michaelmas termpersonal.lse.ac.uk/nava/materials/ec319 handout.pdfec319: auction theory michaelmas...
TRANSCRIPT
EC319: Auction TheoryMichaelmas Term
The first part of the course consists of an introduction to the game theoretical analysis of auctions. It presentsstandard auction formats and strategic behavior in such environments. Auctions will be analyzed both inprivate and interdependent value environments. Fundamental topics such as the revenue equivalence theorem,the optimal auction design problem and the linkage principle will be discussed extensively. Throughout thecourse we will consider departures from the standard model allowing for heterogeneity amongst players,budget constraints and resale. Time permitting a model of bidding rings will be presented. The focus of thecourse is mainly theoretical, but when possible empirical and experimental evidence supporting the formalmodels will be discussed with references to relevant work in the field.
Lecturer: Francesco Nava, [email protected]
Offi ce Hours: Wed 1.30pm-3.00pm, Building 32L, Offi ce 3.20
Course Website: moodle.lse.ac.uk
Time and Location: Thu 9.30-11.00am, Building 32L, Room LG.18
Class Work: The work for each week consists of one or more exercises taken from the textbook and someadditional problems. Answers to exercises will be posted on the website throughout the term. You areexpected to make a reasonable attempt at the classwork in advance of the class. Two written assignmentsare required before class in weeks 6 and 11 of MT. These assignments will be of similar scope and diffi cultyas exam questions. You are expected to be diligent about meeting the deadlines for submission.
Exams: The final exam will determine your grade on the course entirely. The two parts of the course willbe weighted equally. For each part of the course, you will have to answer two short questions and one longproblem. The exam will focus entirely on topics covered in lectures and classes.
Weekly Course Program
1. Review: Order Statistics & Game Theory [KA, KC, KF]
2. Independent Private Value Auctions [K1, K2]
3. Revenue Comparisons and Reserve Prices [K2]
4. Revenue Equivalence Principle [K3]
5. Extensions: Risk Aversion [K4]
6. Extensions: Budget Constraints [K4]
7. Mechanism Design [K5, 3]
8. Optimal Mechanisms [K5, 5]
9. Effi cient and Balanced Budget Mechanisms [K5, 6]
10. Interdependent Values [K6-7, 4]
Auction Theory Nava
Readings
Most materials covered can be found in the main textbook. The slides used in class will be posted. Some ofthe papers below may be suggested to deepen some topics.
Main Textbook: [K] Auction Theory, Vijay Krishna, Academic Press 2010
Related Papers [Not Required]:[0] Athey & Levine, “Comparing Open and Sealed Bid Auctions: Theory and Evidence”, Mimeo 2004
[1]Ausubel, “A Generalized Vickery Auction”, Econometric Society Conference, 2000
[2] Dasgupta & Maskin, “Effi cient Auctions”, Quarterly Journal of Economics, 2000
[3] Krishna & Perry, “Effi cient Mechanism Design”, Mimeo Penn State University, 1998
[4] Milgrom & Weber, “A Theory of Auctions and Competitive Bidding”, Econometrica 50, 1982
[5] Myerson, “Optimal Auction Design”, Journal of Operation Research 6, 1982
[6] Myerson & Satterwaite, “Effi cient Mechanisms for Bilateral Trading”, JET 28, 1983
[7] Reny & Perry, “An Ex-Post Effi cient Auction”, Econometrica 70, 2002
[8] Vickery, “Counterspeculation, Auctions and Competitive Sealed Traders”, Journal of Finance 7, 1961
[9] Wilson, “A Bidding Model of Perfect Competition”, Review of Economic Studies 44, 1977
[10] Zheng, “High Bids and Broken Winners”, Journal of Economic Theory 100, 2001
2
EC319 - Auction TheoryMichaelmas Term
Weekly Course Assignments
Several weekly exercises are taken from Krishna’s textbook. Problems assigned are discussed during classes andlag the topics covered in the lectures by approximately a week. Answers to exercises will be posted on the websitethroughout the term. You are expected to make a reasonable attempt at the classwork in advance of the class. Twoproblem sets have to be handed in as written assignments at the end of weeks 5 and 10 of Michealmas Term. Theseassignments will be of similar scope and diffi culty as exam questions. You are expected to be meet the deadlinesfor submission.
[2MT] Order Statistics
1. Consider n independent draws from the uniform probability distribution function f(x) = 1/k on [0, k].
(a) Compute the cumulative distribution function and probability distribution function of the first and secondorder statistic.
(b) Compute the probability distribution function of the second order statistic conditional on the first orderstatistic being equal to y.
(c) Compute the expected value of the first and the second order statistic.
(d) Compute the expected value of the second order statistic conditional on the first order statistic beingequal to y.
2. Consider 3 independent draws from the exponential probability distribution function f(x) = e−x on [0,∞).
(a) Compute the cumulative distribution function and probability distribution function of the first and secondorder statistic.
(b) Compute the probability distribution function of the second order statistic conditional on the first orderstatistic being equal to y.
[3MT] IPV Auctions
1. K2.1
2. Consider a first price auction with 3 buyers with values distributed according to f(x) = e−x on [0,∞). Findthe unique symmetric equilibrium.
3. [Extra] Consider a Dutch auction in which the auctioneer: starts the auction at a price so high no buyer iswilling to buy; and then lowers the price gradually until a buyer signals his willingness to buy. When thatoccurs, the buyer is sold the object at the last price quoted by the auctioneer. Prove that in an IPV setup thisDutch auction has an equilibrium that is equivalent to the symmetric equilibrium of the first price auction.
[4MT] IPV Auctions
1. Consider a first price auction with 3 buyers with values distributed according to f(x) = e−x on [0,∞).Compute the seller’s expected revenues.
2. K2.2
3. K2.4
[5MT] Revenue and Reserve Prices
1. Consider 3 buyers with values distributed according to f(x) = 1/2 on [0, 2].
(a) Compute the distribution of winning prices in the first and the second price auction.
(b) Compute the mean of the two distributions in part (a). How do these means relate to each other and torevenue in each of the two auctions?
Auction Theory Nava
(c) Compute the variance of the two distributions in part (a). How do these variances relate to each other?Can you explain this in light of the discussion in lecture?
2. Consider a first price auction with 3 buyers with values distributed according to f(x) = e−x on [0,∞).
(a) Compute the seller’s optimal reserve price.
(b) What are the seller’s expected revenues with optimal reserve price in place?
[6MT] IPV Auctions and Revenue Equivalence
4. K3.1
5. K2.5
6. Hand In Problem Set I
[7MT] Revenue Equivalence and IPV Auctions Extensions
1. K3.2
2. K4.1
[8MT] IPV Auctions Extensions and Mechanism Design
1. K4.2
2. Consider three buyers participating in a second price auction. Buyers are financially constrained, and theirvalue-wealth pairs (Xi,Wi) are identically and independently distributed according to a uniform on [0, 1]2.
(a) Show explicitly that bidding according to B(x,w) = min {x,w} is weakly dominant.(b) Compute the probability of winning with value-wealth pair (x,w).
(c) Find an expression for the expected payment of a buyer with value-wealth pair (x,w).
(d) Compute expected revenues of the seller.
3. [Extra] K4.5
[9MT] Revenue Maximizing Mechanisms
1. K5.1
2. K5.2
[10MT] Revenue Maximizing and Effi cient Mechanisms
1. Prove explicitly that if the design problem is regular and symmetric then the optimal mechanism reduces toa second price auction with reserve price.
2. K5.3
3. Consider 3 potential buyers with independent private values for an object. The values of bidders a and b areidentically distributed according to a uniform distribution on the interval [−1, 1], so that Fa(x) = Fb(x) =(x+ 1)/2. The value of bidder c is instead uniformly distributed on the interval [0, 2], so that Fc(x) = x/2.
(a) Define the Vickrey-Clarke-Groves (VCG) mechanism and specialize it to the scenario described above.
(b) Define effi ciency, incentive compatibility, and individually rationality. Show that the VCG mechanismsatisfies these properties.
(c) Compute the revenue of the VCG mechanism.
[11MT] Effi cient and Balanced Budget Mechanisms
1. K5.4
2. [Extra] K5.5
3. Hand In Problem Set II
2
Hand In Problem Set 1 Francesco Nava
EC319: Auction Theory Michaelmas Term
Please give your answers to your class teacher before class in week 6MT. Thank you!
1. Consider the following variant of a 1st price auction in an IPV setup. Sealed bids are collected. The highestbidder pays his bid, but receives the object only if the outcome of a toss of a fair coin is heads. If the outcomeis tails, the seller keeps the object and the high bidders bid. Assume bidder symmetry. [30 Marks]
(a) Find the unique symmetric equilibrium bidding function. [10 Marks]
(b) Do buyers bid higher or lower than in a 1st price auction? [5 Marks]
(c) Find an expression for the seller’s expected revenue. [10 Marks]
(d) Show that the seller’s expected revenue is exactly half that of a standard 1st price auction. [5 Marks]
2. [Harder] In a 2nd price auction with entry fee, buyers have to pay a fee E in order to participate in the auction.Buyers first choose whether to participate and pay the fee and then choose how much to bid. Consider anIPV setup with four buyers with values drawn from a uniform distribution on the interval [0, 1]. Suppose thatonly the seller knows how many buyers pay the fee. [40 Marks]
(a) Find the unique symmetric equilibrium of this auction and write down the expected payment of a buyerwith value x. Hint: Pay the fee only if xG(x) ≥ E. Explain. [15 Marks]
(b) What entry fee maximizes the equilibrium expected revenues of the seller? [15 Marks]
(c) Compare the revenues to those of a standard 2nd price auction. [5 Marks]
(d) Compare the revenues to those of a 2nd price auction with revenue maximizing reserve prices. [5 Marks]
3. Consider three buyers with independent and private values for an object at auction. The value of the objectat sale is uniformly distributed on [0, 1] for each buyer. Individuals know their valuation for the object, butignore that of their competitors. Buyers participate in an auction in which the highest bid wins the objectand in which everyone, but for the lowest bidder, has to pay his bid. [30 Marks]
(a) Compute the probability that a buyer wins the auction if he bids b and everyone follows a strictlyincreasing strategy β : [0, 1]→ R. [5 Marks]
(b) Use the revenue equivalence theorem to derive the symmetric equilibrium bidding strategy. [15 Marks]
(c) Compute the revenues of the auctioneer. [10 Marks]
Hand In Problem Set 2 Francesco Nava
EC319: Auction Theory Michaelmas Term
Please give your answers to your class teacher before class in week 11MT. Thank you!
1. Show that the expected value of the virtual valuation is always zero. [10 Marks]
2. Suppose that the buyers have values independently and identically distributed according to a uniform on [1, 3].Let there be 3 buyers.
(a) Derive the revenue maximizing auction for the seller. [10 Marks]
(b) Compute the seller’s expected revenue. [10 Marks]
3. Consider an environment with three buyers. Buyer 0 has a value that is drawn from a uniform distributionon the interval [0, 2], and every other buyer {1, 2} has a value that is independently drawn from a uniformdistribution on the interval [2, 3].
(a) Find the revenue maximizing mechanism amongst those that are IC, IR. [15 Marks]
(b) Find the revenue maximizing mechanism amongst those that are effi cient, IC, IR. Check that it is IC &IR. [15 Marks]
(c) Is there any mechanism that balances the budget and that is effi cient, IC, IR? If it exists, characterizeit. [15 Marks]
4. Consider a bilateral trade problem. Assume that the value of the object to be traded to the buyer is uniformlydistributed on [1, 3] and that the costs of the producer are uniformly distributed on [0, 2].
(a) Derive the VCG mechanism & check effi ciency IC & IR. [15 Marks]
(b) Is there any mechanism that balances the budget and that is effi cient, IC, IR? Explain. [10 Marks]
Auction
Theory
EC
319
–L
ectu
reN
otes
0
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
AuctionTheory
MichaelmasTerm
1/7
Motivation
Inh
isto
ryau
ctio
ns
hav
eb
een
use
dex
ten
sive
lyto
trad
ego
od
s
To
day
they
are
use
dto
:
1tr
ade
com
mo
dit
ies
[tob
acco
,fl
ower
s,go
ld]
2pr
ivat
ize
pu
blic
com
pan
ies
3se
lllo
ng-
term
secu
riti
es[U
Str
easu
ries
]
4se
llth
eri
ghts
tou
sen
atu
ral
reso
urc
es
5pr
ocu
rese
rvic
es
6se
llgo
od
thro
ugh
the
inte
rnet
[eB
ay].
..
Au
ctio
ns
can
be
use
dfo
rd
iffer
ent
pu
rpos
es:
1R
even
ue
Max
imiz
atio
n
2E
ffici
ency
3..
. Nava
(LSE)
AuctionTheory
MichaelmasTerm
2/7
Not
es
Not
es
Objectivesof
theCourseare:
1U
nd
erst
and
how
bu
yers
ough
tto
bid
ind
iffer
ent
auct
ion
form
ats
2Q
uan
tify
the
reve
nu
esof
the
selle
rin
diff
eren
tau
ctio
nfo
rmat
s
3U
nd
erst
and
how
the
inte
rdep
end
ence
inva
luat
ion
s
and
the
corr
elat
ion
inin
form
atio
naff
ects
beh
avio
r
4D
esig
nre
ven
ue
max
imiz
ing
selli
ng
mec
han
ism
s
5D
esig
neffi
cien
tse
llin
gm
ech
anis
ms
6U
nd
erst
and
how
info
rmat
ion
alfr
icti
ons
mig
ht
limit
trad
e
Nava
(LSE)
AuctionTheory
MichaelmasTerm
3/7
General
Inform
ation
Lec
ture
r:F
ran
cesc
oN
ava
Em
ail:
f.n
ava@
lse.
ac.u
k
Co
urs
eW
ebsi
te:
mo
od
le.l
se.a
c.u
k
Tim
ea
nd
Lo
cati
on
:T
hu
9.30
am-1
1.00
am,
32L
.LG
.18
Offi
ceH
ou
rs:
Wed
1.30
pm
-3.0
0pm
,32
L.3
.20
Nava
(LSE)
AuctionTheory
MichaelmasTerm
4/7
Not
es
Not
es
Grades&
Assignm
ents
Cla
ssW
ork
:
Eac
hw
eek
you
will
be
assi
gned
afe
wex
erci
ses
(tex
tbo
okor
extr
a)
An
swer
sto
exer
cise
sw
illb
ep
oste
dso
met
ime
afte
rcl
asse
s
You
are
exp
ecte
dto
atte
mp
tth
ecl
assw
ork
bef
ore
you
rcl
ass
Inw
eeks
5and10
ofMT
written
assign
ments
arerequired
You
are
exp
ecte
dto
be
dili
gen
tab
out
mee
tin
gth
ed
ead
lines
Exa
ms:
Th
efi
nal
exam
will
det
erm
ine
you
rgr
ade
onth
eco
urs
een
tire
ly
Sec
tion
1re
qu
ires
you
toan
swer
4sh
ort
qu
esti
ons
onM
T&
LT
Sec
tion
2re
qu
ires
you
toan
swer
1lo
ng
qu
esti
ons
onM
T
Sec
tion
3re
qu
ires
you
toan
swer
1lo
ng
qu
esti
ons
onLT
Nava
(LSE)
AuctionTheory
MichaelmasTerm
5/7
WeeklyCourseProgram
1R
evie
w:
Ord
erS
tati
stic
s&
Gam
eT
heo
ry[K
A,
KC
,K
F]
2In
dep
end
ent
Pri
vate
Val
ue
Au
ctio
ns
[K1,
K2]
3R
even
ue
Eq
uiv
alen
ceP
rin
cip
le[K
3]
4E
xten
sion
s:R
esal
e&
Bu
dge
tC
onst
rain
ts[K
4]
5E
xten
sion
s:R
isk
Ave
rsio
n[K
4]
6O
pti
mal
Au
ctio
nD
esig
n[K
5,3]
7O
pti
mal
Au
ctio
nD
esig
n[K
5,5]
8E
ffici
ent
Mec
han
ism
sV
CG
[K5,
6]
9In
terd
epen
den
tV
alu
es[K
6,4]
10
Lin
kage
Pri
nci
ple
[K7,
4]
Nava
(LSE)
AuctionTheory
MichaelmasTerm
6/7
Not
es
Not
es
Materials
Ma
inT
extb
oo
k
Au
ctio
nT
heo
ry,
Vij
ayK
rish
na,
Aca
dem
icP
ress
2010
[su
gges
ted
][K
]
Slid
es
Incl
ud
eal
lm
ater
ials
req
uir
edfo
rth
eev
alu
atio
ns
un
less
oth
erw
ise
spec
ified
Slid
esw
illb
ep
oste
db
efor
eea
chcl
ass
Pa
per
s
Som
eof
the
pap
ers
onth
esy
llab
us
may
be
sugg
este
dto
the
inte
rest
edre
ader
Alt
ern
ati
veT
extb
oo
k
Pu
ttin
gA
uct
ion
Th
eory
toW
ork,
Pau
lM
ilgro
m,
Cam
brid
geU
niv
ersi
tyP
ress
2004
Nava
(LSE)
AuctionTheory
MichaelmasTerm
7/7
Not
es
Not
es
Ord
erS
tati
stic
sE
C31
9–
Lec
ture
Not
esI
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Order
Statistics
MichaelmasTerm
1/7
Ord
erS
tati
stic
s
We
wis
hto
ord
era
list
ofra
nd
omd
raw
s.L
et:
X1,.
..,X
nb
ein
dep
end
ent
dra
ws
from
ad
istr
ibu
tion
Fd
ensi
tyf
.
Rea
rran
geth
ese
dra
ws
soth
atY1≥
Y2≥
...≥
Yn.
Hen
cefo
rin
stan
ceY1=
max{X
1,.
..,X
n}.
Th
era
nd
omva
riab
leYi
isre
ferr
edto
asth
eirhorder
statistic.
Den
oteFi
the
dis
trib
uti
onan
df i
the
den
sity
ofYi.
Nava
(LSE)
Order
Statistics
MichaelmasTerm
2/7
Not
es
Not
es
Fir
stan
dS
econ
dO
rder
Sta
tist
ics
To
com
pu
teth
ed
istr
ibu
tion
ofth
efi
rst
two
ord
erst
atis
tics
not
ice
that
:
Th
eev
entY1≤
yre
qu
ires
Xi≤
yfo
ral
li.
Th
us,
F1(y)=
F(y)n
,
f 1(y)=
nf(y)F
(y)n−1.
Th
eev
entY2≤
yre
qu
ires
one
oftw
om
utu
ally
excl
usi
veev
ents
toh
old
:
1Xi≤
yfo
ral
li;
2Xi≤
yfo
ral
li6=
jan
dXj>
yso
mej.
Th
eref
ore,
F2(y)=
F(y)n
+nF(y)n−1(1−
F(y))
=
=nF(y)n−1−(n−
1)F
(y)n
,
f 2(y)=
n(n−
1)(
1−
F(y))f(y)F
(y)n−2.
Nava
(LSE)
Order
Statistics
MichaelmasTerm
3/7
Exa
mpl
e:F
irst
Ord
erS
tati
stic
Den
ote
byG
n(y)
the
CD
Fof
the
firs
tor
der
stat
isti
cin
asa
mp
leof
size
n.
Ass
um
eth
atF
isu
nif
orm
on[0
,1].
Ifso
:
Nava
(LSE)
Order
Statistics
MichaelmasTerm
4/7
Not
es
Not
es
Rel
atio
nsam
ong
Diff
eren
tS
ampl
eS
ize
For
con
ven
ien
ce,
lab
el:
theith
ord
erst
atis
tic
ina
sam
ple
ofsi
zen
byY
(n)
i;
its
dis
trib
uti
onan
dd
ensi
tyby
F(n)
ian
df(n)
i.
Rec
allF(n)
1(y)=
F(y)n
andF(n)
2(y)=
nF(y)n−1−(n−
1)F
(y)n
.
Th
enob
serv
eth
at F(n)
2(y)=
nF(n−1)
1(y)−(n−
1)F
(n)
1(y),
f(n)
2(y)=
nf(n−1)
1(y)−(n−
1)f
(n)
1(y),
E(Y
(n)
2)=
nE(Y
(n−1)
1)−(n−
1)E
(Y(n)
1).
Ad
dit
ion
ally
,it
mu
stb
eth
at
f(n)
2(y)=
n(1−
F(y))f(n−1)
1(y).
Nava
(LSE)
Order
Statistics
MichaelmasTerm
5/7
Join
tD
istr
ibut
ions
Th
ejo
int
den
sity
ofY
=(Y
(n)
1,.
..,Y
(n)
n)
isgi
ven
by
f(n)
Y(y
1,.
..,y
n)=
{ n!f(y
1)f(y
2).
..f(y
n)
ify 1≥
y 2≥
...≥
y n0
ifot
her
wis
e
wh
ich
sim
ply
amou
nts
ofth
epr
obab
ility
of(x
1,.
..,x
n)
adju
sted
byth
en
um
ber
ofp
ossi
ble
per
mu
tati
ons.
Fro
min
tegr
atio
nfo
ry 1≥
y 2w
eal
soge
tth
at
f(n)
1,2(y
1,y
2)=
n(n−
1)f(y
1)f(y
2)F
(y2)n−2.
For
inst
ance
wh
enn=
3:
f(3)
1,2(y
1,y
2)=∫ y 2 0
3!f(y
1)f(y
2)f(y
3)dy 3
=6f(y
1)f(y
2)F
(y2).
Nava
(LSE)
Order
Statistics
MichaelmasTerm
6/7
Not
es
Not
es
Con
diti
onal
Dis
trib
utio
ns
Th
ed
ensi
tyof
Y(n)
2co
nd
itio
nal
onY
(n)
1=
yth
us
sati
sfies
fory≥
z
f(n)
2(z|Y
(n)
1=
y)=
f(n)
1,2(y
,z)
f(n)
1(y)
=(n−
1)f(z)F
(z)n−2
F(y)n−1
.
Sim
ilarl
yth
ed
ensi
tyof
Y(n−1)
1co
nd
itio
nal
onY
(n−1)
1<
ysa
tisfi
es
f(n−1)
1(z|Y
(n−1)
1<
y)=
f(n−1)
1(z)
F(n−1)
1(y)=
(n−
1)f(z)F
(z)n−2
F(y)n−1
.
Th
ese
coin
cid
en
otby
coin
cid
ence
s.T
hin
kab
out
it.
Nava
(LSE)
Order
Statistics
MichaelmasTerm
7/7
Not
es
Not
es
Gam
eT
heor
yR
evie
wE
C31
9–
Lec
ture
Not
esII
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
1/16
Com
pleteInform
ationGam
es
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
2/16
Not
es
Not
es
Com
plet
eIn
form
atio
n
Aco
mp
lete
info
rmat
ion
gam
eG
={ N, {
Ai,ui}
i∈N
} con
sist
sof
:
Nth
ese
tof
pla
yers
inth
ega
me;
Ai
pla
yeri’
sac
tion
set;
An
acti
onpr
ofile
ais
anel
emen
tth
ese
tA=×
j∈NAj;
For
con
ven
ien
ced
efin
eA−i=×
j∈N\iAj;
ui
:A→
Rp
laye
ri’
su
tilit
yfu
nct
ion
;
Map
pin
gac
tion
profi
les
top
ayoff
s.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
3/16
Nas
hE
quili
bria
Defi
niti
on(N
ash
Equ
ilibr
ium
NE
)
A(p
ure
stra
tegy
)N
ash
equ
ilibr
ium
ofa
gam
eG
con
sist
sof
ast
rate
gypr
ofile
a=
(ai,a −
i)∈A
such
that
ui(a)≥
ui(a′ i,
a −i)
for
anya′ i∈Ai
&i∈N
.
Pro
per
ties
:
stra
tegy
profi
les
are
ind
epen
den
t;
stra
tegy
profi
les
com
mon
know
led
ge;
pu
rest
rate
gyN
ash
equ
ilibr
iam
ayn
otex
ist.
Exa
mp
leon
eb
uye
rw
ith
valu
e2$
and
one
selle
r:
B\S
sale
no
sale
2$0,
20,
01$
1,1
0,0
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
4/16
Not
es
Not
es
IncompleteInform
ationGam
es
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
5/16
Inco
mpl
ete
Info
rmat
ion
An
inco
mp
lete
info
rmat
ion
gam
eΓ={N
, {Ai,
Xi,ui}
N,f}
con
sist
sof
:
Nth
ese
tof
pla
yers
inth
ega
me;
Ai
pla
yeri’
sac
tion
set;
Xi
pla
yeri’
sse
tof
pos
sib
lesi
gnal
s;
Apr
ofile
ofsi
gnal
sis
anel
emen
tx∈
X=×
j∈N
Xj;
For
con
ven
ien
ced
efin
eX−i=×
j∈N\i
Xj;
f∈
∆(X
)a
dis
trib
uti
onov
erth
ep
ossi
ble
sign
als;
ui
:A×
X→
Rp
laye
ri’
su
tilit
yfu
nct
ion
,ui(a|x).
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
6/16
Not
es
Not
es
Bel
iefs
and
Str
ateg
ies
Info
rmat
ion
stru
ctu
re:
Xi
den
otes
the
sign
alas
ara
nd
omva
riab
le;
Pla
yeri
obse
rves
only
Xi;
Con
dit
ion
alob
serv
ingXi=
x ih
isb
elie
fsar
e
f(X−i|x
i)=
f(X−i,x i)
∑x −
i∈X−if(x−i,x i)∈
∆(X−i)
.
Str
ateg
yP
rofi
les:
Ast
rate
gyco
nsi
sts
ofa
map
from
info
rmat
ion
toac
tion
sαi
:Xi→
Ai;
Ast
rate
gypr
ofile
isa
map
α(X
)=
(α1(X
1),
...,
αN(X
N))
.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
7/16
Pay
offs
&D
omin
ance
Th
eex
-post
,ex
-ante
andinterim
expec
tedutility
ofst
rate
gypr
ofile
αar
ere
spec
tive
lyd
efin
edby
:
ui(
α(x)|x)
:X→
R,
E[u
i(α(X
)|X)]=
∑Xui(
α(x)|x)f(x)∈
R,
E[u
i(α(X
)|X)|Xi=
x i]=
∑X−iui(
α(x)|x)f(x−i|x
i):X
i→
R.
Wea
kDominance
:
Str
ateg
yαi
wea
kly
dom
inat
esα′ i
iffo
ran
ya −
ian
dx
ui(
αi(x i),a −
i|x)≥
ui(
α′ i(x i),a −
i|x)
[str
ict
som
ewh
ere]
.
Str
ateg
yαi
isd
omin
ant
ifit
dom
inat
esan
yot
her
stra
tegy
α′ i.
Str
ateg
yαi
isu
nd
omin
ated
ifn
ost
rate
gyd
omin
ates
it.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
8/16
Not
es
Not
es
Dom
inan
tS
trat
egy
and
Bay
esN
ash
Equ
ilibr
ia
Defi
niti
ons
(Dom
inan
tS
trat
egy
Equ
ilibr
ium
DS
E)
ADominantStrategyeq
uilibrium
ofan
inco
mp
lete
info
rmat
ion
gam
eΓ
isa
stra
tegy
profi
leα
that
for
anyi∈N
,x∈
Xan
da −
i∈A−i
sati
sfies
ui(
αi(x i),a −
i|x)≥
ui(
α′ i(x i),a −
i|x)
for
any
α′ i
:Xi→
Ai.
Defi
niti
ons
(Bay
esN
ash
Equ
ilibr
ium
BN
E)
ABayes
Nash
equilibrium
ofan
inco
mp
lete
info
rmat
ion
gam
eΓ
isa
stra
tegy
profi
leα
that
for
anyi∈N
andx i∈
Xi
sati
sfies
E[u
i(α(X
)|X)|Xi=
x i]≥
E[u
i(a i
,α−i(X−i)|X
)|Xi=
x i]
for
anya i∈Ai.
Ifα
isa
DS
Eth
enit
isa
BN
E
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
9/16
Equ
ilibr
ium
Pro
per
ties
BN
Ere
qu
ires
forinterim
optimality.
Ex-ante
equ
ilibr
ium
isa
stra
tegy
profi
les.
t.fo
ran
yα′ i
:Xi−→
∆i
E[u
i(α(X
)|X)]≥
E[u
i(α′ i(Xi)
,α−i(X−i)|X
)].
Ex-post
equ
ilibr
ium
isB
NE
s.t.
for
anya i∈Ai
andx∈
X
ui(
α(x)|x)≥
ui(a i
,α−i(x −
i)|x).
DS
E⇒
Ex-
pos
t⇒
BN
E=
Inte
rim⇒
Ex-
ante
[by
inte
grat
ing]
.
Ex-
ante
opti
mal
ity⇒
Inte
rim
opti
mal
ity
alm
ost
sure
ly.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
10/16
Not
es
Not
es
BN
EE
xam
ple
I:E
xcha
nge
Ab
uye
ran
da
selle
rw
ant
totr
ade
anob
ject
:
Bu
yer’
sva
lue
for
the
obje
ctis
3$;
Sel
ler’
sva
lue
isei
ther
0$or
2$b
ased
onth
esi
gnal
,X
S={L
,H};
Bu
yer
can
offer
eith
er1$
or3$
top
urc
has
eth
eob
ject
;
Sel
ler
cho
ose
wh
eth
eror
not
tose
ll.
B\S
.Lsa
len
osa
leB\S
.Hsa
len
osa
le3$
0,3
0,0
3$0,
30,
21$
2,1
0,0
1$2,
10,
2
Th
isga
me
for
any
prio
rf
has
aB
NE
inw
hic
h
αS(L)=
sale
,αS(H
)=
no
sale
,αB=
1$.
Sel
ling
isw
eakl
yd
omin
ant
forS
.L.
Off
erin
g1$
isw
eakl
yd
omin
ant
for
the
bu
yer.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
11/16
BN
EE
xam
ple
II:
Dis
cret
eS
econ
dP
rice
Auc
tion
Tw
ob
uye
rsw
ant
tob
uy
obje
ctat
seco
nd
pric
eau
ctio
n:
Bu
yeri’
sva
lue
isei
ther
1$or
2$b
ased
onh
issi
gnal
,X
i={L
,H};
Bu
yers
cho
ose
wh
eth
erto
bid
1$or
2$;
Th
eh
igh
est
bid
win
sth
eob
ject
(tie
sar
ebr
oken
wit
hfa
irlo
tter
y);
Th
ese
con
dh
igh
est
bid
isp
aid
byth
ew
inn
er.
1.L\2
.L2$
1$1.L\2
.H2$
1$2$
-0.5
,-0.
50,
02$
-0.5
,00,
01$
0,0
0,0
1$0,
10,
0.5
1.H\2
.L2$
1$1.H\2
.H2$
1$2$
0,-0
.51,
02$
0,0
1,0
1$0,
00.
5,0
1$0,
10.
5,0.
5
Th
ega
me
has
aD
SE
inw
hic
hαi(L)=
1$an
dαi(H)=
2$fo
r∀i∈N
.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
12/16
Not
es
Not
es
BN
EE
xam
ple
III:
Dis
cret
eF
irst
Pri
ceA
ucti
on
Tw
ob
uye
rsw
ant
tob
uy
obje
ctat
seco
nd
pric
eau
ctio
n:
Bu
yeri’
sva
lue
isei
ther
1$or
2$b
ased
onh
issi
gnal
,X
i={L
,H};
Bu
yers
cho
ose
wh
eth
erto
bid
1$or
2$;
Th
eh
igh
est
bid
win
sth
eob
ject
(tie
sar
ebr
oken
wit
hfa
irlo
tter
y);
Th
eh
igh
est
bid
isp
aid
byth
ew
inn
er.
1.L\2
.L2$
1$1.L\2
.H2$
1$2$
-0.5
,-0.
5-1
,02$
-0.5
,0-1
,01$
0,-1
0,0
1$0,
00,
0.5
1.H\2
.L2$
1$1.H\2
.H2$
1$2$
0,-0
.50,
02$
0,0
0,0
1$0,
-10.
5,0
1$0,
00.
5,0.
5
Th
ega
me
has
aD
SE
inw
hic
hαi(L)=
αi(H)=
1$fo
r∀i∈N
.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
13/16
Extra:ContinuousGam
esExistence
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
14/16
Not
es
Not
es
Ext
ra:
Con
tinu
ous
Gam
es&
BN
EE
xist
ence
Are
fere
nce
for
pro
ofs
isA
they
Eco
nom
etri
ca20
01.
Pla
yers
’ac
tion
sets
Ai⊆
Ran
dar
eco
nve
x.
Pla
yers
’ty
pes
sets
Xi⊆
Ran
dar
eco
nve
x.
Let
Ui(a i
,α−i(S)|x i)=∫ x −
i∈Sui(a i
,α−i(x −
i)|x)f−i(x −
i|xi)dx −
i.
Let
Ui(a i
,α−i|x
i)=
Ui(a i
,α−i(
X−i)|x
i).
Defi
niti
on(R
egul
arit
yC
ondi
tion
sR
C)
Typ
esh
ave
ajo
int
den
sity
fw
hic
his
bou
nd
edan
dat
omle
ss.
Ui(a i
,α−i(S)|x)
exis
tsan
dis
fin
ite
for
anyS
con
vex
and
any
non
-dec
reas
ing
fun
ctio
nαj
:Xj→
Aj
forj6=
i.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
15/16
Sin
gle
Cro
ssin
g&
Exi
sten
ceB
NE
Defi
niti
on(S
ingl
eC
ross
ing
ofIn
crem
enta
lR
etur
nsS
CP
IR)
Th
efu
nct
ionh
:R2→
Rsa
tisfi
essi
ngl
ecr
ossi
ng
prop
erty
ofin
crem
enta
lre
turn
sif
for
anya>
aan
dx>
x
h(a
,x)>
h(a
,x)⇒
h(a
,x)≥
h(a
,x).
Defi
niti
on(S
ingl
eC
ross
ing
Con
diti
onS
CC
)
Th
esi
ngl
ecr
ossi
ng
con
dit
ion
for
gam
esw
ith
inco
mp
lete
info
rmat
ion
ism
etif
for
anyi∈N
,w
hen
ever
αj
:Xj→
Aj
isn
on-d
ecre
asin
gfo
ran
yj6=
i,p
laye
ri’
sob
ject
ive
fun
ctio
nUi(a i
,α−i|x
i)sa
tisfi
esS
CP
IRin
(ai,x i).
The
orem
(Exi
sten
ceof
BN
E)
Assumethat
SCCandRChold.Supposethat
ui(a|x)iscontinuou
sin
aforanyi∈N.Then
thereexists
aBNEin
non
-decreasingstrategies.
Nava
(LSE)
GameTheory
Review
MichaelmasTerm
16/16
Not
es
Not
es
Inde
pen
dent
Pri
vate
Val
ueA
ucti
onE
C31
9–
Lec
ture
Not
esII
I
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
1/29
The
Environment
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
2/29
Not
es
Not
es
Eco
nom
y
Con
sid
erth
efo
llow
ing
envi
ron
men
t:
1au
ctio
nee
r;
Np
oten
tial
bu
yers
;
1in
div
isib
leob
ject
own
edby
the
auct
ion
eer;
Bu
yer
i’s
valu
atio
nXi
for
the
obje
ct;
Xi
ind
epen
den
tan
did
enti
cally
dis
trib
ute
d;
Xicd
f ∼F
con
tin
uou
san
din
crea
sin
gon
[0,ω
];
Fad
mit
sa
con
tin
uou
sd
ensi
tyf
and
has
full
sup
por
t.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
3/29
Info
rmat
ion
&B
udge
ts
Pri
vate
Va
lues
How
mu
chp
laye
rva
lues
the
goo
dis
ind
epen
den
tof
oth
ers.
Info
rma
tio
n
Bu
yer
ikn
ows:
x ith
ere
aliz
atio
nof
Xi;
Xjcd
f ∼F
ind
epen
den
tly
acro
ssj6=
i.
Bu
dg
ets
No
bu
yer
face
san
yb
ud
get/
liqu
idit
yco
nst
rain
t.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
4/29
Not
es
Not
es
First
Price
Auction
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
5/29
Fir
stP
rice
Sea
led-
Bid
Auc
tion
Mec
ha
nis
m
All
pla
yers
sub
mit
seal
edb
ids
bi.
Hig
hes
tb
idw
ins
the
obje
ctan
dp
ays
the
amou
nt
bid
.
For
any
profi
leof
bid
sb={b
j}j∈
Np
ayoff
sar
e
Πi=
{ x i−
bi
ifbi>
max
j6=ibj
0if
bi<
max
j6=ibj
Tie
-bre
ak
ing
Ru
le
Eq
ual
prob
abili
tyof
gett
ing
obje
ct&
pay
ing
amon
gst
hig
hes
tb
idd
ers.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
6/29
Not
es
Not
es
FP
AE
quili
bria
&S
hadi
ng
Ast
rate
gypr
ofile{β
j}j∈
N,
βj
:[0,
ω]→
R+
any
j,is
sym
met
ric
if
βj=
βi=
βfo
ral
li,
j.
Wh
atis
the
opti
mal
bid
bby
bu
yer
iif
:
j6=
iad
opt
asy
mm
etri
c,in
crea
sin
g,d
iffer
enti
able
stra
tegy
β;
his
priv
ate
valu
eis
Xi=
x.
Not
eth
at:
Ab
idb=
xse
cure
sa
pay
off0;
An
yb
idb>
β(ω
)is
sub
opti
mal
;
β(0)=
0.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
7/29
FP
AE
quili
bria
&S
hadi
ng
Defi
ne
Y1=
max
j6=iXj=⇒
Y1cd
f ∼G
=FN−1.
Sin
ceβ
isin
crea
sin
gm
axj6=
iβ(X
j)=
β(Y
1).
Pla
yer
iw
ins
the
obje
ctif
b>
β(Y
1)⇔
β−1(b)>
Y1.
Exp
ecte
dp
ayoff
ofi
wh
ench
oos
ing
bth
us
amou
nts
to
G(β−1(b))(x−
b).
Max
imiz
ing
wit
hre
spec
tto
byi
eld
s
g(β−1(b))
β′ (
β−1(b))(x−
b)−
G(β−1(b))
=0.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
8/29
Not
es
Not
es
FP
AE
quili
bria
&S
hadi
ng
As
we
saw
,F
OC
amou
nt
to
g(β−1(b))
β′ (
β−1(b))(x−
b)−
G(β−1(b))
=0.
Su
chco
nd
itio
nat
asy
mm
etri
ceq
uili
briu
m,
b=
β(x),
bec
omes
d dx(G
(x)β(x))
=xg
(x).
Sin
ceβ(0)=
0,in
tegr
atio
nyi
eld
s
β(x)=
1
G(x)
∫ x 0yg
(y)d
y=
E[Y
1|Y
1<
x].
Nec
essa
ryn
otsu
ffici
ent
for
sym
met
ric
equ
ilibr
ium
exis
ten
ce.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
9/29
FP
AS
ymm
etri
cE
quili
bria
The
orem
Sym
met
ric
equ
ilibr
ium
stra
tegi
esin
afi
rst
pric
eau
ctio
nar
e:
β1(x)=
E[Y
1|Y
1<
x]
for
Y1
the
hig
hes
tof
N−
1in
dep
end
entl
yd
raw
nva
lues
.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
10/29
Not
es
Not
es
Pro
of.
Su
pp
ose
that
all
bid
der
sb
ut
one
follo
wth
est
rate
gyβ=
β1.
Let
xb
eh
isva
lue
and
bh
isb
idan
dd
efin
ez=
β−1(b)
[in
vert
ible
].O
nly
bid
sb≤
β(ω
)ar
eop
tim
al.
Exp
ecte
dpr
ofits
for
such
bid
der
are
Π(x
,b)=
G(z)[
x−
β(z)]=
=G(z)[
x−
E[Y|Y
<z]]=
=G(z)x−∫ z 0
yg(y)d
y=
=G(z)[
x−
z]+
∫ z 0G(y)d
y.
Th
us
rega
rdle
ssof
x≤
zor
x≥
zit
mu
stb
e
Π(x
,β(x))−
Π(x
,β(z))
=G(z)[
z−
x]−
∫ z xG(y)d
y≥
0.
Th
eref
ore
this
issy
mm
etri
ceq
uili
briu
m,
sin
ceb>
β(ω
)is
sub
opti
mal
.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
11/29
FP
AS
ymm
etri
cE
quili
bria
Plo
t
Th
ep
lot
bel
owsh
ows
wh
yit
mu
stb
eth
at
Π(x
,β(x))−
Π(x
,β(z))
=G(z)[
z−
x]−
∫ z xG(y)d
y≥
0.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
12/29
Not
es
Not
es
Deg
ree
ofS
hadi
ngan
dth
eU
nifo
rmC
ase
Sh
ad
ing
β1(x)=
1
G(x)
∫ x 0yd
G(y)=
=1
G(x)
[ [yG(y) ]x 0−∫ x 0
G(y)d
y
] =
=x−∫ x 0
G(y)
G(x)
dy=
x−∫ x 0
[ F(y)
F(x)
] N−1d
y.
Exa
mp
le
IfF(x)=
xis
un
ifor
mon
[0,1]:
β1(x)=
x−
1
xN−1
[ yN N
] x 0
=N−
1
Nx
.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
13/29
Uni
form
Exa
mpl
e
For
inst
ance
wit
h2
bid
der
s,ev
eryo
ne
bid
sh
alf
ofth
epr
ivat
eva
lue.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
14/29
Not
es
Not
es
SecondPrice
Auction
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
15/29
Sec
ond
Pri
ceS
eale
d-B
idA
ucti
on
Mec
ha
nis
m
All
pla
yers
sub
mit
seal
edb
ids
bi.
Hig
hes
tb
idw
ins
the
obje
ctan
dp
ays
the
seco
nd
hig
hes
t.
For
any
profi
leof
bid
sb={b
j}j∈
Np
ayoff
sar
e
Πi=
{ x i−
max
j6=ibj
ifbi>
max
j6=ibj
0if
bi<
max
j6=ibj
Tie
-bre
ak
ing
Ru
le
Eq
ual
prob
abili
tyof
gett
ing
obje
ct&
pay
ing
amon
gst
hig
hes
tb
idd
ers.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
16/29
Not
es
Not
es
Equ
ilibr
ium
Beh
avio
r
Lem
ma
Bid
din
gac
cord
ing
toβ2(x)=
xis
aw
eakl
yd
omin
ant
stra
tegy
ina
seco
nd
pric
eau
ctio
n.
Pro
of.
Let
pi=
max
j6=ibj.
Pla
yer
iby
bid
din
gx i
gets
ap
ayoff
Πi=
{ x i−
pi
ifx i
>pi
0if
x i<
pi
Ifh
ew
ere
tob
idbi<
x ih
ew
ould
still
earn
Πi=
{ x i−
pi
ifbi>
pi
0if
bi<
pi
Bu
tP
r(x i
>pi)>
Pr(
bi>
pi)
.S
imila
rly
bi>
x i.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
17/29
Revenue
Com
parisons
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
18/29
Not
es
Not
es
Rev
enue
Com
pari
son
Ina
1st
&2n
dpr
ice
auct
ion
the
exp
ecte
dp
aym
ent
ofb
idd
erw
ith
valu
ex
:
m1(x)=
Pr(
win)∗
1st b
id=
G(x)E
(Y1|Y
1<
x),
m2(x)=
Pr(
win)∗
E(2
ndb
id|1
stb
id=
x)=
G(x)E
(Y1|Y
1<
x).
Th
eex
pec
ted
ex-a
nte
pay
men
tby
ap
arti
cula
rp
laye
ris
for
k∈{1
,2}
E(m
k)=∫ ω 0
[ ∫ x 0yg
(y)d
y
] f(x)d
x=∫ ω 0
[ ∫ ω yf(x)d
x
] yg(y)d
y=
=∫ ω 0
[1−
F(y) ]
yg(y)d
y=∫ ω 0
y[1−
F(y) ]
f(N−1)
1(y)d
y.
Th
eref
ore
the
exp
ecte
dex
-an
tere
ven
ues
der
ived
byth
eau
ctio
nee
rar
e
E(R
k)=
N∗
E(m
k)=∫ ω 0
yf(N
)2
(y)d
y=
E(Y
(N)
2).
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
19/29
Rev
enue
Com
pari
son
Rev
enu
eE
qu
ival
ence
firs
tta
ke:
Lem
ma
Inan
IPV
setu
pth
eex
pec
ted
reve
nu
esfr
omb
oth
a1s
tan
da
2nd
pric
eau
ctio
nar
eth
esa
me.
Rev
enu
esin
bot
hfo
rmat
sh
ave
the
sam
em
ean
,b
ut
diff
eren
tva
rian
ce.
Mea
n-p
rese
rvin
gS
prea
d(M
PS
):
Defi
niti
on
Let
Xcd
f ∼F
and
Z|X
=x
cdf ∼
H(x)
soth
atE(Z|X
=x)=
0.If
Y=
[X+
Z]cd
f ∼G
,th
enw
esa
yth
atG
isa
mea
n-p
rese
rvin
gsp
read
ofF
.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
20/29
Not
es
Not
es
Rev
enue
Com
pari
son
Lem
ma
Inan
IPV
setu
pth
ed
istr
ibu
tion
ofeq
uili
briu
mpr
ices
ofa
2nd
pric
eau
ctio
nis
aM
PS
ofth
ed
istr
ibu
tion
ofeq
uili
briu
mpr
ices
ofa
1st
pric
eau
ctio
n.
Pro
of.
Rev
enu
esar
eas
ran
dom
vari
able
sre
spec
tive
lyR1=
β(Y
(N)
1)
&
R2=
Y(N
)2
thu
sit
mu
stb
eth
at
E(R
2|R
1=
p)=
E(Y
(N)
2|Y
(N)
1=
β−1(p))
=
=E(Y
(N−1)
1|Y
(N−1)
1<
β−1(p))
=β(β−1(p))
=p
.
Th
us
ther
eex
ists
Zsu
chth
atth
ed
istr
ibu
tion
ofR2
isth
esa
me
asth
atof
R1+
Zan
dE(Z|R
1=
p)=
0.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
21/29
Rev
enue
Com
pari
son
Exa
mpl
e
Au
nif
orm
dis
trib
uti
onex
amp
lew
ith
N=
2an
dω
=1.
FP
Ais
less
vola
tile
than
SP
A:
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
22/29
Not
es
Not
es
Reserve
Prices
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
23/29
Res
erve
Pri
ces
SP
A
Res
erve
Pri
ce
Sel
ler
rese
rves
the
righ
tn
otto
sell
ifw
inn
ing
pric
eis
bel
owr.
Th
ere
serv
epr
ice
isco
mm
onkn
owle
dge
amon
gst
pla
yers
.
2nd
Pri
ceA
uct
ion
Win
ner
pay
sre
serv
epr
ice
ifse
con
db
idis
bel
ow.
Mak
esn
od
iffer
ence
inb
ehav
ior
still
wea
kly
dom
inan
tto
bid
own
valu
e.
Th
eex
pec
ted
pay
men
tfo
rb
idd
erw
ith
valu
er
isrG
(r).
Ifx>
rth
en
m2(x
,r)=
rG(r)+∫ x r
yg(y)d
y.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
24/29
Not
es
Not
es
Res
erve
Pri
ces
FP
A
1st
Pri
ceA
uct
ion
:
Not
eth
at:
(a)
x<
rn
ever
gain
and
(b)
β(r)=
r.
Ifx≥
ras
bef
ore
we
can
get
that
β(x)=
E(m
ax{Y
1,r}|
Y1<
x)=
rG(r)
G(x)+
1
G(x)
∫ x ryg
(y)d
y.
Th
eex
pec
ted
pay
men
tb
ecom
es
m1(x
,r)=
rG(r)+∫ x r
yg(y)d
y.
Aga
inex
pec
ted
reve
nu
ein
bot
hfo
rmat
sco
inci
de.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
25/29
Res
erve
Pri
ceE
xam
ple
FP
A
Au
nif
orm
dis
trib
uti
onex
amp
lew
ith
N=
2an
dω
=1.
Bid
din
gis
mor
eag
gres
sive
wit
hre
serv
epr
ices
than
wit
hou
t:
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
26/29
Not
es
Not
es
Rev
enue
Eff
ects
and
Opt
imal
Res
erve
Pri
ces
Th
eex
-an
teex
pec
ted
pay
men
tis
for
k∈{1
,2}
E(m
k(r))
=r(
1−
F(r))
G(r)+∫ ω r
y(1−
F(y))
g(y)d
y.
Ifse
ller
atta
ches
valu
ex 0∈[0
,ω)
toth
eob
ject
and
sets
rese
rve
pric
er
his
exp
ecte
dp
ayoff
is
Π0(r)=
N∗
E(m
k(r))
+F(r)N
x 0.
For
λ(x)=
f(x)/
(1−
F(x))
(haz
ard
rate
),op
tim
alit
yre
qu
ires
dΠ
0
dr
=N[1−
F(r)−
rf(r)]
G(r)+
NG(r)f(r)x
0
=N[1−(r−
x 0)λ
(r)](1−
F(r))
G(r)=
0.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
27/29
The
Exc
lusi
onP
rinc
iple
Ifx 0
>0
then
dΠ
0(x
0)/
dr>
0th
us
r>
x 0.
Ifx 0
=0
then
dΠ
0(0)/
dr=
0,b
ut
it’s
loca
lm
inim
um
thu
sr>
x 0.
Hig
her
pay
offby
sett
ing
r>
x 0[E
xclu
de
som
etim
esto
rais
eb
ids]
.
Fir
stor
der
Nec
essa
ryco
nd
itio
nre
du
ces
to
(r∗−
x 0)λ
(r∗)
=1.
Ifλ
isin
crea
sin
g,th
eco
nd
itio
nal
sosu
ffici
ent.
En
try
fees
can
be
chos
ento
mim
icth
eeff
ects
ofre
serv
epr
ices
.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
28/29
Not
es
Not
es
Effi
cien
cyvs
Rev
enue
Co
mm
ents
on
Effi
cien
cy
Th
ough
are
serv
epr
ice
max
imiz
esre
ven
ues
toth
ese
ller
itm
ayb
ein
effici
ent.
Ifx 0
=0
and
r>
0th
ere
isa
pos
itiv
epr
obab
ility
that
the
selle
rre
tain
sth
eob
ject
.
Th
isis
ineffi
cien
tu
nle
ssal
lp
laye
rd
on
otva
lue
the
obje
ct(z
ero
prob
abili
ty).
Co
mm
itm
ent
Th
ere
nee
ds
tob
eco
mm
itm
ent
tore
serv
epr
ice
and
no
ex-p
ost
resa
le.
Sec
ret
rese
rve
valu
esaff
ect
beh
avio
r.
Nava
(LSE)
Indep
enden
tPrivate
ValueAuction
MichaelmasTerm
29/29
Not
es
Not
es
Rev
enue
Equ
ival
ence
Pri
ncip
leE
C31
9–
Lec
ture
Not
esIV
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
1/12
Revenue
Equivalence
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
2/12
Not
es
Not
es
Mai
nR
esul
t
Defi
niti
on
An
auct
ion
isstandard
ifth
ep
erso
nw
ho
bid
sth
eh
igh
est
amou
nt
isaw
ard
edth
eob
ject
.
Commen
t:A
lott
ery
inw
hic
hth
epr
obab
ility
ofw
inn
ing
dep
end
son
the
amou
nt
bid
isn
otst
and
ard
.
The
orem
(Rev
enue
Equ
ival
ence
)
Ifvalues
areiid
andbiddersrisk
neutral,then
anysymmetricincreasing
equilibrium
ofanystandardauctionsuch
that
theexpectedpaymentof
abidder
with
0valueis
0,yieldsthesameexpectedrevenue.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
3/12
Pro
of.
Fix
any
stan
dar
dau
ctio
nfo
rmat
and
asy
mm
etri
ceq
uili
briu
mβ
.
Let
m(x)
be
the
exp
ecte
dp
aym
ent
wh
enth
eva
lue
isx
.
Let
βb
eso
that
m(0)=
0.S
up
pos
eth
atal
l,b
ut
1p
laye
rfo
llow
β.
Ifh
eb
ids
β(z)
he
win
sifz≥
Y1.
His
exp
ecte
dp
ayoff
is
Π(z|x)=
G(z)x−
m(z).
Inte
rim
opti
mal
ity
req
uir
esth
at
dΠ(z|x)/
dz=
g(z)x−
m′ (z)=
0.
Ifβ
iseq
uili
briu
mth
eeq
ual
ity
hol
ds
atz=
x.
Th
usm′ (y)=
g(y)y
and
m(x)=
m(0)+
∫ x 0yg
(y)dy=
G(x)E
(Y1|Y
1<
x).
Th
us
exp
ecte
dp
aym
ents
do
not
dep
end
onth
eau
ctio
nfo
rmat
.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
4/12
Not
es
Not
es
Exa
mpl
e:U
nifo
rmR
even
ues
Un
ifor
md
istr
ibu
tion
defi
ned
on[0
,1].
F(x)=
xan
dG(x)=
xN−1
andg(x)=
(N−
1)x
N−2.
Exp
ecte
dp
aym
ents
amou
nt
to
m(x)=
∫ x 0(N−
1)y
N−1dy=
N−
1
NxN
.
Exp
ecte
dre
ven
ues
are
E(R
)=
N∫ 1 0
m(x)dx=
∫ 1 0(N−
1)x
Ndx=
N−
1
N+
1.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
5/12
Applications
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
6/12
Not
es
Not
es
Thi
rdP
rice
Auc
tion
IfXiiid ∼F
and
max
Xi=
xw
eh
ave
that
:
F(N−1)
2(y|Y
(N−1)
1<
x)=
F(y)N−1+(N−
1)(F(x)−
F(y))F(y)N−2
F(N−1)
1(x)
,
f(N−1)
2(y|Y
(N−1)
1<
x)=
(N−
1)(F(x)−
F(y))f(N−2)
1(y)
F(N−1)
1(x)
.
Ifβ3
isa
sym
met
ric
BN
Eby
reve
nu
eeq
uiv
alen
cem
3(x)=
m2(x)
soth
at
F(N−1)
1(x)E
[β3(Y
(N−1)
2)|Y
(N−1)
1<
x]=
∫ x 0yg
(y)dy
,∫ x 0
β3(y)(N−
1)(F(x)−
F(y))f(N−2)
1(y)dy=
∫ x 0yf
(N−1)
1(y)dy
,∫ x 0
β3(y)(F(x)−
F(y))f(N−2)
1(y)dy=
∫ x 0yf(y)F
N−2(y)dy
.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
7/12
Thi
rdP
rice
Auc
tion
Th
us
ifβ3
isin
crea
sin
gth
efo
llow
ing
con
dit
ion
hol
ds
∫ x 0β3(y)(F(x)−
F(y))f(N−2)
1(y)dy=
∫ x 0yf(y)F
(N−2)
1(y)dy
.
Tak
ing
der
ivat
ives
wit
hre
spec
tto
xim
plie
sth
at
0+
f(x)∫ x 0
β3(y)f
(N−2)
1(y)dy=
xf(x)F
(N−2)
1(x).
Sim
plif
yin
gf(x)
and
diff
eren
tiat
ing
agai
nim
plie
sth
at
β3(x)f
(N−2)
1(x)=
F(N−2)
1(x)+
xf(N−2)
1(x).
So
that
the
equ
ilibr
ium
bid
din
gst
rate
gym
ust
sati
sfy
β3(x)=
F(N−2)
1(x)
f(N−2)
1(x)+
x.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
8/12
Not
es
Not
es
Unc
erta
inN
umb
erof
Bid
ders
N={1
,...
,N}
be
the
set
ofp
oten
tial
bid
der
s.
Sym
met
ric
prio
rpn=
Pr(|A|=
n+
1)
for
all
bid
der
s.
Con
sid
era
stan
dar
dau
ctio
nan
da
sym
met
ric
&in
crea
sin
geq
β.
Con
sid
era
bid
der
wit
hva
luex
wh
ob
ids
β(z)
inst
ead
.
Ifh
efa
cesn
bid
der
sh
ew
ins
ifz≥
Y(n)
1w
ith
prob
Gn(z)=
F(z)n
.
Th
epr
obab
ility
ofw
inn
ing
bec
omes
G(z)=
∑N−1
n=0pnGn(z).
Th
eex
pec
ted
pay
offth
us
bec
omes
Π(z|x)=
G(z)x−
m(z).
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
9/12
Unc
erta
inN
umb
erof
Bid
ders
2ndPrice
Auction
Sti
llw
eakl
yd
omin
ant
tob
idon
e’s
own
valu
e.
Exp
ecte
dp
aym
ent
isgi
ven
by
m2(x)=
∑N−1
n=0pnGn(x)E
(Y(n)
1|Y
(n)
1<
x).
1stPrice
Auction
Exp
ecte
dp
aym
ent
isgi
ven
bym
1(x)=
G(x)β(x).
By
the
reve
nu
eeq
uiv
alen
cem
1(x)=
m2(x)
and
thu
s
β(x)=
∑N−1
n=0pnGn(x)
G(x)E(Y
(n)
1|Y
(n)
1<
x).
Wei
ghte
dav
erag
eof
the
bid
sin
the
know
nn
um
ber
case
.
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
10/12
Not
es
Not
es
All
Pay
Auc
tion
orL
obby
ing
Hig
hes
tb
idw
ins
obje
ct,
all
bid
der
sp
ayth
eir
bid
.
Su
pp
ose∃β
Asy
mm
etri
ceq
uili
briu
msu
chth
atm
A(0)=
0.
Bec
ause
bid
der
sal
way
sp
ayan
dby
reve
nu
eeq
uiv
alen
ce
βA(x)=
mA(x)=
∫ x 0yg
(y)dy
.
To
solv
eth
eas
sign
edpr
oble
msh
owth
at:
Su
chst
rate
gyis
aB
NE
ofth
eau
ctio
n
Ch
eck
that
nob
od
yca
nb
enefi
tfr
omu
nila
tera
ld
evia
tion
s
... N
ava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
11/12
Sol
ving
Rev
enue
Equ
ival
ence
Pro
blem
s
To
solv
esu
chpr
oble
ms
and
corr
ectl
yw
rite
dow
nex
pec
ted
pay
offs
itis
inst
ruct
ive
toth
ink
ofth
efo
llow
ing
plo
tfo
rN
=3:
Nava
(LSE)
Reven
ueEquivalence
Principle
MichaelmasTerm
12/12
Not
es
Not
es
IPV
Ext
ensi
ons
EC
319
–L
ectu
reN
otes
V
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
1/26
Intr
odu
ctio
n
Ass
um
pti
ons
un
der
lyin
gth
ere
ven
ue
equ
ival
ence
prin
cip
le:
Ind
epen
den
ce
Ris
kn
eutr
alit
y
No
Bu
dge
tC
onst
rain
ts
Sym
met
ry
No
Res
ale
Exp
lore
con
seq
uen
ces
ofre
laxi
ng
the
last
fou
ras
sum
pti
ons.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
2/26
Not
es
Not
es
RiskAversion
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
3/26
Ris
kA
vers
ion
Ris
kn
eutr
alit
y⇒
pay
offlin
ear
inp
aym
ents⇒
reve
nu
eeq
uiv
alen
ce.
Ris
kav
ersi
on:
u:R
+→
Rsa
tisf
yin
gu(0)=
0,u′>
0&
u′′<
0.
Lem
ma
Inan
IPV
setu
pif
bid
der
sar
eri
sk-a
vers
ean
dsy
mm
etri
c,th
enth
eex
pec
ted
reve
nu
ein
a1s
tpr
ice
auct
ion
isgr
ater
than
ina
2nd
pric
eau
ctio
n.
Pro
of.
In2n
dpr
ice
auct
ion
bid
din
gon
e’s
valu
ere
mai
ns
wea
kly
dom
inan
t.T
he
exp
ecte
dpr
ice
isth
us
un
chan
ged
.
In1s
tpr
ice
auct
ion
ifal
lfo
llow
incr
easi
ng
(diff
)st
rate
gyγ
wit
hγ(0)=
0.O
pti
mal
ity
for
ab
idd
erw
ith
valu
ex
agai
nd
icta
tes
x∈
arg
max z
G(z)u(x−
γ(z))
.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
4/26
Not
es
Not
es
Ris
kA
vers
ion
Pro
of.
[Pro
ofco
nti
nu
ed]
Th
efi
rst
ord
erco
nd
itio
nev
alu
ated
atx
req
uir
es
γ′ (
x)=
u(x−
γ(x))
u′ (
x−
γ(x))
g(x)
G(x)
.
Wer
ep
laye
rsri
skn
eutr
alu(x)=
axth
isw
ould
amou
nt
to
β′ (
x)=
(x−
β(x))[g(x)/
G(x)]
.
Sin
ceu(0)=
0&
u′′<
0w
eh
ave
that
[u(y)/
u′ (
y) ]>
y&
thu
s
γ′ (
x)>
(x−
γ(x))[g(x)/
G(x)]
,
wh
ich
imp
lies
that
ifβ(x)>
γ(x)
then
γ′ (
x)>
β′ (
x).
Bu
tsi
nce
γ(0)=
β(0)=
0w
eh
ave
that
γ(x)>
β(x).
Exp
ecte
dp
aym
ent
in1s
tpr
ice
incr
ease
sw
hile
un
chan
ged
in2n
dpr
ice!
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
5/26
Exa
mpl
e:C
RR
AU
tilit
yan
d2
Bid
ders
Su
pp
ose
that
N=
2an
dth
atu
tilit
ysa
tisfi
esu(x)=
xα.
Ifso
,re
lati
veri
skav
ersi
onam
oun
tsto
xu′′ (
x)/
u′ (
x)=
1−
α.
Ob
serv
eth
atF
α=
F1
/α
isal
soa
cdf
&d
enot
eit
sp
df
byf α
=fF
1/
α−1/
α.
Th
e1s
tpr
ice
equ
ilibr
ium
stra
tegy
sati
sfies
γ(0)=
0an
d
γ′ (
x)=
(x−
γ(x))(f(x)/
αF(x))
⇒γ′ (
x)F
(x)1
/α+
γ(x)f(x)F
(x)1
/α−1/
α=
xf(x)F
(x)1
/α−1/
α
⇒γ(x)F
α(x)=∫ x 0
yfα(y)d
y.
Eq
uiv
alen
tto
risk
neu
tral
case
bu
tfo
rF
αin
stea
dof
F.
Sin
ceF
α≤
Fth
eex
pec
ted
reve
nu
eis
grat
erw
ith
risk
aver
sion
.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
6/26
Not
es
Not
es
Exa
mpl
e:C
RR
AU
tilit
yan
d2
Bid
ders
Inth
epr
evio
us
exam
ple
for
α=
1/
2an
dF(x)=
xw
ege
tth
at
γ(x)=
2x/
3≥
x/
2=
β(x).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
7/26
BudgetConstraints
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
8/26
Not
es
Not
es
Bud
get
Con
stra
ints
Con
sid
erth
efo
llow
ing
sett
ing:
Age
nt
typ
eco
nsi
sts
ofa
valu
e-b
ud
get
pai
r(X
i,W
i).
Bid
der
wit
hty
pe(x
,w)
can
nev
erb
idm
ore
than
w.
Ifh
ed
oes
and
def
ault
s,a
big
pen
alty
isap
plie
d.
Age
nt
typ
esar
eiid
and(X
,W)pdf ∼
fon
[0,1]2
.
Ab
idco
nsi
sts
ofa
map
B: [
0,1]2→
R.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
9/26
Sec
ond
Pri
ceA
ucti
on
Lem
ma
Ina
2nd
pric
eau
ctio
nit
isa
wea
kly
dom
inan
tst
rate
gyto
bid
BII(x
,w)=
min{x
,w}.
Pro
of.
(a)
BII(x
,w)>
wis
dom
inat
ed.
Wh
enon
ew
ins:
ifth
ese
con
dh
igh
est
bid
isab
ove
w,
def
ault
and
pen
alty
follo
w,
ifth
ese
con
db
idd
oes
not
exce
edw
,b
idd
ing
ww
ins
asw
ell.
(b)
Ifx≤
wb
idd
eris
un
con
stra
ined
.
Th
esa
me
pro
ofof
clas
sn
otes
IIsh
ows
that
BII(x
,w)=
x.
(c)
Ifx>
wsa
me
argu
men
tsh
ows
that
BII(x
,w)<
wis
dom
inat
ed.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
10/26
Not
es
Not
es
SP
AC
ompu
ting
Rev
enue
Nex
tle
tu
sco
mp
ute
exp
ecte
dre
ven
ue:
For
x′′=
min{x
,w}⇒
BII(x
,w)=
BII(x′′ ,
1).
Th
ese
tof
pla
yers
bid
din
gle
ssth
anx′′
isgi
ven
by
LII(x′′ )
={ (X
,W)|
BII(X
,W)<
BII(x′′ ,
1)} .
Th
epr
obab
ility
that
typ
e(x′′ ,
1)
outb
ids
one
bid
der
is
FII(x′′ )
=∫ L
II(x′′ )
f(X
,W)d
Xd
W.
Th
epr
obab
ility
ofw
inn
ing
thu
sis
:G
II(x′′ )
=(F
II(x′′ ))N−1.
Exp
ecte
du
tilit
yfo
rty
pe(x′′ ,
1)
wh
enb
idd
ing
BII(z
,1)
is
GII(z)x′′−
mII(z
,1).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
11/26
SP
AC
ompu
ting
Rev
enue
Exp
ecte
dp
aym
ents
inth
ese
con
dpr
ice
auct
ion
thu
sam
oun
tto
mII(x
,w)=
mII(x′′ ,
1)=∫ x′′ 0
ydG
II(y).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
12/26
Not
es
Not
es
Fir
stP
rice
Auc
tion
Su
pp
ose
that
for
anin
crea
sin
gfu
nct
ion
βth
eeq
uili
briu
mst
rate
gyis
BI(x
,w)=
min{β
(x),
w}.
Itm
ust
be
that
β(x)<
xor
bid
der
x<
ww
ould
dev
iate
.
To
com
pu
tere
ven
ue,
we
pro
ceed
just
asb
efor
e:
For
x′
such
that
β(x′ )=
min{β
(x),
w}⇒
BI(x
,w)=
BI(x′ ,
1).
Th
ese
tof
pla
yers
bid
din
gle
ssth
anx′
isgi
ven
by
LI(x′ )={ (X
,W)|
BI(X
,W)<
BI(x′ ,
1)} .
Th
epr
obab
ility
that
typ
e(x′ ,
1)
outb
ids
one
bid
der
is
FI(x′ )=∫ L
I(x′ )
f(X
,W)d
Xd
W.
Th
epr
obab
ility
ofw
inn
ing
thu
sis
:G
I(x′ )=
(FI(x′ ))N−1.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
13/26
FP
AC
ompu
ting
Rev
enue
Exp
ecte
dp
aym
ents
inth
efi
rst
pric
eau
ctio
nth
us
amou
nt
to
mI(x
,w)=
mI(x′ ,
1)=∫ x′ 0
ydG
I(y).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
14/26
Not
es
Not
es
BC
Rev
enue
Com
pari
son
Lem
ma
Ifb
idd
ers
are
fin
anci
ally
con
stra
ined
and
the
1st
pric
eau
ctio
nh
asan
equ
ilibr
ium
ofth
efo
rmB
I(x
,w)=
min{β
(x),
w},
then
the
exp
ecte
dre
ven
ue
inth
e1s
tpr
ice
auct
ion
exce
eds
that
ofth
e2n
dpr
ice
auct
ion
.
Pro
of.
Sin
ceβ(x)<
xfo
ran
yx
,it
mu
stb
eth
atLI(x)⊂
LII(x)
FI(x)≤
FII(x)
for
any
x∈(0
,1)⇒
FI
sto
chas
tica
llyd
omin
ates
FII
Exp
ecte
dre
ven
ues
are:
E(R
I)=
E(Y
I(N)
2)>
E(Y
II(N
)2
)=
E(R
II)
Sin
ceE(Y
a(N
)2
)=∫ 1 0
ma(x
,1)f
a(x)d
xfo
ran
ya∈{I
,II }
Bu
dge
tco
nst
rain
tsar
eso
fter
inth
e1s
tpr
ice
case
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
15/26
BC
Rev
enue
Com
pari
son
Th
efo
llow
ing
plo
tcl
earl
yd
isp
lays
the
reve
nu
era
nki
ng:
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
16/26
Not
es
Not
es
AsymmetricBidders
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
17/26
Asy
mm
etri
cB
idde
rsT
wo
Pla
yers
Uni
form
In2n
dpr
ice
bid
din
gon
e’s
valu
ere
mai
ns
opti
mal⇒
effici
ency
.
In1s
tpr
ice
stra
tegy
chan
ges
and
outc
ome
may
be
ineffi
cien
t.
1st
Pri
ceA
uct
ion
Xicd
f ∼Fi=
U[0
,ωi]
for
ap
laye
ri∈{1
,2}.
Str
ateg
ies
are
incr
easi
ng
&d
iffer
enti
able
fun
ctio
ns
βi
[φi=
β−1
i].
Itm
ust
be
that
βi(
0)=
0an
dβi(
ωi)=
bfo
ran
yi∈{1
,2}.
Exp
ecte
dp
ayoff
ofb
idd
eri6=
jof
bid
din
gb
is
Πi(
b,x
i)=
(xi−
b)F
j(φj(
b))
=(x
i−
b)φ
j(b)/
ωj.
Op
tim
alit
yan
dx i
=φi(
b)
imp
ly
φ′ j(
b)(
φi(
b)−
b)=
φj(
b).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
18/26
Not
es
Not
es
Asy
mm
etri
cF
PA
By
rew
riti
ng,
add
ing
and
inte
grat
ing
for
φi(
0)=
0,
(φ′ j(
b)−
1)(
φi(
b)−
b)=
φj(
b)−
φi(
b)+
b,
d db(φ
1(b)−
b)(
φ2(b)−
b)=
2b,
(φ1(b)−
b)(
φ2(b)−
b)=
b2.
Th
ela
steq
uat
ion
imp
lies
that
b=
ω1ω
2/(ω
1+
ω2)
byφi(
b)=
ωi
φ′ i(
b)=
φi(
b)(
φi(
b)−
b)/
b2
byop
tim
alit
y
For
φi(
b)=
ξ i(b)b
+b
this
bec
omes
ξ′ i(
b)b
+ξ i(b)+
1=
ξ i(b)(
ξ i(b)+
1)
⇒ξ i(b)=
1−
k ib2
1+
k ib2
sin
ceξ′ i(
b)
ξ i(b)2−
1=
1 b.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
19/26
Asy
mm
etri
cF
PA
Th
eref
ore
φi(
b)=
2b/(1
+k i
b2)
and
k i=
(1/
ω2 i)−(1
/ω
2 j).
Inve
rtin
gon
ege
tsth
at
βi(
x)=
(1−√ 1−
k ix2)/
k ix
.
Ifω
1>
ω2
not
eth
atβ1(x)<
β2(x).
Th
ew
eake
rp
laye
rb
ids
mor
eag
gres
sive
ly.
Lem
ma
(Effi
cien
cy)
Wit
has
ymm
etri
cally
dis
trib
ute
dva
lues
a2n
dpr
ice
auct
ion
alw
ays
allo
cate
sth
eob
ject
effici
entl
y.A
1st
pric
eau
ctio
nw
ith
pos
itiv
epr
obab
ility
do
esn
ot.
Sin
ceβ1(x
+ε)
<β2(x)
for
ε>
0an
dsm
all
enou
gh.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
20/26
Not
es
Not
es
Asy
mm
etri
cR
even
ueC
ompa
riso
n
Con
sid
eran
exam
ple
wit
hω
1=
1/(1−
α)
and
ω2=
1/(1
+α).
Th
ed
istr
ibu
tion
ofse
llin
gpr
ices
in2n
dpr
ice
auct
ion
is,
for
p∈[0
,ω2],
LII α(p)=
Pr(
min{X
1,X
2}≤
p)=
F1(p)+
F2(p)−
F1(p)F
2(p)
=2p−(1−
α2)p
2
Th
ed
istr
ibu
tion
ofse
llin
gpr
ices
in1s
tpr
ice
auct
ion
is,
for
p∈[0
,1/
2],
LI α(p)=
Pr(
max{β
1(X
1),
β2(X
2)}≤
p)=
F1(φ
1(p))
F2(φ
2(p))
=(1−
α2)(
2p)2
/[ 1−
α2(2
p)4] .
LII α(p)
incr
ease
sw
ith
α,
LI α(p)
dec
reas
esw
ith
α&
EI 0(p)=
EII 0(p).
Th
us
EI α(p)>
EII α(p)
for∀α
>0.
[Un
ran
ked
inge
ner
al].
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
21/26
Resale
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
22/26
Not
es
Not
es
Res
ale
Tw
oA
sym
met
ric
Bid
ders
Will
bid
der
sal
way
sal
loca
teth
eob
ject
effici
entl
yif
resa
leis
per
mit
ted
?
Info
rma
tio
n:
the
win
ner
&al
lb
ids
are
reve
aled
bef
ore
resa
le.
Res
ale
:w
inn
erm
akes
ata
keit
orle
ave
itoff
er.
Ass
um
eth
atF16=
F2
bot
hon
[0,ω
].
Ifβ1,
β2
are
inve
rtib
le,
x 1<
x 2an
dβ1(x
1)>
β2(x
2).
Th
en1
rese
llsth
eob
ject
to2
atpr
ice
p=
x 2⇒
Effi
cien
cy.
Th
en
ext
lem
ma
show
sth
at:
1β1,
β2
are
not
inve
rtib
le;
2th
e1s
tpr
ice
auct
ion
wit
hre
sale
isin
effici
ent.
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
23/26
Lem
ma
Wit
has
ymm
etri
cIP
V,
the
1st
pric
eau
ctio
nfo
llow
edby
resa
led
oes
not
resu
ltin
effici
ency
.
Pro
of.
[Pro
of[1
/3]]
Rec
all
that
βi(
0)=
0an
dβi(
ω)=
bfo
ran
yi∈{1
,2}.
Su
pp
ose
that
βi
isin
vert
ible
and
that
2u
ses
β2⇒
Effi
cien
cy.
Pla
yer
1ex
pec
ted
pay
men
tif
he
bid
sas
ifh
isva
lue
was
z 1is
m1(z
1)=
β1(z
1)F
2(φ
2β1(z
1))−∫ φ 2
β1(z
1)
z 1m
ax{z
1,x
2}d
F2(x
2).
Sec
ond
term
isp
osit
ive
ifz 1
<φ2β1(z
1)
and
neg
ativ
eif
z 1>
φ2β1(z
1).
Typ
ex 1
exp
ecte
dp
ayoff
ifh
eb
ids
asif
his
valu
ew
asz 1
is
x 1F2(z
1)−
m1(z
1).
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
24/26
Not
es
Not
es
Pro
of.
[Pro
of[2
/3]]
Ifβ1
isin
vert
ible
opti
mal
ity
atz 1
=x 1
req
uir
esth
at
m′ 1(x
1)=
x 1f 2(x
1).
Op
tim
alit
yan
dm
1(0)=
0im
ply
that
m1(x
1)=∫ x 1 0
x 2d
F2(x
2).
Effi
cien
cy⇒
[1st
pric
e+
resa
le]∼ =
[2nd
pric
e]!
Ifβ1(ω
)=
b⇒
φ2β1(ω
)=
ω⇒
b=
E[X
2]
sin
ce
m1(ω
)=
bF2(ω
)=∫ ω 0
x 2d
F2(x
2).
Sw
itch
ing
pla
yers
one
gets
b=
E[X
1]6=
E[X
2].
Aco
ntr
adic
tion
!
Th
us
β1
&β2
can
not
be
bot
hst
rict
lyin
crea
sin
g!
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
25/26
Pro
of.
[Pro
of[3
/3]]
Ass
um
eb
oth
con
tin
uou
s,n
on-d
ecre
asin
g&
β2(x
2)=
b2
for
∀x2∈[x′ 2,
x′′ 2].
Cla
im:∃x
1∈(x′ 2,
x′′ 2)
such
that
β1(x
1)≥
b2.
Oth
erw
ise
x′′ 1=
[min
xx
s.t.
β1(x)=
b2]⇒
x′′ 1≥
x′′ 2
.
Als
ox′′ 1>
b2
byop
tim
alit
yof
typ
ex′′ 2
wh
ich
win
sat
mos
tx′′ 1
.
For
2ε<
x′′ 1−
b2,
typ
ex′′ 1−
εw
ould
profi
tby
bid
din
gb2+
ε.
At
infi
nit
esim
alco
stw
inag
ain
stan
yx 2∈[x′ 2,
x′′ 2]⇒
Aco
ntr
adic
tion
!
Th
us
let
x 1∈(x′ 2,
x′′ 2)
be
such
that
β1(x
1)≥
b2.
Ifx 2∈(x
1,x′′ 2]
bid
der
1w
ins
auct
ion
ineffi
cien
tly!
Op
tim
alit
yof
1in
resa
leim
plie
sth
atpr
ice
p>
x 1.
Th
us
wit
hp
osit
ive
prob
abili
typ>
x 2>
x 1⇒
Ineffi
cien
t!
Nava
(LSE)
IPV
Exten
sions
MichaelmasTerm
26/26
Not
es
Not
es
Mec
hani
smD
esig
nIP
VE
C31
9–
Lec
ture
Lec
ture
VI
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
1/29
Intr
odu
ctio
n
How
tod
esig
nse
llin
gm
ech
anis
ms
wit
hd
esir
able
prop
erti
es:
Rev
enu
eM
axim
izat
ion
Effi
cien
cy
Bal
ance
dB
ud
get
For
the
mom
ent
reta
inth
eIP
Vse
tup
:
Np
oten
tial
bu
yers
,1
auct
ion
eer
and
1in
div
isib
leob
ject
.
Bu
yeri’
sva
luat
ionXicd
f ∼Fi
on[0
,ωi]=
Xi
ind
epen
den
tly.
Th
eau
ctio
nee
rva
lues
the
obje
ctat
zero
.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
2/29
Not
es
Not
es
Mec
hani
sms
Rev
elat
ion
Pri
ncip
lean
dR
even
ueE
quiv
alen
ce
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
3/29
Mec
hani
sms
Am
ech
an
ism
con
sist
sof
trip
le(B
,π,µ
):
1M
essa
gesp
aces
Bi;
2A
nal
loca
tion
rule
π:B→
∆N={ p∈
RN +| ∑
N i=1pi≤
1} ;
3A
pay
men
tru
leµ
:B→
RN
.
Ina
1st
or2n
dpr
ice
auct
ion
,ex
cep
tfo
rti
es,
sati
sfy:
Bi=
Xi;
πi(b)=
I(b
i>
max
j6=ibj)
;
µI i(b)=
I(b
i>
max
j6=ibj)bi
for
FP
A;
µII i(b)=
I(b
i>
max
j6=ibj)
max
j6=ibj
for
SP
A.
Ast
rate
gyis
am
apβi
:Xi→
Bi.
Eq
uili
briu
mif
βi(x i)
max
imiz
esp
ayoff
sgi
ven
β−i
for
anyx i
andi.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
4/29
Not
es
Not
es
Rev
elat
ion
Pri
ncip
le
Ad
irec
tm
ech
an
ism
con
sist
sof
pai
r(Q
,M):
1A
nal
loca
tion
rule
Q:X→
∆N
;
2A
pay
men
tru
leM
:X→
RN
.
The
orem
(Rev
elat
ion
Pri
ncip
le–
Mye
rson
)
For
any
mec
han
ism
and
any
equ
ilibr
ium
for
that
mec
han
ism
,th
ere
exis
tsa
dir
ect
mec
han
ism
inw
hic
h:
1it
isan
equ
ilibr
ium
for
each
bu
yer
tore
por
th
isva
lue
tru
thfu
lly;
2ou
tcom
esar
eth
esa
me
asin
equ
ilibr
ium
ofth
eor
igin
alm
ech
anis
m.
Pro
of.
Defi
neQ(x)=
π(β(x))
andM(x)=
µ(β(x))
for
anyx∈
X.
Th
enco
nd
itio
ns
1an
d2
are
met
byd
efin
itio
nof
equ
ilibr
ium
.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
5/29
Ince
ntiv
eC
ompa
tibi
lity
Defi
niti
on
Con
sid
era
dir
ect
mec
han
ism
(Q,M
).
Th
ew
inn
ing
prob
abili
tyan
dth
eex
pec
ted
pay
men
tw
hen
rep
orti
ngz i
,if
oth
ers
are
sin
cere
,am
oun
tto
qi(z i)=∫ X−i
Qi(z i
,x−i)dF−i(x −
i),
mi(z i)=∫ X−i
Mi(z i
,x−i)dF−i(x −
i).
Th
eex
pec
ted
pay
offof
typ
ex i
isgi
ven
by:
x iqi(z i)−
mi(z i).
Th
em
ech
anis
mis
said
tob
ein
cen
tive
com
pa
tib
le(I
C)
if
x iqi(x i)−
mi(x i)=
max
z i∈X
i[xiq
i(z i)−
mi(z i)]≡
Ui(x i).
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
6/29
Not
es
Not
es
Ince
ntiv
eC
ompa
tibi
lity
Equ
ival
ence
Lem
ma
Ad
irec
tm
ech
anis
mis
ince
nti
veco
mp
atib
leif
and
only
ifU′ i(x i)=
qi(x i)
andU′′ i(x
i)≥
0fo
ral
lx i
and
alli.
Pro
of.
If:
IfIC
hol
ds
we
mu
sth
ave
that
Ui(z i)≥
z iqi(x i)−
mi(x i)=
Ui(x i)+(z
i−
x i)q
i(x i).
Th
esa
me
ineq
ual
ity
mu
sth
old
swit
chin
gx i
wit
hz i
and
thu
s
(zi−
x i)q
i(z i)≥
Ui(z i)−
Ui(x i)≥
(zi−
x i)q
i(x i).
(1)
Div
idin
gby
(zi−
x i)
and
taki
ng
limit
sas
z i→
x iyi
eld
s
limz i→
x i
Ui(z i)−
Ui(x i)
(zi−
x i)
=U′ i(x i)=
qi(x i).
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
7/29
Ince
ntiv
eC
ompa
tibi
lity
Equ
ival
ence
Pro
of.
Con
dit
ion
(1)
also
imp
lies
that
(zi−
x i)(qi(z i)−
qi(x i))≥
0,
wh
ich
iseq
uiv
alen
tto
q′ i(x i)≥
0,an
dth
usU′′ i(x
i)≥
0.
On
lyIf
:IfU′ i(x i)=
qi(x i),
the
exp
ecte
dp
ayoff
dep
end
sju
ston
Qsi
nce
:
Ui(x i)=
Ui(
0)+∫ x i 0
qi(t)dt.
Th
us,
for
anyx i
and
anyz i
,w
ege
tth
at
Ui(z i)−
Ui(x i)=∫ z i x i
qi(t)dt≥
(zi−
x i)q
i(x i),
wh
ere
the
ineq
ual
ity
hol
ds
asU′′ i(x
i)=
q′ i(x i)≥
0.B
ut
this
isIC
.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
8/29
Not
es
Not
es
Ince
ntiv
eC
ompa
tibi
lity
Com
men
ts
Th
eif
par
tof
the
lem
ma
sim
ply
amou
nte
dto
the
En
velo
pe
Th
eore
m.
Th
eex
pec
ted
chan
geis
pay
offUi(x i)−
Ui(z i)
isin
dep
end
ent
ofM
.
Pay
offUi(x i)
isco
nve
xas
itis
the
up
per
-en
velo
pe
oflin
ear
fun
ctio
ns:
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
9/29
Rev
enue
Equ
ival
ence
The
orem
(Rev
enue
Equ
ival
ence
–M
yers
on)
Ifth
ed
irec
tm
ech
anis
m(Q
,M)
isIC
,th
enfo
ran
yi
andx i
the
exp
ecte
dp
aym
ent
is
mi(x i)=
mi(
0)+
x iqi(x i)−∫ x i 0
qi(t)dt.
Pro
of.
Sin
ceUi(x i)=
x iqi(x i)−
mi(x i)
andUi(
0)=−m
i(0),
ICim
plie
sth
at
x iqi(x i)−
mi(x i)=−m
i(0)+∫ x i 0
qi(t)dt.
Th
us
exp
ecte
dp
aym
ents
inan
ytw
oIC
mec
han
ism
sw
ith
the
sam
eal
loca
tion
rule
are
equ
ival
ent
up
toco
nst
ant.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
10/29
Not
es
Not
es
Indi
vidu
alR
atio
nalit
y
Am
ech
anis
mis
said
tob
ein
div
idu
ally
rati
on
al
(IR
)if
Ui(x i)≥
0fo
ral
lx i
.
Wh
enIC
hol
ds,
IRsi
mp
lifies
tore
qu
irin
g
Ui(
0)≥
0.
Intu
rnth
isis
equ
ival
ent
to
mi(
0)≤
0.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
11/29
Opt
imal
Mec
hani
sms
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
12/29
Not
es
Not
es
Exp
ecte
dR
even
ue
Am
ech
anis
mis
op
tim
al
ifit
max
imiz
esre
ven
ues
sub
ject
toIC
&IR
.
Ifth
ese
ller
use
sa
dir
ect
mec
han
ism
exp
ecte
dp
aym
ents
are
E(m
i(Xi))=
mi(
0)+∫ ω i 0
x iqi(x i)dFi(x i)−∫ ω i 0
∫ x i 0qi(t)dtdFi(x i).
Not
ice
that ∫ ω i 0
∫ x i 0qi(t)dtdFi(x i)=∫ ω i 0
∫ ω i tdFi(x i)q
i(t)dt=
=∫ ω i 0
(1−
Fi(x i))qi(x i)
f i(x
i)dFi(x i).
Th
eref
ore,
the
selle
r’s
exp
ecte
dre
ven
ues
are
E(R
)=
∑i∈
N
[ mi(
0)+∫ X
[ x i−
1−
Fi(x i)
f i(x
i)
] Qi(x)dF(x)] .
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
13/29
Vir
tual
Val
uati
ons
Defi
ne
the
virt
ua
lva
lua
tio
nas
ψi(x i)=
x i−
1−
Fi(x i)
f i(x
i)=
x i−
1
λi(x i)
.
Vir
tual
valu
atio
nca
nb
ein
terp
rete
das
mar
gin
alre
ven
ue.
Als
oob
serv
eth
at E(ψ
i(Xi))=∫ ω i 0
[ y−
1−
Fi(y)
f i(y)
] dFi(y)
=E(X
i)−∫ ω i 0
∫ ω i yf i(x)dxdy
=E(X
i)−∫ ω i 0
∫ x 0dyf
i(x)dx
=E(X
i)−∫ ω i 0
xfi(x)dx=
0.
Ad
esig
npr
oble
mis
reg
ula
rif
ψ′ i>
0.[λ′ i>
0⇒
ψ′ i>
0]Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
14/29
Not
es
Not
es
Opt
imal
Mec
hani
sms
Th
eo
pti
ma
lm
ech
anis
mso
lves
max
(Q,M
)∑
i∈Nm
i(0)+∫ X
[ ∑i∈
Nψi(x i)Q
i(x) ]dF(x)
s.t.
(IC
)q′ i(x)≥
0an
d(I
R)
mi(
0)≤
0.
Am
ech
anis
mis
opti
mal
ifth
efo
llow
ing
two
con
dit
ion
sh
old
{ Qi(x)>
0⇔
ψi(x i)=
max
jψj(x j)≥
0(1
)M
i(x)=
Qi(x)x
i−∫ x i 0
Qi(t,x −
i)dt
(2)
Ob
serv
eth
at:
(1)
+re
gula
r⇒
IC,
(2)⇒
mi(
0)=
0⇒
IR,
(1-2
)m
axim
ize
reve
nu
e.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
15/29
Opt
imal
Mec
hani
sms
(for
Reg
ular
Pro
blem
s)
Defi
ne
the
smal
lest
valu
eth
atw
ins
agai
nst
x −i
by:
y i(x−i)=
inf{ z i|
ψi(z i)≥
max{m
axj6=
iψj(x j),
0}} =
=m
ax{m
axj6=
iψ−1
i(ψ
j(x j))
,ψ−1
i(0)}
.
Lem
ma
(Opt
imal
Mec
hani
sm–
Mye
rson
)
Ifth
ed
esig
npr
oble
mis
regu
lar
ther
eis
opti
mal
mec
han
ism
sati
sfyi
ng
Qi(x)=
{ 1if
x i>
y i(x−i)
0if
x i<
y i(x−i)
Mi(x)=
{ y i(x−i)
ifQ
i(x)=
10
ifQ
i(x)=
0
Pro
of.
By
sim
plf
yin
gth
epr
evio
us
expr
essi
ons
and
sin
ce∫ x i 0
Qi(t,x −
i)dt=
{ x i−
y i(x−i)
ifx i
>y i(x−i)
0if
x i<
y i(x−i)
.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
16/29
Not
es
Not
es
Opt
imal
Mec
hani
sms
Sym
met
yim
plie
s:ψi=
ψan
dy i(x−i)=
max{m
axj6=
ix j
,ψ−1(0)}
.
Lem
ma
(Sym
met
ric
OM
)
Ifth
ed
esig
npr
oble
mis
regu
lar
and
sym
met
ric
then
ase
con
dpr
ice
auct
ion
wit
hre
serv
epr
icer=
ψ−1(0)
isan
opti
mal
mec
han
ism
.
Th
eo
pti
ma
lm
ech
an
ism
isin
effici
ent
bec
ause
:
1ob
ject
isu
nso
ldw
hen
max
jψj(x j)<
0;
2x i
>y i(x−i)
do
esn
’tim
ply
x i≥
max
j6=ix −
j.
OM
favo
rsd
isad
van
tage
dp
laye
rsto
mot
ivat
eot
her
sto
bid
hig
her
.
IfF1
haz
ard
rate
dom
inat
esF2
[F1
sto
chas
tica
llyd
omin
ates
F2]
λ1(x)≤
λ2(x)⇒
ψ1(x)≤
ψ2(x).
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
17/29
Effi
cien
tM
echa
nism
s
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
18/29
Not
es
Not
es
Pro
per
ties
:B
alan
ced
Bud
get
&E
ffici
ency
Con
sid
erva
lues
inX
i=
[αi,
ωi]⊂
R.
Ap
aym
ent
rule
Mis
said
tob
eb
ala
ced
bu
dg
et(B
B)
if
∑i∈
NM
i(x)=
0.
An
allo
cati
onru
leQ
issa
idto
be
effici
ent
if
Q(x)∈
arg
max
π∈
∆N
∑j∈
Nπjx
j.
For
Qeffi
cien
td
efin
eth
em
axim
ized
soci
alva
lues
by
W(x)=
∑j∈
NQ
j(x)x
j,
W−i(x)=
∑j∈
N\iQ
j(x)x
j.
A2n
dpr
ice
auct
ion
wit
hou
tre
serv
epr
ice
iseffi
cien
tw
hen
αi=
0.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
19/29
VC
GE
ffici
ent
Mec
hani
sm
Defi
niti
on(V
icke
ryC
lark
eG
rove
s)
AV
CG
mec
han
ism
(Qv
,Mv)
con
sist
sof
aneffi
cien
tal
loca
tion
rule
and
ofa
pay
men
tru
leth
atsa
tisfi
es
Mv i(x)=
W(α
i,x −
i)−
W−i(x).
VC
Gis
ICsi
nce
tru
th-t
ellin
gis
opti
mal
byeffi
cien
cyof
Qv
,
Qv i(z
i,x −
i)x i−
Mv i(z
i,x −
i)
=Q
v i(z
i,x −
i)x i+
W−i(z i
,x−i)−
W(α
i,x −
i)
=∑
j∈NQ
v j(z
i,x −
i)x j−
W(α
i,x −
i)≤
W(x)−
W(α
i,x −
i).
VC
Gex
pec
ted
pay
offin
equ
ilibr
ium
isin
crea
sin
gan
dis
Uv i(x
i)=
E[W
(xi,X−i)−
W(α
i,X−i)].
VC
Gis
IRsi
nce
Uv i
isin
crea
sin
gan
dU
v i(α
i)=
0.Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
20/29
Not
es
Not
es
VC
GM
axim
izes
Rev
enue
s
Lem
ma
(Kri
shna
Per
ry)
Am
ong
all
mec
han
ism
for
allo
cati
ng
anob
ject
that
are
effici
ent,
ICan
dIR
,th
eV
CG
mec
han
ism
max
imiz
esth
eex
pec
ted
pay
men
tof
each
agen
t.
Pro
of.
Su
pp
ose
also
(Q∗ ,M∗ )
sati
sfies
the
thre
epr
oper
ties
.
Effi
cien
cyre
qu
ires
that
Q∗=
Qv
.
Sin
ceb
oth
mec
han
ism
sar
eIC
reve
nu
eeq
uiv
alen
ceth
eore
mh
old
s.
Th
us
pay
offs
diff
erby
no
mor
eth
ana
con
stan
tc i=
U∗ i(x
i)−
Uv i(x
i).
By
IRc i≥
0si
nce
Uv i(α
i)=
0.
Bu
tU∗ i(x
i)≥
Uv i(x
i)an
dQ∗=
Qv
imp
lyth
at
E[M∗ i(x
i,X−i)]≤
E[M
v i(x
i,X−i)].
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
21/29
AG
VB
alan
ced
Bud
get
Mec
hani
sm
Defi
niti
on(A
rrow
d’A
spre
mon
tG
erar
d-V
aret
AG
V)
An
AG
Vm
ech
anis
m(Q
a,M
a)
con
sist
sof
aneffi
cien
tal
loca
tion
rule
and
ofa
pay
men
tru
leth
atsa
tisfi
es
Ma i(x)=
1N−1
∑j6=
iE[W−j(x j
,X−j)]−
E[W−i(x i
,X−i)].
AG
Vex
pec
ted
pay
offin
equ
ilibr
ium
isin
crea
sin
gan
dis
Ua i(x
i)=
E[W
(xi,X−i)]−
1N−1
∑j6=
iE[W−j(X)]
.
AG
Vb
alan
ces
the
bu
dge
tsi
nce
∑i∈
N1
N−1
∑j6=
iE[W−j(x j
,X−j)]=
∑i∈
NE[W−i(x i
,X−i)].
AG
Vis
ICfo
rth
esa
me
reas
onV
CG
was
∑j∈
NQ
a j(z
i,x −
i)x j≤
∑j∈
NQ
a j(x)x
j=
W(x).
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
22/29
Not
es
Not
es
Effi
cien
t,B
B,
IC,
IRM
echa
nism
Lem
ma
(Kri
shna
Per
ry)
Th
ere
exis
tsan
effici
ent,
ICan
dIR
mec
han
ism
that
isB
Biff
the
VC
Gm
ech
anis
mre
sult
sin
ap
osit
ive
exp
ecte
dsu
rplu
s.
Pro
of.
Pre
viou
sle
mm
ash
ows
that
VC
G’s
pos
itiv
esu
rplu
sis
nec
essa
ry.
Su
ffici
ency
:B
yIC
&effi
cien
cy⇒
reve
nu
eeq
uiv
alen
ce⇒
form∈{v
,a}:
cm i=
E[W
(xi,X−i)]−
Um i(x
i).
Su
pp
ose
that
VC
Gru
ns
surp
lus
E[ ∑
i∈NM
v i(X
) ]≥
E[ ∑
i∈NM
a i(X
) ]=
0.
Th
isim
plie
s∑
i∈Ncv i≥
∑i∈
Nca i.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
23/29
Pro
ofC
onti
nued
Pro
of.
Defi
nedi=
ca i−
cv i
fori>
1an
dd1=−
∑i>
1di.
Not
ice
that
d1=
∑i>
1(c
v i−
ca i)≥
(ca 1−
cv 1).
Con
sid
erth
eB
Bp
aym
ent
rule
M∗
defi
ned
by
M∗ i(x)=
Ma i(x)−
di.
IfQ∗=
Qa,
the
mec
han
ism
(Q∗ ,M∗ )
iseffi
cien
t.
Itis
also
ICsi
nce
pay
offs
diff
erfr
om(Q
a,M
a)
byad
dit
ive
con
stan
t.
Ad
dit
ion
ally
(Q∗ ,M∗ )
isIR
bec
ause
U∗ i(x
i)=
Ua i(x
i)+
di≥
Ua i(x
i)+
ca i−
cv i=
Uv i(x
i)≥
0.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
24/29
Not
es
Not
es
Bila
tera
lT
rade
Pro
blem
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
25/29
Bila
tera
lT
rade
Pro
blem
1b
uye
rw
ith
valu
eV∼
FV
on[v
,v].
1se
ller
wit
hco
stsC∼
FC
on[c
,c].
Su
pp
osev<
can
dc≤
v[t
rad
eis
som
etim
esb
enefi
cial
].
Ifpr
od
uct
ion
&tr
ade
occ
ur
bu
yer
pay
sM
Van
dse
ller
rece
ives
MC
.
Inth
issc
enar
iop
ayoff
sb
ecom
eV−
MV
andM
C−
C.
BB
req
uir
esM
V=
MC
.
Effi
cien
cyre
qu
ires
trad
eto
take
pla
cew
hen
ever
V>
C.
Lem
ma
(Mye
rson
Sat
tert
hwai
te)
Inth
eb
ilate
ral
trad
epr
oble
m,
ther
eex
ists
no
mec
han
ism
that
iseffi
cien
t,IC
,IR
&B
B.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
26/29
Not
es
Not
es
Bila
tera
lT
rade
Pro
blem
Pro
of.
Con
sid
erth
eeffi
cien
tp
ayoff
sd
eriv
edfo
ran
ypr
ofile
x=
(v,−
c):
QV(x)=
1Q
C(x)=
1
uV(x)
v0
uC(x)
−c
0
To
char
acte
rize
the
VC
Gm
ech
anis
m,
not
eth
ateffi
cien
cyre
qu
ires
:
Qv V(x)=
{ 1if
v−
c>
00
ifv−
c<
0=
I(v
>c)
Qv C(x)=
{ 0if
v−
c>
01
ifv−
c<
0=
I(v
<c)
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
27/29
Pro
of.
For
such
defi
nit
ion
s,so
cial
wel
fare
and
resi
du
also
cial
wel
fare
amou
nt
to:
W(x)=
(v−
c)Q
v V(x)+
0Qv C(x)=
(v−
c)I
(v>
c)=
max{v−
c,0}
W−V(x)=−cQ
v V(x)
&W−C(x)=
Qv V(x)v
Hen
ce,
the
VC
Gal
loca
tion
and
pay
men
tssa
tisf
y:
QV(v
,c)=
I(v
>c)=
1−
QC(v
,c)
MV(v
,c)=
W(v
,−c)−
W−V(x)=
(v−
c)Q
v V(v
,−c)+
cQv V(x)
=(v−
c)I
(v>
c)+
cI(v
>c)=
I(v
>c)
max{c
,v}
MC(v
,c)=
W(v
,−c)−
W−C(x)=
(v−
c)Q
v V(1
,−c)−
vQv V(x)
=(v−
c)I
(v>
c)−
vI(v
>c)=−
I(v
>c)
min{v
,c}
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
28/29
Not
es
Not
es
Pro
of.
Th
eV
CG
exp
ecte
dp
ayoff
ofth
eb
uye
rof
clai
min
gto
bez
amou
nts
to:
E[(v−
max{C
,1})
I(z
>C)]=∫ z 0
vdFC(c)−∫ z 0
max{c
,v}d
FC(c)
=∫ z 0
vdFC(c)−∫ v 0
vdFC(c)−∫ z v
cdFC(c)
=vF
C(z)−
vFC(v)−∫ z v
cdFC(c).
Itis
effici
ent.
Itis
ICsi
nce
tru
thte
llin
gis
opti
mal
,
E[(v−
max{C
,v})
I(z
>C)]=
vFC(z)−
vFC(v)−∫ z v
cdFC(c)
[FO
C](v−
z)fC(z)≤
0⇒
v=
z.
Itis
IRsi
nce
Uv V(v)≥
0an
dU
v C(c)≥
0fo
ran
yv
&c
.
Itis
not
BB
sin
cem
ax{c
,v}−
min{v
,c}<
0ifv>
c.
Sin
ceV
CG
isre
ven
ue
max
imiz
ing
amon
gst
effici
ent,
ICan
dIR
mec
han
ism
s,th
ecl
aim
hol
ds
(pro
veas
inK
rish
na
Per
ryfi
rst
lem
ma)
.
Nava
(LSE)
Mechanism
DesignIPV
MichaelmasTerm
29/29
Not
es
Not
es
Inte
rdep
ende
ntV
alue
Auc
tion
sE
C31
9–
Lec
ture
Not
esV
II
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
1/24
Interdependent
Value
Models
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
2/24
Not
es
Not
es
Inte
rdep
ende
ntV
alue
s
Eac
hb
idd
erh
asso
me
info
rmat
ion
abou
tth
eva
lue
ofth
eob
ject
.
Th
ein
form
atio
nof
any
pla
yer
affec
tsva
luat
ion
sfo
ral
lot
her
s.
Let
Xi∈[0
,ωi]
den
ote
pla
yer
i’s
sign
alan
dX
=(X
1,.
..,X
N).
Pla
yer
i’s
valu
eVi
dep
end
son
all
sign
als
Vi=
v i(X
).
Ass
um
ev i
isn
on-d
ecre
asin
gin
Xj,
incr
easi
ng
inXi
&2-
diff
eren
tiab
le.
Inge
ner
aln
otal
lin
form
atio
nre
veal
edby
sign
als
&va
lues
are
E[V
i|X=
x]=
v i(x).
Ass
um
ev i(0
,...
,0)=
0,E(V
i)<
∞an
dri
skn
eutr
alit
y.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
3/24
Com
mon
Val
ues
Inon
eex
trem
eca
seof
the
mo
del
valu
esar
epr
ivat
eif
v i(X
)=
Xi.
Inth
eop
pos
ite
extr
eme
valu
atio
ns
are
com
mon
(CV
)if
v i(X
)=
v(X
)=
V.
Th
eex
-pos
tva
lue
isco
mm
onto
all
and
un
know
nto
anyo
ne.
Asp
ecia
lca
seco
mm
only
use
dh
as:
V∼
F,
Xi|V
iid ∼
H(V
),
E[X
i|V=
v]=
v.
Th
ism
od
elis
use
dto
des
crib
eoi
ld
rilli
ng
leas
eau
ctio
ns.
[Min
eral
Rig
hts
]
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
4/24
Not
es
Not
es
The
Win
ner’
sC
urse
Con
sid
erth
esp
ecia
lca
seof
the
CV
mo
del
.
Ass
um
eb
idd
ers
pla
ya
1st
pric
eau
ctio
n.
i’s
esti
mat
eof
valu
eb
efor
eth
eau
ctio
nis
E[V|X
i=
x i].
Bu
tif
he
isan
nou
nce
dw
inn
erh
ises
tim
ate
bec
omes
E[V|X
i=
x i,Y
1<
x i]<
E[V|X
i=
x i].
Th
ough
each
sign
alis
un
bia
sed
E[X
i|V=
v]=
v,
the
max
isn
ot
E[m
axXi|V
=v]>
max
E[X
i|V=
v]=
v.
Th
ew
inn
er’s
curs
eco
nsi
stof
the
failu
reto
acco
un
tfo
rth
iseff
ect.
Sh
adin
gth
eb
ids
prev
ents
this
from
hap
pen
ing
ineq
uili
briu
m.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
5/24
Affi
liate
dS
igna
ls
Let
x∨
x′={m
ax{x
i,x′ i}}
i∈N
and
x∧
x′={m
in{x
i,x′ i}}
i∈N
.
Sig
nal
s(X
1,.
..,X
N)
are
dis
trib
ute
dac
cord
ing
toa
join
td
ensi
tyf
.
(X1,.
..,X
N)
are
affi
liate
dif
fsa
tisfi
esfo
ran
yx
,x′∈
X
f(x∨
x′ )
f(x∧
x′ )≥
f(x)f(x′ )
.
Affi
liati
onim
plie
sth
at∂2
lnf
/∂x
i∂x j≥
0if
i6=
jan
dth
at:
1(X
1,Y
1,.
..,Y
N−1)
are
affilia
ted
for
Yk=
(k-h
igh
est
vari
able
inX−1);
2fo
rY1|X
1∼
G(X
1)
ifx 1
>x′ 1
g(y|x1)/
G(y|x1)≥
g(y|x′ 1)
/G(y|x′ 1)
;
3if
γis
anin
crea
sin
gfu
nct
ion
and
x 1>
x′ 1
E[γ(Y
1)|
X1=
x 1]≥
E[γ(Y
1)|
X1=
x′ 1]
.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
6/24
Not
es
Not
es
Sym
met
ric
Mo
del
Sta
rtth
ean
alys
isof
inte
rdep
end
ent
valu
esby
assu
min
gsy
mm
etry
:
Val
uat
ion
map
isco
mm
onto
all
pla
yers
v i(X
)=
u(X
i,X−i)
.
and
sym
met
ric
inX−i.
Th
us
for
any
per
mu
tati
onπ
X−i
ofX−i
u(X
i,X−i)=
u(X
i,π
X−i)
.
Join
td
ensi
tyf
issy
mm
etri
cin
Xon
[0,ω
]Nan
dsi
gnal
sar
eaffi
liate
d.
Exp
ecte
dva
lue
tob
idd
eri
wit
hsi
gnal
xfa
cin
ga
hig
hes
tsi
gnal
ofy
is
v(x
,y)=
E[u(X
i,X−i)|X
i=
x,
Y1=
y].
Itis
non
-dec
reas
ing
iny
,in
crea
sin
gin
xan
dv(0
,0)=
0=
u(0).
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
7/24
SecondPrice
Auction
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
8/24
Not
es
Not
es
Sec
ond
Pri
ceA
ucti
on
Lem
ma
(Milg
rom
Web
er)
Sym
met
ric
equ
ilibr
ium
stra
tegi
esin
2nd
pric
eau
ctio
nar
eβ
II(x)=
v(x
,x).
Pro
of.
Let
any
j6=
1u
seβ=
βII
,p
ayoff
sof
pla
yer
1ty
pe
xof
bid
din
gb
are
Π(b|x)=∫ β−
1(b)
0[v(x
,y)−
β(y)]
dG(y|x).
For
Y1=
max
j6=1
Xj,
Y1|X
1∼
G(X
1)
and
β(x)=
v(x
,x).
For
ψ(b)=
β−1(b),
FO
Cfo
rb
imp
lyth
at
[v(x
,ψ(b))−
v(ψ
(b),
ψ(b))]g(ψ
(b)|
x)ψ′ (
b)=
0.
Wh
ich
intu
rnim
plie
sx=
ψ(b)
oreq
uiv
alen
tly
b=
β(x).
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
9/24
Rem
arks
onS
econ
dP
rice
Sym
met
ric
Equ
ilibr
ium
βII
issu
chth
atb
idd
er1
wit
hva
lue
xbr
eaks
even
byw
inn
ing
agai
nst
ah
igh
est
bid
βII(x)
sin
ce
E(V
1|X
1=
Y1=
x)=
v(x
,x)=
βII(x).
βII
isn
ota
wea
kly
dom
inan
tst
rate
gy.
βII
isth
eu
niq
ue
sym
met
ric
equ
ilibr
ium
.
Bu
tth
ere
may
be
diff
eren
tas
ymm
etri
ceq
uili
bria
[eve
nin
the
sym
met
ric
case
].
IPV
:2n
dpr
ice
and
En
glis
hau
ctio
ns
are
stra
tegi
cally
equ
ival
ent.
Not
anym
ore
sin
ced
rop
pin
gou
tre
veal
sin
form
atio
nab
out
valu
atio
ns.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
10/24
Not
es
Not
es
EnglishAuction
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
11/24
Eng
lish
Auc
tion
[Op
enA
scen
ding
Pri
ce]
Th
efo
rmat
ofE
ngl
ish
Au
ctio
nco
nsi
der
edh
ere
has
:
the
auct
ion
eer
rais
ing
the
pric
egr
adu
ally
star
tin
gat
0;
bid
der
sle
avin
gth
ero
omw
hen
they
stop
par
tici
pat
ing
inth
eau
ctio
n;
bid
der
sn
otal
low
edto
reen
ter
the
auct
ion
once
they
qu
it;
all
bid
der
sin
the
room
obse
rvin
gat
wh
ich
pric
eex
its
occ
ur;
auct
ion
end
ing
wh
enon
eb
idd
erre
mai
ns;
the
pric
ep
aid
bei
ng
that
ofth
ela
stex
it.
Th
eeq
uili
briu
md
iscu
ssed
nex
tis
rob
ust
toa
chan
gein
form
atth
atal
low
sp
laye
rsto
call
out
pric
es[A
very
1998
].
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
12/24
Not
es
Not
es
Str
ateg
ies
inan
Eng
lish
Auc
tion
Str
ateg
yte
llw
hen
tod
rop
out
dep
end
ing
onse
vera
lva
riab
les:
1In
div
idu
alsi
gnal
;
2N
um
ber
ofp
laye
rsin
the
room
;
3P
rice
sat
wh
ich
the
rem
ain
ing
pla
yers
left
the
room
.
Con
sid
erth
efo
llow
ing
stra
tegy
for
k<
N:
βN(x)=
u(x
,...
,x),
βk(x
,pk+1,.
..,p
N)=
u(x
,...
,x,x
k+1,.
..,x
N),
(1)
for
x k+1
defi
ned
byβk+1(x
k+1,p
k+2,.
..,p
N)=
pk+1.
E.g
.βN−1(x
,pN)=
u(x
,...
,x,x
N)
for
βN(x
N)=
pN
.
Wh
ich
can
be
don
esi
nce
∂βk/
∂x>
0.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
13/24
Sym
met
ric
Equ
ilibr
ium
ofE
nglis
hA
ucti
on
Lem
ma
(Milg
rom
Web
er)
Sym
met
ric
ex-p
ost
equ
ilibr
ium
stra
tegi
esin
the
En
glis
hau
ctio
nsa
tisf
y(1
).
Pro
of.
Con
sid
erb
idd
er1
wit
hsi
gnal
X1=
xfa
cin
gY1=
y 1>
...>
YN−1=
y N−1.
Ifx>
y 1on
ew
ins
the
obje
ctan
dp
ays
the
pric
eof
the
last
dro
pou
t:
p2=
β2(y
1,p
3,.
..,p
N)=
u(y
1,y
1,.
..,y
N−1).
On
eca
nn
otaff
ect
pric
e⇒
he
pref
ers
win
nin
gsi
nce
x>
y 1im
plie
s
u(x
,y1,.
..,y
N−1)−
u(y
1,y
1,.
..,y
N−1)>
0.
Ifx<
y 1h
elo
ses
bu
tis
bet
ter
offsi
nce
the
reve
rse
ineq
ual
ity
hol
ds.
So,
βis
ind
epen
den
tof
f,
and
isan
ex-p
ost
equ
ilibr
ium
.Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
14/24
Not
es
Not
es
First
Price
Auction
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
15/24
Fir
stP
rice
Auc
tion
Lem
ma
(Milg
rom
Web
er)
Sym
met
ric
equ
ilibr
ium
stra
tegi
esin
1st
pric
eau
ctio
nar
e
βI (
x)=∫ x 0
v(y
,y)d
L(y|x)
for
L(y|x)=
exp
( −∫ x y
g(t|t)
G(t|t)
dt) .
Pro
of.
[Pro
of1/
3](1
)L(x)
isa
dis
trib
uti
onon
[0,x]:
By
affilia
tion
g(t|t)/
G(t|t)≥
g(t|0)/
G(t|0)
and
−∫ x 0
g(t|t)
G(t|t)
dt≤−∫ x 0
g(t|0)
G(t|0)
dt=−∫ x 0
dln
G(t|0)
dt
dt=−
∞.
Tak
ing
exp
onen
tial
s⇒
L(0|x)=
0an
dn
otic
eL(x|x)=
exp(0)=
1.
Fin
ally
∂L(y|x)/
∂y=
L(y|x)g(y|y)/
G(y|y)≥
0
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
16/24
Not
es
Not
es
Pro
of.
[Pro
of2/
3](2
)β
Iis
incr
easi
ng
inx
:
Not
ice
that
∂L(y|x)/
∂x=−
L(y|x)g(x|x)/
G(x|x)≤
0.
Ad
dit
ion
ally
∂v(y
,y)/
∂y>
0is
incr
easi
ng
byas
sum
pti
on.
Inte
grat
ing
byp
arts
βI
get
βI (
x)=∫ x 0
v(y
,y)d
L(y|x)=
v(x
,x)−∫ x 0
L(y|x)d
v(y
,y).
Th
eref
ore:
∂β
I (x)
∂x=−∫ x 0
∂L(y|x)
∂x
∂v(y
,y)
∂yd
y≥
0.
(3)
β=
βI
isop
tim
alif
follo
wed
byot
her
s:
Sin
ceβ
isin
crea
sin
gth
ep
ayoff
ofa
bid
β(z)
afte
rob
serv
ing
xis
Π(z|x)=∫ z 0
(v(x
,y)−
β(z))
dG(y|x).
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
17/24
Pro
of.
[Pro
of3/
3]F
irst
ord
erco
nd
itio
ns
req
uir
e
∂Π(z|x)
∂z=
(v(x
,z)−
β(z))
g(z|x)−
β′ (
z)G
(z|x)=
0.
Not
ice
that
β′ (
x)=−∫ x 0
∂L(y|x)
∂xd
v(y
,y)=
g(x|x)
G(x|x)
∫ x 0L(y|x)d
v(y
,y)
=g(x|x)
G(x|x)[v(x
,x)−
β(x)]
.
Th
us
x>
z⇒
v(x
,z)>
v(z
,z)
&g(z|x)/
G(z|x)>
g(z|z)/
G(z|z)
∂Π(z|x)
∂z>
[ (v(z
,z)−
β(z))
g(z|z)
G(z|z)−
β′ (
z)] G
(z|x)=
0.
Ifx<
zsi
mila
rar
gum
ent
show
sth
at∂
Π(z|x)/
∂z<
0
Th
eref
ore
opti
mal
ity
req
uir
esth
atx=
zNava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
18/24
Not
es
Not
es
Revenue
Com
parisons
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
19/24
Rev
enue
s:E
nglis
hvs
Sec
ond
Pri
ceA
ucti
on
Lem
ma
Rev
enu
esin
anE
ngl
ish
auct
ion
are
not
smal
ler
than
thos
eof
a2n
dpr
ice.
Pro
of.
Fir
stn
otic
eth
atby
affilia
tion
ifx>
y
v(y
,y)=
E[u(X
)|X1=
Y1=
y]=
E[u(Y
1,Y
1,.
..,Y
N−1)|
X1=
Y1=
y]
≤E[u(Y
1,Y
1,.
..,Y
N−1)|
X1=
x,Y
1=
y].
Th
us
reve
nu
esin
2nd
pric
eau
ctio
nar
ew
eakl
ysm
alle
r
E[R
II]=
E[β
II(Y
1)|
X1>
Y1]=
E[v(Y
1,Y
1)|
X1>
Y1]
≤E[E[u(Y
1,Y
1,.
..,Y
N−1)|
X1=
x,Y
1=
y]|X
1>
Y1]
=E[u(Y
1,Y
1,.
..,Y
N−1)|
X1>
Y1]
=E[β
2(Y
1,.
..,Y
N−1)]=
E[R
En].
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
20/24
Not
es
Not
es
Rev
enue
s:F
irst
vsS
econ
dP
rice
Auc
tion
Lem
ma
Rev
enu
esin
a2n
dpr
ice
auct
ion
are
not
smal
ler
than
thos
eof
a1s
tpr
ice.
Pro
of.
Con
sid
erth
ed
istr
ibu
tion
K(y|x)=
G(y|x)/
G(x|x).
Not
eth
aty<
xim
plie
sL(y|x)≥
K(y|x)
byaffi
liati
onsi
nce
−∫ x y
g(t|t)
G(t|t)
dt≥−∫ x y
g(t|x)
G(t|x)
dt=−∫ x y
dln
G(t|x)
dt
dt=
ln
( G(y|x)
G(x|x)
) .
Pro
bab
ility
ofw
inn
ing
con
dit
ion
alon
xis
the
sam
ein
1st
&2n
dpr
ice.
Bu
tex
pec
ted
pay
men
tsco
nd
itio
nal
onx
diff
er[i
nte
grat
ion
byp
arts
]:
E[β
II(Y
1)|
Y1<
X1=
x]−
βI (
x)=∫ x 0
v(y
,y)d[K
(y|x)−
L(y|x)]=
=∫ x 0
[L(y|x)−
K(y|x)]
dv(y
,y)≥
0.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
21/24
Equ
ilibr
ium
Bid
ding
and
the
Win
ner’
sC
urse
Cor
olla
ry
Wit
hsy
mm
etri
c,in
terd
epen
den
tva
lues
and
affilia
ted
sign
als
exp
ecte
dre
ven
ues
can
be
ran
ked
asfo
llow
s:E[R
En]≥
E[R
II]≥
E[R
I ].
In1s
tpr
ice
auct
ion
pla
yers
shad
eth
eir
bid
sto
acco
un
tfo
rw
inn
er’s
curs
e,
βI (
x)=∫ x 0
v(y
,y)d
L(y|x)≤∫ x 0
v(y
,y)d
K(y|x)
<∫ x 0
v(x
,y)d
K(y|x)=
E[V
1|Y
1<
X1=
x].
Ina
2nd
pric
eau
ctio
np
laye
rssh
ade
toac
cou
nt
for
win
ner
’scu
rse,
E[β
II(Y
1)|
Y1<
X1=
x]=∫ x 0
v(y
,y)d
K(y|x)<
E[V
1|Y
1<
X1=
x].
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
22/24
Not
es
Not
es
Effi
cien
cy
Au
ctio
ns
stu
die
dga
veob
ject
toh
igh
est
sign
aln
oto
hig
hes
tva
lue.
Bu
tth
isca
nb
ein
effici
ent.
Eg:
v i(x)=
x i+
2xj
for
i6=
j∈{1
,2}.
Val
uat
ion
ssa
tisf
yth
esi
ng
lecr
oss
ing
con
dit
ion
(SC
C)
iffo
ran
yi6=
jan
dan
yx
∂vi(
x)
∂xi≥
∂vj(
x)
∂xi
.
Inth
esy
mm
etri
cca
sev i(x)=
u(x
i,x −
i).
Let
u′ j
den
ote
the
par
tial
wrt
jth
argu
men
t.
Th
us
SC
Cre
du
ces
tou′ 1≥
u′ 2
bysy
mm
etry
assu
mp
tion
.
SC
Cen
sure
sth
ator
der
ofth
esi
gnal
sre
flec
tsth
atof
the
valu
es.
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
23/24
Lem
ma
(Milg
rom
Web
er)
Wit
hsy
mm
etri
cin
terd
epen
den
tva
lues
and
affilia
ted
sign
als,
ifS
CC
hol
ds
then
all
3au
ctio
nfo
rmat
sh
ave
aneffi
cien
tsy
mm
etri
ceq
uili
briu
m.
Pro
of.
Con
sid
erx i
>x j
and
let
α(t)=
t(x i
,xj,
x −ij)+(1−
t)(x
j,x i
,x−ij).
By
the
fun
dam
enta
lth
eore
mof
calc
ulu
san
dS
CC
,
u(x
i,x j
,x−ij)=
u(x
j,x i
,x−ij)+∫ 1 0∇
u(α(t))
α′ (
t)d
t.
By
SC
Can
dth
ed
efin
itio
nof
grad
ien
t,
∇u(α(t))
α′ (
t)=
u′ 1(α(t))(x
i−
x j)+
u′ 2(α(t))(x
j−
x i)
=(u′ 1(α(t))−
u′ 2(α(t)))(
x i−
x j)≥
0.
Th
eref
ore,
x i>
x jim
plie
su(x
i,x j
,x−ij)≥
u(x
j,x i
,x−ij).
Nava
(LSE)
Interdep
enden
tValueAuctions
MichaelmasTerm
24/24
Not
es
Not
es
The
Linkage
Principle
EC
319
–L
ectu
reN
otes
VII
I
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
1/10
Linkage
Principle
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
2/10
Not
es
Not
es
The
Linkage
Principle
Let
valu
esb
ein
terd
epen
den
tan
dsy
mm
etri
c.
Let
Ab
ea
stan
dar
dau
ctio
nw
ith
asy
mm
etri
ceq
uili
briu
mβA
.
WA(z|x)=
exp
ecte
dp
aym
ent
ofw
inn
erw
ith
sign
alx
&b
idβA(z).
Let
WA j(z|x)
the
par
tial
der
ivat
ive
wrt
j∈{z
,x}.
Not
ice
that
∂WA(x|x)/
∂x=
WA x(x|x)+
WA z(x|x).
Theorem
(Linkage
Principle–Milgrom
Weber)
Let
AandB
betwoauctionsin
whichthehighestbidder
winsandon
lyhe
paysapositiveam
ount.
Supposethat
each
has
asymmetricandincreasingequilibrium
such
that:
(i)W
A x(x|x)≥
WB x(x|x)foranyxand(ii)W
A(0|0)=
WB(0|0)=
0.
Then
theexpectedrevenuein
Aisnot
smallerthan
that
ofB.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
3/10
Proof.
Inau
ctio
nc∈{A
,B}
bid
der
1w
ith
sign
alx
max
imiz
es
∫ z 0v(x
,y)dG(y|x)−
Wc(z|x)G
(z|x).
Op
tim
alit
yat
z=
xre
qu
ires
that
v(x
,x)g(x|x)−
Wc(x|x)g(x|x)−
Wc z(x|x)G
(x|x)=
0
⇒W
c z(x|x)=
[v(x
,x)−
Wc(x|x)]g(x|x)
G(x|x)
.
Let
∆(x)=
WA(x|x)−
WB(x|x)
and
not
ice
that
∆′ (x)=
[WA x(x|x)−
WB x(x|x)]+[W
A z(x|x)−
WB z(x|x)]
=[W
A x(x|x)−
WB x(x|x)]−
∆(x)[g(x|x)/
G(x|x)]
.
So,
∆(0)=
0&
[∆(x)≤
0⇒
∆′ (x)≥
0]
imp
ly∆(x)≥
0an
yx
.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
4/10
Not
es
Not
es
Rem
arks
ontheLinkage
Principle
Exa
mp
le:
LP
imp
liesE[R
II]≥
E[R
I ]si
nce
byaffi
liati
on
WI (z|x)=
βI (z)
WII(z|x)=
E[β
II(Y
1)|X1=
x,Y
1<
z]
⇒W
II x(z|x)≥
WI x(z|x)=
0.
Corollary
IfLPassumption
sholdandsign
alsareindependentlydistributedthen:
E[R
A]=
E[R
B].
Proof.
Ifsi
gnal
sar
ein
dep
end
ent
only
the
bid
det
erm
ines
the
pay
men
t
WA x(x|x)=
WB x(x|x)=
0.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
5/10
Applications
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
6/10
Not
es
Not
es
Pub
licInform
ation
Su
pp
ose
selle
rh
asso
me
info
rmat
ionS
affec
tin
gva
luat
ion
s
Vi=
v i(S
,X).
Ass
um
ev i(0)=
0an
dv i(S
,X)=
u(S
,Xi,X−i)
sym
met
ric
inX−i.
Als
oas
sum
e(S
,X1,.
..,X
N)
tob
eaffi
liate
d.
Ifin
form
atio
nis
not
reve
aled
agai
nle
t
v(x
,y)=
E[V
1|X
1=
x,Y
1=
y].
Ifin
form
atio
nis
reve
aled
inst
ead
let
v(s
,x,y)=
E[V
1|S
=s,X1=
x,Y
1=
y].
Cle
arly
vis
incr
easi
ng
inal
lva
riab
les
andv(0)=
0.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
7/10
Pub
licInform
ationin
aFirst
Price
Auction
Ifin
form
atio
nis
not
reve
aled
agai
nW
I (z|x)=
βI (z).
Ifin
form
atio
nis
reve
aled
,th
ere
exis
tssy
mm
etri
ceq
uili
briu
m
βI (s,x)=∫ x 0
v(s
,y,y)d
L(y|s
,x)
&L(y|s
,x)=
exp( −∫
x yg(t|s,t)
G(t|s,t)dt) .
Th
eex
pec
ted
pay
men
tin
this
scen
ario
bec
omes
WI (z|x)=
E[β
I (S
,z)|X1=
x].
Th
us
byaffi
liati
onW
I x(z|x)≥
WI x(z|x)=
0.
LP
imp
lies
that
exp
ecte
dre
ven
ues
are
hig
her
ifin
fois
reve
aled
.
Sim
ilar
argu
men
tsfo
rE
ngl
ish
and
2nd
pric
eau
ctio
ns
wor
k.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
8/10
Not
es
Not
es
Alternative
Linkage
Principle(LPA)
MA(z|x)=
exp
ecte
dp
aym
ent
ofp
laye
rw
ith
sign
alx
&b
idβA(z)
Theorem
(AllPay
Linkage
Principle–Krishna
Morgan)
Let
AandB
betwoauctionsin
whichthehighestbidder
wins.
Supposethat
each
has
asymmetricandincreasingequilibrium
such
that:
(i)M
A x(x|x)≥
MB x(x|x)foranyxand(ii)M
A(0|0)=
MB(0|0)=
0.
Then
theexpectedrevenuein
Aisnot
smallerthan
that
ofB.
Proof.
Inau
ctio
nc∈{A
,B}
bid
der
1w
ith
sign
alx
max
imiz
es∫ z 0
v(x
,y)dG(y|x)−
Mc(z|x).
Op
tim
alit
yat
z=
xre
qu
ires
Mc z(x|x)=
v(x
,x)g(x|x).
∆(x)=
MA(x|x)−
MB(x|x)⇒
∆′ (x)=
[MA x(x|x)−
MB x(x|x)]≥
0.
Th
us
∆(0)=
0im
plie
s∆(x)≥
0an
yx
.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
9/10
AllPay
Auctionsor
Lobbying
LP
Ad
oes
not
req
uir
eth
atlo
sers
do
not
pay
.
Inan
all
pay
auct
ion
ever
yon
ep
ays
his
bid
MA
P(z|x)=
βA
P(z).
Ina
firs
tpr
ice
auct
ion
inst
eadM
I (z|x)=
βI (z)G
(z|x).
Th
us,
sin
ceby
affilia
tion
Gx(z|x)≤
0,
MI x(x|x)≤
MA
Px
(x|x)=
0.
By
theo
rem
LP
Ath
us:
E[R
AP]≥
E[R
I ].
Th
esy
mm
etri
ceq
uili
briu
mof
the
all-
pay
auct
ion
is
βA
P(x)=∫ x 0
v(y
,y)g(y|y)dy
.
Nava
(LSE)
TheLinka
gePrinciple
MichaelmasTerm
10/10
Not
es
Not
es
Bid
ding
Rin
gsE
C31
9–
Lec
ture
Not
esIX
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
BiddingRings
MichaelmasTerm
1/11
Intr
odu
ctio
n
Bid
din
gri
ngs
are
cart
els
tom
anip
ula
teth
eou
tcom
esof
the
auct
ion
.
Inve
stig
atin
gco
llusi
onin
auct
ion
sis
sub
stan
tial
anti
tru
stac
tivi
ty.
Bas
icS
etup
Con
sid
erth
eas
ymm
etri
cIP
Vm
od
elfo
rXi∼
Fi
on[0
,ω].
Let
I⊆
Nb
eth
ese
tof
bu
yers
inth
eb
idd
ing
rin
g.
and
N\I
be
the
set
ofb
uye
rsn
otin
the
bid
din
gri
ng.
Let
YS 1=
max
i∈S
Xi∼
GS=
∏i∈
SFi
for
any
sub
set
ofp
laye
rsS⊆
N.
Th
eri
ng
wan
tsth
eob
ject
togo
toth
em
emb
erva
luin
git
the
mos
t.
Nava
(LSE)
BiddingRings
MichaelmasTerm
2/11
Not
es
Not
es
Collusion
inaSecondPrice
Auction
Nava
(LSE)
BiddingRings
MichaelmasTerm
3/11
Col
lusi
onin
aS
econ
dP
rice
Auc
tion
Ass
um
eth
eri
ng
wor
ksan
dm
emb
ers
reve
alin
form
atio
n.
Rin
gpr
od
uce
spr
ofits
tom
emb
ers
byre
du
cin
gpr
ice
com
pet
itio
n.
Rin
gsu
bm
its
one
bid
toth
eau
ctio
nee
r[Y
I 1=
Xi
som
ei∈
I].
Con
dit
ion
alon
rin
gw
inn
ing
the
pay
men
tb
ecom
essm
alle
r
PI=
max{ Y
N\I
1,r} ≤
max{ Y
N\i
1,r} =
Pi.
Exp
ecte
dp
aym
ents
con
dit
ion
alx
are
smal
ler
wh
enth
eri
ng
isin
pla
ce.
Su
rplu
sm
ade
byri
ng
bu
yer
i∈
Ico
nd
itio
nal
onva
lue
xis
δ i(x)=
rGN\i(r)+∫ x r
ydG
N\i(y)−[r
GN\I(r)+∫ x r
ydG
N\I(y)]
GI\i (
x).
δ i(x)=
mi(
x)−
mi(
x)≥
0si
nce
pay
men
tsar
eal
way
slo
wer
.
Nava
(LSE)
BiddingRings
MichaelmasTerm
4/11
Not
es
Not
es
Gai
nsan
dL
osse
sfr
omC
ollu
sion
Th
eex
-an
teex
pec
ted
profi
tsfr
omth
eri
ng
bec
ome
δ I=
∑i∈
IE[δ
i(x i)]
.
Bid
der
sn
otin
the
rin
gd
on
otb
enefi
tor
suff
ersi
nce
for
i∈
N\I
:
Pi=
max{ Y
I 1,Y
N\I\i
1,r} =
max{ Y
N\i
1,r} =
Pi.
So
δ i(x)=
0an
yi∈
N\I
[rin
gex
erts
no
exte
rnal
ity
onou
tsid
ers]
.
Th
us
gain
sby
the
rin
gco
rres
pon
dto
loss
esby
the
selle
r.
Incr
easi
ng
the
rin
g’s
size
toJ⊃
Ib
enefi
tsm
emb
ers
atex
pen
seof
selle
rsi
nce
PJ=
max{ Y
N\J
1,r} ≤
max{ Y
N\I
1,r} =
PI.
Nava
(LSE)
BiddingRings
MichaelmasTerm
5/11
Effi
cien
tC
ollu
sion
An
effici
ent
collu
sio
nm
ech
anis
m:
I.al
loca
tes
the
righ
tto
repr
esen
tth
eri
ng
effici
entl
yan
dis
IC;
II.
det
erm
ines
pay
men
tsm
ade
byth
eri
ng
[pos
sib
lyn
egat
ive]
.
Lem
ma
(Mai
lath
Zem
sky)
Ina
2nd
pric
eau
ctio
nw
ith
rese
rve
pric
e,th
eex
pec
ted
gain
sfr
omeffi
cien
tco
llusi
onof
any
rin
gm
emb
erd
epen
don
the
iden
tity
,b
ut
not
onh
isva
lue.
Mor
eove
rth
eri
ng
do
esn
otaff
ect
outs
ider
s.
Pro
of.
Allo
cati
onin
2nd
pric
eau
ctio
nsa
tisfi
esQ
i(X)=
Pr(
Xi>
max{Y
N\i
1,r})
.
Wit
heffi
cien
tco
llusi
on( Q
,M) al
loca
tion
rem
ain
sth
esa
me
Q=
Q.
Sin
ceb
oth
mec
han
ism
sar
eIC
and
Q=
Q,
reve
nu
eeq
uiv
alen
ceim
plie
s:
δ i=
mi(
x i)−
mi(
x i)≥
0.
Nava
(LSE)
BiddingRings
MichaelmasTerm
6/11
Not
es
Not
es
Pre
-Auc
tion
Kno
ckou
ts(P
AK
T)
Rin
gco
nd
uct
san
auct
ion
tod
eter
min
ew
ho
repr
esen
tsth
eri
ng
[tic
ket]
.
Mem
ber
ssu
bm
itb
ids,
the
hig
hes
tis
awar
ded
the
tick
et.
Con
dit
ion
alon
win
nin
gth
eau
ctio
nh
ep
ays
toth
eri
ng
τ i=
Pi−
PI.
Rin
gm
akes
pay
men
tsof
δ i(d
efin
edab
ove)
toea
chb
idd
er.
Su
chm
ech
anis
mis
effici
ent,
ex-p
ost
IC,
IR,
BB
,b
ut
no
tex
-pos
tB
B.
Itis
n’t
ex-p
ost
BB
sin
ceit
mak
esp
aym
ents
δ iev
enw
hen
itlo
ses.
IfV
CG
isco
nd
uct
edto
allo
cate
tick
etit
run
sa
surp
lus.
Th
us
ther
eex
ists
aneffi
cien
t,IC
,IR
,ex
-pos
tB
Bm
ech
anis
m.
Su
chm
ech
anis
mis
no
tex
-pos
tIC
how
ever
[Kri
shn
aP
erry
Lem
ma]
.
Nava
(LSE)
BiddingRings
MichaelmasTerm
7/11
Collusion
andReserve
Prices
Nava
(LSE)
BiddingRings
MichaelmasTerm
8/11
Not
es
Not
es
Res
erve
Pri
ces
whe
nfa
cing
Col
lusi
on
Ifse
ller
know
sof
rin
gI
wh
atre
serv
epr
ice
shou
ldh
ese
t?
Th
ese
ller
face
sN−
I+
1b
uye
rsw
ith
valu
es{ Y
I 1,X
I+1,.
..,X
N
} .
ZIcd
f ∼H
Id
enot
esth
ese
con
dh
igh
est
ofsu
chva
lues
.
IfY
N 1>
rth
ese
ller
rece
ives
P=
max{ Z
I ,r} .
Th
eref
ore
the
exp
ecte
dse
llin
gpr
ice
is
r(H
I(r)−
GN(r))
+∫ ω r
ydH
I(y).
Op
tim
alit
yof
the
rese
rve
pric
erI
req
uir
es
HI(r
I)−
GN(r
I)−
rIgN(r
I)=
0.
Nava
(LSE)
BiddingRings
MichaelmasTerm
9/11
Lem
ma
(Mai
lath
Zem
sky)
Th
ead
dit
ion
ofa
bid
der
toa
bid
din
gri
ng
cau
ses
the
opti
mal
rese
rve
pric
efo
rth
ese
ller
tow
eakl
yin
crea
se.
Pro
of.
Con
sid
era
rin
gw
ith
J=
I+
1m
emb
ers.
Th
ese
ller
face
sN−
Ib
uye
rsw
ith
valu
es{ Y
J 1,X
I+2,.
..,X
N
} .
IfZ
Jcd
f ∼H
Jis
seco
nd
hig
hes
tof
such
valu
esth
enZ
J6=
ZI
iff
(1)
ZI=
XI+
1>
YN\J
1⇒
ZI>
YN\J
1=
ZJ,
(2)
ZI=
YI 1<
XI+
1⇒
ZI>
YN\J
1=
ZJ.
Bu
tZ
J≤
ZI
imp
lies
HJ(r)≥
HI(r)
for
any
ran
dth
us
HJ(r
I)−
GN(r
I)−
rIgN(r
I)≥
0,
wh
ich
imp
lies
that
rJ≥
rI.
Nava
(LSE)
BiddingRings
MichaelmasTerm
10/11
Not
es
Not
es
Uni
form
Exa
mpl
e:B
iddi
ngR
ing
Tw
ob
idd
ers
wit
hva
lues
un
ifor
mly
dis
trib
ute
din
[0,1].
Wit
hou
tri
ng
the
opti
mal
rese
rve
pric
eis
NFN−1(r)(
1−
F(r))−
rNf(r)F
N−1(r)=
0
⇒2r(1−
r)−
2r2=
0⇒
r=
1/
2.
Wit
han
all-
incl
usi
veri
ng
reve
nu
esof
the
selle
rb
ecom
e
r(1−
FN(r))
=r(
1−
r2).
Op
tim
alit
yth
enre
qu
ires
1−
3r2=
0⇒
r=
1/√
3>
1/
2.
Ina
1st
pric
eau
ctio
nco
llusi
onis
less
trac
tab
lean
dn
ot
self
-en
forc
ing
,b
ut
resu
lts
exis
tfo
rth
esy
mm
etri
cca
se.
Nava
(LSE)
BiddingRings
MichaelmasTerm
11/11
Not
es
Not
es
Mec
hani
smD
esig
nw
ith
Inte
rdep
ende
ntV
alue
sE
C31
9–
Lec
ture
Not
esX
Nav
a
LSE
Mic
hae
lmas
Ter
m
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
1/13
Intr
odu
ctio
n
Con
sid
erth
ese
tup
wit
hin
terd
epen
den
tva
lues
and
affilia
ted
sign
als.
Rev
elat
ion
prin
cip
leap
plie
sal
soin
this
con
text
.
Val
uat
ion
sv
sati
sfy
thesingle
crossingco
ndition,
if∀i6=
j
∂vi(
x)/
∂xi>
∂vj(
x)/
∂xi
atan
yx
such
that
v i(x)=
v j(x)=
max
k∈N
v k(x).
Ad
irec
tm
ech
anis
mco
nsi
sts
oftw
ofu
nct
ion
s
Q:X→
∆N
&M
:X→
RN
.
Th
ean
alys
ispr
oce
eds
asfo
llow
s:
1E
ffici
ent
dir
ect
mec
han
ism
s;
2R
even
ue
max
imiz
ing
dir
ect
mec
han
ism
s.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
2/13
Not
es
Not
es
Effi
cien
tM
echa
nism
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
3/13
Wit
hout
SC
Cno
Mec
hani
smm
ayb
eE
ffici
ent
Con
sid
era
scen
ario
inw
hic
hN
=2
and
v 1(x)=
x 1an
dv 2(x)=
x2 1.
Su
pp
ose
X1=
[0,2]
and
X2=
c[w
log]
.
Cle
arly
SC
Cis
viol
ated
sin
ce
∂v1(1
,c)/
∂x1=
1<
2=
∂v2(1
,c)/
∂x1.
Su
pp
ose
that
mec
han
ism
(Q,M
)is
effici
ent.
Effi
cien
cyan
dIC
for
pla
yer
1re
qu
ire
that
aty<
1<
z
y−
M1(y
,c)≥
0−
M1(z
,c),
0−
M1(z
,c)≥
z−
M1(y
,c).
Wh
ich
imp
lies
y≥
za
con
trad
icti
on!
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
4/13
Not
es
Not
es
SC
Cis
nece
ssar
yfo
rE
ffici
ency
Su
pp
ose
ther
eis
aneffi
cien
tm
ech
anis
mw
ith
anex
-pos
teq
uili
briu
m
[⇒th
ere
isa
dir
ect
effici
ent
mec
han
ism
wit
han
ex-p
ost
equ
ilibr
ium
].
Giv
enx −
iif
pla
yer
iw
ins
orlo
ses
rega
rdle
ssof
x ith
enS
CC
hol
ds.
Bu
yer
iispivotal
atx −
iif
ther
eex
ists
yan
dz
such
that
v i(y
,x−i)>
max
j6=iv j(y
,x−i)
,
v i(z
,x−i)<
max
j6=iv j(z
,x−i)
.
Bu
yer
i’s
ex-p
ost
ICat
yan
dz
req
uir
esth
at
v i(y
,x−i)−
Mi(
y,x−i)≥−
Mi(
z,x−i)
,
−M
i(z
,x−i)≥
v i(z
,x−i)−
Mi(
y,x−i)
.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
5/13
SC
Cis
nece
ssar
yfo
rE
ffici
ency
Rew
riti
ng
imp
lies
v i(y
,x−i)≥
Mi(
y,x−i)−
Mi(
z,x−i)≥
v i(z
,x−i)
.
Ex-
pos
tIC
imp
lies
mec
han
ism
ism
onot
onic
inva
lues
[if
iw
ins
wit
hsi
gnal
x ih
ew
ins
wit
han
ysi
gnal
big
ger
than
x i].
Effi
cien
cy⇒
i’s
valu
ere
mai
ns
the
hig
hes
tif
sign
alx i
incr
ease
s.
Th
us
ifv i(x)=
v j(x)
ach
ange
∂xi
ben
efits
mor
ei
than
j,
∂vi(
x)/
∂xi≥
∂vj(
x)/
∂xi.
Th
eref
ore
SC
Cis
nec
essa
ry.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
6/13
Not
es
Not
es
Gen
eral
ized
VC
GM
echa
nism
Con
sid
era
dir
ect
mec
han
ism
(Qv
,Mv)
sati
sfyi
ng
Qv i(x)=
{ 1if
v i(x)>
max
j6=iv j(x)
0if
v i(x)<
max
j6=iv j(x)
Mv i(x)=
{ v i(y
i(x −
i),x−i)
ifQ
v i(x)=
10
ifQ
v i(x)=
0
for
y i(x−i)=
inf{ z|v
i(z
,x−i)≥
max
j6=iv j(z
,x−i)} .
Su
chm
ech
anis
maw
ard
sob
ject
effici
entl
y&
gen
eral
izes
VC
G.
Itin
du
ces
tru
thte
llin
gby
chan
gin
gp
aym
ents
v i(y
i(x −
i),x−i)
.
Lem
ma
(Cre
mer
McL
ean)
Ifva
luat
ion
sv
sati
sfy
SC
C,
tru
thte
llin
gis
anex
-pos
teq
uili
briu
mof
the
VC
Gm
ech
anis
m(Q
v,M
v).
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
7/13
SC
Can
dth
eG
ener
aliz
edV
CG
Cra
mer
McL
ean
Pro
of.
Su
pp
ose
all
pla
yers
bu
ti
are
tru
thfu
l.
(I)
Ifv i(x)>
max
j6=iv j(x)⇒
x i>
y i(x−i)
.
Ifi
rep
orts
asi
gnal
z i>
y i(x−i)
he
win
san
dp
ays
v i(y
i(x −
i),x−i)<
v i(x).
sin
ceby
SC
Cv i(z
i,x −
i)>
max
j6=iv j(z
i,x −
i).
Ifi
rep
orts
az i≤
y i(x−i)
his
surp
lus
isze
roby
SC
C.
No
dev
iati
onz i6=
x iis
profi
tab
le.
(II)
Ifin
stea
dv i(x)<
max
j6=iv j(x)⇒
x i<
y i(x−i)
.
For
ito
win
byS
CC
he
wou
ldn
eed
tob
idz i>
y i(x−i)>
x i.
Bu
tp
aym
ents
wou
ldb
ecom
ev i(y
i(x −
i),x−i)>
v i(x).
Aga
inn
od
evia
tion
ispr
ofita
ble
.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
8/13
Not
es
Not
es
Com
men
tson
Gen
eral
ized
VC
G
Wit
hpr
ivat
eva
lues
this
red
uce
sto
ord
inar
y2n
dpr
ice
auct
ion
.
Wit
htw
ob
uye
rsto
the
effici
ent
equ
ilibr
ium
ofth
eE
ngl
ish
Au
ctio
n.
Th
eV
CG
mec
han
ism
isn
otd
etai
lfr
ee!
Itpr
esu
mes
that
the
auct
ion
eer
know
sth
eva
luat
ion
map
s.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
9/13
Opt
imal
Mec
hani
sm
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
10/13
Not
es
Not
es
Opt
imal
Mec
hani
sms
Ifb
uye
rs’
info
isco
rrel
ated
they
are
un
able
toga
inan
yre
nts
from
that
info
!
For
sim
plic
ity
con
sid
era
setu
pw
ith
dis
cret
esi
gnal
s
Xi={0
,∆,.
..,(
t i−
1)∆}.
Πd
enot
esth
ejo
int
prob
abili
tyd
istr
ibu
tion
onX
.
Πi=
[π(x−i|x
i)]
isi’
sb
elie
fsm
atri
xw
ith
t iro
ws
and×
i6=jtj
colu
mn
s.
Ifsi
gnal
sar
ein
dep
end
ent
all
row
sar
eeq
ual
ran
k(Π
i)=
1.
Aga
inas
sum
ev i(0)=
0an
dv i(x
j+
∆,x−j)≥
v i(x)
stri
ctif
i=
j.
SC
Cin
this
setu
pb
ecom
esfo
ran
yi6=
j
v i(x)≥
v j(x)⇒
v i(x
i+
∆,x−j)≥
v j(x
i+
∆,x−j)
.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
11/13
The
orem
(Cre
mer
McL
ean)
Ifsi
gnal
sar
ed
iscr
ete,
valu
atio
ns
vsa
tisf
yS
CC
and
Πi
isof
full
ran
kfo
ran
yi,
then
ther
eex
ist
am
ech
anis
min
wh
ich
tru
thte
llin
gis
the
effici
ent
ex-p
ost
equ
ilibr
ium
inw
hic
hal
lb
uye
rsge
tan
exp
ecte
dp
ayoff
ofze
ro.
Pro
of.
Con
sid
erth
etr
uth
telli
ng
exp
ecte
dp
ayoff
ofth
eV
CG
mec
han
ism
Uv i(x
i)=
∑x −
i∈X−i[v
i(x)Q
v i(x)−
Mv i(x)]
π(x−i|x
i).
Uv i
den
otes
the
t i-v
ecto
r,si
nce
Πi
isof
full
ran
k∃c
i=
[ci(
x −i)] x−i∈
X−i:
Πic
i=
Uv i
.
Con
sid
erm
ech
anis
m(Q
v,M
c)
wit
hp
aym
ents
Mc i(x)=
Mv i(x)+
c i(x−i)
.
Tru
thte
llin
gis
ex-p
ost
ICas
inV
CG
sin
cep
ayoff
sd
iffer
byco
nst
ant.
and
inte
rim
exp
ecte
dp
ayoff
sar
eze
rofo
ral
lb
uye
rsby
con
stru
ctio
n.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
12/13
Not
es
Not
es
Com
men
tsO
nC
rem
erM
cLea
nM
echa
nism
Inte
rim
IRal
way
sh
old
sw
ith
equ
alit
ysi
nce
Uc i(x
i)=
∑x −
i∈X−i[v
i(x)Q
v i(x)−
Mv i(x)−
c i(x−i)]π(x−i|x
i)=
0.
Ex-
pos
tIR
isn
otm
etsi
nce
Qv i(x)=
0&
Mc i(x)>
0ca
nh
old
aton
ce.
CM
inte
rim
exp
ecte
dp
aym
ent
isV
CG
’sp
lus
that
ofa
lott
ery
one
mu
stp
arti
cip
ate
inc i(X−i)
.
Ifva
lues
are
priv
ate
CM
pay
men
tis
2nd
pric
ep
aym
ent
plu
sc i(x−i)
.
(I)
Op
tim
alit
yre
sult
hin
ged
onth
eco
rrel
atio
nam
ongs
tsi
gnal
s,b
ut
not
onin
terd
epen
den
cein
valu
es.
(II)
Effi
cien
cyre
sult
hin
ged
onth
ein
terd
epen
den
cein
valu
es,
bu
tn
oton
corr
elat
ion
amon
gst
sign
als.
Nava
(LSE)
Mechanism
DesignwithInterdep
enden
tValues
MichaelmasTerm
13/13
Not
es
Not
es
Note 1: Entry Fees Francesco Nava
EC319: Auction Theory Michaelmas Term
This note explains the comparison between entry fees and reserve prices. On page 24 of the textbook, it isclaimed that a reserve price r is equivalent to an entry fee:
e = rG(r)
For sake of simplicity consider a second price auction. In such auction it remains dominant to bid one’svaluation even if some players are not participating in the auction, because conditional on winning playerscannot affect the trading price. Thus, if only players with value x ≥ r participate the expected payment atauction of a type x ≥ r is:
m(x) =
∫ x
rydG(y)
Therefore a player with value x = r is just indifferent about whether to participate in the auction if theentry fee is e = rG(r), since:
u(r) = rG(r)−m(r) = rG(r) = e
Finally, buyers with value x′ < r strictly prefer not to participate in the auction:
u(x′) = x′G(r)−m(x′) < rG(r) = e
and buyers with value x′′ > r strictly prefer to participate:
u′(x′′) = G(x′′) > 0
Therefore an entree fee of e = rG(r) is equivalent to a reserve price of r since in both cases the interimexpected payoffs prior to the auction satisfy:
U(x) =
{xG(x)− rG(r)−
∫ xr ydG(y) if x ≥ r
0 if x < r
Note 2: Budget Constraints Francesco Nava
EC319:Auction Theory Michaelmas Term
This is an explicit proof of the argument made in the Extensions slides on page 10.
(1) Recall that since β(x) < x ≤ 1:
LI(x) = {(X,W )|min{β(X),W} < min{β(x), 1}} == {(X,W )|min{β(X),W} < β(x)}
LII(x) = {(X,W )|min{X,W} < min{x, 1}} == {(X,W )|min{X,W} < x}
I want to show that LI(x) ⊂ LII(x). To show this it suffi ces to argue that:
min{β(X),W} < β(x) ⇒ min{X,W} < x
But notice that since β′ > 0:
min{X,W} < x ⇔ β(min{X,W}) < β(x)
min{X,W} < x ⇔ min{β(X), β(W )} < β(x)
But since β(W ) < W :
min{β(X),W} < β(x) ⇒ min{β(X), β(W )} < β(x)
Which concludes the argument.
(2) The additional consideration on the revenues in the inequality. Notice that for any a ∈ {I, II}:
Ya(N)2 ∼ F a(N)2 (y) = F a(y)N +N(1− F a(y))F a(y)N−1
= NF a(y)N−1 − (N − 1)F a(y)N
We want to show that F I(x) < F II(x) implies F I(N)2 (x) < FII(N)2 (x) for any for any x ∈ (0, 1).
To do so notice that the expression NkN−1 − (N − 1)kN is strictly increasing in k for k ∈ (0, 1) since:
∂
∂k
[NkN−1 − (N − 1)kN
]= N(N − 1)[kN−2 − kN−1] > 0
Therefore since 0 < F I(x) < F II(x) < 1:
FI(N)2 (x)− F II(N)2 (x) = [NF I(x)N−1 − (N − 1)F I(x)N ]− [NF II(x)N−1 − (N − 1)F II(x)N ] < 0
Which in turn implies from integration by parts that:
E(RI) = E(YI(N)2 ) =
∫ 1
0ydF
I(N)2 (y) >
∫ 1
0ydF
II(N)2 (y) = E(Y
II(N)2 ) = E(RII)
Note 3: Regularity and IC Francesco Nava
EC319: Auction Theory Michaelmas Term
In the mechanism design slides on page 9, it is claimed that condition (1) and regularity imply incentivecompatibility. To prove this it suffi ces to show that condition (1) and regularity imply that q′i ≥ 0, sincethat was proven to imply IC. Notice that without ties condition (1) is equivalent to:
Qi(x) = 1⇔ ψi(xi) ≥ max{maxj 6=i ψj(xj), 0} (i)
1. For I(·) the indicator function, condition (i) implies that:
qi(xi) =
∫X−i
Qi(xi, x−i)dF−i(x−i)
=
∫X−i
I(ψi(xi) ≥ max{maxj 6=i ψj(xj), 0})dF−i(x−i)
= I(ψi(xi) ≥ 0)∫X−i
∏j 6=i I(ψi(xi) ≥ ψj(xj))dF−i(x−i)
= I(ψi(xi) ≥ 0)∏j 6=i
∫XjI(ψi(xi) ≥ ψj(xj))dFj(xj)
= I(ψi(xi) ≥ 0)∏j 6=i
∫XjI(ψ−1j ψi(xi) ≥ xj)dFj(xj)
= I(ψi(xi) ≥ 0)∏j 6=i Fj(ψ
−1j ψi(xi))
where in the second to last equality we used that by regularity ψ′j > 0 and thus ψj is invertible.
2. Notice that if ψi(xi) < 0 then q′i(xi) = 0. If instead ψi(xi) ≥ 0 we have that:
q′i(xi) = ψ′i(xi)∑k 6=i
fk(ψ−1k ψi(xi))
ψ′k(ψ−1k ψi(xi))
[∏j 6=i,k Fj(ψ
−1j ψi(xi))
]≥ 0
since any term in the summation is positive given regularity, ψ′k > 0, and since fk ≥ 0.
Which proves q′i(xi) ≥ 0 and therefore IC.
Note 4: Optimal Mechanism Francesco Nava
EC319: Auction Theory Michaelmas Term
Some more details on the example developed in class. So we had 2 players {a, b}. Player a with valuesuniformly distributed on [0, 2] and player b with uniform values on [1, 3]. In class we had shown:
ψa(xa) = 2xa − 2 and ψb(xb) = 2xb − 3
Therefore, the highest competing virtual valuations are:
ya(xb) = inf {za|ψa(za) > max {ψb(xb), 0}} == inf {za|za > xb − 0.5 ∩ za > 1} == max {xb − 0.5, 1}
yb(xa) = inf {zb|ψb(zb) > max {ψa(xa), 0}} == inf {zb|zb > xa + 0.5 ∩ zb > 1.5} == max {xa + 0.5, 1.5}
Thus, for these expressions as desired we get that OM satisfies for i ∈ {a, b}:
Qi(x) =
{1 if xi > yi(x−i)0 if xi < yi(x−i)
Mi(x) =
{yi(x−i) if xi > yi(x−i)0 if xi < yi(x−i)
Now with the correct expressions it can be easily noted that this mechanism can be ineffi cient for tworeasons:
1. The object is not sold. For instance, if (xa, xb) = (0.5, 1.25) since:
xa = 0.50 < ya(xb) = max {xb − 0.5, 1} = 1xb = 1.25 < yb(xa) = max {xa + 0.5, 1.5} = 1.5
2. The object is sold to the low value player. For instance, if (xa, xb) = (1.5, 1.75) since:
xa = 1.50 > ya(xb) = max {xb − 0.5, 1} = 1.25xb = 1.75 < yb(xa) = max {xa + 0.5, 1.5} = 2
Note 5: AGV Shifts & BB Francesco Nava
EC319: Auction Theory Michaelmas Term
Due to the algebra mishaps today during lectures let me establish again how to construct the shifts on theAGV mechanism that insure that the mechanism is IR.
1. First notice that by revenue equivalence the expected utility of both the VCG & the AGV mechanismscan differ at most by a constant. For m ∈ {v, a}, let:
cvi = E[W (xi, X−i)]− Uvi (xi)cai = E[W (xi, X−i)]− Uai (xi)
Suppose that VCG runs surplus. If so:
E[∑
i∈NMvi (X)
]≥ E
[∑i∈NM
ai (X)
]= 0 ⇔
⇔∑i∈N E [M
vi (X)] ≥
∑i∈N E [M
ai (X)] = 0 ⇔
⇔∑i∈N [c
vi − E[W−i(xi, X−i)]] ≥
∑i∈N [c
ai − E[W−i(xi, X−i)]] = 0 ⇔
⇔∑i∈N c
vi ≥
∑i∈N c
ai ⇔
⇔∑i>1(c
vi − cai ) ≥ (ca1 − cv1)
2. Define di = cai − cvi for i > 1 and d1 = −∑i>1 di. Notice that d1 ≥ (ca1 − cv1), by the last observation.
3. Consider the BB payment rule M∗ defined by:
M∗i (x) =Mai (x) + di
Such rule is trivially BB, since:∑i∈NM
∗i (X) =
∑i∈NM
ai (X) +
∑i∈N di = 0
4. If Q∗ = Qa the mechanism (Q∗,M∗) is effi cient. It is also IC since payoffs differ from (Qa,Ma) byadditive constant. Additionally (Q∗,M∗) is IR because:
U∗i (xi) = Uai (xi) + di ≥ Uai (xi) + cai − cvi = E[W (xi, X−i)]− cvi = Uvi (xi) ≥ 0
Note 6: Expectation and Maximum Francesco Nava
EC319: Auction Theory Michaelmas Term
In the interdependent value slides on page 4, it is claimed that for the pure common value model:
E[maxXi|V = v] > maxE[Xi|V = v] = v
Recall that the special case has:
• V ∼ F
• Xi|Viid∼ H(V )
• E[Xi|V = v] = v
Therefore we have that:
E[maxXi|V = v] =
∫ ωi
0ydH(y|v)N
maxE[Xi|V = v] = E[Xi|V = v] =∫ ωi
0ydH(y|v) = v
But notice G(y|v) = H(y|v)N is also a distribution and that it stochastically dominates H(y|v), sinceG(y|v) < H(y|v) for any y ∈ (0, 1). Thus we have that:
E[maxXi|V = v] =∫ ωi
0ydG(y|v) ≥
∫ ωi
0ydH(y|v) = maxE[Xi|V = v]
In general since the max is a convex operator we have that:
E[maxXi|V = v] ≥ maxE[Xi|V = v]
by Jensen inequality. That the max is convex can be seen from this plot:
0 1 20
1
2
x1
x2
The plot show that a lottery amongst two bundles having the same maximum is preferred to the maximumof the lottery. Indeed consider:
(x1, x2) =
{(2, 1) with probability 1/2(1, 2) with probability 1/2
Then immediately we have that:
E[maxXi] = 2 ≥ maxE[Xi] = 1.5