c Paul Klemperer 2004-2022

A1. (i) See section 1.6, section 4.2.1 page 128, and Appendix 1.C.

(ii) [Appendix 1.D pp. 55-57 contains a common-values example closely related
to the private-values example described here.]
(a) Bidders bid up to their value, so the price is the expected second
2
highest actual value, i.e. + ¡1
+1
(see Appendix 1.E). Pro¯ts are 2 ¡ +
.
(b) Each bidder bids for some (since the di®use prior means she
has no knowledge of how high or low her value is relative to others).
A bidder who deviates to bid + would earn if she won, and wins
¡1
with probability ( + ) and, so she would expect to earn
1
1 ¡1 (+ ) ( ¡ )
() = =0
( )( + ) = ( ) = [(1 + ) ] =
=0
( ) [(1+ ) ¡ ] ¡ ¡1 ¡1
= + [ (1 + ) ].
( )
In equilibrium = 0 at = 0, i.e. = 1.
1
The highest bidder (with expected value + +1
) wins so pro¯ts are +1
=
2¡ ¡1
2+ , con¯rming part (i)'s result that sealed-bid pro¯ts are below ascending
pro¯ts when bidders' private signals (here ) are a±liated.
(iii)(a) Ascending auction behaviour is una®ected.
(b) If is common knowledge, bidders' private information are now , and
2¡
independent, so revenue equivalence with (i) applies and pro¯t is again 2+ .
The di®erence between (ii)b and (iii)b illustrates the linkage principle (see section
1.6).
(iv)(a) An equilibrium is for ascending auction behaviour to be una®ected. But
it is not clear why bidders who know they do not have the highest value will
bother to bid; if they don't, then pro¯t will be much lower.
(b) The equilibrium is for the highest-value bidder to bid (and win) at
(just above) the second-highest value. There is the same di±culty that low-value
bidders may not bother to play.
(v) If the are also a±liated we have
(iva) = (ivb) (assuming the equilibria described above),
(iiia) (iiib) (a±liation),
(iiib) (iib) (linkage), and
(iva) = (iiia)= (iia) (trivially).
Summarising: (iia) = (iiia) = (iva) (ivb) (iiib) (iib).
A2. (i) See Appendix 1.A.
The key is equation (1): ( ) ( )+( ) ( ) This assumes private values.
[It's not hard to do the common values case.]

(ii) [Equation numbers in the below all refer to Appendix 1.A; the Oxford
examination candidates were not expected to write all of what follows!]

(a) (1) fails because a type who deviates to behave as if he had type
does not earn the right hand side of (1) { because type expects to face
competitors with values that are conditional on her having type , while type
expects to face competitors with values that are conditional on her having type .
Appendix 1.C describes informally how to proceed, though candidates were not
expected to go into this. A candidate might note that revenue equivalence still
holds between the ¯rst-price sealed-bid and Dutch auctions, and also between
the second-price sealed-bid and ascending auctions if either values are private or
there are just two bidders.

(b) (1) fails because it is generally no longer true that

( )= ( ) 's expected payment.
[One rather special case is that ( )= ( ) ( 's payment)
i.e., the bidder is risk-averse with respect to money, but the prize is not (equivalent
to) money. In this case (1) holds as usual, hence ( ) is pinned down by ( ) and
( ) as usual, so (type 's payment) is equivalent across auctions under the
usual revenue equivalence conditions. (And since risk-averse buyers are equally
well o®, a risk-neutral seller prefers the auction that stabilises their payments in
order to give them a given expected utility most cheaply. So all-pay auctions
raise more money than ¯rst-price auctions, which themselves raise more money
than second-price auctions.)]

Again, of course, the ¯rst-price sealed-bid and Dutch are revenue equiva-
lent, as are the second-price and ascending.

(c) This depends on the model of collusion { how do bidders share infor-
mation and collude? But if, for example, all the bidders agree to (jointly) win
the object at a price of zero, and then share the object among themselves aµ la
Cramton, Gibbons and Klemperer (1987) then (1) holds and revenue equivalence
follows as usual | it's just that each bidder's surplus is higher by the same
amount ( ) 0.

A special case of CGK is McAfee and McMillan's (1992) suggestion that

the colluders allocate the object among themselves using a ¯rst-price\knockout"
auction, the winner of which pays 1/ th of his bid to all colluders (including
himself). This is incentive compatible (since a loser's payo® does not depend
on his bid, each colluder makes his usual bid) and budget balanced ex post. Or
the colluders can allocate the object among themselves in dominant strategies
by running a second-price auction, the winner of which pays a risk-neutral ring-
center who previously paid all the colluders 1/ th of the expected second-highest
of their values (again a loser's payo® does not depend on his bid, so each
colluder makes his usual bid), but this is only budget balanced in expectation
{ Graham and Marshall (1987). (The GM result extends to cases when not all
bidders collude if the main auction is a second-price auction; if the knockout
winner wins the main auction, he pays the ring-center any excess of the knockout
price over the main auction price, and the ring-center previously pays all the
colluders 1/ th of the expected value of this.)

(Note, of course, that a colluder may want to cheat at the main auction.
However, collusion is easier to sustain in second-price and ascending auctions
than in ¯rst-price auctions, because in the former the designated winner can bid
in¯nitely high and other colluders have no incentive to cheat and try to win the
auction (Robinson, 1985). Also, (tacit) collusion is easier in ascending (multi-unit
or repeated) auctions because bidders can use bids to signal (Chapters 3, 4, of my
book; Klemperer, 2000; Brusco and Lopomo, 2002; Cramton and Schwartz, 2000).
With non-independent values, an ascending auction allows colluding bidders to
induce non-colluding opponents to bid less aggressively, by having some colluders
drop out at a low price, thus signalling a low valuation (Pagnozzi, 2003). See
also Section 1.9.)

(d) If each bidder is restricted to a single unit only, then equations (1) {
(5), and hence revenue equivalence, hold as usual. The only change is that the
( ) function is di®erent to re°ect the higher probability of winning that more
units provides.

If a bidder's value is linear in the number of units won, we can just write
( ) = (expected number of units wins) and (1) { (5), and revenue equivalence,
again hold as before.

With multiple identical indivisible units, and if bidders' values are not
necessarily linear in the number won, a generalisation of the usual argument
applies: let ( 1 ) represent the type whose marginal value of winning a
unit is , and ( ) ( 1 ( ) ( )) be the vector of type 's equilibrium
probabilities of winning at least units (i.e., his value of winning exactly units
is =1 and his probability of doing so is ( ) +1 ( )). Then equation
(1) becomes
( ) ( )+( ) ( ) (10 )
(in which ( ) ( ) is the dot product of ( ) and ( )) and the standard
kind of argument now applies.1

Even more generally, if the units are not identical, let ( 1 )

be the type whose valuation of the possible allocation of the units is ,
0
and we again get equation (1 ). (So with distinct indivisible units and
+1
bidders, = if the auctioneer can retain units; a bidder's valuations may
be di®erent for allocations which di®er only in the assignments to other bidders, so
externalities between bidders can also be taken into account in this formulation.)
(See Engelbrecht-Wiggans, 1988 and Krishna and Perry, 2000.)

So revenue equivalence applies fairly generally in the multi-unit context,

in the sense that any two auctions that allocate the objects in the same way (i.e.,
have the same ( ) function) and give the same surplus to a particular type, are
revenue-equivalent under the usual kinds of conditions.

However, it is important to note that standard auction forms will not in

general result in the same allocation, ( ). In particular, while most standard
auctions achieve the e±cient allocation and are therefore revenue equivalent in
the symmetric single-unit case, this is not true in the multi-object case because
the demand-reduction and other problems both a®ect them and a®ect them dif-
ferently (see Section 1.10 and the Afterword to Chapter 1).
 (e) Equation (1) holds for types that exist with positive probability, but
(4) (i.e., ( ) ) may not be de¯ned everywhere. However, the argument will go
through, and so revenue equivalence will apply, unless the distribution is neither
strictly increasing nor atomless, and fails only if there is an atom at the edge of
a \gap":

When the distribution of bidders' feasible valuations has a gap from ¡

to + (i.e., the distribution is not strictly increasing in this range), the equation
( + )¡ ( ¡ )
after (3) becomes: ( + ) +¡ ¡ ( ¡)
+ ¡ + ¡ ¡ ¡ + ¡ +
= ( ) [ ( )+( ) ( ) ( )+( ) ( )]

When there is a gap but not an atom, ( + ) = ( ¡ ), so ( + ) is

determined by ( ¡ ) (and everywhere else it is determined by (4): ( ) = ( )),
so revenue equivalence holds.

Similarly, the surplus function is pinned down when there is an atom but
not an adjacent gap. Suppose there is an atom at . Then there will be a
discontinuity in the ( ) function at : bidders with values slightly higher than
win against the positive measure set of bidders with value , whereas bidders
with values slightly lower than do not. When there is not also an adjacent gap,
all this means is that there is a discontinuity in ( ) and hence a kink in the
( ) function at , so the entire surplus function is pinned down, and revenue
equivalence again holds.

When there is an atom at the edge of a gap, however, it is impossible to

pin down the surplus function: ( ¡ ) ( + ), so ( + ) is not determined by
( ¡ ), and revenue equivalence fails. See (the solution to) Exercise 2 of Auctions:
Theory and Practice, which we discuss below.2

Nevertheless the standard auction forms generally remain revenue equiv-

alent even when there are atoms at the edges of gaps (and hence also mixed
strategy equilibria in, for example, ¯rst-price auctions).

For example, consider the case of symmetric bidders each of which has
a private value independently drawn from a ¯nite set of types, 1
¡1
¡1
with probabilities 1 respectively. Let ( ) = =1 (so ( ) is
the probability that a bidder with value has a strictly higher value than all his
1 competitors). In a ¯rst-price auction, type 1 bids 1 , while other types
 randomise. By standard arguments there can be no atoms in the randomisation,
nor can there be gaps (since the lowest type above a gap would lower his bid),
nor can the ranges of di®erent types overlap (since an overlap would imply that a
higher and a lower type of the same bidder are both happy choosing these di®erent
strategies that give the higher type a lower probability of winning3 ). So type is
indi®erent about mimicking type ¡1 by bidding at the top of type ¡1 's range
(which is the bottom of type 's range). But in an ascending auction, also, type
is indi®erent about (almost) mimicking type ¡1 by dropping out at (just
above) ¡1 . So in both these auctions

( ) = ( ¡1 ) + ( ¡1 ) ( )
= ( ) = ( 1) + =1 ( ¡1 ) ( ) (50 )
That is, bidder surplus is pinned down and equal across the standard auctions in the
\discrete" (¯nite number of types) case, just as in the \continuous" case (in which the
distribution from which bidders' types are drawn is strictly increasing).

The point is that these auctions have the property that the highest type
always wins so (50 ) approximates (5) in a continuous approximation to the discrete case.
Although we can ¯nd examples of auctions in the discrete case that are e±cient but
not revenue equivalent to the standard auctions (and may raise more money than the
standard auctions), these examples don't yield e±ciency in a continuous approximation
to the discrete model that \¯lls in the gaps".

To illustrate all these issues, consider Exercise 2's analysis of e±cient mech-
anisms which always assign a unit to just one of two bidders who each have value
with probability and value with probability . E±ciency implies types and
win with probabilities ( ) = 2 + and ( ) = 2 respectively, and (1)
=
( ) ( )+( ) ( ) (1¤ )
and ( ) ( )+( ) ( ) (1¤¤ )
giving a range of solutions from the seller-optimum characterised by (1*) holding with
equality, to the buyer-optimum characterised by (1**) holding with equality. (See
solution to Exercise 2.)

Now consider a continuous approximation of this discrete case in which most

types are very close to or with a tiny density everywhere else. Revenue equiva-
lence holds, of course, for this distribution for any given allocation. Furthermore, for
the types ¡ (the lowest of the types who are very close to ) and + (the highest of
the types who are very close to ), for example
( ¡) ( + ) + ( + )( ¡ +
) (1~ )
 and ( ¡ ) ( + ) + ( ¡ )( ¡ +
) (1~~ )
Of course, ( ¡ ) ( ) and ( + ) ( ) (using equation (1), since ¡
0
+
and 0). But pinning down the allocation in the discrete case (all 's beat
all 's) does not pin down a unique allocation in the continuous analog: an auction
can treat types such as ¡ and + in several di®erent ways while still corresponding to
the (symmetric) e±cient allocation of the discrete case.

To ¯nd the seller-optimum in the continuous case, note that the marginal
revenue is initially just below , then turns very negative at ¡ when the density
suddenly becomes very thin, then becomes barely positive again at + . So to maximise
revenue, all types with a value of ¡ or below are pooled and given an equal chance
of winning.4 (See Bulow and Roberts (1989); as always, we cannot give higher-value
types lower chances of winning than lower-value types.) So ( ¡ ) = ( + ) = ( ).
(Of course, ( ) and ( ) are unchanged from the discrete case.) So (1~ ) and (1~~ )
both hold with equality, and therefore ( ) = ( ) + ( ) ( ), just as in the
seller-optimum of the discrete case.

The \seller-pessimum" of the continuous model, by contrast, minimises rev-

enue by pooling all the negative marginal revenues in with the high marginal revenue
bidders at the top of the distribution, i.e., pools everyone at + or above. (Again, we
cannot give higher-value types lower chances of winning than lower-value types.) So
now ( + ) = ( ¡ ) = ( ) and (1~ ) and (1~~ ) again hold with equality, but in this
case ( ) = ( ) + ( ) ( ) as in the buyer optimum of the discrete model.

Finally consider the continuous model for any standard auction (e.g., as-
cending or ¯rst-price sealed-bid) in which the highest-value bidder wins. In this case
( ¡) ( +) = ( ) (recall ( ) = 2 ) since the intermediate types beat
all 's whereas a beats another only half the time. So (1~ ) and (1~~ ) both hold
with (approximate) equality and ( ) = ( ) + ( ) . Again this corresponds
exactly to a standard auction such as a ¯rst-price or ascending auction in the discrete
model in which, as noted above: (a) a type is just indi®erent about (almost) mim-
icking a , and (b) a type who (almost) mimicks a beats all 's and no 's
so wins with probability , so again ( ) = ( ) + ( ) .
 The way I think of all this, then, is that revenue equivalence holds in the discrete
case over mechanisms that would induce the same allocation across all types including
those types that do not actually exist (but would exist if we ¯lled the gaps with a
continuous approximation): if a type ¡ were to exist in our discrete example he
would pool with the 's in the seller-optimum; if a type + existed he would pool with
the 's in the seller-pessimum; but either of these types would beat the 's and lose
to the 's in a \standard" auction.

References
Brusco, S. and Lopomo, G. (2002) Collusion via Signalling in Simultaneous Ascend-
ing Bid Auctions with Heterogeneous Objects, With or Without Complementarities.
Review of Economic Studies, 69, 407-436.
Bulow, J. and Roberts, J. (1989) The Simple Economics of Optimal Auctions,
Journal of Political Economy, 97, 1060-1090.
Cramton, P., Gibbons, R., and Klemperer, P. D. (1987) Dissolving a Partnership
E±ciently, Econometrica, 55, 615-632.
Cramton, P. and Schwartz, J. A. (2000) Collusive Bidding: Lessons from the FCC
Spectrum Auctions. Journal of Regulatory Economics, 17, 229-252.
Engelbrecht-Wiggans, R. (1988) Revenue Equivalence in Multi-Object Auctions.
Economics Letters, 26, 15-19.
Graham, D. A and Marshall, R. C. (1987) Collusive Bidder Behavior at Single-
Object Second-Price and English Auctions. Journal of Political Economy, 95, 1217-
1239.
Klemperer, P. D. (2000) What Really Matters in Auction Design. Working paper,
Nu±eld College, Oxford (revised version reprinted as Chapter 3 of Auctions: Theory
and Practice).
Krishna, V. and Perry, M. (2000) E±cient Mechanism Design. Working paper,
Pennsylvania State University and Hebrew University of Jerusalem.
Maskin, E. S. and Riley, J. G. (1985) Auction Theory with Private Values. Amer-
ican Economic Review, 75, 150-155.
McAfee, R. P. and McMillan, J. (1992) Bidding Rings. American Economic Review,
82, 579-599.
Pagnozzi, M. (2003) Bids as a Vehicle of (Mis)Information: Collusion in English
Auctions with A±liated Values. Mimeo, Oxford University.
Robinson, M. S. (1985) Collusion and the Choice of Auction. The RAND Journal
of Economics, 16, 141-145.
A3. [A closely related question is Problem 9. See also Problem 1 (and Part 3 of my
¯rst-year course notes).]
(i) By symmetry will win if .
So will quit at = + =2
[ is just indi®erent about quitting or not at this point. It is easy to see that if
were to win later she would lose money (assuming is following her equilib-
rium strategy) and quitting earlier would mean quitting when 's value exceeds
the asking price (again assuming is following her equilibrium strategy). See
Appendix 1.D (or section 3.2 of my ¯rst-year course notes).]

(ii) If has type , then conditional on winning, is uniformly distributed

as [0 ] So on average will quit at 2 2 conditional on winning.
i.e. 's expected payments are conditional on winning.
By the Revenue Equivalence Theorem, the expected payment of type is the
same in the English and Dutch auctions, so it is the same conditional on winning
(since type has the same probability of winning either auction).
So type bids in the Dutch auction.

2
(iii) Type 's expected payments are unconditional on winning in the
English (and Dutch) auction.
By the Revenue Equivalence Theorem, the expected payment of type is the
2
same in the English (and Dutch) and All-Pay auctions, so type bids in the
All-Pay auction.

(iv) We look for a symmetric equilibrium in which type bids ( ).

If bids ( ), she beats types , i.e. with average signal 2 . So conditional
on winning type 's value is + 2 , and she wins with probability , so type 's
expected surplus from bidding ( ) is = ( + 2 ) ( ).
0
= +2+ 2
( )
= 0 at = in equilibrium
0
= ( ) = 2 with boundary condition (0) = 0 (since = 0 never wins in
equilibrium). The solution is ( ) = 2 which agrees with part (iii).

(v) See Klemperer (1998) (or section 3.5 of my ¯rst-year course notes):
bidder 2 quits at 2 in the ascending auction (if she bothers to enter at all); the
Dutch auction solution is almost una®ected from part (ii).
For revenue equivalence between two auctions to hold, we need that the prob-
ability that any given type of bidder will win is the same in the two auctions.
 See Appendix 1.A. (It would be su±cient for the purposes here that the object
always went to the bidder with the highest signal.) But here type of bidder
1 wins an ascending auction with probability 1, but wins a Dutch auction with
probability only (slightly above) , see section 1.7.2 of the book or Klemperer
(EER 1998) (or section 3.5 of my ¯rst-year course notes).

A4. (i)(a) Initially each bidder stays in until it would be just indi®erent if it suddenly
found itself a winner, i.e., if two others quit simultaneously. So at the point at which
a bidder with value quits the price must be

= + [2 + ( )] (1)

(i.e., if two others were to quit, their signals would be assumed to be and the
remaining signal would be assumed to be somewhere above ). So after the ¯rst quit
at price , the signal of the quitter is inferred to be the that satis¯es (1). Call this
signal (4) . Remaining bidders with signals then quit at prices such that they would
be just indi®erent about ¯nding themselves winners, i.e., at prices

= + (4) + + ( ) (2)

(i)(b) If a bidder with value were suddenly to ¯nd himself a winner at the point at
which he was himself about to quit he would infer the quitter's signal would also be ,
and the two bidders excluded from his auction had signals ( ). i.e. he quits at price

= + [ + 2 ( )] (3)

(i)(c) In the single-auction case, (a), expected revenue = 2 (3) + (4) + (3) + ( (3) )
using (2). For the uniform distribution, ( ) = 5 , so expected revenue = 2( 25 + 15 +
5¡

2 3 12
5
+ 5
= 3 25 . And in the two-auction case, (b), expected revenue = 2
) (2 of 2) +

(2 of 2) + 2 ( ) in which (2 of 2) is the second-highest of two signals, using (3), so for

the uniform distribution expected revenue = 2( 13 + 13 + 2( 12 ) ) = 3 13 .

(ii) (a) In the single-auction case the expected revenue is ( (1) + (2) ). In
the two auction case the two high-signal bidders are in the same auction one-third of
the time (in which case the winner are the highest and third-highest signal bidder), so
2 1
the expected revenue is ( (1) + 3 (2) + 3 (3) ). So revenue is higher in the
single-auction case if ( (2) ) ( (3) ).

(b) (I) For ( ) = ()= + 1 so (2) (3) always, so the single-

auction case is more pro¯table.

(II) For ( ) = 1 ¡2 , ()= 2

so (2) (3) always, so the two-auction
sales mechanism is more pro¯table.5

(III) Since the problem is pure common values, any allocation is equally e±cient, so
bidders' preferences are precisely the opposite of the seller's.

1¡ ( )
(IV) In the private-value case, = , marginal revenue = ( )
is \down-
¡2
ward sloping" for both ( ) = ( ( )=2 1) and ( ) = 1 ( ( ) = 2 ),
so the seller prefers the single auction for either distribution.

(c) A good exam answer would just be that non-decreasing is much more likely
with common than with private values, so it's much more likely that separate auctions
are more pro¯table with common than with private values.
A good answer might also point out why: with pure common values (our prob-
lem) non-decreasing just means ( 1¡( () ) ) is non-increasing in which is ob-
1¡ ( )
viously much easier than having ( ( )
) non-increasing. We could note that
non-decreasing is even easier in common-value contexts if values are non-additive
(e.g., we could discuss the \maximum game" see Bulow and Klemperer \Prices and
the Winner's Curse", Rand Journal, 2002). With private values, by contrast, we know
(2) = ( () (2) ) so ( (2) ) = ( (1) ) and, of course, (2) (2) , so
( (1) ) ( (2) ) which is at least suggestive of the di±culty of having su±-
ciently non-decreasing for split auctions to be more pro¯table (but see below).

I did not expect any examinee to write any of what follows:

 One might even wonder whether separate auctions could ever be more pro¯table
in the private case since always ( (1) ) ( (2) ). However, it is not always
true that ( (2) ) ( (3) ), because (3) = ( () (3) ) so ( (3) ) =
1
2
( (1) + (2) ) (since (3) is the price, i.e. average revenue, obtained from selling
to everyone above (3) who on average are (1) and (2) ), so ( (2) ) = (2 (3) (2) ) =
( (3) )
A simple example which give separate auctions being more pro¯table with private
values is a two-point support: probability 10% of ; probability 90% of , so
Prob ( (4) = ) = 0 0001, Prob ( (3) = ) = 0 0037, Prob ( (2) = ) = 0 0523,
etc.6 For this private-values example, it's actually easier to note directly that expected
revenue from a single auction is (2 (3) ) while that from two separate auctions is
( 13 ( (2) + (4) + 23 ( (3) + (4) )) so the separate auctions win i® ( (2) (3) ) 3 ( (3) (4) )
(Of course it's easy to check that the MR formula yields this too.7 )
(1) + (2)
Returning to private values vs common values: since ( (3) ) = 2
while ( (2 of 2) ) = ( 16 ( (2) + (4) )+ 13 ( (3) + (4) )) we ¯nd (from part (i)) that two separate
auctions are more pro¯table in our common value case i® 2 ( (2) (3) ) 3 ( (3) (4) )
which is clearly easier to satisfy than the private-value condition.
It's also easy to look at bidders' preferences in the private-value case. Bidder
surplus is welfare less revenue, i.e., of the winning bidders (in expectation).
So the bidders prefer the single auction i®

( (2) (2) ) ( (3) (3) )

1¡ ( )
(2) 1¡ (
(3) )
or ( (2) ) ( (3) )
Note that this is precisely the condition for the
seller to prefer the two auctions in our additive common-value model. With linear
demand this is false (because is twice as steep as demand everywhere) but with
¡2
CE -2 demand ( ( ) = 1 ) it is true (because is half as steep as demand
everywhere). With CE -2 demand, bidders with private values dislike the split auction
because the e±ciency losses outweigh the small reduction in payments to the seller.
Those who are truly fascinated by this problem may wish to look at Ellison, Fu-
denberg and MÄobius' paper "Competing Auctions", Journal of the European Economic
Association (2004) though this paper is somewhat di®erently focused. (MÄobius is a
former Oxford Economics M.Phil student who is now (2006) a Harvard Professor).
A5. SKETCH ANSWER

The bare bones of an answer are discussed in the book (Auctions: Theory and Prac-
tice, Klemperer (2004)) see e.g., pages 33 (including note 91) and 63-4 (including note
7). Other useful references are Rothkopf et al (JPE 1990), Milgrom's 2005 Clarendon
Lectures (given in Oxford this academic year), and Ausubel and Milgrom "The Lovely
but Lonely Vickrey Auction" (Chapter 2 of Combinatorial Auctions, Cramton, Shoham
and Steinberg,eds.(2006)).

The e±ciency properties in the private-values context are standard.

Problems that apply even in the single-unit context include

- others can see the winner's value (so the winner may be vulnerable to ex-post ex-
ploitation (Rothkopf, JPE, 1990), or auctioneer exploitation (the auctioneer may have
a confederate place a bid just less good than the winner's to leave the winner (almost)
zero surplus), or even absent foul play the result may be a PR disaster (as in the
celebrated New Zealand auction, pg 110 of the book)
- the e±ciency of the auction can facilitate entry deterrence by strong players and be
disastrous for revenue (see the general discussion of entry in 1st vs 2nd price auctions
in section 4.3.1 of the book and elsewhere)
- disadvantages of 2nd price auctions vs 1st price auctions more generally can be
discussed (revenue properties with risk-aversion, asymmetries, budget constraints, etc;
see book)
- e±ciency properties are limited outside the private-value context.

In the multi-unit context there are more problems including:

- the auction is \non-transparent", that is, payments are hard to understand for non-
experts, as is the mechanism itself for many people
- di®erent bidders pay di®erent amounts for identical objects and often those who value
an object more pay less
- in multi-unit contexts the mechanism \favours larger bidders over smaller bidders"
(e.g. relative to a standard ascending auction)
- when there are complementarities, there are curious and di±cult-to-guard-against
collusive possibilities (even for bidders who would all lose absent collusion)
- collusion may be facilitated even absent complementarities (though ascending auctions
are worse for collusion -see the book)
- demerging, or setting up new bidders, can help a bidder
- revenues can be very low, when goods are complements (and, relatedly, adding bidders
can sometimes reduce revenues) as discussed in, for example, Milgrom's Clarendon
lectures.
 A6. [I have ignored subscripts.]
(i) Total bidder surplus is the expected high value of 2/3, since expected transfers
have to sum to zero. So surplus per bidder is on average 1/3 versus 1/6 in a \regular"
auction.
(ii) The revenue equivalence theorem applies exactly, and for the same reasons, as
\normal". So
( ) = (0) + ( )
0

Clearly, ( ) = Also, given the symmetry of the problem, (0) = 1 6 because

average surplus is 1 6 greater than in the \regular" auction (1 3 versus 1 6) and we
know from the revenue equivalence theorem that the di®erence is a constant type-by-
type since = ( ) (the same amount) in both games, for all types. So
2 2
1
( ) = (0) + = +
2 6 2
(If necessary, candidates could work out that (0) = 16 by using the fact that
( ( )) = 13 )
(iii) In the sealed-bid auction the bidder will pay his bid when he is highest and
receive the competitor's bid when that bid is higher. For a bidder with value we have
1
2 2
1 6+ 2= ( )= ( )+ ( )

since the right-hand side computes expected surplus as value times probability of buy-
ing, less expected payments when a buyer, plus expected receipts when a seller. We
can now see directly which linear function solves this equation, or we can di®erentiate
both sides to obtain
=2 ( ) ( )

which yields = 1 2 ( ) Since the question makes clear that we will ¯nd is a
constant it is easy to see that a solution is ( ) = 3 (It is also clear directly from the
formula for ( ) that (1) = 1 3 since a bidder with value 1 must be a winner.)
(iv) In the ascending auction the analogous surplus equation is:

2 2
1 6+ 2= ( )= ( ) + (1 ) ( )
0

which produces (through di®erentiation of both sides)

=2 2 ( ) + (1 )

so = 2 1¡
( )¡
The formula immediately yields (0) = 1 6 and the solution is therefore
1
( )= 6+3
 Bidders go beyond their value, and take the risk of buying at a price that exceeds
their value, in the hope of selling out at a higher value (as in Bulow, Huang and
Klemperer (JPE 1999) though in a private value context).
(v) Type of a partner with share earns 1 6+ 2 2 , so it is simple to compute
2
that the worst o® type of a partner with share has value and earns 1 6 2 (or
1 24 for = 1 2 | up to now we are assuming = 1 2, but we will want general
in the next part.)
(vi) Nothing is di®erent in the analysis of general shares. If we are to run an e±cient
mechanism, so ( ) = , incentive compatibility requires us to maintain ( ) = (0)+
2
2
All we can do is to alter (0) by having the smaller partner pay the larger partner
for participating in the mechanism.
Given shares and (1 ), the worst-o® types of the two partners earn surpluses of
2 2
1 6 2 and 1 6 (1 ) 2 respectively. Therefore, we can have the smaller partner
1
pay 4 2 (i.e., half the di®erence between these two surpluses) to the larger partner,
2
so that both worst-o® types make (the same) non-negative surplus if (1 6 2) +
2
(1 6 (1 ) 2) 0 | i.e., [1 2 3 6 1 2+ 3 6]. For outside this interval,
e±cient individually-rational dissolution is impossible if the partners know their own
values before agreeing to a mechanism. (This con¯rms Myerson and Satterthwaite's
(JET 1983) result on the impossibility of e±cient bilateral trading when one player
owns 100% of the good to be traded.)
See Cramton, Gibbons, and Klemperer (Econometrica 1987) for the general case;
note also that the problem is identical to one in which the two bidders collude in buying
from a seller, by bidding against each other in a pre-auction knock-out for the right to
buy the asset from the seller for zero.
 A7. [Sketch Answer]
(i) [Paul Klemperer was quoted in the Economist 9/9/06, p77.] He was probably
suggesting that auctioning larger blocks of spectrum might have worked better in this
particular case.
(ii) concern about "exposure" among some bidders. That is, bidders who had a
strong preference for complete coverage of a region were reluctant to bid on smaller
blocks for fear that they would be "stranded" winning some but not all parts of the
region. The result could be that bidders who either wanted only small blocks, or were
willing to risk winning only some parts of the region could have picked up those areas
cheaply. The fear of stranding would be reinforced if resale was not permitted, or was
hard for some other reason. Larger bidding increments are perhaps slightly more likely
to create this problem.
(iii) a Vickrey auction would sell private-value objects e±ciently at disparate prices
assuming a ¯xed number of non-colluding bidders, private values, and no budget con-
straints.
A sequential auction (by English/Japanese, Dutch, ¯rst-price sealed bid, second-
price sealed bid, etc) or a discriminatory auction of identical objects is, in theory,
e±cient assuming symmetric bidders.
Perfect price discrimination is e±cient, and imperfect price discrimination may be
more e±cient than a ¯xed price.
(iv) if there are no complementarities at all between the smaller areas, then the
objects should be sold independently, but if the smaller areas are perfect complements
(individually useless to an owner without the whole region) then they should be sold
together. More generally, selling a small number of large licences makes life very
di±cult for bidders who only want smaller areas, especially if resale is hard. So it
depends upon whether it is more e±cient to encourage bidders who are interested in
winning smaller areas (these bidders may be new entrants) or bidders interested in
winning larger regions.
(v) in a package auction, a bidder can specify a bid for a group of properties and then
wins all of these if his bid beats the sum of the best bids for the individual properties.
This still makes life very di±cult for smaller bidders who have di±culty coordinating
their bids to beat the package bidder, and also face an "free-rider" problem (each
small bidder wants the other small bidders to bid higher prices that will contribute to
defeating the package bidder). There are other problems: a bidder who wants to bid
for a package as well as for individual units is bidding against himself; if bids must be
left on the table in subsequent rounds then bidders may be reluctant to make o®ers,
but if bids need not be left on the table coordination is particularly hard, etc.
 A8. [Sketch of a possible answer.]

1. The candidate might begin by showing that under the appropriate conditions,
all auctions make the same expected pro¯ts
(see sec 2 of my ¯rst-year class notes on the course website)

2. But the pro¯ts are only the same in expectation,

and under the assumptions that:
- The number of bidders is ¯xed
- They play noncooperatively
- There is a single indivisible unit (or if there are multiple identical objects, bidders
want a single indivisible unit each)
- Bidders' values are independent
- They are risk neutral
- There are no budget constraints
- There are no externalities between bidders
- Bidders are drawn from symmetric distributions, and the auction is a standard
one, so that the highest actual type is the winner (or some alternative assumptions so
that the the lowest type's payo® is constant (e.g., zero), and a given type of a given
bidder wins any auction with equal probability).
Note that setting a reserve price a®ects types' probabilities of winning { so setting
a reserve price is an important part of auction design, even if all the other assumptions
above hold.
(Not all these issues were discussed in the ¯rst-year lectures, but many were)

3. The candidate might then discuss how auctions' expected pro¯ts di®er if these
assumptions are relaxed. (With the exception of independence, probably externalities,
and in some cases symmetry, relaxing the assumptions above generally favours ¯rst
price over second price mechanisms.)
(only some of this was discussed in the ¯rst-year lectures)

4. The auctioneer's objective may not be (only) revenue. (In particular, e±ciency
considerations tend to favour second price mechanisms.)
 A9. (i) See Appendix 1.C of Auctions: Theory and Practice. See also Section 1.6.
(ii) References, and the bare bones of an answer are in Section 1.6.
(iii) See Section 1.6, Appendix 1.C and also, e.g., Section 4.2.
See also question 17 of the book and its sketch solution.

A10. (i) A. We run a "second marginal-revenue auction". The woman's demand

curve is 80 40 , so her marginal revenue is 80 80 = 80 80((80 ) 40) = 2 80,
while the man's demand is 40 40 , so his =2 40. So we ask the bidders to
state their values, and give the object to the bidder with the higher marginal revenue,
provided that marginal revenue exceeds the seller's valuation, at a price equal to the
lowest value at which that bidder could have won given the other bidder's report (and
given the seller's cost).
B. Since the woman's marginal revenue beats the man if and only if her value
exceeds his by at least 20, we could give the man a discount of 20 conditional on him
being the winner, and simply run a standard ascending auction with a reserve price of
45.
C. An optimal auction maximises expected marginal revenue, just as a price dis-
criminating monopolist maximises the sum of marginal revenues. Because the man is
drawn from a distribution that is lower by a constant, he has a higher marginal revenue
at any given value.
(ii) A. This is the problem described in Bulow and Roberts (1989), pages 1078-
81, but with 20 deducted from all the valuations. The demand curve for 20 is
80 200 , so = 80 400 = 2 80 (as before); demand for 20 is 50 50 , so
= 50 100 = 2 50. Because marginal revenue is not downward sloping, selling
to the highest value bidder does not sell to the highest marginal-revenue bidder, so the
seller would do better to "iron" the marginal revenue curve and can do this best by
treating all bidders with values between 35 and 50 equally, as if they all had marginal
revenues of 20 (since the triangular areas cut o® by the = 20 line are then equal
in a diagram like Bulow and Roberts' ¯gure 3). The reserve price is 30, since a bidder
with value 30 has = 10.
The higher-value bidder is sold to at the price of the lower-value bidder except that
the randomisation (if it is necessary) is at a price of 35, and that if the winner's value
exceeds 50 and the runner-up's value is between 35 and 50, the winner pays 42.5 (so
that if the winner had a value of 50, as was just necessary to guarantee victory, it would
have been indi®erent between its outright victory and participating in the lottery).
 [The question did not ask this, but this outcome could be achieved by running a
simple ascending auction, but skipping all the prices between 35 and 42.5. In fact,
no-one will then quit between 35 and 50; any bidder with a value between 35 and 50
will quit at 35{if its competitor quits simultaneously, it prefers a fair lottery at 35 to
winning for sure at 42.5, and if its competitor doesn't quit simultaneously, it is doomed
anyway. (With bidders still in the auction at a price of 35, the auctioneer would need
to jump the price to 50 (15 ).)]
B. It doesn't matter whether or not we treat the man and woman fairly, provided
that for each individual we make it equally likely that all values between 35 and 50
are equally likely to win, so that the expected of the winner is as in the case of
"fair" allocation (and we must choose prices so that this outcome is achieved given the
incentive compatibility constraints).
If we always favour the man in the "ironing" range, then the higher-value bidder
is sold to at the price of the lower-value bidder (as before), except that: if both values
are between 35 and 50 the man wins at a price of 35; if the man's value exceeds 50 and
the woman's value is between 35 and 50, the man pays 35 (so if the man had a value of
50, he would have been indi®erent between stating that value and stating a somewhat
lower value); and if the woman's value exceeds 50 and the man's value is between 35
and 50, the woman pays 50 (so if the woman had a value of 50, she would have been
indi®erent between stating that value and stating a somewhat lower value).
 A11. (i) see part 2 of the course notes on the web (or Appendix 1A of my book).

(iia) 2 (= expected second highest signal).

pro¯t = 4 (= expected di®erence between highest and second highest
signal).
[Recall that the expected value of the highest of independent draws
from a uniform distribution on [ ] equals +[( +1 ) ( +1)]( )
from the course notes on the web, or Appendix 1A of my book, or this is
not hard to compute directly, for = 3 ¯rms.]
(iib) 2, and pro¯t = 4 by revenue equivalence theorem with ascending auc-
tion.
[Recall revenue equivalence applies to expected bidders' surpluses as
well as to expected total payments.]
(iiia) the contest for the uncommitted consumers is a sealed bid auction which
yields ( ) 4, by revenue equivalence with ascending auction. All three
¯rms also make (1 ) automatically on their locked in customers, and
average costs are 2, so expected industry pro¯ts are 3 (1 2) + (
) 4
(iiib) prize is ( )(1 ), independent { of course { of the bids. Payment
is (1 ) for winner, (1 )for losers. (Any ¯rm with the highest
possible cost would choose = 1, hence making expected surplus of zero
as required for the revenue equivalence theorem to hold.) So this auction
yields ( ) 4 by direct analogy with (iib). In addition, all three ¯rms
also make (1 ) automatically. So expected industry pro¯t = 3 (1
2) + ( ) 4.
(iv) losing the ability to price discriminate means ¯rms no longer rip-o® their
locked in consumers so much - but it also dulls competition for the uncom-
mitted consumers. In this example the e®ects cancel. More generally, either
could dominate.
[See Klemperer (Review of Economic Studies, 1995, pp. 515-539).]

The model of the question is from appendix B of Bulow and Klemperer (Brookings
Papers: Microeconomics, 1998, pp. 323-394) which applies this model to the cigarette
market.
 A12. (Note that those in charge of the exam slightly altered the wording of the
question.)
i. a. A is indi®erent about winning at price if +( )(1 ) = or =
( + 1) so A stays in the bidding until the price reaches = in
which = ( + 1) and then quits.
b. When = 1, we have pure private values so bids up to its private value.
When 1, we have common-value elements to valuations, so faces a winner's curse
and therefore bids less, the more aggressively that bids (i.e., the larger is ). [One
way of thinking about this equilibrium is that it illustrates the "strategic substitutes"
and "strategic complements" analysis developed in Bulow, Geanakoplos and Klemperer
(JPE, 1985). With pure private values, your bidding is una®ected by how aggressive
your opponent is. But with a common values element, the more aggressively your
opponent bids, the worse is your winner's curse, so the less aggressively you will want
to bid. That is, we have strategic substitutes. And this e®ect is greater, the greater is
the common value element, that is the smaller is ]
c. Clearly, = ( + 1) It is easy to see that = = 1 is a Nash
equilibrium.
[there are imperfect equilibria in general (one bids to , the other bids 0); assuming
the 's have positive density everywhere, there are other perfect Bayesian equilibria
only in the pure common values case, = 1 2 ]
ii. a. A is indi®erent about winning at price if +( )(1 ) = or
= ( + ) so A stays in the bidding until the price reaches =
in which = ( + ) and then quits. [If (1 ), A bids ]
b. As before, = ( + 1) Substituting the claimed equilibrium values of
and into this expression for and into the expression for that was derived
in the previous part, con¯rms the claimed equilibrium.
c. = ; = ; i.e., stays in forever; bids up only to the minimum
possible value it can have conditional on its own signal. The reason is that it is
common knowledge that 's value exceeds 's, so it can never be rational for to
buy if is unwilling to.
iii a. As (1 ), quits at so pro¯t = ( ). At = 1, pro¯t
= min( ), which is greater if and are independent draws from the same
uniform distribution, and 2 3.
b. Run a ¯rst price auction; use a reserve price; give a toehold or options or
other subsidy to the weaker player (e.g., counting 's bid as times what it actually
is{equivalently giving a discount equal to a fraction ( 1) of its bid{yields
symmetric behaviour, so pro¯t = (( + 1) 2) min( ), since the auctioneer has
to pay the subsidy half the time.)
 A13.(sketch answer)
A student might begin by discussing possible designers' objectives, before addressing
the two most obvious
1. e±ciency:
In theory, at least, any standard auction is ¯ne if bidders are symmetric. [This
assumes there is only a single indivisible unit on o®er, or bidders just want a single
indivisible unit|my answer will stick to this case, since that is all that those answering
the question had been taught about.]
More generally, an ascending auction does well, but needs to be \open ascending"
(i.e., not just a 2nd price) if there are common value components. (With pure private
values, a 2nd price auction is e±cient.)
2. revenue
Begin by invoking the Revenue Equivalence Theorem (section 2.1 of 2011 revision of
class notes) under the conditions of which auction design is irrelevant. (One approach
would be to spend some time of explaining this result.)
Note that the independent signals assumption is less natural for common values
and for private values. It's not clear how much this matters. (I hadn't discussed
a±liation with those answering the question, but in any case it is not clear that its
a®ect is quantitatively large either in practice or in theory.)
With either private or common values, an ascending auction faces the "almost-
certain winners" problem. I emphasise this (and it is perhaps particularly important)
for the \almost common value case" (section 3.5 of 2011 revision of class notes). It
should be explained why the RET fails in this case|the reason is that a bidder's
probability of winning (as a function of its \type") di®ers greatly across auctions.

A14. (i) Each object is sold by a separate ascending auction. (Objects may di®er
from each other.)
All auctions take place simultaneously.
No auction closes until no one wants to bid again in any auction.
A bidder who is the current high bidder on an object cannot remove his bid unless
his bid is topped by another bidder.
There may be additional rules about the number of objects you can be the high
bidder on at any one time; and/or about the amount by which any new bid must top
any standing high bid; and/or about either excluding a bidder, or reducing the number
of objects that it is eligible to win, if it has not been su±ciently active in the bidding.
 (ii) An SAA is e±cient when bidders (a) want at most a single unit each, and
have (b) private values, and (c) no budget constraints. Because assuming (a) and (b),
"straightforward" bidding is a Nash equilibrium. ("Straightforward bidding" means
bidding in every round as if the current round is the last.) And assuming (b), (c) and
"straightforward" bidding, we are guaranteed the socially e±cient outcome.
(a) is necessary, because otherwise a bidder may gain from demand reduction
to end the auction at a lower price than if he continued to bid "straightforwardly" .
(b) is necessary, because otherwise a bidder may gain from distorting its bidding
to hide information from rivals
(c) is necessary, because otherwise a bidder may not be able to outbid a lower-
value but less-budget-constrained rival (just as in an ordinary auction for a single
object).
[A candidate could comment , perhaps in part (iii), that straightforward bid-
ding is easily understood to be an optimal strategy (assuming (a), (b), and others
do likewise) so bidders' behaviour is likely to approximate this, and hence to achieve
e±ciency in practice (not merely in theory).]
(iii) If there are either common-value components to valuations, or complementari-
ties, there may be ine±ciencies. However, even though the auction may not work well,
it may nevertheless work better than alternatives in these circumstances { because the
multi-round aspect allows bidders to infer common value information from others, and
also to work out what complementary bundles they may be able to win.
Perhaps most important is that multiunit demand can create a variety of prob-
lems. First, it permits demand reduction ("oligopsony behaviour"), and facilitates
collusion and/or predation through signalling behaviour. Second, if there are comple-
ments, bidders may face "exposure" problems.
Problems may be created by the time taken by the auction, and (relatedly)
the costs of running and of participating in the auction, which might also reduce entry
into the auction. Weaker bidders may be deterred from entering, as by any ascending
auction design.
[A candidate could comment on the implications of revenue being an objec-
tive. There is a general presumption that e±cient allocation across bidders is likely to
increase revenue by creating a larger pie to share out (more precisely, greater e±ciency
tends to increase the likelihood of selling to bidders with higher marginal revenues).
An (ine±cient) reserve price might also increase the SAA's
revenues. One could make some standard comments about the roles of risk aversion,
a±liation etc., in ascending vs. sealed-bid style auctions. With multiunit demand, a
seller interested in revenue maximisation may do better by bundling.]
[Given the way I taught the course, candidates may choose to talk about cir-
cumstances under which core-selecting and/or product-mix auctions might perform
better. But I (deliberately) wrote the question in a way that did not ask for this.]
 A15. (sketchy answer)
With private values and no budget constraints, it is a dominant strategy to \tell
the truth" and the outcome is then fully e±cient!
But there ARE incentives for \dishonest" bidding if there are budget constraints,
common values, or \wider-game" issues (you want to keep your information secret for
reasons beyond the current auction (Rothkopf et al, JPE, 1990).
Collusion may be much easier since with complements it requires only two
participants to pro¯tably collude (by contrast with most industrial organisation context
where collusion only is very e®ective if all market participants engage in it). Relatedly,
there are unusual incentives not only to merge, but also to demerge (with complements),
and shill-bid (Ausubel-Milgrom, in Cramton et al (ed), 2006).
Revenue may be lower than in other auctions.
Prices can seem odd, and \unfair" (e.g., more-aggressive/higher-value bidder
pays less). Relatedly, and crucially, the auction is hard for participants to understand.
Levin and Skrzypacz's, and Janssen and Karamychev's, current working papers
develop the point that since your own bids only a®ect other people's payments (not
your own), there is a problem if bidders care about others' payments. It is a worse
problem if the Vickrey auction is the ¯nal stage of multistage process (such as the
recent CCAs) so the e®ects feed back into early-stage incentives.
For a single unit, in which case it's a second price auction, and very like an as-
cending auction, it may do ¯ne{especially if the bids can be kept secret. For multiple
homogeneous substitute objects, it may also be ¯ne{and could if desired be imple-
mented as in Ausubel (AER 2004). More generally, for auctions with substitutes it is
perhaps okay (at least when the Levin-Skrzypacz-Janssen- Karamychev critique is not
too problematic).
For an auction with complements there are many more problems{a sealed-
bid core selecting package auction may be an improvement (Day and Milgrom, Intl
Jnl of Game Theory, 2008; Erdil-Klemperer, Jnl of European Econ Assoc, 2010; etc.),
perhaps after a clock auction as a prior stage (see, e.g., Maldoom, Working Paper, 2007;
Cramton, Rev. Ind. Org., 2013)|though this may accentuate the Levin-Skrzypacz-
Janssen- Karamychev problems.
 A16. (i) 1
(ii) prob of winning = [(( 1 5) 10)2 ]; expected gain conditional on winning=
[( 1 5) 3]
Expected gain =[(( 1 5) 10)2 ][( 1 5) 3] = [( 1 5)3 ] 300
(iii) risk neutrality (this su±ces for us to be able to apply the Revenue Equivalence
Theorem)
[might some students talk about rational behaviour, etc? This would be reasonable
as to (iii) why di®ers from (i)|even more so if they also worried about this in part
(ii){so maybe partial credit]
(iv) 1 [( 1 5) 3] = 5 + (2 3)( 1 5) (by RET, gain is the same conditional on
winning)
(v) risk averse bidders all bid higher (to take a larger probability of a smaller gain)
and so gain less; risk seeking bidders all bid less aggressively and gain more
(this was not discussed in class, but was done in the problem set).
A17.
+
(a) A bidding strategy for player is a function : [0 1] that assigns a
bid for each valuation [0 1]
(b) The expected utility for bidder with value is:
( ( ¡ )) = Pr( ¡ ( ¡ ) ( )) ( ( ))
(assuming that draws have zero probability.)
(c) Suppose bidder = uses a linear bidding strategy (bids according to
( )= ). The expected utility of bidder given value and bid is:
( ( ¡ )) = Pr( ¡ ( ¡ ) ) ( ) = ( ) if

FOC for the optimal (which maximizes the expected payo®):

1
0= ( ) = ( ) =
( )(1¡ ) 1+
If bidder will win the object for sure and should therefore never bid
anything larger than . So the equilibrium strategy is:
( ) = min
1+
(d) Part C implies that there is a Bayesian-Nash equilibrium in which for each
player : ( ) = if = because each player is making a best
1+
reply to his opponent's strategy in this case.
(e) W.l.o.g. let 1 2. Then ( ) =
is not an equilibrium, be-
1+
cause Part C implies that player 1 would do strictly better by reducing
his bid to min 1+1 1 1+1 2 (The question did not ask this, but ( ) =
min 1+ 1+1 is also not an equilibrium if = : if, w.l.o.g., 1 2
player 2 bidding 1+ 2 is a best reply to 1 ( 1 ) = 1+1 1 , but is not a best re-
2

ply to 1 ( 1 ) = min 1+1 1 1+1 2 The reason is that player 1 would then be
bidding exactly 1+1 2 with strictly positive probability, and player 2 would
then do better to move any bids that would have been just below 1+1 2 to
just above 1+1 2 )
(f) In order to justify the solution of part (d), it su±ces that the distribution
of values is common knowledge among the players, and, in addition, each
player anticipates that the other player ( ) is choosing a linear strategy
1
( )= for some 1+
(not necessarily the equilibrium strategy of
part (d)).
(I believe this is the only time that an Oxford MPhil exam question on auc-
tions has not been set by me. The original version was °awed, and the
substitute version on this website is included so that students with access to
the original understand the error.)
 A18. [sketch of possible answer]
In both cases bidders bid demand schedules. [The auctioneer computes the ag-
gregate demand, and ¯nds the price at which its supply equals aggregate demand.
All demand above that price is satis¯ed. Winning bids are ¯lled at that price in the
uniform-price auction, but at the prices actually bid in a discriminatory auction (which
is therefore also sometimes called a "pay-your-bid" auction).]
A uniform-price auction bidder's demand schedule starts at the bidder's actual
valuation for his ¯rst unit (exactly as in a second-price auction, and for the same
reason), but a bidder for any larger quantity than a single unit recognises that buying
an additional unit drives up the market price, and so drives up the price it pays for
all the other units it buys. It therefore cuts back the price it bids for larger quantities
exactly as a monopsonist does. (Note, however, that a bidder for a su±ciently small
quantity recognises that it will not signi¯cantly a®ect the market price, so can simply
bid its actual value for each unit it demands - as in a second-price auction.)
By contrast, a bidder's discriminatory demand schedule starts below (typically
well below) the bidder's actual valuation for his ¯rst unit. Exactly as in a ¯rst-price
auction, bids for high-valued units will be shaded down a long way towards the expected
market clearing price. However, also as in a ¯rst-price auction, bids for low-valued units
will hardly be shaded at all.
Indeed the pressure towards °at demand schedules is even greater for (multi-
unit) discriminatory auctions than in (single-unit) ¯rst-price auctions. A bidder recog-
nises that winning a second unit requires beating more opponents' bids than winning
just one unit, so the bidder faces stronger competition for the second unit than for the
¯rst. (Indeed, if the bidder's "true demand" is fairly °at (i.e., it has a fairly constant
marginal valuation of units), it might even want to bid more for the second unit if the
rules of the auction permitted that. However, we assume the auctioneer takes bids in
descending order.)
So discriminatory demands will generally be °atter than uniform-price de-
mands, and uniform-auction bids are more dispersed.
[A good answer could provide more detail. I handed out, and discussed in the
lecture, a series of pictures illustrating these points, and those below.]
(i) At least in theory, a uniform-price auction can sustain implicitly collusive strate-
gies as a one-shot Nash equilibrium. This may particularly be a problem in frequently-
repeated settings, with small numbers of bidders, di±cult entry, and very stable de-
mand (e.g. electricity). (This is probably not a problem if supply can be varied by the
auctioneer, or even just if supply is elastic.) The discriminatory auction is in principle
unproblematic, though if it does discourage entry it may facilitate collusion simply by
reducing the number of participants.
(ii) A uniform-price auction is attractive to small bidders, because of the way in
which large bidders reduce their demand. Furthermore a small bidder is protected
by the fact that everyone pays the same price. (A possible exception is that weak
bidders may ¯nd entering a uniform-price auction unattractive, just as weak bidders
¯nd ascending auctions of small numbers of units in which bidders can win at most
one unit each unattractive.)
By contrast, in a discriminatory auction, bidders are playing "Guess the price"
{ that is, information about the likely market clearing price is very valuable, and
collecting and analysing information is disproportionately costly for small bidders. So
the uniform-price auction will increase participation. [This is especially true with
common values, but the question speci¯ed private values in order to reduce the number
of issues to think about.]
A good answer could note that this problem can be resolved by allowing non-
competitive bidding in a discriminatory auction.
(iii) E±ciency will be impacted in obvious ways by participation and collusion
(discussed above).
So let's assume similar participation and no collusion: either kind of auction
can then be more e±cient than the other, because di®erent kinds of bidders shade
their bids downwards by di®erent amounts (see ¯rst part), and these amounts are also
di®erent for di®erent units.
If all bidders have °at true demands (constant marginal valuation of units) for
the same number of units, the bids in a discriminatory auction are °at. (Each bidder
then faces the same competition for the ¯rst unit it would win, or the second unit,
or the third unit, etc. ... so bids the same for each unit.) So if their valuations are
drawn from the same distribution, then just like in an ordinary single-unit auction the
person with the highest valuation bids the most, in Nash equilibrium, so we have full
e±ciency! By contrast, a uniform price auction is not fully e±cient since each bidder
bids a set of decreasing prices for units it values the same.
However, if each bidder has downward sloping demand, the discriminatory auc-
tion may be less e±cient than the uniform price auction even if everybody's valuations
are drawn from the same distribution. (In a discriminatory auction for units, a bid-
der's bid to win an th unit depends upon its th-highest valuation, which is generally
drawn from a di®erent distribution than the distribution of the + 1th opponent's
valuation (which determines the bid it is competing against).)
(Ausubel, Cramton, Pycia, Rostek, and Weretka "Demand Reduction and In-
e±ciency in Multi-Unit Auctions", REStud, 2014 [http://pycia.bol.ucla.edu/ausubel-
cramton-pycia-rostek-weretka-auctions.pdf], provide an example (Ex IV, p1376) in
which the uniform price auction is more e±cient whenever bidders' values of win-
ning a second unit are a su±ciently small fraction of their values for winning a ¯rst
unit.)
Moreover, the discriminatory auction is not e±cient for bidders for a single
unit if their values are drawn from di®erent distributions.
 By contrast, the uniform-price auction is perfectly e±cient when bidders only
demand a single unit, and generally also close to this when bidders are all "small".
More generally, if bidders are not too large, bidding is probably easier in the
uniform-price auction, so actual bidding may more closely approximate the theoretical
equilibrium, and outcomes may be more e±cient.
A very impressive answer might observe that some other issues may a®ect
e±ciency more broadly than in the immediate outcome of this auction: if bidders'
demands are not too large relative to the auction size, uniform bidding is likely to be
more informative; on the other hand, because information is less valuable to bidders in a
uniform auction, bidders may collect more information before a discriminatory auction,
so bidding and allocation, especially in the longer run, may be more e±cient (ignoring
resources "wasted" collecting privately-useful but socially-useless information, and the
fact bidders may strategically withhold information so any pre-market will be thinner
and less informative ...). (There may be other issues too: a uniform price auction
makes a short squeeze harder; possibly e®ects of budget constraints, risk-aversion,
asymmetries, etc.)
It could also be commented that none of this might matter very much to
e±ciency if there is a very good resale market.

A19. [Sketch of a possible answer; a good answer was one that was relevant to the
context given, rather than just discussing Vickrey auctions in general.]

(a &b) With private values (as the question assumed) and no budget constraints,
etc., it is a dominant strategy to \tell the truth" and the outcome is then fully e±cient.
However, there are incentives for \dishonest" bidding if there are \wider-game"
issues (bidders want to keep their information secret for reasons beyond the current
auction, see, e.g.,Rothkopf et al, JPE, 1990), and/or there are budget constraints.8 In
this very simple case, a bidder with a value exceeding its budget constraint will simply
bid its budget constraint, so in that sense bidding is honest, but the outcome may not
be socially e±cient.
 Revenue may be lower than in other auctions|especially in this complements set-
ting.
Prices can seem odd, and \unfair",9 and the auction may be hard for participants
and observers to understand, even in a setting this simple.
So the merits of a Vickrey auction depend on the auctioneer's objectives, includ-
ing the extent to which it cares about participants' and observers' perceptions and
behaviour.
Collusion may be much easier than with other designs or in other settings
since, in this complements context, the Vickrey auction allows only two participants to
pro¯tably collude (by contrast with most contexts where collusion only is very e®ective
if all market participants engage in it). Relatedly, there are unusual incentives not
only to merge, but also to demerge [more precisely, in this case, to stay de-merged]
(since we have complements), and shill-bid (see, e.g., Ausubel-Milgrom, in Cramton
et al (ed), 2006). [My lecture gave examples relevant to the setting of this problem.]
These problems are likely to be particularly severe if the bidders know the number
and identities and interests of the other bidders, and if co-operation and/or (implicit
or explicit) collusion is feasible{so these factors are crucial to the merits of a Vickrey
auction, since there is such a tiny number of bidders.

(c) Since we don't have common values, and there's also no coordination issue in a
problem this simple, there would be limited bene¯t to any ¯rst-stage dynamic process
here. Indeed a ¯rst-stage may actually aggravate problems (e.g., it might facilitate
collusion or even allow someone to risk pushing up price on the object it does not want
to exhaust the complement-bidder's budget).
But a sealed-bid core-selecting package auction may be an improvement. See, es-
pecially, Day-Milgrom, in Neeman et al (eds.), 2013 [this paper updates and corrects
their earlier Intl Jnl of Game Theory, 2007 paper]; smaller contributions include Erdil-
Klemperer, Jnl of European Econ Assoc, 2010; etc.
(There are, of course, other possibilities too (e.g., a revenue-maximising auctioneer
might seek to exploit risk-averse bidders), but it is worth noting that unless we choose
some kind of package format, optimal, or even "sensible", bidding will be very hard for
the complements bidder.)
 A20. (i) In the single-unit case, the auctions are strategically identical.
Even if they are strategically identical, they may be played di®erently in prac-
tice. (The Nash equilibrium is not dominant strategy.)
[Although we did not discuss this in the lectures, it could be mentioned that
with multiple units they are not strategically identical (and nor are equilibrium out-
comes identical, though they may|-under tight conditions|be revenue equivalent).]
(ii) If the ascending auction is a Japanese auction for a single unit, the auctions
are strategically identical if there are just two players. [A good answer would explain
the rules of a Japanese auction, and how the ability to use complicated strategies, make
jump bids, etc makes a free-form ascending auction di®erent.]
With private values (and any number of bidders), a Japanese auction OR an
ascending auction (free-form, but ignoring the bidding increment) is identical in equi-
librium outcome (and likely to have the same outcome, because the Nash equilibrium
is in dominant strategies), but is not strategically identical, to a second-price auction.
[Since I asked about second-price auctions, I would not expect students to
address multi-unit issues (and, anyway, we did not discuss them in the lectures), but
it could be added that (with private values) a ( +1th) price auction for units (with
bidders who want 1 unit each) is identical in equilibrium outcome (and likely to have
the same outcome), but is not strategically identical, to a Japanese auction or an
ascending auction.]
A good answer might note that in an ascending auction the winning bidder
will not (in general) reveal her actual value. In practice, bidders may be unwilling to
bid their true values in a second-price auction for fear of revealing information that
may be useful to others (such as regulators, competitors, or those they might have to
bargain with such as suppliers) after the auction.
(iii) When the assumptions of the revenue equivalence theorem hold they all
yield the same expected payment by each bidder and the same expected revenue for
the auctioneer, in Nash equilibrium.
There are various forms of this theorem. This was the one I proved in the lecture
notes:
Revenue Equivalence Theorem (IPV Case): Assume each of risk-neutral
potential buyers of an object has a privately known value independently drawn from
a common distribution ( ) that is strictly increasing and atomless on [ ] Any
auction mechanism in which (i) the object always go to the buyer with the highest
value, and (ii) any bidder with value expects zero surplus, yields the same expected
revenue, and results in a buyer with value making the same expected payment.
[This one (also in the lecture notes) would be even better:
Assume each of risk-neutral potential buyers of an object has a privately known
signal independently drawn from a common distribution ( ) that is strictly increasing
and atomless on [ ] Any auction mechanism in which (i) the object always go to
the buyer with the highest signal, and (ii) any bidder with signal expects zero surplus,
yields the same expected revenue, and results in a buyer with signal making the same
expected payment.]
The key assumptions are risk-neutrality, independent signals (or values), and com-
mon distribution (i.e., symmetry). (Also given number of bidders, and non-cooperative
equilibrium bidding.)
sketch of proof of IPV case
Consider any mechanism for allocating the object; let ( ) be the expected surplus
that bidder will obtain in equilibrium; let ( ) be her probability of receiving the
object in the equilibrium.
Then ( ) ( )+( ) ( )
The right-hand side is the surplus that would obtain if she had type but devi-
ated from equilibrium behaviour, and instead followed the strategy that type of is
supposed to follow in the equilibrium of the game induced by the mechanism. That is,
if type exactly mimics what type would do, then makes the same payments and
wins the object as often as would. So gets the same utility that would get ( ( )),
except that in states in which would win the object (which happens with probability
( )) type values the object at ( ) more than does, and so obtains an extra
( ) ( ) more surplus in all.
So, since type mustn't want to mimic type +
we have ( ) ( + )+( ) ( + )
And since + mustn't want to mimic type
we have ( + ) ( )+( ) ( )
( + )¡ ( )
Reorganising, ( + ) ( )
and taking the limit as 0 we obtain = ( )
Integrating up, ( ) = ( ) + = ( )
So any mechanisms which have the same ( ) and the same ( ) for all and for
every , have the same ( ). So any given type of player makes the same expected
payment in each of the mechanisms (since ( ) = ( ) (payment by type of
player ) since the bidder is risk-neutral). So 's expected payment averaged across
her di®erent possible types, , is the same for the mechanisms. Since this is true for
all bidders, the mechanisms yield the same expected revenue for the auctioneer.
In particular any mechanism (including all four of our auctions) which always gives
the object to the highest-value bidder in equilibrium (so ( ) = the probability that all
( 1) other bidders have lower values than ), and give a bidder of the lowest feasible
type no surplus (so ( ) = 0) yield the same expected payment by each bidder and
the same expected revenue for the auctioneer.
 A.21 (i) A,B,C bid 5,7,10 for the packages they want, so A,B win and pay 3,5
(ii) (a) Vickrey-nearest: same winners, but they pay 4,6 (since total must be 10);
(b) Reference rule (Erdil and Klemperer, JEEA, 2010): 3 13 , 6 23 , since objects are
identical and B wins two (the objects are identical, so the prices should be as near as
possible to identical, but with total=10.
[other rules could be used, if justi¯ed by the examinee]
(iii) A,B',B",C bid 5,3,4,10 so A,B',B" win. They pay 3,1,2 in Vickrey;
they pay 4 13 , 2 13 , 3 13 in Vickrey-nearest; and they pay 3 12 , 3, 3 12 in Reference rule
(since we can't charge more than 3 to B', so have to charge 3 12 to A,C)
(iv) it may be natural to (a) compare the Vickrey rule against non-Vickrey core-
selecting-package alternatives in general; then (b) discuss distinctions between the
non-Vickrey alternatives:
(a) The alternatives are intended to reduce the Vickrey auction's problems that
revenue may be very low (as in the example|since a loser was prepared to pay 10),
collusion is easier than with other designs in complements contexts such as this one
(since it allows only two participants to pro¯tably collude, by contrast with most
contexts where collusion only is very e®ective if all market participants engage in it),
and the related unusual incentives to merge, or demerge (as in the example) (or stay
de-merged), and/or shill-bid (see, e.g., Ausubel-Milgrom, in Cramton et al (ed), 2006).
[My lecture gave examples relevant to the setting of this problem.] (The question did
not ask for a general discussion of the merits of a Vickrey auction.)
However, it is very hard to understand how to bid sensibly in any core-selecting
rule even in many simple settings. And if bidders can ¯gure out how to bid, Jacob
Goeree and Yuanchuan Lien (Theoretical Economics 2016, On the Impossibility of
Core-Selecting Auctions) shows that with equilibrium bidding, core-selecting auctions
may well result in lower than Vickrey revenues, and in ine±cient outcomes that are on
average further from the core than Vickrey outcomes! (They argue that no Bayesian
incentive-compatible core-selecting auction exists.)
(b) Reference rules seem fairer and more comprehensible and Erdil-Klemperer
showed they are likely to give lower marginal incentives to bidders to deviate from
truth-telling (and probably su®er less from \odd" incentives to merge, demerge, col-
lude and shill than Vickrey-nearest (as well as than Vickrey)). However, the choice of
reference prices may be contentious if objects are not identical. Erdil-Klemperer give
a simple example in which Vickrey-nearest is ex-ante welfare maximising among all
minimum revenue core-selecting rules, but Larry Ausubel and Oleg Baranov's (mimeo
2013, Core-Selecting Auctions with Incomplete Information) results suggest the op-
posite. More generally, Ausubel-Baranov looks at four di®erent core-selecting auction
formats suggested in the literature, and seem to like the proxy rule of Ausubel-Milgrom
which is a form of Erdil-Klemperer reference rule, for revenues and e±ciency, especially
as correlations among bidders' values increase.
 A.22 [N.B.: I do not expect students to know (or write) anything like as much
detail as I have given here.]

0. This is a \short" question, so I don't think a good answer requires descriptions of

the di®erent auctions. But a student with less to say might tell us what they are: Both
SMRAs and PMAs transact multiple potentially-di®erent objects. For an auction that
is selling goods, a SMRA is a simultaneous ascending auction, and is described in the
answer to my 2012 question (auctions to buy are the opposite). PMAs are described
at www.nu®.ox.ac.uk/users/klemperer/productmix.pdf.

1. Given the way the question is posed it would not be necessary, but it might be
natural, to explain that when bidders (a) want at most a single unit each, (b) have
private values, and (c) have no budget constraints, "straightforward" bidding is a Nash
equilibrium of both the SMRA and the PMA. (Strictly speaking we also need that the
bidding increment is in¯nitesimal in the case of the SMRA.) That is, in the SMRA
each bidder bids in every round as if the current round is the last; in the PMA each
bidder simply expresses his true values.
Moreover, \straightforward" bidding is also optimal in both the SMRA and the
PMA for any bidder who wants multiple objects but thinks his demand is too small to
a®ect market prices, and valuations are strong substitutes (i.e., all units of all goods
are mutual substitutes), for all bidders.
So in these cases the SMRA and the PMA are equivalent (and e±cient). (The
PMA, but not the SMRA, is also e±cient for a wider class of problems.)

2. The distinctions between the auctions °ow from the fact that the SMRA is
dynamic (multi-round) and the PMA is static (single-round):

(a) advantages of the SMRA:

(i) If there are either common-value components to valuations, or complementarities
(including any non-strong-substitutabilities), neither the SMRA nor a simple PMA
allows bidders to accurately express their preferences in general, and neither is e±cient
in general. (\Simple" PMAs in which bidders bid ordinary \dot bids" restrict bidders
to expressing strong-substitute preferences, so force bidders to approximate any other
preferences by strong-substitute preferences.)
With common-value components, the SMRA may work better than the PMA |
because the SMRA's multiple rounds allows bidders to infer common-value information
from others. (However, the PMA could be extended to allow bidders with \common
values" to revise their bids based on reported \interim" auction prices.)
The SMRA's multiple rounds can also allow bidders to learn something about which
objects they are likely to win and then focus their attention on the values of those,
without having to waste resources analysing their valuations for objects that they would
have no chance of winning.
 For the same reason, the SMRA helps \package discovery", that is, helps bidders
work out what complementary bundles they may be able to win, and so might sensibly
bid for.
(ii) Another possible advantage of the SMRA over the PMA is that winning bidders
need not (in general) reveal their actual values. This is the same advantage as the
single-unit ascending auction has over a single-unit second-price auction{bidders may
be unwilling to bid their true values in PMA for fear of revealing information that may
be useful to others, such as regulators, competitors, or suppliers they might have to
bargain with after the auction. (However, a PMA can be run in a way that no-one sees
the winning bids.)
(iii) Bidders with budget constraints will generally ¯nd bidding easier (and so out-
comes maybe more e±cient) in an SMRA than in a standard PMA. (However, there
is a version of the PMA for budget constrained bidders, which may work better than
either an SMRA or a standard PMA in this context.)

(b) advantages of the PMA:

(i) A potentially major advantage of the PMA is that the multi-round aspect of the
SMRA may allow bidders to coordinate their bidding, possibly facilitated by the ability
to signal. Bidders in an SMRA may be able to implement collusion (as illustrated, for
example in the lectures' examples of the Austrian and German auctions), or some
bidders may be able to follow predatory strategies (see, for example, the lectures'
example of an early US auction).
Recent versions of the SMRA are more robust than earlier versions, but SMRAs
remain vulnerable to this kind of behaviour. (In principle, implicit collusion might be
a Nash equilibrium of a PMA, just as it is in an ordinary uniform price auction, but
this seems even less likely than in most uniform-price-auction settings, because of the
extreme di±culty of coordinating behaviour across di®erentiated goods.)
[Less important, for bidders who want more than one unit, it is generally individ-
ually rational (i.e., ignoring coordination possibilities) to engage in demand reduction.
In experiments in auctions of a single variety, they do this much less than equilibrium
analysis would imply (at least if there are more than two bidders and/or more than two
units available) but they also do this much more in a dynamic auction than in a static
auction. So non-coordinated demand reduction may be less in a PMA, which would
make it more e±cient (although non-coordinated demand reduction is not always a
signi¯cant problem in an SMRA either). For various reasons competitive equilibrium
seems a good assumption in the Bank of England's PMA, and likely more generally.]
(ii) The fact that the PMA's outcome is found by solving a static problem based on
stated preferences has several advantages relative to the SMRA's approach. The latter
relies on bidders revealing the necessary information over time in response to monoton-
ically increasing prices and also requires activity rules. (Absent activity rules, bidders
in an SMRA typically have incentive to play "snake in the grass" strategies of waiting
for others to reveal their hands ¯rst; even absent common values complementarities,
and budget constraints, a bidder may gain from waiting for reason (aii) above.)
Importantly, when rationing is permitted, versions of the PMA can, unlike the
SMRA, ¯nd the e±cient solution when participants have any substitutes valuations
(these are represented using \asymmetric dot bids")|the software at pma.nu®.ox.ac.uk
does this. (The SMRA is only e±cient for "strong substitutes" because of its need for
an activity rule|see the results in Milgrom, 2000.) Moreover, the PMA can also handle
some kinds of complements valuations. (The Bank of England's 2014 implementation
includes complementarities on the seller's side.)
Second, the PMA allows the auctioneer to easily express its own preferences about
how the mix, and the total quantity, of objects sold should depend on the bidding, and
the PMA also extends naturally to permitting multiple sellers as well as buyers. (The
SMRA could be modi¯ed to do these things, but I am not aware of this having been
suggested prior to Klemperer (2008). Moreover, the PMA's approach to incorporating
auctioneer preferences is simpler and more intuitive than a modi¯ed SMRA would be.)
Third, a PMA can easily implement objectives di®erent than e±ciency, e.g., auc-
tioneer revenue maximisation, and it is possible to make other modi¯cations such as
varying the pricing rule. (These things don't seem possible in anything that would be
recognisable as a version of an SMRA.)
(iii) The speed of the PMA makes it less costly for participants. It may therefore also
increase participation. (The SMRAs multiple-round bidding may also be impractical
in, e.g., ¯nancial markets with rapidly changing conditions.)
 A.23 [sketch of possible answer]
0. In a discriminatory auction winning bids are ¯lled at the prices actually bid{
exactly as in a ¯rst-price auction.
In a uniform price auction, winning bids are ¯lled at a price where demand equals
supply. In economists' models, this is usually the highest losing bid{exactly as in a
second-price auction. (In reality the price chosen is usually the lowest winning bid, but
with many units, this is usually approximately, or exactly, the same price.)

1. Assuming a ¯xed number of bidders who play Nash equilibrium strategies,

a uniform-price auction bidder's demand schedule starts at the bidder's actual val-
uation for her ¯rst unit (exactly as in a second-price auction, and for the same reason),
but a bidder for any larger quantity than a single unit recognises that buying an ad-
ditional unit drives up the market price, and so drives up the price it pays for all the
other units it buys. It therefore cuts back the price it bids for larger quantities exactly
as an oligopsonist does.
By contrast, a bidder's discriminatory demand schedule starts below the bidder's
actual valuation for her ¯rst unit. As in a ¯rst-price auction, bids for high-valued units
will be shaded down towards the expected market clearing price. However, also as in
a ¯rst-price auction, bids for low-valued units will hardly be shaded at all. Indeed the
pressure towards °at demand schedules is even greater for (multi-unit) discriminatory
auctions than in (single-unit) ¯rst-price auctions. A bidder recognises that winning a
second unit requires beating more opponents' bids than winning just one unit, so the
bidder faces stronger competition for the second unit than for the ¯rst. (Indeed, if the
bidder's "true demand" is fairly °at (i.e., it has a fairly constant marginal valuation of
units), it might even want to bid more for the second unit if the rules of the auction
permitted that. However, we assume the auctioneer takes bids in descending order.)
So:

1a. If bidders want only a single unit each, the analogy is a very good one:
Bidders "tell the truth" in the uniform case, just as in a second-price auction, but
shade below their values in the discriminatory case, just as in a ¯rst price auction.
And uniform price auctions will generally be more e±cient than discriminatory
auctions for the same reason (neither the discriminatory auction nor the ¯rst-price
auction is e±cient for bidders for a single unit if their values are drawn from di®erent
distributions).
Discriminatory demands will generally be °atter than uniform-price demands, and
uniform-auction bids are more dispersed (just as second-price bids will be more dis-
persed than ¯rst priced bids in a single unit auction).
Moreover, in this special case, revenue equivalence holds in the multiunit case as in
the single unit case.
 1b. The analogy remains pretty good for bidders whose demands are "small" rela-
tive to the total demand. (A bidder for a su±ciently small quantity in a uniform-price
auction recognises that it is very unlikely to signi¯cantly a®ect the price [so it can
simply bid its actual value for each unit it demands, as in a second-price auction]; and
a bidder for a su±ciently small quantity in a discriminatory expects to face a similar
strength of competition across the range of quantities it might win.)

1c. When bidders are not "small", analogies with the single unit auction cases
become less helpful, as the discussion above makes clear.
Either kind of auction can then be more e±cient than the other, because di®erent
kinds of bidders shade their bids downwards by di®erent amounts, and these amounts
are also di®erent for di®erent units.10
Moreover, at least in theory, a uniform-price auction can sustain implicitly collusive
strategies as non-cooperative equilibria (see below). The discriminatory auction is in
principle unproblematic, in this respect.

2. If we worry about participation, and coordination/collusion possibilities,

that is, we drop the assumption of a ¯xed number of bidders who play Nash equi-
librium strategies, the analogies become much less good:

2a. Participation: A uniform-price auction is attractive to small bidders, because

of the way in which large bidders reduce their demand11 . Furthermore a small bidder
is protected by the fact that everyone pays the same price. By contrast, in a discrimi-
natory auction, bidders are playing "Guess the price" { that is, information about the
likely market clearing price is very valuable, and collecting and analysing information is
disproportionately costly for small bidders. So the uniform-price auction will increase
participation (especially with common values). (This "Guess the price" problem can
be resolved in a discriminatory auction by allowing non-competitive bidding, but that
is rather like introducing a uniform price component into the discriminatory auction.)
By contrast, a ¯rst-price auction for a single unit creates more uncertainty about
who might win, and so is more likely to attract entrants, than a second-price auction
(especially in some common-value settings).

2b. Coordination/collusion possibilities: Since, at least in theory, a uniform-price

auction can sustain implicitly collusive strategies as non-cooperative equilibria, it may
not be too hard for bidders to sustain co-operation in more frequently-repeated settings,
with small numbers of bidders, di±cult entry, and very stable demand (e.g. electricity).
(This is probably not a problem if supply can be varied by the auctioneer, or even just if
supply is elastic.) The discriminatory auction is less problematic in this respect (though
if it does discourage entry it may facilitate collusion simply by reducing the number of
participants). So here there is a closer analogy with the single-unit auction formats,
where it is usually argued that a second-price auction makes collusion somewhat easier.
 A24. (a) Marginal Revenues (MRs) are uniform on [0,40] and [0,80], so run a 2nd
MR auction (as in Bulow-Roberts, JPE, 1989), but never take an MR less than your
value (10). (The question did not require it, but you could comment that the auction
corresponds to running an auction with a minimum price of $45, and promising the
man a $20 discount if he is the winner.)
(b) MR goes from 80 (at 0) to 0 (at 12 ), then linearly (as if from 60 (at 0)) to) 20
(at 12 ) to -20 (at 1)
[Reason: Demand curve goes from 80 to 40 (at 12 ) then from 40 to 20 (at 1).
So ¯rst half of MR curve is straightforward. MR goes from 80 (at 0) to 0 (at 12 ).
For second half, note that value of MR is determined locally, so is same for this
linear segment as if the curve were linear starting at 0, in which case the intercept
would be 60 and demand would go from 60 to 20 over the range 0 to 1.
So the MR goes linearly (as if from 60 (at 0) to 20 (at 12 ) and) from 20 (at 12 ) to
-20 (at 1).
(Alternatively, demand = 1-F(v) = ¯rst (80-v)/80 and then = (1/2) + (40-v)/40,
so MR (= v-(1-F)/f) = ¯rst 2v-80, then 2v-60) ]
So the MR curve crosses the cost curve (10) more than once, with the better solution
clearly being the larger|so there is an optimal take it or leave it o®er of 35 (i.e, MR=10
on 2nd part of MR curve).
[The larger crossing yields pro¯t=5/8 x (35-10) = 250/16 . The lower crossing gets
7/16 x (45-10) = 245/16.]
(c) The MR curve jumps up (at 12 ) so you will pool bidders in a range. The question
only asks for the qualitative features, so it would be su±cient to explain that if the
optimal auction will involve rationing in this range if the optimal reserve price is lower,
so that in that case the auction will be ascending, then involve rationing, then be
ascending again. (If the optimal reserve price is higher than the pooling range, then it
will just be an ascending auction for the women, starting at their reserve price.)
[Quantitatively, to ¯nd the range assume x is the fraction of women above the
\ironing" and y is the fraction of men in the ironing, and solve 80x=20-40y (=critical
MR at which ironing starts and ends] and (80x)(x/2)=(40y)(y/2) assuming as is true
here the ironing takes place at an MR 10. The critical MR is approx.11.7, or price of
approx. 35.85 for men and 45.85 for women. Run an ascending auction starting at 35
(i.e., MR=10 on the man's MR curve) and rising to 35.85, then jump the price to 40.85
(the point where any woman with value 45.85 is just indi®erent about not jumping
and winning with probability 12 conditional on the other bidder also being below 45.85,
and jumping and winning with probability 1 conditional on the other bidder also being
below 45.85). If both bid 40.85, then price keeps going up, but no one will drop out
before 45.85. (Conditional on two above 35.85, this raises more money from those
above 45.85, when there is exactly one of them, than the money lost when there are
none (so both between 35.85 and 45.85).)]
 A25. (a) bidder with value 12 would pay the least to enter. if this bidder does enter,
she wins with probability 12 and earns surplus 12 in that event, so gets expected surplus
= 14 . So the seller can charge 14 .
(b) entry fees are 14 on average; 3/4 time we get 0 or 1 entrant and price= 0;
1/4 time we get 2 entrants, and expected price = 2/3 [you learnt in the lectures
that the expected kth highest of n signals drawn from uniform distribution on [a,b] is
a + (n+1-k)(b-a)/(n+1) ; alternatively it is not too hard to work this out directly ]
so expected Revenue= 1/4 + 1/4 x 2/3 = 5/12
(c) a very good student will observe that this mechanism is revenue equivalent to
the previous one (bidder's winning probabilities are the same, and the surplus of type
0 (and all types up to 12 ) are the same in both mechanisms
students who are uncertain about this can compute that 1/4 time we get 2 entrants,
and expected price = 2/3; 1/2 time we get 1 entrant and price= 12 ;1/4 time we get 0
entrants, and price = 0, so expected Revenue= 1/4 x 2/3 + 12 x 12 = 5/12
(d) the revenue equivalence theorem tells us that bidders, as well as the auctioneer,
are indi®erent between the two mechanisms
I want students to say this, and not just say that with the entry fee the bidder pays
1
4
, but gains 12 if the other bidder doesn't show (with probability 12 ), so is indi®erent
(e) (iii) (ii) (i) for seller's expected revenue
[By revenue equivalence the three options are equally pro¯table if the bidders are
risk neutral
Risk averse bidders bid the same way as risk neutral ones IF they enter an ascending
auction. So the pro¯tability of (ii) is una®ected by risk-aversion.
However, fewer bidders enter (i) than if the bidders were risk-neutral (a bidder
with value v=1/2 would no longer enter, so the cut-o® for entry is higher), so the
pro¯tability of (i) is reduced by risk-aversion.
Risk-aversion has no e®ect on which bidders enter (iii), and risk averse bidders
bid more than risk neutral ones (to gain a larger chance of a smaller prize), so the
pro¯tability of (iii) is increased by risk-aversion.]
 A26. This question was open book format (unlike any question previous to this
year). Even so, it was a hard question. We expected only outstanding students would
achieve a perfect answer, and marked accordingly.

(1) We expect to see truthful bidding in the Vickrey auction. The resulting allocation
is the e±cient allocation (1 2). Bidder A's Vickrey payment equals 9 8 5 = 0 5.
Bidder B's Vickrey payment is 10 6 = 4.

(2) The following table states their truthful demands.

Truthful clock demands
Clock price Demand bidder A Demand bidder B
0 3 3
1 3 2
2 2 2
3 1 2
4 1 1

The clock ends at a price of 3 with market clearing.

(3) Bidders do not submit additional sealed bids. Hence, bidder A bids 3 for 3 units,
4 for 2 units, and 3 for 1 unit. Bidder B bids 0 for 3 units and 6 for 2 units. Note
that the bidding functions are not increasing in the quantity. The ¯nal allocation
is the e±cient allocation (1 2). Bidder A's payment is 6 6 = 0. Bidder B's
payment is 4 3 = 1.
Comment: It could be that students analyze the sealed-bid phase after the clock
ending at = 4. Without any additional sealed bids, bidder A then bids 3 for 3
units, 4 for 2 units, and 4 for 1 unit (instead of 3). Bidder B bids 0 for 3 units,
6 for 2 units, and 4 for 1 unit (instead of 0). These bidding functions still lead
to the e±cient allocation (1 2). Bidder A's payment is 6 6 = 0 (unchanged).
Bidder B's payment is 4 4 = 0 (lower than above).

(4) Bidder A's strategy leads to the clock ending at the clock price = 4 with the
demands (2 1). Hence, bidder A's bids in the sealed-bid phase are 9 for 3 units,
8 for 2 units, and 0 for 1 unit. Bidder B's bids in the sealed-bid phase are 0 for
3 units, 6 for 2 units, and 4 for 1 unit. The ¯nal allocation is (2 1), which is
ine±cient. Bidder A's payment is 6 4 = 2. Bidder B's payment is 9 8 = 1.

(5) The Vickrey auction leads to the e±cient allocation under truthful bidding.
Truthful bidding in the clock phase also leads to the e±cient allocation.
The payments in (1) and (3) are quite di®erent, however. The payments are
much lower in the CCA, as are the bids implied by the clock phase bids.
 In (4) we see that demand expansion can be an optimal reaction to the lack
of additional sealed bids. Bidder A's demand expansion leads to a surplus of
8 5 2 = 6 5, which is higher than the surplus of 6 from truthful bidding.
Thus, the (expected) behaviour in the sealed-bid phase has implications for bid-
ding in the clock phase. One has to study both phases together in equilibrium
to see how deviations from truthful bidding impact payments and revenue.
Levin and Skrzypacz (2016, AER) call not submitting additional bids \quiet"
bidding. Levin and Skrzypacz show that the less bidders raise their sealed bids
in a symmetric equilibrium (i.e., the more quiet they bid), the lower revenue.
The revenue ranking is ambiguous in asymmetric equilibria.

(6) When the clock phase ends with market clearing, the typical activity rules imply
that the permitted sealed-bid phase bids are such that the ¯nal clock allocation
is the ¯nal allocation. (This is the case with \¯nal cap" activity rules: ¯nal bids
must satisfy revealed preference with respect to the last clock bids|see Levin
and Skrzypacz, 2016.) Hence, the sealed bids do not determine the allocation.
Moreover, with Vickrey pricing, they do not determine one's own payment|they
only a®ect the payments of the other bidders. If bidders are indi®erent about the
payments of the other bidders, and if the clock phase ends with market clearing,
it is a reasonable assumption that no additional bids will be submitted. However,
if bidders are not indi®erent, they will submit additional bids to alter the other
bidders' payments. Moreover, the clock phase may not end with market clearing.
(See Levin and Skrzypacz, 2016, for more discussion.)

(7) In the SMRA we typically expect demand reduction and not demand expansion
as in (4). Both types of behaviour can lead to ine±ciencies.
 A27. This question was open book format (unlike any question previous to this
year). Even so, and even though the general approach was explicitly discussed in
the lectures and in the revision class and was spelt out in the notes, the question
was too hard. Some candidates made almost no progress, and none gave a perfect
solution. However, we compensated by marking the question very generously, and
discrimination between candidates was possible, because there was a wide range
in the quality of attempts.

Conjecture a symmetric equilibrium bidding function in which a bidder with signal

bids ( ).
Consider the bidder's pro¯t from bidding as if her signal was :
1
(i) ( ) = =0 ( ( ) ( ) 2) ( ) + = ( ( ) 2) ( )
( ) =( ( ) ( ) 2) ( ) ( 0 ( ) 2) ( ) + ( ( ) 2) ( )
( ) = 0 at = implies 0 ( ) = 2 ( ) ( ) ( )
with solution ( ) = =0 [2 ( ) ( ) ( )]
assuming (0) = 0 (in fact ( ) can be anything) and SOC apply (they do).
(ii) easiest to write for the maximum of the other signals, in which case the
derivation above applies, with the exception that payments are ( ) not ( ) 2
So ( ) = ( ¡1) ( ) = ( 1) ( ¡2) , and ( ) = [ + +( 2)( 2)] since
conditional on being the maximum of the other ( 1) signals, the average value of
the remaining ( 2) signals is ( 2)
Solution ( ) = =0 [ ( ) ( ) ( )] = =0 [(( + 2) 2)(( 1) )] =
( + 2)( 1) 2
(iii) A bidder's bid only a®ects her payo® when she wins. So risk-averse bidders
will bid more to win smaller amounts with higher probabilities.
 A28 a. (i) = 1; (ii) = 0; (iii) =1; (iv) = 0
[note in (iii) there are only 2 bidders, so cannot condition on when a bidder quits]
b. Bidder wins with probability .
Conditional on winning at price ( ), the opponent's value is uniform on (0, )
so the opponent quits uniformly on the price range (0, ).
So with probability (1 ) the winner drops out at without ¯nding out its
opponent has already quit.
And with probability the winner wins when it ¯nds out its opponent has already
quit, and this price is uniform on (0 ). (Or we can say that with probability
the winner wins when it ¯nds out its opponent has already quit, and in this case the
opponent's value is uniform on (0 ), and so the opponent drops out uniformly on
(0 ), and the winner ¯nds out at (1 ) of the drop-out price.)
So conditional on winning,
expected payment = (1 ) + ( 2) = (2 )( 2)
So expected payment made by a bidder with value = (2 )( 2)
c. (i) The conditions for the Revenue Equivalence Theorem apply, so the auction
is revenue equivalent to a standard ascending auction ( = 1), so expected payment is
the same as for a standard auction.
For an ascending auction the expected payment of type is easily seen to be ( 2)
(since bidder wins with probability , at average price of 2).
(ii) So (2 )( 2) = ( 2) , so = 1 (2 ) (and bid = (2 )).
d. If 1, raising bids gives a larger chance of a smaller payo® which risk-averse
bidders prefer, so risk aversion increases bids and revenue.
If = 1, risk aversion has no e®ect.