need to start from introducing the
Gale and Shapley’s deferred acceptance algorithm (1962). The student
proposing deferred acceptance (SP-DA) algorithm is one of the main
contributions studied by them and it can be explained as: each student s 2 S
proposes to her most preferred university. Meanwhile, each university u 2 U
ranks its applicant(s), if any, in a list depends on student academic perfor-
mance. Specifically, u temporarily keeps the best qu 2 N applicants and
rejects the rest if the number of applications exceeds qu , and it keeps all
acceptable offer(s) if the number of applications does not exceed qu . Each
student s 2 S who gets rejected in the last step proposes to her next choice,
and each university again tentatively accepts at most qu applicant(s) based
on the priority from both the new applicant(s) and previous ones, and re-

11
jects the rest. Student who gets rejected will not apply to the same university
again, and student will be matched to the no-university option if she has been
rejected by all the universities she finds acceptable, and this algorithm con-
tinues progressing until no students make further proposals. This algorithm
yields a student optimal stable matching that is preferred by students,
and it is the worst stable matching for universities (Roth 1985). Note that,
a matching µ 2 M is optimal if for each student s 2 S is at least as well off
under the matching µ as she would be in any other stable matching.

3.2 Compromise Algorithm

As mentioned in the previous section, the SP-DA algorithm generates opti-
mal stable matching for students, while it is an “adverse” stable matching to
universities. In order to take both students’ and universities’ preferences into
account, we design the compromise algorithm. The algorithm starts from
running a SP-DA algorithm, and yields a student optimal stable matching.
Secondly, for those universities who filled their capacities can reject their least
preferred applicants, and the algorithm runs the SP-DA algorithm again, and
we check if it generates a stable matching. The process continues iterating
two preceding steps until the unstable matching µ 2 M appears, and the al-
gorithm selects the stable matching that previously adjacent to this unstable
matching µ. The compromise algorithm is as follows.

Step 1. Set k = 0 as an initial point, and run the SP-DA algorithm and
get the student optimal stable matching µ0 .
Step 2. Each university such that, |µk (u)| = qu , rejects its least preferred
student among the students assigned to it at µk , that is, for u 2 U , s 2 µk (u)
is rejected if, for each s0 2 µk (u), s0 Rulex s.
Step 3. Each rejected student continues with her application, and the
application procedure follows the SP-DA algorithm and generates matching
µk+1 .

12
 Step 4. If the matching µk+1 is unstable, then select µk and the algorithm
terminates. Otherwise, increase k by 1, and return to step 2.5

The following example is constructed to illustrate our algorithm progres-

sively.
EXAMPLE 1. Let S = {s1 , s2 , s3 , s4 , s5 , s6 , s7 }, U = {u1 , u2 , u3 }, q =
(qu1 , qu2 , qu3 ) = (3, 2, 1), F = {f1 , f2 , f3 }. The threshold matrix and grade
matrix are as follows:

2 3 2 u1 u 3 2 3
gfu1 gf1 gf21 gfu31 83 85 80
6 7 6 7 6 7
GT = 4gfu2 5 = 4gfu12 gfu22 gfu32 5 = 481 82 805
gfu3 gfu13 gfu23 gfu33 84 83 85

2 3 2 s 3 2 3
gfs1 gf11 gfs21 gfs31 95 87 76
6 s2 7 6 s2 7 6 7
6 g f 7 6 g f1 gfs22 gfs32 7 687 85 907
6 s7 6 s 7 6 7
6g 3 7 6g 3 gfs23 gfs33 7 6 837
6 f 7 6 f1 7 692 85 7
6 7 6 7 6 7
G = 6gfs4 7 = 6gfs14 gfs24 gfs34 7 = 685 77 807
6 s7 6 s 7 6 7
6g 5 7 6g 5 gfs25 gfs35 7 6 907
6 f 7 6 f1 7 685 93 7
6 s6 7 6 s6 7 6 7
4 g f 5 4 g f1 gfs26 gfs36 5 483 87 955
gfs7 gfs17 gfs27 gfs37 78 95 90

And students’ and universities’ preference relations are as follows:6

5
We use “asterisk” to denote the iteration process and the number of asterisks equals
to the number of times of iteration, e.g. Step 2⇤ represents repeating step 2 once.
6
Table 1.c displays the preferences orderings only for acceptable students, i.e. omitting
those students who acquire insufficient grades in comparison with thresholds for each
university.

13
 Pulex
1
Pulex
2
Pulex
3
s7 s1 s6
Ps 1 Ps2 Ps 3 Ps4 Ps 5 Ps6 Ps 7 Pu1 Pu2 Pu3 s5 s3 s2
u1 u1 u1 u1 u2 u2 u2 f2 f1 f3 s1 s2 s5
u3 u2 u2 u3 u3 u1 u3 f1 f2 f1 s6 s5 s7
u2 u3 u3 u2 u1 u3 u1 f3 f3 f2 s3 s4
(a) Students’ Preferences on U (b) Universities’ s2 s6
Preferences on F
(c) Universities’ Lexico-
graphic Preferences on
S

Table 1: PREFERENCES FOR EXAMPLE 1

Step 1: In the light of the SP-DA algorithm, the admission procedure

implements as follows. Firstly, each student applies to her most preferred
university, for example, s1 , s2 , s3 , s4 propose to u1 , and s5 , s6 , s7 propose
to u2 . Secondly, each university is limited by its capacity qu and it declines
the unacceptable applications, for instance, u1 rejects s4 since gfs4 < gfu1 , u2
rejects s7 since gfs7 < gfu2 . Thirdly, s4 and s7 move to their next preferred
universities, i.e. s4 and s7 both apply to u3 ; u3 rejects s4 , owing to s4 is
not qualified for u3 either, meanwhile, s7 is tentatively accepted by u3 . s4
continues applying to her last option u2 , s6 is then rejected by u2 , due to
s5 Pulex
2
s4 Pulex
2
s6 and qu2 = 2; and s6 moves to her next preferred university
u1 , and u1 rejects s2 , since s1 Pulex
1
s6 Pulex
1
s3 Pulex
1
s2 and qu1 = 3. s2 then goes
to her next preferred choice u2 . Lastly, s4 is rejected by u2 , because she has
been rejected by all universities, s4 is assigned to the no university option,
i.e. µ0 (s4 ) = s4 . The rest students are assigned to their tentative positions.
Therefore, Step 1 generates student optimal stable matching µ0 (µk , and k =
0).
!
u1 u 2 u3
µ0 =
s1 , s3 , s6 s2 , s5 s7

14
 Step 2: Each university u 2 U such that, |µk (u)| = qu , rejects its least
preferred applicant at µ0 . For example, u1 rejects s3 and u2 rejects s5 , since
s1 Pulex
1
s6 Pulex
1
s3 and s2 Pulex
2
s5 , u3 rejects s7 given qu3 = 1.
Step 3: Each rejected student continues with her application, and the
application procedure obeys the SP-DA algorithm. For example, s3 applies
to u2 , s5 proposes to u3 and s7 proposes to u1 . And this generates the
following matching µ1 .
!
u1 u2 u3
µ1 =
s1 , s6 , s7 s2 , s3 s5
Step 4: The matching µ1 is not blocked by any individual and there is no
blocking pair exists, hence, µ1 is a stable matching. The algorithm continues
and starts from µ1 (increase k by 1), and it returns to step 2.
Step 2⇤ : Each university u 2 U such that, |µk (u)| = qu , rejects its least
preferred applicant at µ1 . For example, u1 rejects s6 and u2 rejects s2 , since
s7 Pulex
1
s1 Pulex
1
s6 and s3 Pulex
2
s2 , u3 rejects s5 given qu3 = 1.
Step 3⇤ : Each rejected student continues with her application, and the
application procedure follows the SP-DA algorithm. For instance, s6 and s5
propose to u3 and u1 , respectively; and s2 has no university to apply, because
she has been rejected by all universities. And this generates the matching
µ2 .
!
u1 u 2 u3
µ2 =
s1 , s5 , s7 s3 s6
Step 4⇤ : The matching µ2 is unstable with one blocking pair (s2 , u2 ), i.e.
s2 Pulex
2
u2 and u2 Ps2 s2 . Therefore, the algorithm terminates and we select µ1
such that, M(PS , PUlex , q) = µ1 .

PROPOSITION 1. The compromise algorithm generates a stable match-

ing that falls in between the student optimal stable matching and university

15
optimal stable matching.
In the following three sections, we will introduce three important theo-
rems for the compromise algorithm, i.e. the equivalence theorems. Based on
these equivalence theorems, the algorithm can be converted into two alter-
native versions. And Section 4 will demonstrate the first possible option.

4 WORSE
In this section, we demonstrate the first equivalence theorem, and this the-
orem interprets the compromise algorithm in an alternative way. Before
illustrating the theorem, we can define the relevant concept of this theorem.
The university will not reject its current least preferred applicant if the re-
jection will replace with an even less preferred applicant in the next round
for the university, otherwise, the university is worse o↵ . Suppose a stu-
dent s 2 µk (u) is the least preferred student among all students that are
matched to u at round k, then university u 2 U is not going to reject s if a
less preferred student s0 is assigned to u at next round k + 1, otherwise, the
university u is worse off.
DEFINITION 4 (Worse) A university u is worse off at µk+1 than at µk if,

• there exists s0 2 µk+1 (u) such that, for all s 2 µk (u), sPulex s0 , or

• µk+1 (u) ( µk (u)

In view of the compromise algorithm, for any u 2 U , and for some s0 2 µk (u)
is the least preferred student at µk (u) implies s0 62 µk+1 . If µk+1 assigns u
a less preferred student than µk , then u is worse off at µk+1 than at µk . In
addition, µk+1 (u) ( µk (u) corresponds the case where student s who has
passed the threshold grades of u, and u does not get any new applicant at
µk+1 , then u is worse off at µk+1 if s is no longer an assigned applicant for

16
u, i.e. for some (g s )s2S > g u , µk+1 (u) = µk (u)\{s}. We should note that
µk+1 (u) ( µk (u) implies the matching µk+1 is wasteful.
The definition of worse is intimately related to the compromise algorithm,
and it is one of the important building blocks for the following three theorems.

THEOREM 1. Given µ0 , µ1 , . . . , µk is stable for k 0, µk+1 is stable if

and only if no university is worse off at µk+1 than at µk .

P roof of T heorem 1
Part 1 : Given the conditions from theorem 1, prove that if µk+1 is stable,
then no university is worse off at µk+1 than at µk .
Assume µk+1 is stable, by contradiction, suppose u 2 U is worse off at
µk+1 than at µk . By definition of worse, we have the following two cases.
case 1 :
1
there exists s0 2 µk+1 (u) such that,
1
f or each s 2 µk (u), sPulex s0 .
Let s be the least preferred student at µk (u), i.e. µk (s) = u, and |µk (u)| =
qu . By the design of algorithm, s proposes to a less preferred university until
she is not rejected.
) uPs µk+1 (s),
Combine with the definition of case 1, i.e. sPulex s0 .
) (s, u) 2 S ⇥ U is a blocking pair to µk+1 , which contradicts the as-
sumption that µk+1 is stable.
Therefore, if µk+1 is a stable matching then no university is worse off at
µk+1 than at µk .
case 2 :
µk+1 (u) ( µk (u)
) µk (u) = µk+1 (u) [ {s}, since s is the least preferred student at µk (u).
By the design of the algorithm, u is wasteful at matching µk+1 such that,
|µk+1 (u)| < qu . And s proposes to a less preferred university until she is not

17
rejected.
) sRulex ; and uPs µk+1 (s)
) (s, u) 2 S ⇥ U is a blocking pair to µk+1 , which contradicts the as-
sumption that µk+1 is stable.
Again, if µk+1 is a stable matching then no university is worse off at µk+1
than at µk .

Part 2 : Given the conditions from theorem 1, prove that if no university

is worse off at µk+1 than at µk , then µk+1 is stable.
Firstly, introduce the following notations:
s0 := least preferred student at µk+1 (u),
s00 := least preferred student at µk (u).
(I) Assume u is not worse off at µk+1 than at µk ,
) s0 Rulex s00
(II) Suppose µk+1 is not stable, and there is a blocking pair (s, u) 2 S ⇥ U
at µk+1 such that,
) sPulex s0 and uPs µk+1 (s)
(III)From the known conditions of theorem 1, µk is a stable matching.
) for any pair (s, u) 2 S ⇥ U is not a blocking pair at µk
) u¬Ps µk (s), or s¬Pulex s0
, µk (s)Rs u, or s¬Pulex s0
Hence, there are two cases:
case 1 :
µk (s)Rs u
* uPs µk+1 (s) from (II)
) µk (s)Rs uPs µk+1 (s)
By the design of algorithm, and s 2 / µk+1 (s) and s0 2 µk+1 (s),
) s has been rejected by u before µk+1
case 1.1 :
s 2 µk (u)

18
 case 1.1.1 :
s = s00
* s0 Rulex s00 from (I)
) s0 Rulex s
) contradicts sPulex s0 from (II)
case 1.1.2 :
s 6= s00
By the design of the algorithm, s has been rejected because u has received
better proposals and is able to fill its capacity with better student at µk+1 .
) s0 Rulex s
) contradicts sPulex s0 from (II)
case 1.2 :
s2 / µk (u)
By the design of the algorithm, s has been rejected because u has received
better proposals and is able to fill its capacity with better student at µk+1 .
) s0 Rulex s
) contradicts sPulex s0 from (II)
Hence, if no university is worse off at µk+1 than at µk , µk+1 is stable.
case 2 :
s¬Pulex s0
* u is not worse off at µk+1 than at µk , and s¬Pulex s0 holds at µk+1
) (s, u) 2 S ⇥ U is not a blocking pair for µk+1
Again, if no university is worse off at µk+1 than at µk , then µk+1 is stable.
) For k 0, µk is stable, µk+1 is stable if and only if no university is
worse off at µk+1 than at µk . ⇤

19
5 RESPONSIVENESS
Now, we proceed to the second equivalence theorem, similar with the previ-
ous one, it offers an intuitive translation of the compromise algorithm, and
necessary concept is explained in this section. A responsive preference is
applied when university’s capacity is greater than one, and it is a preference
relation orders the sets of students rather than individual students, i.e. for
any university u’s preference is responsive if, only one student is different
under two different matchings, and it prefers the set consists more preferred
student. Let PuLEX denotes the responsive preferences to u’s lexicographic
preference relation, Pulex .
DEFINITION 5 (Responsiveness) For a transitive prefence relation, PuLEX ,
is responsive to Pulex if,

8T ✓ S, s 2 S\T and s0 2 S,

sPulex s0 and sRulex u implies T [ {s}PuLEX T [ {s0 }.

THEOREM 2. For µ 2 U , µk (u)PuLEX µk+1 (u) if and only if u is worse off

at µk+1 than at µk .

P roof of T heorem 2
Part 1 : Given the conditions from theorem 2, prove that µk (u)PuLEX µk+1 (u)
if no university is worse off at µk+1 than at µk .
Assume for each u 2 U is not worse off at µk+1 than at µk . By definition
of worse, we have the following two cases.
case 1[s0 6= ;] :
1
there exists s0 2 µk+1 (u) such that,
1
f or each s 2 µk (u), sPulex s0 .

20
 Let s be the least preferred student at µk (u), i.e. µ(s) = u, and |µk (u)| =
qu .
By the definition of algorithm, and the definition of worse.
) µk (u)\{s} = µk+1 \{s0 }
Let T be the complement of both sets, i.e.
) T = µk (u)\{s} = µk+1 \{s0 }
) T [ {s} = µk (u) and T [ {s0 } = µk+1 (u)
* sPulex s0 and s0 6= ;
) T [ {s}PuLEX T [ {s0 }
, µk (u)PuLEX µk+1 (u)
Hence, µk (u)PuLEX µk+1 (u) if no university is worse off at µk+1 than at µk .
case 2[s0 = ;] :
µk+1 (u) ( µk (u)
) µk (u) = µk+1 (u) [ {s}, since s is the least preferred student at µk (u).
By the design of the algorithm,
) sRulex ;,
By the definition of responsiveness,
) µk+1 (u) [ {s}PuLEX µk+1 (u) [ ;
)µk (u)PuLEX µk+1 (u)
Again, µk (u)PuLEX µk+1 (u) if no university is worse off at µk+1 than at µk .

Part 2 : Given the conditions from theorem 2, prove that u 2 U is worse

off at µk+1 than at µk if µk (u)PuLEX µk+1 (u).
Firstly, introduce the following notations:
s0 := least preferred student at µk+1 (u),
s := least preferred student at µk (u).
Assume µk (u)PuLEX µk+1 (u).
By the design of algorithm, let T = µk (u)\{s} = µk+1 (u).
) T [ {s} = µk (u) and T [ {s0 } = µk+1 (u)
) T [ {s}PuLEX T [ {s0 }

21
 ) sPulex s0
Suppose for each u 2 U is not worse off at at µk+1 than at µk ,
) s0 Rulex s
) contradicts the assumption that µk (u)PuLEX µk+1 (u).
Hence, u is worse off at µk+1 than at µk if µk (u)PuLEX µk+1 (u).
) µk (u)PuLEX µk+1 (u) if and only if u 2 U is worse off at µk+1 than at µk .
⇤

We have now established the first two theorems, and we summarise all
three equivalence theorems in next section.

6 EQUIVALENCE
The theorem 1 and theorem 2 have presented that there exist equivalence re-
lationships among stability, worse and responsiveness. For example, theorem
1 states that µk is stable and µk+1 is not stable is equivalent to saying u 2 U
is worse off at µk+1 than at µk , and theorem 2 implies u is worse off at µk+1
than at µk is equivalent to µk (u)PuLEX µk+1 (u). We can thus infer that there
is an equivalence relationship between stability and responsiveness.

THEOREM 3. Given µ0 , µ1 , . . . , µk is stable for k 0, for each u 2 U ,

µk+1 is not stable if and only if µk (u)PuLEX µk+1 (u).

P roof of T heorem 3
For u 2 U , and based on the compromise algorithm, the following two
theorems have been proved. From theorem 1, given µk is stable for k 0,
µk+1 is stable if and only if no university is worse off at µk+1 than at µk .
Equivalently, given µk is stable for k 0, µk+1 is not stable if and only if for
some u is worse off at µk+1 than at µk . From theorem 2, µk (u)PuLEX µk+1 (u)

22
if and only if for some u is worse off at µk+1 than at µk . Therefore, by the
above two theorems, given µk is stable for k 0, µk+1 is not stable if and
only if µk (u)PuLEX µk+1 (u). ⇤

The above theorems (Theorem 1, Theorem 2, and Theorem 3) are termed

equivalence theorems, and the equivalence theorems offer us alternative ways
to check the stability of the compromise algorithm. In some circumstances,
it is more straightforward to verify if some university is worse off or has re-
sponsive preferences at two adjacent matchings. For example, according to
theorem 1, the compromise algorithm can be interpreted as follows. The com-
promise algorithm runs the SP-DA algorithm first, and produces a student
optimal stable matching. And then we check if each university such that,
|µk (u)| = qu , wants to reject its least preferred applicant in such a way that,
rejection will not make it worse off at µk+1 than at µk . The algorithm thus
can be illustrated as a series of iterative SP-DA processes with a terminating
point such that, not all universities want to reject their least preferred ap-
plicants. We leave to the reader to verify that, by equivalence theorems, we
can terminate the compromise algorithm at µk+1 (u), if µk (u)PuLEX µk+1 (u).

7 STRATEGY-PROOF
7.1 A General Case
A mechanism is strategy-proof if reporting the true preference is a dominant
strategy for everyone (Roth 1985). In consequence of the compromise al-
gorithm, university can benefit from misrepresenting its preferences, but it
is less obvious whether the matching M(PS , PUlex , q) is immune to students’
manipulation. We can test whether the property of strategy-proof is tenable
for compromise algorithm by following example.

23
 EXAMPLE 2. This example is the same as EXAMPLE 1, except that s7
untruthfully reports her preference orderings, and the new preference rela-
tions for s7 is u2 P˜s7 u3 P˜s7 s7 P˜s7 u1 . And students and universities’ preference
relations are as follows:

Pulex
1
Pulex
2
Pulex
3
s7 s1 s6
Ps 1 Ps2 Ps 3 Ps4 Ps 5 Ps6 P˜s7 Pu1 Pu2 Pu3 s5 s3 s2
u1 u1 u1 u1 u2 u2 u2 f2 f1 f3 s1 s2 s5
u3 u2 u2 u3 u3 u1 u3 f1 f2 f1 s6 s5 s7
u2 u3 u3 u2 u1 u3 f3 f3 f2 s3 s4
(a) Students’ Preferences on U (b) Universities’ s2 s6
Preferences on F
(c) Universities’ Lexico-
graphic Preferences on
S

Table 2: PREFERENCES FOR EXAMPLE 2

Step 1: Student s1 , s2 , s3 and s4 propose to university u1 , student s5 , s6 ,

s7 propose to university u2 , and the SP-DA algorithm assigns each student
with at most one university. Therefore Step 1 generates student optimal
stable matching µ0 (µk , and k = 0).7
!
u1 u 2 u3
µ0 =
s1 , s3 , s6 s2 , s5 s7
Step 2: Each university such that, |µk (u)| = qu , rejects its least pre-
ferred applicant at µ0 . For example, u1 rejects s3 and u2 rejects s5 , since
s1 Pulex
1
s6 Pulex
1
s3 and s2 Pulex
2
s5 , u3 rejects s7 given qu3 = 1.
Step 3: Each rejected student applies to her next preferred university, and
the application procedure follows the SP-DA algorithm, and this generates
the following matching µ1 .
7
Student s4 is assigned to the no university option, i.e. µ0 (s4 ) = s4 .

24
 !
u1 u 2 u3
µ1 =
s1 , s6 s2 , s3 s5
Step 4: The matching µ1 is unstable, since u1 is worse off at µ1 than µ0
according to equivalence theorems, i.e. s3 Pulex1
u1 . The algorithm terminates
and we select µ1 such that, M(PS (P˜s7 ), PUlex , q) = µ0 .

The matching results of true and truthless preferences of s7 are presented

in Figure 3. From the results, s7 is able to manipulate results of compro-
mise algorithm by reporting the untruthful preferences. For example, at
M(PS , PUlex , q), s3 , s5 and s7 are assigned to u2 , u3 and u1 , respectively.
while, at M(PS (P˜s7 ), PUlex , q), s3 , s5 and s7 move to u1 , u2 and u3 , respec-
tively. And other students’ matchings are unchanged. From Table 3(b), s3 ,
s5 and s7 are assigned to more preferred choices.

s1 s2 s3 s4 s5 s6 s7
M(PS , PUlex , q) u1 u2 u2 s4 u3 u1 u1
M(PS (P˜s7 ), PUlex , q) u1 u2 u1 s4 u2 u1 u3
(a) Matching Results for Students in EXAMPLE 1 and EXAM-
PLE 2

Ps3 Ps 5 Ps7
u1 u2 u2
u2 u3 u3
u3 u1 u1
(b) Preferences for
Some Students on U

Table 3: MATCHING RESULTS FOR STRATEGY-PROOFNESS

25
7.2 A Special Case
If we restrict the number of fields to 1, that is, the number of fields set
F only contains one element. Based on the structure of the universities’
preferences, all universities have the same preference orderings over students
S. In this case, the student optimal stable matching is equivalent to the
university optimal stable matching, and we can verify this matching is the
unique strategy-proof and stable matching over the preference domain. Al-
calde and Barberà (1994) formally proved this result, and the result states
that if the preferences of one side of agents satisfy the top dominance con-
dition, and the preferences of the other side of agents are unrestricted, then
the matching rule generates a unique stable and strategy-proof matching on
0
admissible preference profiles. For any pair of preferences Pulex , Pulex for uni-
versity u 2 U , and any two students s and s0 that are preferred to u, if s
is the most preferred student at Pulex , and s0 is the most preferred student
0
at Pulex , then this type of preferences orderings is top dominance. When we
limit the number of fields to 1, each university have only one possible order
of preferences, and this order satisfies top dominance, meanwhile, the prefer-
ences of the students are unrestricted. No one can manipulate the matching
results by misrepresenting their preferences when |F | = 1. More generally,
the number of fields equals to total number of possible relations.

PROPOSITION 2. The compromise algorithm is generally not strategy-

proof, however, there is one unique stable and strategy-proof matching if
|F | = 1.

8 RESPECTING IMPROVEMENTS
An algorithm is respecting improvements if no student ever receives a less
preferred assignment on account of she performs better in grades. Balinski
and Sönmez (1999) show that the SP-DA algorithm is the unique stable

26
admissions mechanism respects improvements. It is indispensable to check
the respecting improvements of the compromise algorithm, and we can test if
the compromise algorithm respects improvements by the following example.
EXAMPLE 3. This example is the same as EXAMPLE 1, except that s5
deliberately scores worse in f3 , and gfs˜5 = ( gfs15 gfs25 gfs˜5 ) = ( 85 93 87 ).
3
And students and universities’ preference relations are as follows:8

Pulex
1
Pulex
2
Pulex
3
s7 s1 s6
Ps 1 Ps2 Ps 3 Ps4 Ps 5 Ps6 Ps 7 Pu1 Pu2 Pu3 s5 s3 s2
u1 u1 u1 u1 u2 u2 u2 f2 f1 f3 s1 s2 s7
u3 u2 u2 u3 u3 u1 u3 f1 f2 f1 s6 s5 s5
u2 u3 u3 u2 u1 u3 u1 f3 f3 f2 s3 s4
(a) Students’ Preferences on U (b) Universities’ s2 s6
Preferences on F
(c) Universities’ Lexico-
graphic Preferences on
S

Table 4: PREFERENCES FOR EXAMPLE 3

Step 1: Student s1 , s2 , s3 and s4 propose to university u1 , student s5 , s6 ,

s7 propose to university u2 , and the SP-DA algorithm assigns each student
with at most one university. As a result, Step 1 generates student optimal
stable matching µ0 (µk , and k = 0).9
!
u1 u 2 u3
µ0 =
s1 , s3 , s6 s2 , s5 s7
Step 2: Each university u 2 U such that, |µk (u)| = qu , rejects its least
preferred applicant at µ0 . For example, u1 rejects s3 and u2 rejects s5 , since
8
Table 1.c displays the preferences orderings only for acceptable students, i.e. omitting
those students who acquire insufficient grades in comparison with thresholds for each
university.
9
Student s4 is assigned to the no university option, i.e. µ0 (s4 ) = s4 .

27
s1 Pulex
1
s6 Pulex
1
s3 and s2 Pulex
2
s5 , u3 rejects s7 given qu3 = 1.
Step 3: Each rejected student applies to her next preferred university, and
the application procedure obeys the SP-DA algorithm, and this produces the
following matching µ1 .
!
u1 u 2 u3
µ1 =
s1 , s6 , s7 s2 , s3 s5
Step 4: The matching µ1 is unstable, since u3 is worse off at µ1 than µ0
according to equivalence theorems, i.e. s7 Pulex
1
s5 . The algorithm terminates
and we select µ1 such that, M(PS (gf˜ ), PU , q) = µ0 .
s5 lex

PROPOSITION 3. The compromise algorithm does not respect improve-

ments.
The matching results of respecting improvements are presented in Table
5. From the results, s5 is assigned to her most preferred choice because of
a lower grade in f3 , therefore, the compromise algorithm does not respect
improvements. As a consequence of worse performance of s5 , students s3 ,
s5 and s7 are assigned to more preferred choices than before, e.g. s3 moves
from u2 to u1 , s5 moves from u3 to u2 and s7 prefers new assigned u3 to
u1 (Table 5(b)). The proposition 3 is consisting with the results from the
student placement problem (Balinski and Sönmez 1999), that is, the SP-DA
algorithm is the unique stable admissions mechanism respects improvements.

28
 s1 s2 s3 s4 s5 s6 s7
M(PS , PUlex , q) u1 u2 u2 u0 u3 u1 u1
M(PS (gfs˜5 ), PUlex , q) u1 u2 u1 u0 u2 u1 u3
(a) Matching Results for Students in EXAMPLE 1 and EXAM-
PLE 3

Ps3 Ps 5 Ps7
u1 u2 u2
u2 u3 u3
u3 u1 u1
(b) Preferences for
Some Students on U

Table 5: MATCHING RESULTS FOR RESPECTING IMPROVEMENTS

9 CONCLUSION
The Swedish PhD program is a vital component of national higher education
system, meanwhile, it supports students with salaries. Even the compromise
algorithm is neither strategy-proof (generally) nor respecting improvements,
it produces stable matching that compromise the students and universities’
preferences. The most important desideratum for a matching is stability, and
the compromise algorithm does lead to a stable matching M(PS , PUlex , q).
What makes this stable matching differ from other matchings? We can anal-
yse the matching results produced from EXAMPLE 1 and compare with
typical matchings generated from the Gale-Shapley deferred acceptance al-
gorithm, i.e. the student optimal stable matching and the university optimal
stable matching.

29
 µs = µ0 M(PS , PUlex , q) = µ1 µu
u1 s1 , s3 , s6 s1 , s6 , s7 s1 , s5 , s7
u2 s2 , s5 s2 , s3 s2 , s3
u3 s7 s5 s6
(a) Matching Results of EXAMPLE 1

µs / M(PS , PUlex , q) / µu
(b) A Path of Stability

Figure 1: STABILITY OF COMPROMISE ALGORITHM

We use µs to denote the student optimal stable matching, which is equiv-

alent to µ0 at the compromised algorithm. And µu signifies the university
optimal stable matching, and compromise algorithm generates stable match-
ing M(PS , PUlex , q), i.e. µ1 in EXAMPLE 1. The matching results can be
seen from Figure 1, each university is assigned with different students at
different matchings, and compromise algorithm produces a stable matching
that follows the direction from the SP-DA algorithm to UP-DA algorithm
(Figure 1(b)). Particularly, the SP-DA algorithm creates the stable match-
ing µs that is optimal for students, and the UP-DA algorithm produces the
stable matching µu that is optimal for universities, whereas the compromise
algorithm generates a compromise stable matching that in between µs and
µu .
We should notice that there is one special case of the strategy-proofness.
If the number of fields |F | = 1, then compromise algorithm satisfies the top
dominance condition (Alcalde and Barberà 1994), and it creates a unique
stable and strategy-proof matching that is equivalent to the student optimal
stable matching and the university optimal stable matching. Alternatively,
if |F | = 1, then µs = M(PS , PUlex , q) = µu .
The equivalence theorems are major findings from the compromise algo-
rithm, and they endow the algorithm with alternate options. Essentially, we

30
can determine whether to terminate the algorithm by checking its stability,
worse or responsiveness.
This thresholds factor generalises the student placement problem, it al-
lows each university to rank the no student option prior to some students
at its preference orderings. The related issues of thresholds factor can be
further studied, and one possible implication is students will not apply to
those universities that she is not qualified, since she will be rejected for sure.
Specifically, suppose each student has the symmetric information about the
thresholds of each university, and she knows whether she is qualified for the
basic requirements for each university before the application, and she will
only apply to universities that she is both qualified and desired. Two ques-
tions arise from the above problem: What will change for the compromise
algorithm under the above case? What if students access to asymmetric
information set?
Lastly, this paper tries to design a mechanism that captures attributes
of the real-world Swedish PhD admissions problem, and it aims to offer a
solution such that, both students and universities are “satisfied” from the
matching produced by the algorithm.

31
References
[1] Abdulkadiroğlu, A. and Sönmez, T. School choice: A mechanism design
approach. American Economic Review, 93(3):729–747, 2003.

[2] Alcalde, J. and Barberà, S. Top dominance and the possibility of

strategy-proof stable solutions to matching problems. Economic Theory,
4:417–435, 1994.

[3] Balinski, M. and Sönmez, T. A tale of two mechanisms: Student place-

ment. Journal of Economic Theory, 84:73–94, 1999.

[4] Chen, Y. and Sönmez, T. School choice: an experimental study. Journal

of Economic Theory, 127:202–231, 2006.

[5] Gale, D. and Shapley, L. S. College admissions and the stability of

marriage. American Mathematical Monthly, 69(1):9–15, January 1962.

[6] Gudmundsson, J. Compromises and rewards: Stable and non-

manipulable probabilistic pairing. Working Papers, Department of Eco-
nomics, Lund University, 2015:32, October 2015.

[7] Roth, A. E. The evolution of the labor market for medical interns and
residents: A case study in game theory. Journal of Political Economy,
92:991–1016, 1984.

[8] Roth, A. E. The college admissions problem is not equivalent to the

marriage problem. Journal of Economic Theory, 36:277–288, 1985a.

[9] Roth, A. E. A natural experiment in the organization of entry-level

labor markets: Regional markets for new physicians and surgeons in the
united kingdom. The American Economic Review, 81(3):415–440, 1991.

[10] Roth, A. E. Repugnance as a constraint on markets. Journal of Eco-

nomic Perspectives, 21(3):37–58, 2007.

32
[11] Roth, A. E., Abdulkadiroğlu A. and Pathak, P. A. The new york city
high school match. American Economic Review, 95(2):364–367, 2005.

[12] Roth, A. E., Sönmez T. and Ünver, M. U. Kidney exchange. Quarterly

Journal of Economics, 119(2):457–488, May 2004.

[13] Shapley, L. S. and Scarf, H. On cores and indivisibility. Journal of

Mathematical Economics, 1:23–37, 1974.

33