Games in normal form are static games.

Games in extensive form are dynamic games.

So far, we have assumed that the players know each other’s preferences. We now introduce another component, namely, asymmetric
information in this topic.

ECON0027 Lecture notes

Section 4: Games with incomplete information
Reading: Osborne: Ch 9-10
In many strategic situations, players may be unsure (not informed) about their opponents’
characteristics
Examples:

• bidding in an auction – players don’t know how much opponents value the object

• buying a used car – the seller may know whether the car is a lemon or a peach, but the
buyer may not

Harsanyi provided a framework for modelling such situations—Bayesian games

4.1 Bayesian game: an example

Battle of the sexes with incomplete information
Player 1 would like to match player 2’s action
Player 1 is unsure about player 2’s preferences. There are two states of the world, !1 and !2

!1 : player 2 may like to match player 1’s actions

!2 : player 2 may like to mismatch player 1’s actions

Player 1 does not know the preferences of player 2 (i.e., he does not know the state of
the world). Player 2 knows his own preferences (i.e., he knows the state of the world). Both
statements are common knowly known.

4.1.1 Payo↵s
B S B S Note that player 1’s payoffs in the two
tables are the same because his payoff
B 2,1 0,0 B 2,0 0,2
does not depend on the state w.
S 0,0 1,2 S 0,1 1,0
state !1 , Pr(!1 ) = ⇡ state !2 , Pr(!2 ) = 1 ⇡

• Nature chooses state ! 2 {!1 , !2 }, with prob (⇡, 1 ⇡)

• Player 2 observes the realized state !; player 1 does not.

• The two player’s simultaneously choose actions

• Payo↵s given by the actions chosen and state, as in tables.

Structure of this game is commonly known by both players, e.g., player 2 knows that 1 knows
that Pr(!1 ) = ⇡
Asymmetric information arises because player 1 is not informed about the preferences of player 2, but player 2 is
informed about his own preferences.

1
One way to think about the game above is to view it as a game of three players, namely, player 1, player 2 with state w_1, and player 2
with state w_2. We can denote the three players as N = {1, w_1, w_2}. We introduce a fictitious character, nature, who is going to pair up
player 1 with either w_1 or w_2 with probability π and (1 - π) respectively. When the nature matches the two players together, w_1 or w_2
will know that he is playing with player 1, but player 1 will not know who he is playing with.

We can express the game as follows: player 1 chooses the matrix (action B corresponds to the matrix on the top and action S
corresponds to the matrix on the bottom), player w_1 chooses the row, and player w_2 chooses the column.

The utilities in each cell is calculated as follows:

• Consider the strategy profile (B, B, B). Player 1 will receive a payoff of 2π + 2 * (1 - π) = 2; player w_1 will receive 1; and player w_2 will
receive 0. The corresponding payoffs on the original game are highlighted in yellow.
• Consider the strategy profile (B, B, S). Player 1 will receive a payoff of 2π + 0 * (1 - π) = 2π; player w_1 will receive 1, and player w_2 will
receive 2. The corresponding payoffs on the original game are underlined in blue.
• The payoffs for the other cells can be found using the same process.

The step above shows that we can write the Bayesian game as a game of three players without asymmetric information. We can then
solve this game using the approaches learned before. These NEs are the Bayesian NEs for the original Bayesian game.

This process can be time-consuming and difficult to use when the information structure is more complex. We can solve for a Bayesian
game without going through the process above.
4.2 (Bayesian) Nash equilibrium
The strategy in this game is a function that maps player’s knowledge into actions. This new
definition of strategy allows us to use Nash equilibrium:
Pure strategy equilibrium: action for player 1, a1
A pair of actions for player 2, a2 (!1 ), a2 (!2 ).
Player 1’s action is optimal, i.e. 8a01 :

⇡u1 (a1 , a2 (!1 )) + (1 ⇡)u1 (a1 , a2 (!2 )) ⇡u1 (a01 , a2 (!1 )) + (1 ⇡)u1 (a01 , a2 (!2 ))

(in this particular example, player 1’s payo↵ does not depend upon state)
Player 2’s action is optimal at !1 i.e. 8a02 :

u2 (a1 , a2 (!1 ), !1 ) u2 (a1 , a02 , !1 )

Player 2’s action is optimal at !2 i.e. 8a02 :

u2 (a1 , a2 (!2 ), !2 ) u2 (a1 , a02 , !2 )

4.3 Solution
B S B S
B 2,1 0,0 B 2,0 0,2
S 0,0 1,2 S 0,1 1,0
state !1 , Pr(!1 ) = ⇡ state !2 , Pr(!2 ) = 1 ⇡

Strategy for player 1: a1 2 A1

Strategy for 2: (a2 (!1 ), a2 (!2 )).
We can look at pure strategy profiles and check which one are equilibria.
Is [B; (B, S)] an equilibrium? If 1 plays B, optimal for 2 to play B at !1 and S at !2 . For 1,

E[u1 (B; (B, S))] = ⇡ ⇥ 2 + (1 ⇡) ⇥ 0 = 2⇡.

E[u1 (S; (B, S))] = ⇡ ⇥ 0 + (1 ⇡) ⇥ 1 = 1 ⇡.

Optimal to play B as long as
1
2⇡ 1 ⇡ () ⇡ .
3
If ⇡ 13 , [B; (B, S)] a Nash equilibrium.
If ⇡ < 13 , [B; (B, S)] is not a Nash equilibrium.
Is [S; (S, B)] an equilibrium? Similar caclulation: if ⇡ 23 , [S; (S, B)] is an equilibrium, not
if ⇡ < 23 .
What happens if ⇡ < 13 ? Neither is an equilibrium, so equilibrium must be in mixed strategies.

2
We only need to check for the strategy profiles in which player 2 does not have a profitable deviation. In this game, we see that the
game in state w_1 is a game of coordination and the game in state w_2 is a game of hide and seek. This means that we should look for
strategy profiles where player 2 plays the same strategy as player 1 in state w_1 (i.e. choose B when player 1 chooses B) and plays the
different strategy as player 1 in state w_2 (i.e. choose S when player 1 chooses B). Thus, we only need to check [B, (B, S)] and [S, (S, B)]
because for any other strategy profile, player 2 will have an incentive to deviate.

The implication of the threshold for π is that the probability for state w_1 to occur must be greater than a certain level. In both strategy
profiles, player 2 achieves coordination with player 1 in state w_1 and player 2 manages to avoid player 1 in state w_2. So the intuition is
that player 1 gains from the coordination in state w_1 and loses from the game in state w_2. As a result, the threshold that is required for
the strategy profile to be a NE depends on size of the gains and losses.

Mixed Strategies

Consider the following strategy:

• Player 1 plays B with probability α and S with probability (1 - α).
• Type w_1 of player 2 plays B with probability β and S with probability (1 - β).
• Type w_2 of player 2 plays B with probability γ and S with probability (1 - γ).

Since we have considered all possible strategy profiles in which player 1 plays a pure strategy, we only need to consider the cases where
α is strictly between zero and one.

Consider the case where gamma is strictly between zero and one. This means that the type w_2 of player 2 is indifferent between playing
B and S because this is the only reason for him to mix. Thus, the utility that type w_2 of player 2 gets from playing B must be the same as
the utility that he gets from playing S conditional on player 1 playing the mixed strategy (alpha, 1 - alpha).

The utility that type w_2 of player 2 gets from playing B or S in state w_2 when everyone else plays according to the strategy profile is:

The utilities that type w_1 player 2 gets from playing B and S in state w_1 when everyone else plays according to the strategy profile are:

Since 4/3 is strictly greater than 1/3, player 2 in state w_1 strictly prefers playing S over playing B. This means that β is equal to zero.

As mentioned above, we only need to consider the case in which α is between zero and one. This means that player 1 is indifferent
between playing B and S. It follows that player 1 gets the same utility from playing S and B conditioning on player 2’s actions.

The utility’s that player 1 gets from playing B or S when everyone else plays according to the strategy profile are:

To make sure that γ is a probability, we need to check that for the given value of π, the expression for γ is smaller than or equal to 1:

The process above shows that one mixed Bayesian NE for this game is

We can potentially find other mixed Bayesian NEs by repeating the process above and assuming that β is strictly between zero and one.
For the NE to be mixed, it must be the case that either both β and γ are strictly between zero and one or one of them is between zero
and one. Otherwise the result will be one of the pure strategy NE we have already considered earlier.
4.3.1 Mixed strategy Nash equilibrium
Consider the following strategy profile:

• Player 1 plays B with prob. ↵ and S with a remaining probability.

• Type !1 plays S.

• Type !2 plays B with probability and S with a remaining probability.

Since type !2 is mixing he must be indi↵erent between B and S:

1
1 ↵ = 2↵ () ↵ =
3
Type !1 strictly prefers S over B:
4 1
2(1 ↵) = > =↵
3 3
Finally, player 1 must be indi↵erent between B and S:

0 + 2(1 ⇡) = ⇡ + (1 ⇡)(1 )

Solving this with respect to we get

1
=
3(1 ⇡)

Note that this equilibrium exists for all ⇡ 2 [0, 2/3].

4.4 More complex information structure

In order to model more complex information structure we introduce signals that partially reveal
the state of the world to the players. Consider the following example:
a b a b a b
A 1,1 -1,-1 A 1,-1 -1,1 A 2,2 0,0
B -1,-1 1,1 B -1,1 1,-1 B 0,0 0,0
Pr(!1 ) = 1/2 Pr(!2 ) = 1/3 Pr(!3 ) = 1/6

Assume that the information structure is

!1 !2 !3
Player 1 t11 t11 t12
Player 2 t21 t22 t22

Suppose that player 1 will know if w_3 has been realized, but he cannot distinguish between w_1 and w_2; and suppose
that player 2 will know if w_1 has been realized, but cannot distinguish between w_2 and w_3. Though they do not have
perfect information separately, the two players can distinguish between all three states together. We use the character
nature to model this information structure. First, nature is going to pick a scenario according to the ex ante probabilities
(i.e. 1/2, 1/3, and 1/6). Then, nature is going to send information to the players as shown by the table above (e.g. player 1
3
will receive the signal t11 if w_1 or w_2 is realized and the signal t12 if w_3 is realized). The superscript indicates the player
and the subscript represents the name of the message (i.e. message 1 or message 2).
 Note that unlike the previous game, the players in this game cannot use the ex ante probabilities to compute their expected
utilities. This is because they update their beliefs about the state they are in when they receive additional signals (e.g. when
player 1 receives t12, he knows he is in w_3). The distributions Pr(w_1), Pr(w_2), and Pr(w_3) are the ex ante beliefs about the
state of the world before actually playing the game. The players use the Bayesian rule to update their beliefs after receiving
new information. These ex post or posterior beliefs are the ones featured in the expected payoff expressions.

Let us look at player 1’s signals. If the state of the world is !3 , the player learns it perfectly.
However, if the state of the world is either !1 or !2 , there is residual uncertainty after the
player receives his signal (because it is the same signal in both states). To quantify the residual
uncertainty, we will use Bayesian updating. With posterior beliefs, we can view this game as a game with four players (i.e.
Posterior beliefs for player 1: two types of player 1 and two types of player 2). However, it is dif cult to write
Pr{!1 | t1 } = Pr{!1 }+Pr{!2 } = 6/10 = 3/5, down a game with four players so we are not going to do this explicitly.
1 Pr{!1 }

Pr{!2 | t11 } = 2/5, The strategy for player 1 is going to be two actions:
Pr{!3 | t12 } = 1 One action that player 1 takes after receiving t11.
Posterior beliefs for player 2: One action that player 1 takes after receiving t12.
The same applies to player 2.
Pr{!1 | t21 } = 1,
Pr{!2 | t22 } = 2/3, In Bayesian games, the de nition of a strategy is a mapping from the space of
Pr{!3 | t22 } = 1/3 messages into the space of actions. There is a function that tells what action the
player should choose for every message that the player receive.

4.4.1 Equilibrium
To find one equilibrium in this example, consider the following strategy profile:
Player 1 plays A with probability ↵i after receiving a signal t1i .
Player 2 plays a with probability i after receiving a signal t2i .

4.4.2 Best replies and equilibrium

1. E[u1 (A, s2 , !) | t11 ] =
Pr{!1 | t11 }( 1 (1 1 )) + Pr{!2 | t11 }( 2 (1 2 )) =
3
5
( 1 (1 1 )) + 25 ( 2 (1 2 ))

2. E[u1 (B, s2 , !) | t11 ] = 35 ( 1 + (1 1 )) + 25 ( 2 + (1 2 ))

3. E[u1 (A, s2 , !) | t12 ] = 2 2

4. E[u1 (B, s2 , !) | t12 ] = 0

5. E[u2 (a, s1 , !) | t21 ] = 2↵1 1
6. E[u2 (b, s1 , !) | t21 ] = 1 2↵1
2
7. E[u2 (a, s1 , !) | t22 ] = 3

8. E[u2 (b, s1 , !) | t22 ] = 0

It is easy to check that the following strategy profile does not admit profitable deviations
and therefore is an equilibrium:
• 1 = 1/6
• 2 =1
• ↵1 = 1/2
• ↵2 = 1

4
4.5 Bayesian game: general case
• Set of players N = {1, 2, ..., n}. We are using the term action set instead of strategy set because strategies are instructions
that tell the agents what to do after each possible signal that the nature is going to send to
• Each player has action set Ai the player, thus are more complex to de ne. Here we de ne the action set and build
strategies using these actions.
• Nature chooses state ! 2 ⌦ = {!1 , !2 , ., !k } Ω is the set of the states of the world.
P
• State ! has probability p(!), p(!) = 1
⌦

• Each player i receives a signal regarding the state, ⌧i (!)

τ_i maps the states of the world
• Signal function:⌧i : ⌦ ! Ti , where Ti is the set of signals for player i into signals T_i.

• If at states ! and ! 0 , ⌧i (!) 6= ⌧i (! 0 ), then player i can distinguish these states

• If ⌧i (!) = ⌧i (! 0 ), then i cannot distinguish ! 0 from !

• At signal ti , player i has a belief over the set of states

The parameter a indicates the particular action a player is
going to choose. The vector a is a set of n actions, one for p(!) Players use the Bayesian rules to update
each of the n players. The payoffs depend on each player’s
p(! | ti ) = their beliefs about the state of the world.
Pr(ti )
action and the state of the world. The utility functions maps
these information into a real number.
• Payo↵ of player i is ui (a, !), where a = (a1 , a2 , .., an )

• Pure strategy for a player is a function si : Ti ! Ai A player can have many private personal scenarios in a Bayesian game.
A pure strategy for player i, denoted by s_i, needs to tell the player what
to do after receiving every possible message from the nature. A strategy
4.6 Back to our example is a function that maps information, T_i, into the actions available, A_i.

B S B S
B 2,1 0,0 B 2,0 0,2
S 0,0 1,2 S 0,1 1,0
state !1 , Pr(!1 ) = ⇡ state !2 , Pr(!2 ) = 1 ⇡

States are {!1 , !2 }

player 1 receives signal from set {t1 } (not informative)
2 receives signal from set {t2 , t02 }:
at !1 she gets t2 , at !2 , she gets t02 .
We refer to the messages that the players receive as types of a player (i.e. in
the previous game, we can refer to player 1 as type one of player 1 when he
4.7 Types vs states of the world receives signal t11 and type two of player 1 when he receives signal t21).

In special cases of games in which one agent’s signal contains no information about the other
agent’s signal, we will refer to these signals as agents’ types. Moreover, instead of specifying the
distribution over the states of the world, we will specify distributions over types and state that
these types are independant. From information we can restore the distribution over the states
of the world (although, we would not need to do that to solve the game). One example of such
a game is an auction with independant private valuations discussed in the next section.

5
Games with asymmetric information differs from games with symmetric information in one important aspect, that is, actions taken by
agents can have information contents. In other words, an agent may be able to learn something from the actions that are taken by his
opponents.

Example: Jury

Information Structure
Consider a jury that is trying to decide whether to convict or to acquit a person. This person in question can be one of the two types,
namely, guilty or innocent, with the following probabilities:

The person’s real type is not known by the juror. There are two possible signals that the jurors can receive. The signal ‘g’ is an imperfect
signal in favor of guilt and the signal ‘i’ is an imperfect signal in favor of innocence. Note that both signals are noisy signals. Each juror will
receive one private signal. The probability that a juror receives the signal ‘g’ conditional on the person being guilty is p and the
probability that a juror receives the signal ‘i’ conditional on the person being innocent is q, where both p and q are strictly between 1/2
and 1 (the probabilities has to be smaller than 1 because they are not perfectly revealing). Mathematically, the conditional probabilities
for receiving the signals are

Furthermore, we assume that the signals are received by jurors independently. In other words, the noise components in the signals are
independent across all jurors.

Players and Actions

The players in the game are all the jurors. Note that the person in question is not a player in the game. Each juror has two available
actions, either to acquire or to convict the person in question.

Payoffs
The payoff received by a juror does not depend directly on his action. Instead, each juror’s payoff is going to be a function of the
outcome as listed below:
• The payoff is 0 if guilty is convicted or innocent is acquitted (i.e. no loss if the outcome of the jury matches the person’s true type)
• The payoff is -z if innocent is convicted
• The payoff is -(1 - z) if guilty is acquitted
The numbers z and (1 - z) can be viewed as the costs of making different types of mistakes. We assume that z is between zero and one
such that the payoff received by the jurors is negative when they make mistakes.

Situation 1: Single Juror

Suppose that there is only one juror deciding whether to convict or to acquit the person.

Consider the following decision rule or Bayesian equilibrium strategy:

• The juror votes to convict when he receives the signal ‘g’ and he votes to acquit if he receives the signal ‘i’.

Suppose that the signal the juror receive is X. The payoff the juror is going to get if he convict or acquit is the following:

The juror should convict if and only if

In other words, the juror should convict if the probability of the person being guilty given all available information is greater than the
penalty of convicting an innocent person. Note that this is a general rule, so it applies in the case with more than one juror. In this case
with only one juror, the additional information available is the signal the juror receives.

The probability of G conditional on receiving ‘g’ is

The juror should acquit if and only if

The probability of G conditional on receiving ‘i’ is

The general decision rule in the case with a single juror is that the juror should vote according to his signal whenever the following
inequality hold:

The interpretation is that the design of the information structure allows the juror to create some beliefs about the type of the person in
question. Those beliefs depend on the prior beliefs and the precision of the signals (i.e. the value of p and q). For example, as the values
of p and q go to one, the signals will become perfectly revealing and then the threshold will become 0 ≤ z ≤ 1. Since z is always between
zero and one, the juror should always vote according to his signal in this case. This makes sense intuitively because the the signals are not
noisy and in fact tells the person’s true type. As the signals become less precise, the interval for z will shrink, meaning that the equilibrium
where the juror vote according to the signal he receives will becomes more and more difficult to sustain. When the signals are not
precise, the juror’s decision will depend on the size of the penalties. If the penalty for convicting an innocent is very high, then the juror
may not want to convict because his posterior beliefs of the person being guilty is not strong enough. On the other hand, if the penalty
for acquitting a guilty person is very high, then he might choose to convict instead of acquit.

Situation 2: Multiple Jurors

Suppose that there are n jurors where n is larger than one. Each of the n jurors needs to vote in favor of convicting or acquitting. The rule
is that a unanimous decision is needed in order to convict the person. This means that in order to convict the person, all the n jurors need
to vote for conviction. A key implication of this example is that a player can learn something from the equilibrium strategy profile even if
he does not observe the actions of his opponents directly. In other words, the equilibrium actions that are supposed to be taken by the
players have informational content.

Since unanimity is required to convict someone, a juror’s action will be inconsequential if at least one other player decides to vote in
favor of acquittal. Thus, to any one of the jurors, his decision matters if and only if he is the pivotal voter and his vote changes his payoff.
In other words, a juror’s action matters only if everyone else vote for conviction.

Consider the following strategy profile:

• Each juror votes according to the private signal he receives.

It follows from the idea of unanimity that a juror’s action will be dictated by a scenario where he is the pivotal voter. Under the strategy
profile above, this scenario happens when the juror himself receives some signal s_1 and everyone else receives the signal ‘g’.
Specifically, n - 1 jurors will receive the signal ‘g’ in this scenario and vote for conviction. The final outcome depends now on the last juror
with the signal s_1. This juror should vote for convict if the following general decision rule hold:
• The juror should convict if Pr {G | X} ≥ z.
• The juror should acquit if Pr {G | X} ≤ z.

** One of the inequalities above should be strict, otherwise we will wander into the realm of mixed strategies. However, since we are
focusing only on pure strategies, we can tie break in any direction in the event of a tie.

Following the equilibrium strategy, the juror should acquit when he receives the signal ‘i’, that is, when s_1 = i. The probability of the
person being guilty conditional on all information available is

Note that the information available here include the signal the juror receives (i.e. the signal ‘i’) and the knowledge that all other players
received the signal ‘g’. The second part is indirectly obtained from the fact that the juror’s decision matters only when he is pivotal.

We focus only on the lower bound of the interval for z because this is what matters in this example. We can write the interval for z as
follows:
Note when when n = 1, the lower bound of the interval is the same as the result we obtained in the case with a single juror:

The results shows that when the number of jurors increases, the lower bound of the interval will approach one from below:

This means that sustaining the equilibrium where everyone votes with the signal he receives becomes more difficult as the number of
jurors becomes larger. At some point, the equilibrium is sustainable only when the punishment for convicting an innocent is very large
and the punishment for acquitting the guilty is very small.

In this example, we are trying to construct an equilibrium in which every juror votes with his own signal. The result shows that sustaining
the equilibrium where everyone votes with the signal he receives becomes more difficult as the number of jurors becomes larger. At
some point, the equilibrium is sustainable only when the payoffs are very specifically defined. Specifically, when the penalty for convicting
the innocent is very large and the penalty for acquitting the guilty is very small. The intuition is that when the juror’s own signal is ‘i’ and
when he knows that all other players received ‘g’, he might not want to vote for acquittal because his own signal has only limited
precision and at the same time he is overwhelmed by vast amount of signals ‘g’ received by the other players. The only way for the juror
to continue to stick with his own signal is when the punishment for convicting a innocent is very large.

From the game theoretic perspective, the example is interesting because it shows that a agent can learn about the private information of
everybody else’s without observing their actions directly and before he makes his own action. He obtains information using just the
equilibrium strategy profile and the fact that he is pivotal only in certain cases.
Example: Auction

Common Framework / Environment for Auction Set Up

Suppose that there is one indivisible object for sale and the object is owned by an auctioneer (the auctioneer is a passive character).
There are n customers who want to buy the object. Each bidder i has a value for the object, denoted by v_i. The v_i for all bidders come
from a certain set. In this example, we assume that the valuations are drawn from a interval between zero and one.

Each bidder i wishes to maximize his payoff, which is equal to v_i - b_i, where b_i is the price the bidder pays to the auctioneer. The value
of the object to the bidder is zero if he does not get the object.

The description above gives the underlying environment of an auction. Within this general framework, we can define different games
using different rules for auction.

Asymmetric Information
Assume that v_i will be drawn randomly from some distribution F, where F is a cumulative distribution function. Furthermore, assume that
v_i is i.i.d, meaning that the realization for the valuation for player i is statistically independent from the realizations of the valuations of all
the other players, and these valuations are identically distributed across all players. An implication is that since the valuations are all
drawn from the same distribution, there is a certain amount of symmetry in the game. All bidders are the same ex ante, but may differ ex
post once they draw their own valuations from the distribution. We start with the assumption that F is uniform on the interval [0, 1] (i.e.
F(x) = x on [0, 1]).

The situation we described is referred to as auctions under independent private values (IPV) because the valuation v_i is a private
information of player i. This is the part that gives rise to asymmetric information. However, because this set up is a common knowledge,
each agent knows that his opponents’ valuations come from the distribution F and this gives each player some probabilistic idea about
the valuations. For example, a player will know that the probability for opponent j to have a valuation of above 1/3 is 2/3.

Auction: First-Price Sealed Bid

Rules
Each of the n bidders will simultaneously write down his bid and put the bid in a sealed envelop. The auctioneer will rank the bid in a
descending order and the person with the highest bid will receive the object and pay the bid he wrote down. All bidders except the one
with the highest bid do not need to pay anything.

This is a Bayesian game because all players decide their actions simultaneously and there is asymmetric information in terms of players’
preferences (this depend on both the actions taken by all players and the state of the world, which is a vector of the valuations that
describes everyone’s preferences).

The strategy of a player is a bidding function that maps each of the possible valuations from the interval [0, 1] into a action, which must
be a positive real number. The action set is the set of positive real number. Mathematically, the strategy is

As we did before, we can think of player i as many incarnations of himself with different valuations. To solve the problem we do not need
to optimize player i’s payoff with respect to his entire strategy, instead, we only need to optimize with respect to his realized valuation.

Our aim is to find a symmetric Bayesian NE in linear strategies. We make a guess that the bidding function (i.e. the strategy) in
equilibrium is symmetric and linear. Mathematically, we assume that the equilibrium strategy is

We need to complete two tasks

1. Confirm that a linear strategy is the best reply to a linear strategy.
2. Find the value of the constant k.

Solution
Consider player i whose realized valuation is v_i and the submitted bid is b_i. The expected payoff for player i is

To calculate the probability of winning, we assume that everyone except for player i follows the equilibrium strategy (i.e. the linear
bidding function). Player i wins when b_i is strictly larger than the bids of all other players.
** We do not consider tie breaks here because v_i is a continuous random variable and the probability of having a tie is zero. Thus, we
use strict inequality in the expression above.

Plugging in the expression for the equilibrium strategy gives

Since in equilibrium, the players’ bids are functions of their valuations, which are random variables, the bids must also be random
variables. When dealing with random variables, we want to rearrange the expression such that the random variable is on one side of the
inequality and everything else is on the other side of the inequality.

The expression shows that player i wins if the realized valuations of all his opponents are below the threshold b_i / k, which is a function
of player i’s bid. Then higher b_i is, the higher the threshold is, and therefore the greater the probability of winning. We can simplify the
expression further using the assumption that the valuations are independent across players. Base on this assumption, we can find the
probability of the joint event where v_j is smaller than b_i / k for all j not equal to i. Since all events are independent, the probability of
the joint even is equal to the product of the probability of each individual event. Note that there are n - 1 terms in the expression below.

Then, using the assumption that the events are identically distributed, the probability for each of the individual event is the same across
all j not equal to i. Thus, the probability of player i winning becomes the following:

Lastly, recall that the cumulative distribution function F(x) gives the probability of the random variable lying below x. To simplify the
expression above, we can write the probability as the cumulative distribution function F(.) with b_i / k in place of an argument:

Moreover, because v_i is distributed uniformly over the interval [0, 1], the probability is equal to the following:

Since we are calculating a probability, we need to make sure that the result is always between zero and one. In this case, the probability
will be greater than one if b_i is larger than k. If this happens, we might need to write it as a piece-wise function. However, since we are
never going to be in a area where b_i is above k, we can safely ignore this scenario.
** Keep in mind that corner solutions can occur in certain cases.

To summarize, the probability for player i to win is an increasing function of b_i. This makes sense because the higher the bid that player i
submits, the greater the chance of winning the auction.

Plugging in the probability of winning shows that player i’s expected payoff is

The next step is to maximize player i’s payoff conditional on his realized valuation v_i with respect to b_i.
** In our approach, we are optimizing player i’s payoff with respect to a particular action b_i instead of a entire strategy β_i. In particular,
we are not using the unknown constant k as a control and find the optimal k from the point of view of the agent. We are not doing this
because ???

The graph below illustrates the payoff function for player i. The intuition of the hump shaped function is that when player i’s bid is low,
increasing the bid would increase player i’s chance of winning and thus increase his expected payoff. However, at some point, increasing
the bid would reduce the benefit of winning. Thus, the expected payoff goes down when b_i is high. Eventually, player i would receive a
payoff of zero when v_i = b_i. Thus, players in the auction must find a balance between the chances of winning the auction and the
benefits from winning.
Mathematically, we can find the optimal bid for player i by calculating the first order condition of his expected payoff.

The final result shows that if all player i’s opponents play according to some linear strategy, then type v_i of player i should submit a bid
of b_i*. Since we did not specify v_i during the derivative, the result is true for any value of v_i. In the process above, we have achieve
two tasks:
1. We have proved that when everyone else is playing according to a linear strategy, then player i’s best reply is also a linear strategy
(i.e. b_i* is a linear function of v_i).
2. We have found the value of k, which is equal to [(n - 1) / n]. Since we did not complete the derivation for a specific player, the value of
k is true for all players.

In conclusion, the Bayesian NE in this world is

Intuitions and Observations

First, the resulting b_i* is strictly smaller than v_i. This is nature because submitting a bid that is higher than the valuation would lead to a
negative payoff, which is not optimal. Secondly, as n becomes larger, the value of k approaches one. The reason is that when the number
of bidder increases, the competition becomes more intense. From the game theoretic perspective, as n increases, a player would need to
out-bid more opponents. In particular, the player wants to out-bid the person with the highest valuation across all his opponents. It
follows that the larger the number of opponents (i.e. n - 1), the larger is that valuation in the stochastic sense (this means that the chance
that a opponent will draw a high valuation is greater when the number of opponents is larger). As the competition becomes harder, the
player would need to increase his bid in order to win. In other words, the share of the player’s valuation that he submits as a bid would
go up. In the limit when n is infinity, the bid becomes v_i as k becomes 1. Intuitively, when there are infinitely many players, the
probability of having at least two players with a valuation of 1 happens with a probability of 1. Thus, the only way to win is to submit b_i =
v_i, which is equal to 1. Note that this is not the full proof for first price sealed auction with infinitely many players, but this is the intuition
for why the b_i approaches 1 when n goes to infinity.
4.8 Auctions: Setup
• Single object up for auction
• n bidders
• Independent private values
• Each player has a valuation vi 2 [0, 1]
• Player’s valuation does not depend upon his opponent’s valuation or information
• vi ⇠ F (v); F 0 (v) = f (v)
• Each bidder is risk neutral: payo↵ is

vi I{won the auction} p

where p is the expected payment she makes

4.9 Auctions: Rules

• First price sealed bid auction: highest bid wins object, and pays her own bid
• Second price sealed bid auction: highest bid wins object, and pays second highest bid
• All pay sealed bid auction: highest bid wins object, everyone pays their own bids

4.10 Second price sealed bid auction

Theorem 1 In a second price auction with independent private values, it is a weakly dominant
strategy to bid bi = vi

Proof :
Two cases: (1) player i bids his valuation and wins; (2) player i bids his valuation and loses.
Case 1: Player i gets v_i - p_i > 0 in case (1) if he follows the equilibrium strategy (note that p_i is the second highest
price, which must be smaller than v_i, which is the highest price).
• bidding above vi does not change the payo↵: price remains the same and player i remains
a winner
• bidding below vi does not change the price but lowers the chances of winning. If player
wins he always gets a positive payo↵ and if he loses he gets 0, hence a decrease in winning
probability leads to a decrease in payo↵
Player i cannot make a pro table deviation in any circumstances under case 1.
Case 2:
• bidding below vi does not change anything The probability of player i winning is still zero if he bids below v_i.

• bidding above vi increases the probability of winning the object and paying strictly more
than vi , i.e. it leads to a negative payo↵
Q.E.D.

6
Auction: Second-Price Sealed Bid

Rules
Each of the n bidders will simultaneously write down his bid and put the bid in a sealed envelop. The auctioneer will rank the bid in a
descending order and the person with the highest bid will receive the object and pay the bid submitted by the second highest bidder. All
bidders except the one with the highest bid do not need to pay anything.

Proposition & Corollary

Proposition: bidding the valuation, that is, the bidding function β_i (v_i) = v_i is a weakly dominant strategy.
Corollary: the bidding function β_i (v_i) = v_i is a Bayesian NE.

The proposition leads to the corollary because by the definition of a weakly dominant strategy and by the fact that the proposition
applies to every player, the strategy β_i (v_i) = v_i is a best reply to anything. Therefore, we can assemble these weakly dominant
strategies into a strategy profile, which gives a Bayesian NE.

We are interested in this proposition because from the following corollary, we can conclude that we have found a BNE in which everyone
just writes down his true valuation of the object. An implication is that a second-price auction is a mechanism that recovers the valuation
of every player in the auction.

Proof of the Proposition

Suppose that one player follows the equilibrium strategy above and everyone else follows some arbitrary strategy. We can prove the
proposition by showing that in every possible scenario, the player who follows the bidding function β_i (v_i) = v_i cannot make a
profitable deviation.

See lecture notes.

In every possible scenario, player i cannot do better than bidding b_i = v_i regardless of his opponents’ strategies. In conclusion, in a
second price sealed bid auction, there is a equilibrium with weakly dominant strategy where every player bid his own valuation.

Another Equilibrium
The strategy above is not the unique equilibrium in second price sealed bid auction. The following strategy is also an equilibrium:
• Player 1 bid a number that is above the highest possible valuation in the game.
• Every other player bids zero.

This is a equilibrium because under this strategy profile, player 1 will always win and the only way to outbid player 1 is to submit an bid
that is greater than b_1. However, in this case, the player who now becomes the highest bidder would need to pay the amount that was
submitted by player 1, which is guaranteed to be larger than any possible valuation. This shows that the best reply is to bid zero. At the
same time, playing according to the equilibrium strategy would allow player 1 to obtain the object for free (since the second highest
price is zero) and reducing his bid would change the outcome only if he reduces his bid to zero. However, in this case, player 1 would
lose meaning that reducing his bid is not a profitable deviation.

This equilibrium highly replies on the rationality of the players (while all equilibrium relies on rationality, this equilibrium’s reliance is
especially high). Suppose players pay more attention to costly mistakes and less attention to less costly mistakes. Then, players who are
bidding zero are more likely to make mistakes (i.e. submit a bid that is higher than zero) because based on the common knowledge that
some other bidder (player 1 in this strategy considered) is going to submit a ridiculously high bid, they will not lose anything if they make
some mistakes. However, those mistakes can be costly to player 1, especially when his realized valuation is low. This is because he now
has to pay the second highest price that is not equal to zero due to his opponents’ mistakes. If player 1 thinks that other players will
make mistakes, he will submit a lower bid in order to protect himself from costly mistakes, which can destroy the construction of the
equilibrium. The equilibrium strategy considered is robust to this logic because it is a weakly dominant strategy.
4.11 First price sealed bid auction
As an illustration, we will study this game under the simplifying assumption: F (x) = x (uniform
distribution on [0, 1])
First we will show that there is an equilibrium in which each bidder employs an equilibrium
bidding function (pure strategy) (v) = rv that is strictly increasing (r > 0).

u(bi , vi ) = (vi bi ) Pr {bi (vj ), 8j 6= i}

Since the bids are independent and the values are independent,
Y
Pr {bi (vj ), 8j 6= i} = Pr {bi (vj )}
j6=i

Moreover, since (vj ) = rvj , we have

Pr {bi (vj )} = Pr(rvj  bi )

✓ ◆
bi
= Pr vj 
r
✓ ◆
bi
= F
r
bi
=
r
Plug this into a payo↵ function to get
✓ ◆n 1
bi
ui (bi , vi ) = (vi bi )
r

4.11.1 Equilibrium conditions

✓ ◆n 1
bi
ui (bi , vi ) = (vi bi )
r
bi chosen by i to maximize u(bi , vi )
✓ ◆n 2 ✓ ◆n 1
@ui bi 1 bi
= (vi bi ) ⇥ (n 1) = 0.
@bi r r r
✓ ◆
1 bi
(vi bi )(n 1) = 0.
r r

(n 1)vi = nbi

n 1
b⇤i = vi .
n
This tells us that if i expects everyone else to bid a fraction r of their valuation, it is optimal
for i to bid a fraction nn 1 of his valuation. Therefore if everyone bids a fraction r = nn 1 of their
valuation. This is an equilibrium.

7
Auction: All-Pay Sealed Bid

Rules
Each of the n bidders will simultaneously write down his bid and put the bid in a sealed envelop. The auctioneer will rank the bid in a
descending order and the person with the highest bid will receive the object. Every participant has to pay the bid that he submitted,
irrespectively of whether he won or lost the auction.

Solution
We try to look for a equilibrium in the following class of bidding function based on the assumption that the valuations are distributed
i.i.d. according to the uniform distribution from zero to one:

We need to complete two tasks based on the guess of the functional form of the bidding function above:
1. We need to check that the class of the bidding functions above are best replies to the same bidding functions
2. We need to find the value of the constants g and p

Consider bidder i with valuation v_i and a submitted bid of b_i. Bidder i’s expected payoff conditional on everyone else following the
equilibrium strategy is the following:

The payoff function is hump shaped: when b_i is small, increasing b_i would increase player i’s chance to win; when b_i is big, increasing
b_i further would not be optimal because the payment is going to approach v_i. Players need to balance the effect of increasing b_i on
the chance of winning and the effect on the resulting value of wining.

We can find the optimal b_i by finding the FOC fo the expected utility for player i:

The expression above shows that when everyone else plays according to the equilibrium strategy, bidder i would want to best reply with
the bidding function with the functional form above. We have completed the first task by showing that b_i is in the same class as the
function we guessed above.

To complete the second task, we can equate the expression for the constant to g and the expression for the power to q.

Note that since we have not specified any player or any valuation in the derivation, the value of the constants we obtained above apply
to every player. This means that we have found a symmetric equilibrium where everyone plays the same strategy (i.e. the same bidding
function β). In conclusion, the equilibrium bidding function is

Since every valuation is between zero and one, the expression for β_i is always strictly less than the bidding function we found in the first
price auction. In other words, the bids submitted in the all-pay auction are lower than the bids submitted in the first price auction. This
makes sense intuitively because the bids submitted in all-pay auction are payable regardless of whether the player wins the auction. Thus,
the players would want to scale back and be less aggressive than they are in the first price auction.
4.12 All pay sealed bid auction
We continue using simplifying assumption: F (x) = x (uniform distribution on [0, 1])
We will show that there is an equilibrium in which each bidder employs an equilibrium
bidding function (pure strategy)
(v) = gvjq
that is strictly increasing (g, q > 0).

u(bi , vi ) = vi Pr {bi (vj ), 8j 6= i} bi

Since the bids are independent and the values are independent,
Y
Pr {bi (vj ), 8j 6= i} = Pr {bi (vj )}
j6=i

Since (vj ) = gvjq , we have

Pr {bi (vj )} = Pr(gvjq  bi )

✓ ◆1/q !
bi
= Pr vj 
g
✓ ◆1/q !
bi
= F
g
✓ ◆1/q
bi
=
g

Plug this into a payo↵ function to get

✓ ◆ nq 1
bi
ui (bi , vi ) = vi bi
g

4.12.1 Equilibrium conditions

✓ ◆ nq 1
bi
ui (bi , vi ) = vi bi
g
bi chosen by i to maximize u(bi , vi )
✓ ◆ nq 1 1
@ui (n 1) bi 1
= vi 1 = 0.
@bi q g g
or ✓ ◆n q
gq 1 q q

b⇤i = g vi n 1 q

n 1

8
 Recall that (vj ) = gvjq , hence
q
q= () q = n
n 1 q
and ✓ ◆n q
gq 1 q n 1
g=g () g =
n 1 n
The symmetric equilibrium is
n 1
b⇤i (vi ) = vin
n

4.13 Expected revenue

Expected revenue is " #
X
R=E p⇤i (vi )
i2N

where p⇤i (vi ) is the expected payment made by player i.

All-pay auction:
n 1 n
p⇤i (vi ) = v
n i

4.13.1 Expected payment: First price auction

p⇤i (vi ) = b⇤ (vi ) Pr{i wins | vi }

= b⇤ (vi ) Pr{8j 6= i : b⇤ (vi ) b⇤ (vj )}
= b⇤ (vi ) Pr{8j 6= i : vi vj }
= b⇤ (vi ) (Pr{vi vj })n 1

= b⇤ (vi )vin 1
n 1 n
= v
n i

4.13.2 Expected payment: Second price auction


p⇤i (vi ) = E max{vj }I{8j 6= i : vi vi }
j6=i

Let v (1) = max {vj }. It is called largest order statistics. For agent i, v (1) is a r.v. The cdf
j2N :j6=i
for it is Y
Pr{v (1) < x} = Pr{vj < x} = xn 1

j2N :j6=i

9
 Therefore,

p⇤i (vi ) = E max{vj }I{8j 6= i : vi vi }
j6=i
Zvi
= xd(xn 1 )
0
Zvi
= (n 1)xn 1 dx
0
n 1 vi
= xn
n 0
n 1
= vin
n

4.13.3 Revenue equivalence

The expected payements are the same in all three auctions:
n 1 n
p⇤i (vi ) = v
n i
hence the revenue is the same as well.
Theorem 2 (Revenue equivalence theorem) In the setting with independent private val-
ues, consider two auctions that have the equilibria with the same allocation across two auctions.
If the equilibrium expected payo↵ of the lowest type (a bidder with the lowest possible valuation)
is the same in both auctions, the expected equilibrium payments and revenue are the same as
well.

4.14 How to solve all pay auction for a generic continuous distribu-
tion
Let vi ⇠ F , i.i.d. across i 2 N .
We will look for symmetric BNE in strictly monotone strategies: each bidder employs an
equilibrium bidding function (pure strategy) (v) that is strictly increasing.

u(bi , vi ) = vi Pr {bi (vj ), 8j 6= i} bi

Since the bids are independent and the values are independent,
Y
Pr {bi (vj ), 8j 6= i} = Pr {bi (vj )}
j6=i
1 n 1
= Pr (bi ) vj
⇥ 1
⇤n 1
= F (bi )
where 1 is the inverse bidding function.
Therefore: ⇥ ⇤n
1 1
u(bi , vi ) = vi F (bi ) bi

10
4.14.1 Best reply
Choose bi to maximize payo↵:
⇥ 1
⇤n 1
u(bi , vi ) = vi F (bi ) bi

F.o.c. ⇥ ⇤n
1 2 1 1 0
(n 1)vi F (bi ) f (bi ) ( ) (bi ) 1=0
1
Recall that in equilibrium (vi ) = bi or (bi ) = vi hence the expression above can be
rewritten:
1
(n 1)vi [F (vi )]n 2
f (vi ) 0 (v
1=0
i)

This is a linear di↵erential equation. Solve by integrating both parts:

Zvi
(vi ) = (0) + (n 1)x [F (x)]n 2
f (x) dx
1

First price auction can be solved in the same manner, but the di↵erential equation is more
difficult.

11
Auction: Expected Revenue (Comparison Between Auction Formats)

All-Pay Sealed Bid

The expected payment by bidder i with valuation v_i in a all-pay auction is

Since everyone must pay his submitted bit, there is no uncertainty from the point of view of the bidder.

First-Price Sealed Bid

The expected payment by bidder i with valuation v_i in a first-price auction is

In first price auction, uncertainty arises from the fact that the bidder either pays his submitted bid or zero, depending on whether he
wins. Since only the winner needs to make the payment, we need to multiply the value of the bid with the probability of winning. In
equilibrium, bidder i will win if and only his bid is above everyone else’s bid. Since this is a symmetric equilibrium, player i’s equilibrium
bid is above everyone else’s equilibrium bid if and only if his valuation is above everyone else’s valuation (this follows from the fact that
the bidding function is a function of v_i and the strategy is the same for all players). Based on the assumption that the valuations are
random variables from the uniform distribution on the interval [0, 1], the probability that v_i is greater than v_j for j ≠ i is v_i. Since there
are n - 1 opponents, the joint probability that v_i is higher than all v_j for j ≠ i is (v_i) ^ (n - 1).

Second-Price Sealed Bid

Second price auction involves more uncertainty than in the other two formats because uncertainty can arrive from two components in a
second price auction:
1. Whether or not a player needs to make a payment depend on whether he wins the auction or not;
2. The value of the payment depend on the bid submitted by the second highest bidder.

In this case, the expected payment by bidder i with valuation v_i is the probability of player i winning multiplied by the expected value of
the second highest bid conditional on player i winning the auction.

The probability of player i winning in a second price auction is equivalent to the probability computed in the case of first-price auction:

Now lets consider the conditional expectation of the second highest bid. Player i’s ex ante belief (i.e. player i’s belief before he receives
his valuation) about everyone else’s valuation is uniformly distributed on the interval [0, 1]. Suppose that player i’s valuation is v_i, then,
conditional on player i winning, player i would know that everyone else’s valuation is below v_i. Thus, his posterior belief (i.e. updated
belief) is that everyone else’s valuation is distributed on the interval [0, v_i]. We denote the distribution of the second highest price
conditional on player i winning by the function G(z), where z is a random variable.

Suppose that there are n - 1 random variables that are uniformly distributed on [0, v_1] where each of those random variables represents
the valuation for one of player i’s opponents. Then, the joint probability that the number z is above v_j for all j not equal to i conditional
on v_j being smaller than v_i (i.e. conditional on player i winning) is

Again, G(z) is the distribution of the second highest bid conditional on player i winning. Then, we can write the conditional expected
value of the second highest bid as a integral. The expected payment by bidder i is the following:
Comparing the results found in the three auction formats shows that the expected payments for player i are the same in all three cases.
Although the derivation relies on the assumption that the valuations are uniformly distributed on [0, 1], the result holds in general. This is
referred to as the revenue equivalent theorem.

We can conclude that two different auction formats generate the same expected payments by the bidders and the same expected
revenue for the auctioneer if the following two conditions hold:
1. The equilibrium allocation rule is the same across the two different auction formats;
2. If we fix some type of a player, this player will get the same equilibrium payoff in the two different auction formats.

We can show that both conditions hold for all three auctions we have considered:
• Condition one: equilibrium allocation of the object
All-pay auction: follows from the fact that the equilibrium bidding functions are monotone in v_i, a person with valuation v_i will
get the object with probability v_i ^ (n-1) in equilibrium
First-price auction: follows from the fact that the equilibrium bidding functions are monotone in v_i, a person with valuation v_i
will get the object with probability v_i ^ (n-1) in equilibrium
Second-price auction: follows from the fact that the equilibrium bidding functions are monotone in v_i, a person with valuation v_i
will get the object with probability v_i ^ (n-1) in equilibrium
The equilibrium allocation is the same in all three auction formats
• Condition two: expected payoff of an agent of a certain type
For simplicity, consider a player with a valuation equal to zero
This player has no chance of winning in all three auctions, therefore, he will receive a payoff of zero in all three auctions
Since both conditions are satisfied, we can conclude that all three auctions formats will generate the same expected payments and
therefore the same expected revenue for the auctioneer.
Auction: All-Pay Sealed Bid

In this section, we consider how to solve an all-pay auction without making to much of a guess about the functional form for the bidding
function. We continue with the assumption that valuations are i.i.d., but now we assume some generic distribution F instead of uniform
distribution. Further, we assume that the valuations are continuous random variables. The only guess we make here is that the bidding
function β_i (v_i) is strictly monotone. This is a sensible guess because the higher the valuation, the higher the bid one will submit.

Under this setup, the expected payoff for player i with valuation v_i and bid b_i conditional on everyone else following the equilibrium
strategy is

Our goal is to construct a symmetric equilibrium. Thus, we will assume that the bidding function is the same for every player. Under this
assumption, we can erase the index such that the bidding function and the expected payoff become the following:

Note that while all players follow the same strategy, they can have different bids because the strategy is a mapping from valuations into
bids. Thus, agents with different valuations will submit different bids.

Now, instead of plugging in the functional form as before, we are going to plug in the inverse function of β. Due to the assumption of
monotonicity, the function β is invertible and we denote the inverse of beta_i by β^{-1}. Then, the expression becomes

Using the idea of i.i.d, we can further simplify the equation as following:

Next, we maximize the expression of the expected payoff with respect to b_i:

We can simplify the expression above using the following properties.

Property One
Since the value b_i is player i’s best reply to everyone else’s strategy, it must be the vase that b_i = β(v_i) since β(v_i) is the equilibrium
strategy. It follows that

Property Two
We need to find the derivative of the inverse bidding function. As shown by the diagram below, the inverse function β^{-1} is the mirror
image of the original function around the 45 degree line. Furthermore, the diagram suggests that the derivative of the inverse function,
that is, the slope of the tangent to a particular point on β^{-1} is the mirror image of the slope of the tangent to the corresponding point
on β. Thus, the derivative of the inverse function if one divide by the derivative of the original function. Note that when we use this
property, we need to change the variable in the function β from b_i to v_i. This is because when looking at the mirror image of a point,
the point is b_i on the inverse function but is v_i on the original function.

Using the two properties above, we can simplify the expression to the following:
We are illustrating the derivation using all-pay auction because with all pay auction, the differential equation we get is relatively simple.
Specifically, we have a derivate of β on the LHS and nothing with β on the RHS. We can solve such a differential equation by intergrating
both sides. If we follow the same step in the setting of a first-price auction, we will end up with a differential equation with some
derivative of β on the LHS and some function of β on the RHS, which makes process more complex.

The last element to decide is the value of the bounds of the integration. Let the support of the distribution be from v lower bar to v
upper bar. Mathematically,

Then, the optimal bounds to use is to set the the lower bound on the support of the distribution, that is, v lower bar, as the lower bound
of the integral and set v_i as the upper bound. We do this because with such a bound, we can obtain the following expression after
taking the integral:

In equilibrium, beta (v lower bar), which is the bid submitted by the player with the lowest valuation, will be equal to zero because with a
strict monotone bidding function, the player with the lowest valuation will have no chance of winning the auction. In other words, the
probability for this person to win when he follows the equilibrium strategy is zero. Therefore, submitting any positive bid in an all-pay
auction would not be optimal as he can make a profitable deviation by bidding zero. Then, the expression becomes the following, where
the bidding function is equal to the integral on the RHS. The RHS can be solved once the distribution F is given.

The process above shows how to solve auctions without assuming too much about the shape of the bidding function.