Matching Markets with Two-Sided Preferences

Indranil Chakraborty

National University of Singapore

1 / 33
Babylon, 5th Century BC

Figure: Painting by Edwin Long, 1875

2 / 33
Babylon Auction

Key takeaways
It solved an allocation/matching problem through sequential
auctions
It did not put a financial burden (subsidy) on the society
The market was one-sided – the women’s preferences did not
matter

A better design would

remove money from transactions
consider women’s preference, too

3 / 33
The Different Matching Approach

More in line with what is called a direct mechanism

Agents submit type and preferences
System matches based on type and preferences expressed
There are no side transfers

Example of Direct Mechanism: VCG mechanism

Submit your true values for objects
An algorithm then decides which and how many objects you
receive
Algorithm also decides your payment
The mechanism is strategy-proof, has other nice properties

4 / 33
Marriage Market

Marriage Market: (M, W , (≻m )m∈M , (≻w )w ∈W )

Set of men M and set of women W

Man m ∈ M has a strict preference ordering (no

indifference) over women and remaining single ≻m over
W ∪ {m} (m means remaining single)

Woman w ∈ W has a strict preference ordering ≻w over men

and remaining single M ∪ {w } (w means remaining single)

Not every partner is acceptable: some men/women may be

worse than remaining single!

5 / 33
Marriage Market

Marriage Market: (M, W , (≻m )m∈M , (≻w )w ∈W )

M: set of men, W : set of women
Example 1: M = {m1 , m2 } and W = {w1 , w2 }

Example 2: M = {m1 , m2 , m3 } and W = {w1 , w2 }

6 / 33
Marriage Market

Marriage Market: (M, W , (≻m )m∈M , (≻w )w ∈W )

Man m ∈ M has a strict preference ordering (no
indifference) over women and remaining single ≻m over
W ∪ {m} (m means remaining single)

Example 1:
mi ’s preferences defined over W ∪ {mi } = {w1 , w2 , mi }
w1 ≻m1 w2 ≻m1 m1 ; w2 ≻m2 w1 ≻m2 m2

Example 2:
mi ’s preferences defined over W ∪ {mi } = {w1 , w2 , mi }
w1 ≻m1 w2 ≻m1 m1 w1 ≻m2 w2 ≻m2 m2 w2 ≻m3 w1 ≻m3 m3

7 / 33
Marriage Market

Marriage Market: (M, W , (≻m )m∈M , (≻w )w ∈W )

Woman w ∈ W has a strict preference ordering ≻w over men

and remaining single M ∪ {w } (w means remaining single)

Example 1:
wi ’s preferences defined over M ∪ {wi } = {m1 , m2 , wi }
m1 ≻w1 m2 ≻w1 w1 , m1 ≻w2 w2 ≻w2 m2

Example 2:
wi ’s preferences defined over M ∪ {wi } = {m1 , m2 , m3 , wi }
m1 ≻w1 m2 ≻w1 m3 ≻w1 w1 , m2 ≻w2 m1 ≻w2 m3 ≻w2 w2

8 / 33
Matching
In a matching an individual is either matched with another
individual from opposite sex or remains single.

Definition. A matching µ is a one-to-one function from W ∪ M

to W ∪ M such that for all m ∈ M, w ∈ W , we have
1 µ (m) = w if and only if µ (w ) = m,
2 µ (w ) ∈ W implies µ (w ) = w , and
3 µ (m) ∈ M implies µ (m) = m.

E.g., W = {w1 , w2 }, M = {m1 , m2 }

µ(w1 ) = m2 , µ(w2 ) = m1
µ(m1 ) = w2 , µ(m2 ) = w1
Not a match:
µ(w1 ) = m2 , µ(w2 ) = w2
µ(m1 ) = w2 , µ(m2 ) = w1

9 / 33
Application: Partner matching websites

To make a successful partner matching application

you need to place clients on the map of characteristics
you need to figure clients’ preferences over the map of
potential partners
the better and more successful the matches the more the
traffic to your site
“better” translates to pareto optimality
“successful” translates to stability

10 / 33
Stability

Definition. A matching µ is Pareto efficient if there is no other

matching ν such that ν(i) ⪰i µ(i) for all i ∈ M ∪ W and
ν(i) ≻i µ(i) for some i ∈ M ∪ W .

Example. M = {m1 , m2 } and W = {w1 , w2 }

w1 ≻m1 w2 ≻m1 m1 ; w2 ≻m2 w1 ≻m2 m2
m1 ≻w1 m2 ≻w1 w1 , m1 ≻w2 w2 ≻w2 m2

µ(w1 ) = w1 , µ(w2 ) = w2 µ(m1 ) = m1 , µ(m2 ) = m2

ν(w1 ) = m1 , ν(w2 ) = w2 ν(m1 ) = w1 , ν(m2 ) = m2

11 / 33
Stability

Definition. A matching µ is blocked by or unacceptable to an

individual if i ≻i µ(i) for i ∈ M ∪ W .
Definition. A matching is individually rational if it is not blocked
by an individual, i.e., acceptable to all individuals.

Example. M = {m1 , m2 } and W = {w1 , w2 }

w1 ≻m1 w2 ≻m1 m1 ; w2 ≻m2 w1 ≻m2 m2
m1 ≻w1 m2 ≻w1 w1 , m1 ≻w2 w2 ≻w2 m2
Blocked/Unacceptable match:
µ(w1 ) = m1 , µ(w2 ) = m2 , µ(m1 ) = w1 , µ(m2 ) = w2
Individually Rational/Acceptable match:
µ(w1 ) = m1 , µ(w2 ) = w2 µ(m1 ) = w1 , µ(m2 ) = m2

12 / 33
Stability
Definition. A matching µ is blocked by a pair (m, w ) ∈ M × W if
they both prefer each other to their partners under µ, i.e.,
w ≻m µ(m) & m ≻w µ(w )

Example. M = {m1 , m2 } and W = {w1 , w2 }

w1 ≻m1 w2 ≻m1 m1 ; w2 ≻m2 w1 ≻m2 m2
m1 ≻w1 m2 ≻w1 w1 , m1 ≻w2 w2 ≻w2 m2
A match blocked by a pair:
µ(w1 ) = m2 , µ(w2 ) = m1 , µ(m1 ) = w2 , µ(m2 ) = w1
⋄{m1 , w1 } block this match
A match that is not blocked pairwise:
µ(w1 ) = m1 , µ(w2 ) = m2 µ(m1 ) = w1 , µ(m2 ) = w2
⋄ This is individually blocked by w2
A match not blocked pairwise or individually:
µ(w1 ) = m1 , µ(w2 ) = w2 µ(m1 ) = w1 , µ(m2 ) = m2

13 / 33
Stability

Definition. A matching is stable if it is not blocked by any

individual or a pair.

Theorem. Stability implies Pareto Efficiency

14 / 33
Proof by Contradiction

Consider a match µ that is stable,

i.e., µ is (i) not blocked individually (ii) not blocked by a pair

Now suppose that µ is not Pareto Efficient.

This means: There is another match ν that makes nobody worse
off but makes some individual better off.
W.l.o.g. suppose that m1 is strictly better off in ν.
(i) µ(m1 ) = wi (ii) ν(m1 ) = wj

Since wj ’s partner has changed she is strictly better off under ν

(since preferences are strict)
m1 is also better off under ν

This means that (m1 , wj ) could pairwise block µ

– A contradition to the fact that µ is stable (not pairwise blocked)
So µ must be Pareto Efficient.

15 / 33
1st Example of Stability

Two men: M = {m1 , m2 }

Two women: W = {w1 , w2 }
Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m1 ≻w1 m2 ≻w1 w1

w2 ≻m2 w1 ≻m2 m2 ; m2 ≻w2 m1 ≻w2 w2

Stable matching: (m1 , w1 ) and (m2 , w2 ) (or formally we can

write µ(m1 ) = w1 , µ(m2 ) = w2 )
Unstable matching: (m1 , w2 ) and (m2 , w1 )

16 / 33
2nd Example of Stability

Two men: M = {m1 , m2 }

Two women: W = {w1 , w2 }
Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m2 ≻w1 m1 ≻w1 w1

w2 ≻m2 w1 ≻m2 m2 ; m1 ≻w2 m2 ≻w2 w2

Two stable matchings: {(m1 , w1 ) and (m2 , w2 )} and

{(m1 , w2 ) and (m2 , w1 )}
First one is better for men and second one is better for women

17 / 33
Deferred Acceptance (DA)

Also known as Gale-Shapley Algorithm

Two versions: (i) man-proposing and (ii) woman-proposing
Man-proposing version:
Each man simultaneously proposes to the most preferred
woman
Woman receiving proposal keeps the most preferred man if
acceptable, and rejects the rest
The rejected men propose to the most acceptable woman
next in the list
Continue until there is no more rejections and offers.
Woman-proposing version: the same but women are making
proposals

18 / 33
Important Notes

⋄ A man proposes to the most preferred woman since he can

always propose to the less preferred woman if the more preferred
woman rejects him

⋄ A woman does not reject the most preferred man at a time since
rejecting most preferred man definitely makes her worse off

⋄ A man proposes to nobody if the remaining women are worse

than remaining unmatched

⋄ A woman rejects all proposals if they are all worse than

remaining unmatched

19 / 33
1st example for man-proposing DA

Two men: m1 and m2 ; Two women: w1 and w2

Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m1 ≻w1 m2 ≻w1 w1

w1 ≻m2 w2 ≻m2 m2 ; m2 ≻w2 m1 ≻w2 w2

In man-proposing version, both men propose to w1 .

Woman w1 rejects m2 . Then man m2 propose to w2 .
Woman w2 does not reject m2 . Process ends.
What is the woman-proposing DA?

20 / 33
2nd example for man-proposing DA

Three men: m1 , m2 and m3 ; Two women: w1 and w2

Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m1 ≻w1 m2 ≻w1 m3 ≻w1 w1

w1 ≻m2 w2 ≻m2 m2 ; m2 ≻w2 m1 ≻w2 m3 ≻w2 w2
w2 ≻m3 w1 ≻m3 m3

In man-proposing version, both men m1 and m2 propose to

w1 , and m3 purpose to w2 .
Woman w1 rejects m2 . Then man m2 proposes to w2 .
Woman w2 rejects m3 . Process ends.

21 / 33
Stability of DA

Theorem. Outcomes of DA are stable.

22 / 33
Proof of Stability

Consider the man-proposing DA. (Similar proof for

woman-proposing.)
Stability requires
(1) individual rationality and (2) pair-wise stability
Clearly, it is individually rational (why?)

23 / 33
Individual Rationality for DA

No man ever proposes to an unacceptable woman

No woman accepts an unacceptable man

24 / 33
Pair-wise stability for DA

Algorithm must end in finite number of rounds.

Again, Proof by contradiction! So suppose the outcome is
pairwise blocked (by (m1 , w2 ))
Suppose m1 and w1 are matched but m1 prefers w2 ,
and w2 prefers m1 to her current match, say, m2
At some round, m1 has proposed to w2 but got rejected.
At that round, w2 preferred her then tentative match to m1 .
After that round, w2 can be no worse off. (Only woman can
dissolve tentative match! And she will not dissolve a tentative
match except to be better off.)
This implies w2 prefers her final match to m.
– Contradiction to (m1 , w2 ) being a blocking pair.
There cannot be any blocking pair.

25 / 33
Optimality of DA

A stable matching is man-optimal if every man prefers his current

partner to any partner he could possibly have in a stable matching.

Theorem. Outcome of man-proposing DA is man-optimal.

This matching is also woman-pessimal (each woman gets worst

outcome in any stable matching).

26 / 33
(Man) Optimality of MP-DAA

Two men: M = {m1 , m2 }

Two women: W = {w1 , w2 }
Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m2 ≻w1 m1 ≻w1 w1

w2 ≻m2 w1 ≻m2 m2 ; m1 ≻w2 m2 ≻w2 w2

What is the outcome of the MP-DAA

What is the outcome of the WP-DAA

27 / 33
2nd Example of Stability

Two men: M = {m1 , m2 }

Two women: W = {w1 , w2 }
Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m2 ≻w1 m1 ≻w1 w1

w2 ≻m2 w1 ≻m2 m2 ; m1 ≻w2 m2 ≻w2 w2

Two stable matchings: {(m1 , w1 ) and (m2 , w2 )} and

{(m1 , w2 ) and (m2 , w1 )}
First one is better for men and second one is better for women

28 / 33
Strategy-proofness of DA

So far we took preferences as fixed. Now we will assume that any

preference is possible.

A matching mechanism is a mechanism that maps reported

preferences into an assignment.
A mechanism is strategy-proof if it is always optimal to be
truthful for each participant.

Theorem. Man-proposing DA is strategy-proof for man.

29 / 33
Strategy-proof for women? No

Two men: m1 and m2 ; Two women: w1 and w2

Preferences:

w1 ≻m1 w2 ≻m1 m1 ; m2 ≻w1 m1 ≻w1 w1

w2 ≻m2 w1 ≻m2 m2 ; m1 ≻w2 m2 ≻w2 w2

Under man-proposing DA, we have w1 = µ (m1 ), and

w2 = µ (m2 )
However, if women w1 reports m1 is not acceptable
(m2 ≻w1 w1 ≻w1 m1 ), then

w1 = µ (m2 ) and w2 = µ (m1 )

which is better for w1 ! (also better for w2 )

30 / 33
Strategic Behavior

Theorem. There is no matching mechanism that is strategy-proof

(for all individuals) and that always generates stable outcomes
given reported preferences.
Both versions of DA lead to stable matches
Neither version is strategy-proof for all participants!

31 / 33
Theorem

Theorem. The set of men and women who are unmatched is the
same in all stable matchings.
Those who are unmatched under DA will also be unmatched
in other mechanism that leads to stable matchings.

32 / 33
Proof

M̂ and Ŵ be the sets of matched men and women in the outcome

of man-proposing DA algorithm
The match is man-optimal and woman-pessimal

M ′ and W ′ be the sets of matched men and women in another

stable match

Every m in M ′ is also in M̂ by man-optimality

⇒ M ′ ⊂ M̂ ⇒ |M ′ | ≤ |M̂|

Every w in Ŵ is also in W ′ by woman-pessimality

⇒ Ŵ ⊂ W ′ ⇒ |Ŵ | ≤ |W ′ |

By definition of matching |M̂| = |Ŵ | and |M ′ | = |W ′ |

⇒ |M̂| = |Ŵ | = |M ′ | = |W ′ |

⇒ M ′ = M̂ and Ŵ = W ′

33 / 33