Electricity and Magnetism II Griffiths Chapter 11 Radiation Clicker Questions 11.1.
11.1 School Matching. New Questions
description
Transcript of 11.1 School Matching. New Questions
![Page 1: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/1.jpg)
New questions raised by school choice
• How to do tie breaking?
• Tradeoffs between Pareto optimality, stability, strategy proofness—what are the ‘costs’ of each?
• Evaluating welfare from different points in time
• Restricted domains of preferences?
1
![Page 2: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/2.jpg)
Matching with indifferences
• When we were mostly using matching models to think about labor markets, strict preferences didn’t seem like too costly an assumption
– Strict preferences might be generic
• But that isn’t the case with school choice
– We already saw that one of the first NYC design decisions we faced in 2003 was how to randomize to break ties.
2
![Page 3: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/3.jpg)
3
New Theoretical Issues
• Erdil, Aytek and Haluk Ergin, What's the Matter with Tie-breaking? Improving Efficiency in School Choice , American Economic Review , 98(3), June 2008, 669-689
• Abdulkadiroglu, Atila , Parag A. Pathak , and Alvin E. Roth, " Strategy-proofness versus Efficiency in Matching with Indifferences: Redesigning the NYC High School Match ,'' American Economic Review, 99(5) December 2009, 1954-1978.
• Featherstone, Clayton and Muriel Niederle, “EX ANTE EFFICIENCY IN SCHOOL CHOICE MECHANISMS: AN EXPERIMENTAL INVESTIGATION http://www.nber.org/papers/w14618.pdf.
![Page 4: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/4.jpg)
Other new issues we won’t get to today…
• Pathak, Parag and Tayfun Sönmez “Leveling the Playing Field: Sincere and Strategic Players in the Boston Mechanism” , American Economic Review, 98(4), 1636-52, 2008
• Ergin, Haluk and Tayfun Sonmez, Games of School Choice under the Boston Mechanism ” , Journal of Public Economics , 90: 215-237, January 2006.
• Kesten, Onur On Two Kinds of Manipulation for School Choice Problems March, 2011, forthcoming in Economic Theory.
• Kesten, Onur, “School Choice with Consent,” Quarterly Journal of Economics 125(3), August, 2010; 1297-1348.
• Abdulkadiroglu, Atila, Yeon-Koo Che, and Yosuke Yasuda, " Resolving Conflicting Preferences in School Choice: The “Boston Mechanism” Reconsidered" American Economic Review, February, 2011, 101(1): 399–410.
• Kesten and Unver—random matchings--in progress... 4
![Page 5: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/5.jpg)
Matching with indifferences I: a finite set of students (individuals) with (strict)
preferences Pi over school places. S: a finite set of schools with responsive weak
preferences/priorities Rs over students (i.e. can include indifferences: Ps (≻s ) is the asymmetric part of Rs).
As before: q = (qs)sєS: a vector of quotas (qs ≥ 1, integer). A matching is a correspondence μ: I U S → S U I satisfying: (i) For all i є I : μ(i) є S U {i} (ii) For all s є S : |μ(s)| ≤ qs, and i ∈ m(s) implies μ(i) = s. We’ll mostly concentrate on student welfare and student
strategy, and regard RS as fixed. 5
![Page 6: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/6.jpg)
Matchings and student welfare A matching μ is individually rational if it matches every x ∈ I ∪ S
with agent(s) that is(are) acceptable for x.
A matching μ is blocked by (i, s) if sPiμ(i), and either *|μ(s)| < qs and i ≻s s] or [i ≻s i′ for some i′ ∈ μ(s)+. μ is stable if μ is individually rational and not blocked by any student-school pair (i, s).
A matching μ dominates matching if μ(i)Ri(i) for all i ∈ I, and μ(i)Pi(i) for some i ∈ I. (Weak Pareto domination for students.)
A stable matching μ is a student-optimal stable matching if it is not dominated by any other stable matching.
“A” not “the”: When school preferences aren’t strict, there won’t generally be a unique optimal stable match for each side, rather there will be a non-empty set of stable matches that are weakly Pareto optimal for agents on that side.
6
![Page 7: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/7.jpg)
Example: multiple optimal stable matchings
7
![Page 8: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/8.jpg)
Weak Pareto optimality generalizes…
• Proposition 1. If μ is a student-optimal stable matching, there is no individually rational matching n (stable or not) such that n(i)Piμ(i) for all i ∈ I.
• (terminology: a student optimal stable matching is weakly Pareto optimal because it can’t be strictly Pareto dominated, but the outcome of student proposing deferred acceptance algorithm might not be strongly Pareto optimal, i.e. might not be student optimal, because it can be weakly Pareto dominated)
8
![Page 9: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/9.jpg)
Tie breaking
• A tie-breaker is a bijection r:I→N, that breaks ties at school s by associating Rs with a strict preference relation Ps :
iPs j⇔[(i≻s j) or (i∼s j and r(i) < r(j))].
9
![Page 10: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/10.jpg)
10
• Step 0: arbitrarily break all ties in preferences • Step 1: Each student “proposes” to her first choice. Each
school tentatively assigns its seats to its proposers one at a time in their priority order. Any remaining proposers are rejected.
… • Step k: Each student who was rejected in the previous step
proposes to her next choice if one remains. Each school considers the students it has been holding together with its new proposers and tentatively assigns its seats to these students one at a time in priority order. Any remaining proposers are rejected.
• The algorithm terminates when no student proposal is rejected, and each student is assigned her final tentative assignment.
Basic Deferred Acceptance (Gale and Shapley 1962)
![Page 11: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/11.jpg)
Deferred acceptance algorithm with tie breaking: DAτ
• A single tie breaking rule uses the same tie-breaker rs = r at each school, while a multiple tie breaking rule may use a different tie breaker rs at each school s.
• For a particular set of tie breakers τ=(rs)s∈S, let the mechanism DAτ be the student-proposing deferred acceptance algorithm acting on the preferences (PI,PS), where Ps is obtained from Rs by breaking ties using rs, for each school s.
11
![Page 12: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/12.jpg)
Single and Multiple tie breaking
• The dominant strategy incentive compatibility of the student-proposing deferred acceptance mechanism for every student implies that DAτ is strategy-proof for any τ.
• But the outcome of DAτ may not be a student optimal stable matching.
– We already saw this is true even for single tie breaking.
12
![Page 13: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/13.jpg)
Single versus multiple tie breaking NYC Grade 8 applicants in 2006-07
(250 random draws: simulation standard errors in parentheses)
13
![Page 14: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/14.jpg)
Proposition: For any (PI,RS), any matching that can be produced by deferred acceptance with multiple tie breaking, but not by deferred acceptance with single tie breaking is not a student-optimal stable matching.
14
![Page 15: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/15.jpg)
Dominating stable matchings
• Lemma: Suppose μ is a stable matching, and ν is some matching (stable or not) that dominates μ. Then the same set of students are matched in both ν and μ
15
![Page 16: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/16.jpg)
Proof
• If there exists a student who is assigned under μ and unassigned under ν, then ν(i)=iPiμ(i), which implies that μ is not individually rational, a contradiction. So every i assigned under μ is also assigned under ν.
• Therefore |ν(S)|≥|μ(S)|. If |ν(S)|>|μ(S)| then there exists some s∈S and i∈I such that |ν(s)|>|μ(s)| and ν(i)=s≠μ(i). This implies there is a vacancy at s under μ and i is acceptable for s. Furthermore, sPiμ(i) since ν dominates μ. These together imply that μ is not stable, a contradiction. So |ν(S)|=|μ(S)|.
• Then the same set of students are matched in both ν and μ since |ν(S)|=|μ(S)| and every student assigned under μ is also assigned under ν. 16
![Page 17: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/17.jpg)
Stable Improvement Cycles (Erdil and Ergin, 08)
Fix a stable matching μ w.r.t. given preferences P and priorities R. Student i desires s if sPiμ(i). Let Bs = the set of highest Rs-priority students among those who
desire school s. Definition: A stable improvement cycle C consists of distinct
students i1, . . . , in = i0 (n ≥ 2) such that (i) μ(ik) є S (each student in the cycle is assigned to a school), (ii) ik desires μ(ik+1), and (iii) ik є Bμ(ik+1), for any ) k = 0, . . . , n − 1. Given a stable improvement cycle define a new matching μ’ by: m’(j) = μ(j) if j is not one of {i1, . . . , in} m’(j) = μ(ik+1) if j = ik Proposition: μ’ is stable and it (weakly) Pareto dominates μ.
17
![Page 18: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/18.jpg)
Improving on DAτ
• Theorem (Erdil and Ergin, 2008): Fix P and R, and let μ be a stable matching. If μ is Pareto dominated by another stable matching , then μ admits a stable improvement cycle.
• Algorithm for finding a student optimal matching: start with a stable matching. Find and implement a stable improvement cycle, as long as one exists.
18
![Page 19: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/19.jpg)
Outline of proof
Fix P and R. Suppose μ is a stable matching Pareto dominated by another stable matching n.
Simplifying assumption: Each school has one seat. 1. I’ := ,i є I |n(i)Piμ(i)- = ,i є I |n(i) ≠ μ(i)-. 2. All students in I’ are matched to a school at n. 3. S’ := n(I’)=μ(I’). Hence, I *S+ can be partitioned into two subsets I’ and
I\I’ *S’ and S \ S’+ such that • Those in I \ I’ *S \ S’+ have the same match under μ
and n. • The matches of those in I’ *S’+ have been “shuffled”
among themselves to obtain n from μ.
19
![Page 20: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/20.jpg)
4. For all s є S’: I’s := (i є I’|i desires s at μ, and no j є I’ desires s at μ
and j Ps i) is nonempty;. 5. Construct a directed graph on S’: • For each s є S’, arbitrarily choose and fix is є I’s. • is є Bs: i.e., is desires s at μ, and there is no j є I
who desires s at μ and j Ps i. (from stability of n) • For all s, t є S’, let t →s if t = μ(is). 6. The directed graph has a cycle of n ≥ 2 distinct
schools: s1 → s2 → · · → sn → s1
7. The students is1, is2, . . . , isn constitute a stable improvement cycle at μ
20
![Page 21: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/21.jpg)
How much room is there to improve on deferred acceptance?
• Are there costs to Pareto improvements in welfare?
21
![Page 22: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/22.jpg)
Strategy-proof mechanisms
A direct mechanism φ is a function that maps every (PI ,RS) to a matching.
For x ∈ I∪S, let φx(PI ;RS) denote the set of agents that are matched to x by φ.
A mechanism φ is dominant strategy incentive compatible (DSIC) for i ∈ I if for every (PI ,RS) and every P′i ,
φ i(PI ;RS)Ri φ i(P′i , P−i;RS).
A mechanism will be called strategy-proof if it is DSIC for all students.
22
![Page 23: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/23.jpg)
Pareto improvement and strategy proofness
Fix RS.
We say that a mechanism φ dominates ψ if
for all PI : φi(PI ;RS)Ri ψi(PI ;RS) for all i ∈ I, and
for some PI : φi(PI ;RS)Pi ψi(PI ;RS) for some i ∈ I.
Theorem (Abdulkadiroglu, Pathak, Roth): For any tie breaking rule τ, there is no mechanism that is strategy-proof and dominates DAτ.
23
![Page 24: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/24.jpg)
Proof
• Suppose that there exists a strategy-proof mechanism ϕ and tie-breaking rule r such that ϕ dominates DAτ. There exists a profile PI such that
ϕi(PI;RS)Ri DAτ(PI;RS) for all i∈I, and
ϕi(PI;RS)Pi DAτ(PI;RS) for some i∈I.
Let si=DAiτ(PI;RS) and s’i=ϕi(PI;RS) be i's assignment
under DAτ(PI;RS) and ϕ(PI;RS), respectively, where s’iPisi.
24
![Page 25: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/25.jpg)
…continued
• Consider profile PI′=(Pi′,P-i), where Pi′ ranks s’i as the only acceptable school. Since DAτ is strategy-proof, si=DAi
τ (PI;RS)RiDAiτ(PI′;RS), and
since DAiτ(PI′;RS) is either s’i or i, we conclude
that DAiτ(PI′;RS)=i. Then the Lemma implies
ϕi(PI′;RS)=i.
• Now let (PI′;RS ) be the actual preferences. In this case, i could state Pi and be matched to ϕi(PI;RS)=s’i, which under Pi′ she prefers to ϕ(PI′;RS )=i.
• So ϕ is not strategy-proof. 25
![Page 26: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/26.jpg)
Let’s look at some data
• We can’t tell what preferences would have been submitted with a different (non strategy-proof) mechanism, but we can ask, given the preferences that were submitted, how big an apparent welfare loss there might be due to not producing a student optimal stable matching.
26
![Page 27: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/27.jpg)
Inefficiency in the NYC match (cost of strategy-proofness)
27
![Page 28: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/28.jpg)
Cost of stability in NYC
28
![Page 29: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/29.jpg)
Comparison with Boston
29
![Page 30: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/30.jpg)
Open questions • (Equilibrium) misrepresentation in stable
improvement cycles? (Can potential gains be realized?)
– It appears there will be an incentive to raise popular schools in your preferences, since they become tradeable endowments…
• Restricted domains of preference?
– Manipulation will be easier on some domains than others, and potential welfare gains greater on some domains than others.
30
![Page 31: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/31.jpg)
CAN WE MAKE SCHOOL CHOICE MORE EFFICIENT? AN EXAMPLE
EDUARDO M. AZEVEDO AND JACOB D. LESHNO (2011)
31
![Page 32: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/32.jpg)
Consider the equilibrium of (any) SOSM in which everyone reports truthfully except the two bi who both (mis)report A>S>f (notice that S is popular and the bi’s have priority there…)
• The outcome of the DA-STB for this profile is:
– a: ½ S, ½ A
– z: f
– bi: ¼ A, ¾ S
• SOSM: stable improvement cycles would allow a to trade A for S with a bi
– a: S,
– z: f
– bi: ½ A, ½ S
• None of the students do better under this equilibrium, and some do strictly worse.
32
![Page 33: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/33.jpg)
AZEVEDO AND LESHNO
33
![Page 34: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/34.jpg)
Ex post versus ex ante evaluation?
• E.g. Boston mechanism in uncorrelated environment, where you don’t have to pay the cost for lack of strategy proofness…Featherstone and Niederle 2008
• Recall that DA is strategy-proof (DSIC) while the Boston mechanism is not.
• (The following slides are adapted from F&N’s)
34
![Page 35: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/35.jpg)
Example; correlated preferences (likely the general case…)
35
![Page 36: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/36.jpg)
Boston mechanism in the correlated environment—complex eq. strategies
36
![Page 37: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/37.jpg)
Uncorrelated preferences: (a conceptually illuminating simple environment)
37
Consider a student after he knows his own type, and
before he knows the types of the others. Then (because
the environment is uncorrelated) his type gives him no
information about the popularity of each school. So, under
the Boston mechanism, truthtelling is an equilibrium.
(Note that for some utilities this wouldn’t be true e.g. of the
school-proposing DA, even in this environment.)
• 2 schools, one for Art, one for Science, each with one seat
• 3 students, each iid a Scientist with p=1/2 and Artist with p=1/2. Artists prefer the art school, scientists the science school.
• The (single) tie breaking lottery is equiprobable over all orderings of the three students.
![Page 38: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/38.jpg)
Boston can stochastically dominate DA in an uncorrelated environment Example: 3 students, 2 schools each with one seat
38
![Page 39: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/39.jpg)
Things to note
• The uncorrelated environment let’s us look at Boston and DA in a way that we aren’t likely to see them in naturally occurring school choice.
• In this environment, there’s no incentive not to state preferences truthfully in the Boston mechanism, even though it isn’t a dominant strategy. (So on this restricted domain, there’s no corresponding benefit to compensate for the cost of strategyproofness.)
• Boston stochastically dominates DA, even though it doesn’t dominate it ex-post (ex post the two mechanisms just redistribute who is unassigned)
39
![Page 40: 11.1 School Matching. New Questions](https://reader030.fdocuments.net/reader030/viewer/2022020202/577cc0931a28aba711908c1c/html5/thumbnails/40.jpg)
Recap: New questions raised by school choice
• How to do tie breaking?
• Tradeoffs between Pareto optimality, stability, strategy proofness—what are the ‘costs’ of each?
• Evaluating welfare from different points in time
• Restricted domains of preferences?
40