1: GAMES IN STRATEGIC FORM

MARTIN CRIPPS

The purpose of this topic is to teach you what a game in strategic form is and how
economists analyze them.

1. Games in Strategic Form: Description and Examples

1.1. Defining a Game in Strategic Form. A game in strategic form consists of several
elements. Here we are quite precise about what must be included in the description. There
must be the following three things:
(1) A list of players; i = 1, 2, . . . , I. This list can be finite or infinite. Here the names
of the players are just numbers but they can be any entity you are interested in
studying.
Examples:
(a) In Rock-Paper-Scissors there are two players i =Fred, Daisy.
(b) In an oligopoly with 5 firms there are n = 5 players i =Ford, Honda, Toyota,
VW, Fiat, GM.
(c) In bargaining there is a buyer and a seller, so the list of players is i =buyer,
seller.
(2) A description of all the actions each player can take. The set/list of possible
actions for player i is usually called their pure strategies by game theorists. We will
represent this by the set Si and we write a typical action as si ∈ Si .
Examples:
(a) In Rock-Paper-Scissors, a player has three pure strategies: so
si ∈ {Rock, Paper, Scissors} = Si , for i = 1, 2.
(b) In an oligopoly, each firm, i, will choose a non-negative quantity of output to
produce, 0 ≤ qi the set Si is the set of all positive output quantities.
(c) In bargaining, the buyer proposes a price 0 ≤ p ≤ 1 and the seller makes a
proposal too so the strategy sets are just intervals of prices: Sbuyer = {0 ≤
p ≤ 1}, Sseller = {0 ≤ p ≤ 1}.
When you want to write all the actions that were taken by all of the players
you will write a list or vector s = (s1 , . . . , sI ). This is called an action or strategy
profile. The set of all possible strategy profiles (all possible plays of the game) is
written as S , where s = (s1 , . . . , sI ) ∈ S.
Examples:
1
2 MARTIN CRIPPS

(a) In Rock-Paper-Scissors: an example of a profile is s = (Scissors, Paper) here

player 1 does S and 2 does P . The set of all possible profiles S has 9 elements.
(b) In oligopoly: a profile is a list of quantities for every firm each firm, i. That is
a list (q1 , q2 , . . . , qn ). The set of all possible profiles is Rn+ .
(c) Bargaining: a profile is two prices (pb , ps ) and the set of all possible profiles is
[0, 1]2 .
(3) The final element is the payoffs (or utility or profit) the players get from their
actions or pure strategies. We can write this as a utility/payoff function ui (s) =
ui (s1 , . . . , sI ), that determines player i?s payoff at every possible play of the game.
The payoffs are usually determined by the rules of the game or economic phenom-
enon you are studying.
Examples:
(a) (Rock Paper Scissors): Recall there are two players i = 1, 2 and each has three
actions Si = {R, P, S} and the payoffs can be represented in a table:
R P S
R 0,0 -1,1 1,-1
P 1,-1 0,0 -1,1
S -1,1 1,-1 0,0
(b) (Oligopoly where firms choose quantities): Recall that the firms’ names are
i = 1, 2, . . . , n. The firms? actions are outputs qi > 0. The firms’ payoffs are
their profits, which depends on the demand. Price will depend on total output
so we write P (q1 + q2 + .. + qn ). (For example P = 50 − q1 − q2 ... − qn .) Profit
also depends on a firm’s costs, we assume these only depend on their output
c(qi ). Hence
Profit of Firm i = Revenue − Costs
= Output of i × Price − Costs of i
= qi P (q1 + q2 + .. + qn ) − c(qi )
This completes our initial description of a game in strategic form. We now want to allow
for the possibility that players act randomly. We don’t necessarily think that players are
genuinely randomising. But it may be a very good description of what the players think
the others are doing. You may know exactly how you’re going to play Paper-Scissors-Rock,
but I don’t. From my point of view your action looks random. We call random actions
Mixed Strategies. On mixed action for a player is one probability distribution, so the set
of all the player’s mixed strategies is the set of all probability distributions on their pure
actions. We write the mixed strategy of player i as σi . We will write the profile of mixed
strategies for all players as σ = (σ1 , . . . , σI ).
Example 1: In paper, scissors rock Si = {R, P, S} and σi = (p, q, 1 − p − q) where p is the
probability R is played, q is the probability P is played and 1 − p − q is the probability S
is played.
 1: GAMES IN STRATEGIC FORM 3

Example 2: When firms choose quantities Si = [0, ∞) is the set of possible quantities. A
random choice of a quantity can be represented by a probability distribution over the set of
positive numbers. One way of describing such a distribution is to write down its cumulative
distribution function (cdf) F (x) := P r(Firm’s outputs less than or equal to x).
Payoffs from Mixed Actions: A player’s payoff when mixed actions are played is an expec-
tation (or average) taken over all the payoffs they may get multiplied by the probability
they get them. This expectation is taken assuming the players randomise independently.
Examples of payoffs from mixed actions (Rock Paper Scissors): Suppose in this game you
are the column player and you believe the row player will play action R with probability
p, P with probability q and S with probability 1 − p − q. You are interested in your payoff
from action P. Below we have written out the payoffs and emphasised the relevant numbers
for the column player.
R P S
p 0,0 -1,1 1,-1
q 1,-1 0,0 -1,1
1 − p − q -1,1 1,-1 0,0
If she plays P she expects to get 1 with probability p, 0 with probability q and −1 with
probability 1 − p − q. Thus on average she expects to get
p × 1 + q × 0 + (1 − p − q) × (−1) = 2p + q − 1.
If we repeat these calculations for each column we can work out what each action will give
her in expectation
R P S
p 0,0 -1,1 1,-1
q 1,-1 0,0 -1,1
1−p−q -1,1 1,-1 0,0
1 − p − 2q 2p + q − 1 q − p
2. Dominance
2.1. Strict Dominance.
Definition 1. A mixed strategy σi strictly dominates the pure action s0i for player i, if and
only if, player i’s payoff when she plays σi and the other players play actions s−i is strictly
higher than her payoff from s0i against s−i for any actions s−i the others may play:
ui (σi , s−i ) > ui (s0i , s−i ), ∀s−i .
Consider the following game (we only put in the row player’s payoffs as that is all that
matters for now).
L R
T 3 0
M 0 3
B 1 1
4 MARTIN CRIPPS

And consider two strategies for the row player: play the first two rows with equal probability
σi = (1/2, 1/2, 0), or play the bottom row si = B. We can write the expected payoffs to
these strategies in the following way.
L R L R
1
2 3 0 0 3 0
1
2 0 3 0 0 3
0 1 1 1 1 1
3 3 1 1
2 2

From this you can see that if the column player plays L the strategy σi gives the row player
the payoff of 32 but the strategy si gives the row player the payoff 1. And, if the column
player plays R strategy σi also gives the row player the payoff of 23 but the strategy si gives
the row player the payoff 1. Thus the strategy σi always does better than the strategy
si . To describe this we say si is strictly dominated and we would never expect a rational
player to play this.
However, eliminating strictly dominated actions can allow us to make strong predictions
about what actions the players will use. Consider the following game. . .
L R
U (8,10) (-100,9)
D (7,6) (6,5)
First observe that R is strictly dominated by L for the column player. So a rational column
player will never play R. If the row player knows this then they should play U getting
8 rather than 7. So we predict (U, L) as the outcome of this game. But, some types
of players may be very worried about the −100. If you had some doubts about column
player’s rationality would you be willing to play U ?

2.2. Weak Dominance. The notion of weak domination does not require players to
strictly prefer one strategy to another, it is enough to weakly prefer one to the other.
Definition 2. A mixed strategy σi weakly dominates the pure action si ∈ Si for player i,
if and only if, playing σi is at least as good as playing si whatever the other players do:
ui (σi , s−i ) ≥ ui (si , s−i ), ∀s−i .

L R
T (1,1) (0,0)
M (0,0) (0,0)
B (0,0) (0,0)

T weakly dominates M for the row player, because T is better than M if column plays L
and T is no worse than M is column plays R.
 1: GAMES IN STRATEGIC FORM 5

2.3. Iterative Elimination of Dominated Strategies. If we can repeatedly (or itera-

tively) eliminate weakly or strictly dominated strategies to end up with one prediction on
how the game will be played, then this can be used to solve games.
Example:

L M R
T (3,3) (2,2) (5,2)
M (2,10) (4,4) (6,3)
B (2,1) (2,7) (1,0)
• Step 1: R is strictly dominated by L, so a rational column player will not play R
L M R
T (3,3) (2,2) -
M (2,10) (4,4) -
B (2,1) (2,7) -
• Step 2: In the remaining game 12 T + 12 M strictly dominates B. So if the row player
knows the column player is rational and is rational themselves, then they will not
play B.
L M R
T (3,3) (2,2) -
M (2,10) (4,4) -
B - - -
• Step 3: In the remaining game L strictly dominates M for the column player. So if
the column player is rational, knows that the row player is rational and knows that
the row player knows the column player is rational, then they will not play M.
L M R
T (3,3) - -
M (2,10) - -
B - - -
• Step 4: In the remaining game T strictly dominates M for the row player. So if the
row player is rational and the row knows that the column player is rational, and
the row knows that the column player knows the row player is rational, then they
will not play M.
L M R
T (3,3) - -
M (2,10) - -
B - - -
Summary: the set of strategies that survive the iterated elimination of strictly dominated
actions are those actions that can be played by rational players with common knowledge
of rationality
6 MARTIN CRIPPS

3. Nash Equilibrium
Definition 3. A action profile s∗ = (s∗1 , . . . , s∗I ) is a Nash Equilibrium if s∗i is a best
response to the actions (s∗−i ) of the other players, and this is true for all players i. A
mixed strategy profile σ ∗ is a Nash Equilibrium if for all players i, if the same condition
holds for mixed strategies.

Example: Consider the game:

L R
U (8,10) (-100,9)
D (7,6) (6,5)
If row plays U then the best thing for column to play is L. (We say L is a best response
to U.) And, if column plays L then the best thing for row to play is U. (U is also a
best response to L.) This means that (U, L) satisfies the definition of a Nash equilibrium.
Recall that this game appeared before and this “solution” to the game was found using
the elimination of strictly dominated actions. So there appears to be a link between these
two concepts.
Example: As a second example, consider the game:
L R
U (1,1) (0,0)
D (0,0) (0,0)
Now observe that the action pair (D, R) satisfies the definition of a Nash equilibrium. If
row plays D then R is a best response. And if column plays R then D is a best response.
So although we may believe that this is a very unappealing solution to the game, it is
certainly a Nash equilibrium of the game, because it satisfies the definition above. This
leads us to a new notion of Nash equilibrium.

3.1. Strict Nash Equilibrium.

Definition 4. A Nash equilibrium is said to be strict, if it is an equilibrium in pure
strategies and each player i has a unique best response to the equilibrium action of their
opponents.
Strict equilibria are robust to small perturbations in the players’ payoffs, but do not
always exist. This is particularly the case in games where the Nash equilibria are all mixed
(involve random actions).
Example1: (Rock Paper Scissors)
R P S
R 0,0 -1,1 1,-1
P 1,-1 0,0 -1,1
S -1,1 1,-1 0,0
 1: GAMES IN STRATEGIC FORM 7

This game has a mixed strategy Nash equilibrium where each player plays every action with
probability 13 . If one player plays each action with equal probability, then their opponent
gets zero from any action. Every action they play gives them zero,, so every action is a
best response to this strategy. One best response is to play ( 31 , 13 , 13 ).
Example 2 : (Matching Pennies)

H T
H (-1,1) (1,-1)
T (1,-1) (-1,1)

If the column player plays ( 12 , 12 ) then the row player expects to get zero from H and from
T . Thus, any action is optimal. It follows that ( 21 , 12 ) is a best response by the row player.
(This is a general property of mixed equilibria—one player’s actions at the equilibrium
make their opponent indifferent and therefore willing to randomize.)

4. How to Find a Nash Equilibrium

4.1. Method 1 : Iteratively eliminate dominated strategies. If this works the ac-
tions that remain after you have eliminated all others have to be a Nash equilibrium. You
already have already seen an example of this above. If you only use strict dominance to
eliminate actions, then iterative elimination finds all Nash equilibria, But, if you eliminate
some actions using weak dominance then you may also eliminate a Nash equilibrium. For
example, in the game below the bold outcome is a Nash equilibrium, but plays weakly
dominated actions so it would be eliminated.
L R
U (1,1) (0,0)
D (0,0) (0,0)

You might think that this example is silly because the players ought to be able to avoid this
Nash equilibrium. But suppose we increase the number of players so there are three players
(player 3 chooses the matrix) then there still is a weakly dominated Nash equilibrium.

3 Plays L L R 3 Plays R L R
U (1,1,1) (0,0,0) U (0,0,0) (0,0,0)
D (0,0,0) (0,0,0) D (0,0,0) (0,0,0)

This Nash equilibrium seems much harder to avoid as one of the player (say player 3) now
has to persuade both of the others to change there actions. There are, therefore, sensible
situations where players end up playing weakly dominated actions at a Nash equilibrium.

4.2. Method 2 : The underlining method. This is a process for finding all the pure-
strategy Nash equilibria of a game. You only use it in finite games. Here is an example of
8 MARTIN CRIPPS

the process applied to the following game

L M R
T (3,3) (2,2) (5,2)
M (2,10) (4,4) (6,3)
B (2,1) (2,7) (1,0)
• Step 1: Underline the best thing for the column player in each row. First consider
the column player’s payoffs:
L M R
T (3,3) (2,2) (5,2)
M (2,10) (4,4) (6,3)
B (2,1) (2,7) (1,0)
Then underline the biggest number in each row. (If there are many biggest underline
them all.) Results in:
L M R
T (3,3) (2,2) (5,2)
M (2,10) (4,4) (6,3)
B (2,1) (2,7) (1,0)
• Step 2: Underline the best thing for the row player in each column. So consider
the row players payoffs and find the biggest one in each column:
L M R
T (3, 3) (2, 2) (5,2)
M (2,10) (4,4) (6,3)
B (2,1) (2,7) (1,0)
• Step 3: Any place where two elements are underlined is a NE

4.3. Method 3 : Mixed strategy NE’s in 2 × 2 games. This is a process that will
help you find Nash equilibria in games with 2 players each of which has 2 actions. To find
mixed strategy Nash equilibria is difficult and will usually require a computer for n × n
games. To find a mixed we will consider a specific game below, which is often called the
Battle of the Sexes
L R
T (3,1) (0,0)
B (0,0) (1,3)
Here is the recipe you follow:
• Step 1: Assume that the column player plays a mixed strategy (q, 1−q)and evaluate
the row player’s payoffs to all their actions.
q 1−q
T (3,1) (0,0) = 3q + 0(1 − q)
B (0,0) (1,3) = 0q + 1(1 − q)
 1: GAMES IN STRATEGIC FORM 9

• Step 2: Choose q to make these two payoffs the same.

3q + 0(1 − q) = 0q + 1(1 − q)
This solves to give q = 14
• Step 3: Suppose the row player randomizes (p, 1 − p) and evaluate the column
player’s payoffs to every action.
L R
p (3,1) (0,0)
1−p (0,0) (1,3)
1p + 0(1 − p) 0p + 3(1 − p)
• Step 4: Choose p to make these the same.
1p + 0(1 − p) = 0p + 3(1 − p)
3
solves to give p = 4.
• Summary: Then, p = 3/4 and q = 1/4 is a mixed strategy NE.
4.4. Method 4: The intersection of best responses. This is a direct generalization
of the underlining method. We will apply it to economics models of oligopoly later in this
course. This approach applies to all games, not just oligopoly. To understand how this
method works it is, again, optimal to do an example.
We will consider a game where. Player 1 chooses a number x, where −∞ < x < ∞
and Player 2 chooses a number y, where again −∞ < y < ∞. When (x, y) is chosen they
experience costs (which they would like to make as small as possible. The costs are
Player 1’s Costs = (x − y)2 + x2
Player 2’s Costs = (y − 4)2 + (x − 2y)2
So we now have a well-defined game with payoffs and actions. The main difference here is
that the action sets are an interval so we can use calculus to minimise costs. Here is the
process to find a NE.
• Step 1: Find the optimal x for player 1 if it knows the y chosen by player 2: That is
choose x to minimise C1 = (x − y)2 + x2 for a given y value. Thus we differentiate
dC1
= 2(x − y) + 2x = 0, ⇒ x = y/2.
dx
• Step 2: find the optimal y for player 2 if it knows the x chosen by player 1. Choose
y to minimize C2 = (y − 4)2 + (x − 2y)2 for a given x value.
dC2
= 2(y − 4) − 4(x − 2y) = 0, ⇒ y = 2(2 + x)/5.
dy
• Step 3: Solve where the two lines: (y = 2(2 + x)/5 and x = y/2) intersect. This
gives x = 1/2 and y = 1. This is guaranteed to be a place where player 1’s x is
best for her given y = 1 and player 2’s y is best for her given x = 1/2.
Here are some observations about this process:
10 MARTIN CRIPPS

(1) This method is a generalization of the underlining method to continuous actions.

(2) If a player has many best responses, they should all be included in the best response
function.
(3) It is possible a player does not have a best response (see Bertrand competition).
(4) This generalizes to n players. (There are always as many equations as actions.)
(5) It is possible that there is no intersection. This means that no pure Nash equilibrium
exists.

5. E. Nash Equilibrium Multiplicity

Coordination Games: There are two players who want to meet at location A or location
B but we don’t know which to meet at. This creates a game with the following payoffs

A B
A (1,1) (0,0)
B (0,0) (1,1)

These games generally have huge numbers of equilibria (it has two pure strategy equilibria
and one mixed strategy equilibrium). Sometimes we need a tool for selecting among Nash
equilibria in order to make predictions. In this game it seems virtually impossible, any
justification for the (A, A) equilibrium seems like it ought to be a justification for (B, B)
also. However, in some games we can make choices.
J-J. Rousseau introduced the Stag-Hunt game. He had in mind the following story. Two
hunters can either independently hunt a rabbit (this requires no cooperation), or can try
to coordinate to catch a stag. But hunting a stag fails unless we both commit to hunting
a stag. Here are the usual payoffs given to this game. Another important feature is that
hunting a rabbit gets easier if the other player isn’t also hunting a rabbit.

Stag Rabbit
Stag (9,9) (0,8)
Rabbit (8,0) (7,7)

This game has two Nash equilibria: (S, S) and (R, R). The (S, S) is Pareto efficient, but
quite risky because for me to get 9 I need to be quite sure my opponent is also playing S.
The (R, R) NE is safer, because whatever my opponent does I do OK (I get at least 7),
but not Pareto Efficient. We say that: (S, S) is Pareto dominant NE and (R, R) is a Risk
Dominant NE.
Most games have many Nash equilibria. Some people view this a problem and we have to
find a way of choosing a unique way of predicting how the game will be played. Sometimes
the names or history of the strategy mean that particular action is most prominent and
therefore sensible to coordinate on. This is called a Focal Point. For example, at the end of
Slumdog Millionaire the two characters have no mobile phones and must meet somewhere
 1: GAMES IN STRATEGIC FORM 11

in Mumbai. They face a game that looks like:

x y z u v w ?
x (1,1) (0,0) (0,0) (0,0) (0,0) (0,0) ?
y (0,0) (1,1) (0,0) (0,0) (0,0) (0,0) ?
z (0,0) (0,0) (1,1) (0,0) (0,0) (0,0) ?
.. .. .. .. .. .. ..
. . . . . . .
Where x, y, z,? are potential places to wait for the other person. As it stands this game
has huge number of NE?s. However, only one location, x =Chhatrapati Shivaji Terminus,
is significant so they both meet each other there.
Sometimes communication before play will allow the players to choose a NE. The set of
NE’s of a game is the only set of credible non-binding agreements the players can reach.)
 2: GAMES IN EXTENSIVE FORM

MARTIN CRIPPS

1. Perfect Information Games in Extensive Form

1.1. Game Trees. A game in extensive-form captures information about the order of
moves, that is ignored in the strategic-form model of a game in the previous section. An
extensive form is a diagram that describes the details of the sequence of decisions and
actions the players take when they play a game. To do this we write done a mathematical
object called a decision tree. This is a collection of nodes (or vertices) linked by arrows
(or edges). The tree must have: (i) a unique initial node and (ii) a unique path from the
A. Perfect
initial node to every successor node. Information
Here is Games in Extensive
an example of a Form
tree.
Tree

Here are examples of directed graphs that are not trees. the first violates the unique
A. Perfect Information Games in Extensive Form
path condition and the second violates the unique initial node condition.
Not a tree
A. Perfect Information Games in Extensive Form

Not a tree

The important feature of these graphs (extensive form games) is the way they are la-
belled, because it is this that really explains the game. Here are the rules for labelling:
• Each non-terminal node is labelled with a player’s name and the edges from that
node are actions the player can take at that point in the game.
1
2 MARTIN CRIPPS

• Each terminal node describes a unique history of play. It is labelled with a payoff
to every player in the game for that history.
A. Perfect Information Games in Extensive Form
Here is an example of a tree with such a labelling:
. 1

Fold Raise

2 2

Fold Raise Fold Raise

(-1,1) (1,-1)
(0,0)
1
Raise
Fold

(-1,1) (2,-2)

If you wanted to give a verbal description of what goes on in this game. You might proceed
as follows. . . The game begins with Player 1 choosing between two actions Fold or Raise.
If Player 1 chooses Fold, Player 2 observes 1’s action and then must choose between Fold
and Raise and the game ends. The payoffs the players get are either (0, 0) or (−1, 1). If
Player 1 Raises initially, then Player 2 also gets to choose between Fold or Raise. If Player
2 Folds, the game is over and the payoffs are (1, −1). If Player 2 Raises, then Player 1 gets
another go and can choose between Fold or Raise a second time after which the game ends.
If Player 1 folds, the payoffs are (−1, 1) and if Player 1 Raises the payoffs are (2, −2).
I hope you can see A.#Perfect#Information#Games#in#Extensive#Form6
that the picture is much clearer than the words. Also notice the
convention that we describe the payoffs as a vector giving Player 1’s payoff first.
.6 1

Fold'' Raise'

2 2
Raise'
Fold'' Fold'' Raise'

(0,0)'
1
Raise'
Player'1' Player'2' Fold''

We also want to allow for the possibility that there is some randomness in the game.
(Dealing cards, rolling dice, exogenous uncertainty in the model.) To allow for random
 2: GAMES IN EXTENSIVE FORM 3

moves we introduce a player called Nature with a name zero. Each edge from a zero
node has a probability attached that describes the probabilities of the random move. For
example, A. Perfect Information Games in Extensive Form

Random Moves 1

Fold Raise

2 2

Fold Raise Fold Raise

(-1,1) (1,-1) Cards are

(0,0)
0 compared

1/2
1/2

(-1,1) (2,-2)

Important: There is a reason these pictures are called Perfect Information Games in
Extensive Form, this is because there is very little uncertainty in these games. In a perfect
information extensive form game each player knows exactly where they are in the game
when they take a move/decision. (The only thing they don’t know is how future moves
will be played.) They do know exactly what has happened at every point they make must
make a decision. Examples of perfect information games like this are Chess, Backgammon,
Go, Monopoly. However, games such as Poker, Bridge (where the players have private
information), or Paper-Scissors-Rock (where the players move at the same time) cannot be
described in this way—we will get to them a little later.

1.2. Pure Strategies. Thus far we have been describing a situation the players might
find themselves in. Now we want to describe what the players might do.

Definition 1. A Pure Strategy for a player is an instruction book on how to play the
game. This instruction book must be complete and tell the player what to do at every point
at which he/she must make a move. Because we want to consider what happens when
players make mistakes we must even include apparently redundant instructions.

Here is an example of an instruction book (Pure Strategy) for Player 1 in the game we
have been studying
4 A.#Perfect#Information#Games#in#Extensive#Form6
MARTIN CRIPPS

1’s#Instructions6 1
(Fold,Fold)6
Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
1

Fold'' Raise'

(11,1)' (2,12)'

This instruction book has the instruction Fold at the first decision and Fold at the second
decision. (Clearly, this second instruction is redundant but we still include it.) Here are
some other pure strategies for Player 1 in the game.

A.#Perfect#Information#Games#in#Extensive#Form6

1’s#Instructions6 1
(Raise,Fold)6
Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
1

Fold'' Raise'

(11,1)' (2,12)'

Player 1 Raises at the first decision node and Folds at the second.
 A.#Perfect#Information#Games#in#Extensive#Form6
2: GAMES IN EXTENSIVE FORM 5

1’s#Instructions6 1
(Raise,Raise)6
Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
1

Fold'' Raise'

(11,1)' (2,12)'
A.#Perfect#Information#Games#in#Extensive#Form6
Player 1 Raises at the first decision node and Raises at the second.
1’s#Instructions6 1
(Fold,Raise)6
Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
1

Fold'' Raise'

(11,1)' (2,12)'

Player 1 Folds at the first decision node and Raises at the second.
In this game Player 1 has 4 pure strategies: (Raise, Raise), (Raise, Fold), (Fold, Raise),
(Fold, Fold). Notice that Player 2 also has 4 pure strategies, but none of her instructions
are ever redundant.
1.3. Mixed and Behaviour Strategies. From our study of games in strategic form we
also want to allow for the players to make random moves. There are two ways of doing this.
The first is to imagine a library containing all the possible instruction books for Player 1.
(In this case there would be 4 books (Raise, Raise), (Raise, Fold), (Fold, Raise), (Fold,
Fold).) Then choosing one of the books at random, for example Player 1 could choose the
(Raise,Raise) book with probability 21 and the (Fold, Fold) book with probability 12 . This
is an example of what is called a mixed strategy.
Definition 2. A mixed strategy for a player is a random choice of an instruction book
(i.e. a random choice of a pure strategy).
This is a picture of what’s going on
6 MARTIN CRIPPS
A.#Perfect#Information#Games#in#Extensive#Form6 A.#Perfect#Information#Games#in#Extensive#Form6

1’s#Mixed#strategy6 1 1’s#Mixed#strategy6 1
½#(#Raise,#Raise)6 ½#(#Raise,#Raise)6
Fold'' Raise' Fold'' Raise'
½#(#Fold,#Fold)6 ½#(#Fold,#Fold)6
6 6
2 2 2 2
Fold'' Raise' Fold'' Raise' Raise'
Fold'' Fold'' Raise'

(11,1)' (1,11)' (1,11)'

(0,0)' (0,0)' (11,1)'
1 1

Fold'' Raise' Raise'

Fold''

(11,1)' (2,12)' (11,1)' (2,12)'

Notice that the randomness appears to be occurring only once (at the selection stage)
but from the perspective of the other players in the game watching what their opponent
does it may be that there are many points at which Player 1’s move looks random.
If you wanted to describe a general mixed strategies for Player 1 in this game you would
write down a probability distribution (p, q, r, 1 − p − q − r) where p is the probability (R, R)
is played, q is the probability (R, F ) is played, r is the probability (F, R) is played, and
1 − p − q − r is the probability (F, F ) is played.
The second way of including randomness is to allow for the instructions themselves to
have random elements. That is at each place where the player must move the instruction
book tells the player how to randomize at that point. For example such an instruction
book for Player 1 might be: At first node with probability 12 play Raise. At second node
with probability 14 play Raise. These books with random instructions are called Behaviour
Strategies.
Definition 3. A Behaviour Strategy for a player is an instruction book that has random
instructions. A.#Perfect#Information#Games#in#Extensive#Form6
Here’s our example of this random instruction book in action.
1’s#Behaviour#strategy6 1
6
Fold'½'' Raise'½'

2 2

Fold'' Raise' Fold'' Raise'

(21,1)' (1,21)'
(0,0)'
1

Fold'¾''' Raise'¼''

(21,1)' (2,22)'

If you wanted to write down a general behaviour strategy for Player 1 in this game. You
would write the pairs (p, 1 − p) and (q, 1 − q) where (p, 1 − p) is the randomization at the
first decision Player 1 must take and (q, 1 − q) is the randomization at the second decision
she must take.
 2: GAMES IN EXTENSIVE FORM 7

2. Incomplete Information Games in Extensive Form

The game trees above have the property that everyone playing the game knows every-
thing about the past events in the game—this rules out the following important features
of strategic interaction:
• Players moving simultaneously.
• Or players’ past actions being hidden from others.
• Or players knowing something that other players do not.
We need to have a way of representing this in the trees and the device we have to do
this is called an information set. These are things that we add on to the graphs above to
describe what players know or don’t know. In particular we the nodes in an information
set represent positions in the game that the player is unable to tell apart. Of course, if a
player is unable to tell the positions apart she must have the same actions and the same
name:
Definition 4. An Information Set is a collection of nodes with the property that: Each
node in the set has the same player’s name. Each node in the set has the same actions
available.
Consider the simplest kind of game where two players must each choose two actions
B.#Incomplete#Information#Games#in#Extensive#
simultaneously (such as matching pennies or battle of the sexes or the prisoners’ dilemma).
The picture below is inadequate, because PlayerForm7
2 knows what Player 1 has done before
she chooses her move.
2#sees#1’s#action7 1

Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
(0,0)'
B.#Incomplete#Information#Games#in#Extensive#
Form7
To get over this we make the following addition to the picture.
2#does#not#see#1’s#action71
Informa7on'Set'
(players#move#simultaneously)7
Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(11,1)' (1,11)'
(0,0)'
(0,0)'
8 MARTIN CRIPPS

This indicates that 2 does not know what 1 has done when she chooses between Raise and
Fold.
Often we will use dashed lines to indicate information sets—this is done in all the
following examples.
Information sets can be used to describe situations where players move simultaneously.
But there are many more subtle things that they can be used to represent
(1) Information Sets can be used to describe situations where a player’s past actions
are hidden from others (this is like the simultaneous move tree).
(2) Information Sets B.#Incomplete#Information#Games#in#Extensive#
can be used to describe situations where a player has private
information. For example, Player 2 knows the outcome of a coin toss but Player 1
Form7
does not. Hidden#information:#player#2#knows#the#outcome#of#a#coin#
toss#but#player#1#does#not7
0

1/2%% 1/2%

2 2

Fold%% Raise% Raise%

Fold%%
1 1
(0,0)%
(1,21)%
Fold%% Raise%%

(21,1)%
(21,1)% (2,22)% (21,1)%

(3) Information Sets can be used to describe situations where a player forgets what
B.#Incomplete#Information#Games#in#Extensive#
they knew previously. For example, Player 2 knows the outcome of a coin toss and
Form7
then forgets it. Hidden#information:#player#2#forgets#they#the#outcome#of#a#
coin#toss7
0

1/2%% 1/2%

2 2

Fold%% Raise% Raise%

Fold%%
2 2
(0,0)%
(1,21)%
Fold%% Raise%%

(21,1)%
(21,1)% (2,22)% (21,1)%

When we introduce information sets we need to change the definition of what a pure
strategy is:
 2: GAMES IN EXTENSIVE FORM 9

Definition 5. A pure strategy for an incomplete information extensive form game is just
an instruction for every information set where it must make a move. (NOT instructions
for every node.)
If a player does not know which node at an information set they actually are at they are
forced to choose the same action at every node in that set. For example, in the simultaneous
move game above Player 2 only has two pure strategies Fold or Raise. But in the game
without the information set Player 2 has four pure strategies (Fold,Fold), (Fold,Raise),
(Raise,Fold), (Raise,Raise), that is, they can act in a way that depends on what Player 1
did.

3. Nash Equilibrium
We have already defined what a Nash equilibrium is: At a Nash equilibrium each player’s
strategy is a best response to the other players’ strategies. This seems to be a good notion
of equilibrium in strategic form games, but (as we will see) it is not such a great notion
of equilibrium in extensive form games. Remember the test for a Nash equilibrium is:
fix everyone else’s strategy then check whether my strategy is optimal for me given that
everyone else is committed to playing their strategy.
To understand why a Nash equilibrium might not be a good way of describing equilibrium
in these games let us look at two Nash C.#Nash#Equilibrium0
equilibria of the same game. Here is a simple two-
stage game and one Nash equilibrium of it.
A#Nash#Equilibrium.0 1

Fold'' Raise'

2 2

Fold'' Raise' Fold'' Raise'

(0,0)' (11,1)' (1,11)'

(12,12)'

If we fix Player 2’s strategy (in green) you can see that changing 1’s strategy will reduce her
payoff to −2 from −1, so certainly 1 is doing the best she possibly can given 2’s actions.
Now let us test 2’s strategy. If we change what 2 does at the right decision it has no
effect on her payoff (she gets 1), so any action here passes the optimality test of a Nash
equilibrium. If we change 2’s action at her left decision her payoff goes down to zero from
1, so again she is acting optimally here. I hope this convinces you that this is a Nash
equilibrium.
Now we will argue that the above is not a particularly sensible equilibrium. Consider
again why Fold is optimal for Player 1. This is because she believes she will get −2 if she
plays Raise. Is this a sensible belief? The answer to this is yes under the assumption that
Player 2 plays the green actions, but is it reasonable to expect that Player 2 will play the
green action? The answer to this is no. Player 2 is threatening to play Raise if Player 1
10 MARTIN CRIPPS

Raises. However, this threat is not something Player 2 would actually want to carry out—it
is a non-credible threat. Because if 1 actually played Raise, the best thing for Player 2 to
do would be to Fold (they would get −1 rather than −2). Thus the problem with this NE
is that Player 2 is threatening to do something that they are not actually going to do and
as a result of the NE assumption Player 1 believes this threat.
We have just argued that this is a bad kind of equilibrium, so is there another NE of
this game that does not have this problem? Here is another NE of the same game that
does not rely on a non-credible threat.
C.#Nash#Equilibrium0

Credible#Threat.0 1
0
Fold'' Raise' This'then'changes'1’s'
0 behaviour.'
0
2 2
0
Fold'' Raise' Fold'' Raise'
0
0
(0,0)' (11,1)' (1,11)' (12,12)'
0
Once#player#1#realizes#this#threat#is#non=credible#she#is#
be>er#off#raising#too0
This is another Nash equilibrium and at this Nash equilibrium no non-credible threats
are made. Can we always find such a Nash equilibrium? The answer is yes and what
is more there is a very nice process to find these good Nash equilibria and it is called
backwards induction. This always works in games of perfect information (i.e. games
without information sets), so in these games we can always find Nash equilibria that only
rely on credible threats.

4. Backwards Induction in Finite Perfect Information Games

The process called backwards induction works for all finite, perfect information games
and it can be used to find the Nash equilibria that do not rely on non-credible threats.
Here we explain how you do backwards induction by working through an example. First
you consider the very last decision that gets taken in the game and the player who takes
it. You make the optimal decision for that node.
 2: D.#Backwards#Induction2
GAMES IN EXTENSIVE FORM 11

Example2 1
2
Fold'' Raise'
2
2
2 2
2
Fold'' Raise' Fold'' Raise'
2
2
(11,1)' (1,11)'
2 (0,0)'
1
Start#at#the#last#move#and#figure2 Raise'
Fold''
out#what#is#optimal#there2
2
(11,1)' (2,12)'

(As −1 < 2 Player 1 prefers Raise at the last node.) Then you determine the optimal
decisions at all the other final nodes.

D.#Backwards#Induction2

Example2 1
2
Fold'' Raise'
2
2
2 2
2
Fold'' Raise' Fold'' Raise'
2
2
(1,11)'
2 (0,0)' (11,1)'
1
Do#all#last#moves2 Raise'
Fold''
2

(11,1)' (2,12)'

(As 1 > 0 Player 2 prefers Raise at this node.)

Once the last decisions have been determined it is possible to figure out the second last
decision. We assume the two last decisions cannot now be changed. (Notice that we have
ensured there are no non-credible threats by making all the last decisions optimal for the
Players that have to take them.)
12 D.#Backwards#Induction2
MARTIN CRIPPS

Example2 1
2
Fold'' Raise'
2
2
2 2
2
Fold'' Raise' Fold'' Raise'
2
2
(0,0)' (11,1)' (1,11)'
2 1
Now#do#the#second#last#move.2 Raise'
Fold''
2

(11,1)' (2,12)'

(As −1 > −2 Player 2 prefers Fold at the second last decision.)

D.#Backwards#Induction2
Now we can finally determine the first decision

Example:2 1
2
Fold'' Raise'
2
2
2 2
2
Fold'' Raise' Fold'' Raise'
2
2
(0,0)' (11,1)' (1,11)'
2 1
Finally#do#the#first#move2 Raise'
Fold''
2

(11,1)' (2,12)'

(Player 1 prefers Raise as 1 > −1.)

We have now created an instruction book (or pure strategy) for every player in the game.
These pure strategies are a Nash equilibrium, because no change can improve any players’
payoff. But also they do not rely on non-credible threats. These equilibria are given a
special name, they are called Subgame Perfect Equilibria. However, the notion of subgame
perfect equilibrium can be applied more broadly to all extensive form games, not just those
with perfect information.
Notes:
• Backwards induction proves the existence of a NE in finite perfect information
games.
• This result is also called Zermelo’s Theorem and it implies that games such as
Chess, Draughts, Backgammon, Go, etc. must have Nash equilibria.
• Backwards induction is a generalization of dynamic programming.
 2: GAMES IN EXTENSIVE FORM 13

• Backwards induction will not work in games with information sets, so it can only
work in perfect information games.
• Reinhart Selten generalized the idea of backwards induction to create the idea of
subgame perfection and got the Nobel prize for it (we will study this in the next
section).

4.1. Economic Applications of Backwards Induction. There are many Macroeco-

nomic applications of backwards induction. Here are two microeconomic applications.

4.1.1. Investment by multinational firms. In this model we think of a firm first deciding
whether to invest overseas or not. Then the government of the destination country de-
D.#Backwards#Induction#2
cides whether to expropriate the investment or not. The country would like to commit
to not expropriating but, in the simple version of this game below, this is not a credible
undertaking. Given#expropriation#it#is#optimal#not#to#invest.2

(50,50)
DC
MNF
Expropriate
(-20,80)
Do not invest
(0,0)

(The first payoff is whatThe#purpose#of#bankruptcy#&#contract#law#is#to#avoid#this#

the multinational firm (MNF) gets and the second payoff is what
the developing country problem#that#investors#assets#get#stolen.2
(DC) gets.) The backwards induction solution, illustrated on the
figure, has no investment and no expropriation.
100
4.1.2. Entry deterrence. Here we imagine there is a monopolist in an industry and it is
trying to deter another firm from entering D.#Backwards#Induction#2
the industry and competing with it. The entrant
moves first and decides whether to enter or not and then if entry has occurred the incumbent
Entry#is#therefore#optimal.#(An#incumbent#would#prefer#the#
gets to decide whether toentrant#not#to#enter#but#cannot#back#up#its#threat#to#fight#entry.)2
fight the entrant.
#2
#2 (40,40)
Incumbent
#2
#2 Entrant Fight entry
#2
(-20,30)
#2 Do not enter
# market (0,100)
2
Suppose#the#incumbent#builds#a#large#plant#which#is#loss#making#
(The first number is theunless#it#has#a#large#share#of#the#market.##This#reduces#the#
entrant’s profit the second is the incumbents.) Again the in-
incumbentCs#payoff#to#sharing#the#market#to#20.##Then#what#does#
cumbent would like to commit to fighting the entrant, so entry did not occur. But this
the#incumbent#want#to#do?2
is a non-credible threat and the backwards induction solution is for entry to occur. One
102 a large plant which is loss making
possible solution would be for the incumbent to build
unless it has a large share of the market. This reduces the incumbent’s payoff to sharing
the market to 20. Then what does the incumbent want to do?
14 MARTIN CRIPPS

5. Subgame Perfect Equilibrium

Subgame perfect equilibrium is a way of generalising backwards induction to games that
have information sets. It has the good properties of backwards induction—every threat is
credible.
Subgame perfect equilibrium divides the game tree up into subgames (that is, parts of
the tree that can be considered separately). It requires that the players’ strategies are a
Nash equilibrium on every subgame, that is, the players behave in a credible way on all
the parts of the game tree that can be considered separately. Again a subgame perfect
equilibrium can be found by working backwards, taking the last independent game and
finding strategies that are a Nash equilibrium. Then taking the next last independent
game and so on. (Backwards induction always finds a subgame perfect equilibrium.)
Let us begin with an example of a game and a bad Nash equilibrium.
E.#Subgame#Perfection3

A#(Bad)#Nash#Equilibrium31
Fold'' Raise'

(1,1)' 2

L'' R'

1 1

a'' b' a'' b'

(2,2)' (11,1)' (0,1)' (1/2,0)'

In this game Player 1 folds because they believe they will get a payoff of 0 if they raise.
Player 2’s action genuinely has no effect on their payoffs so R is optimal for them. Hence
the situation above passes the test for a Nash equilibrium. However, if Player 2 got to
actually make a move in this game, then going R gets them 1 while going left gets them
2. It looks like Player 2 is making a threat they wouldn’t actually want to carry out. We
would like to rule this situation out by doing something like backwards induction. But
that’s not going to work in this game as we cannot work out what Player 1 would do at
the very last move—it might be optimal for 1 to do action a if she thought she was at the
left node but it might be optimal for her to do action b if she thought she was at the right
node.
Reinhard Selten suggested that we treat the events after Player 1 moves Raise as a
separate independent game, or a subgame, of the original game. He suggested that what
players do on this separate independent game must be a Nash equilibrium of that game.
In this example, this is a game where Player 2 and Player 1 simultaneously move. The
 2: GAMES IN EXTENSIVE FORM 15

strategic form for this subgame is given below.

L R
a (2,2) (0,1)
b (-1,1) ( 12 ,0)

What the players are doing at the bad NE of this game is (a, R). If you were given just
this game to analyse you would predict (a, L) because it is the unique Nash equilibrium
and can be found by iteratively eliminating strictly dominated actions. Selten said that
we should expect the players to play a NE of this game, so if we replace the (a, R) actions
with the Nash equilibrium of the subgame we get the following picture.
E.#Subgame#Perfection3

3 1
3
Fold'' Raise'
3
3
(1,1)' 2
Once3
L'' R'
we#have#solved#this3
last#game#it#is#easy3 1 1

to#figure#out3 a'' b' a'' b'

what#player#1#will3
do.3 (2,2)' (11,1)' (0,1)' (1/2,0)'

Here there is also a dashed line around the subgame we solved separately (and some text
I have been unable to crop). Once we have solved this subgame we can easily figure out
what Player 1 will do at the initial move and find a subgame perfect equilibrium of the
entire game. If Player 1 plays Fold they will get 1 but if they play Raise and the NE of
the subgame is then played, then Player 1 will get 2. Thus this SPE has player 1 Raising.
Summary: The way of finding a subgame perfect equilibrium is to find a Nash equilibrium
on the smallest subgames and then work backwards through the treat finding a Nash
equilibrium on subgames as you go.
There is a problem with subgame perfect equilibria in that they may not be unique.
This occurs because the subgames may have many Nash equilibria. To see this consider a
slightly different version of our game.
16
E.#Subgame#Perfection3
MARTIN CRIPPS

3 1
3
Fold'' Raise'
3
3
(2,2)' 2

L'' R'

1 1

a'' b' a'' b'

(3,1)' (0,0)' (0,0)' (1,3)'

The subgame of this extensive form can be represented as the strategic form below.

L R
a (3,1) (0,0)
b (0,0) (1,3)

This strategic form has 2 pure Nash equilibria at (3, 1) and (1, 3). So which one do we
choose to put in the game to find the subgame perfect equilibrium (SPE)? Well we can
choose either of these.
If we choose the (3, 1) NE E.#Subgame#Perfection3
on the subgame we get the SPE below.

3 1
3
Fold'' Raise'
3
3
(2,2)' 2

L'' R'

1 1

a'' b' a'' b'

(3,1)' (0,0)' (0,0)' (1,3)'

If we choose the (1, 3) NE on the subgame we get the SPE below.

 E.#Subgame#Perfection3
2: GAMES IN EXTENSIVE FORM 17

3 1
3
Fold'' Raise'
3
3
(2,2)' 2

L'' R'

1 1

a'' b' a'' b'

(3,1)' (0,0)' (0,0)' (1,3)'

It is clear that one can choose the NE to get very different behaviour from Player 1 at
the first move. But these are both perfectly fine SPE’s.
 3: MONOPOLY

MARTIN CRIPPS

In this section we resit some models you probably studied last year so we can warm up
for the study of more complex models of firm behaviour.

1. The Basics of Profit Maximization

We now revise the simplest possible case where there is one firm making a pricing/output
decision. A monopolist is faced by a collection of customers who have a demand curve
p = P (q). It also has a cost function (that describes the costs of making each output level)
that depends on input prices and output C(q, w). With these elements we can write down
its profit

Profit = Revenue − Costs

= qP (q) − C(q, w)

To maximize profit (and choose the optimal price-output combination) we differentiate and
set the slope of profit equal to zero.

P = qP (q) − C(q, w)
 
dP dp dC
=q + P (q) − =0
dq dq dq
Marginal Revenue − Marginal Costs =0

Notes: the formula for marginal revenue does not equal the price. This is because to sell
more a monopolist has to cut the price on all units it sells not just the additional unit.
Thus
dP
MR = q + P (q) < P (q)
dq

(The inequality is true as the first term is negative.)

Thus we have derived the first rule for profit maximisation M R = M C. This is informa-
tive about the firm’s optimal quantity (its output choice) but not the optimal price. Here
is the usual kind of picture we draw and it has pretty good information about quantities
on it, but how price is determined is unclear.
1
2 MARTIN CRIPPS
• Now%we%can%figure%out%profit%per%unit1

1 MC

Optimal Price
AC

Demand

MR
Optimal Quantity

15

(The arrow indicates the profit per unit.) As the demand moves around it is not clear how
the price will change. It is possible that an increase in demand reduces the price.

An#increase#in#demand#can#reduce#price.#/
/
/
/ New
New
/ Demand
MR
/
/
/ MC
/
/
AC
/
/
/
/
/ Old MR Old Demand
Note:#Here#the#new#demand#is#much#fla8er#and#quantity#has#gone#up./
/ 17
/

You see this in DVD’s where low demand DVD’s often are quite expensive but very popular
family films are much cheaper.
It is also possible that an increase in demand reduces the price.
 3: MONOPOLY 3
An#increase#in#demand#can#increase#price.#.
.
New
. Demand
. New
MR
.
.
.
MC
.
. AC
. Old Demand
.
. Old MR
Note:#Here#the#new#demand#is#steeper..
. 19
You see this in fashion when some new fad becomes popular and simultaneously expensive.
To understand what’s going on here we need to take the condition for profit maximisation
and reinterpret it in terms of pricing:
dp dC
q + P (q) =
dq dq
 
q dp dC
P (q) +1 =
p dq dq
" #
p dq −1

dC
P (q) ( ) +1 ==
q dp dq
Hence we get
dC
dq MC
P (q) =  −1 = 1
p dq 1+
q ( dp ) +1 Price Elasticity
It is this rule that tells us about pricing:
• As marginal cost goes up so do prices.
• If demand becomes more elastic (the elasticity moves from −3 to −4 for example)
so prices go down.
This results in the following typical properties: Steeper demand curves have higher prices
and flatter demand curves have lower prices. This is what is going on the in the previous
examples.

1.1. Monopoly Power. Pure monopoly is rare but firms do face downward sloping de-
mand curves. As one firm varies its price there will be a change in the demand the firm
experiences.
The individual firm’s demand is more responsive to price than the market demand, but
as long as demand does not entirely vanish when it raises its price the firm enjoys some
monopoly power.
4 MARTIN CRIPPS

One way of measuring monopoly power is to consider the share of the price that goes in
profit.

Price − MC
Lerner’s Measure of Monopoly =
Price

If the firm is maximising profits and we substitute in for the profit maximising price from
above into this measure we get that Lerner’s measure is equal to = −1/Elasticity of demand.
So the question of what determines a firm’s monopoly power amounts to the question:
What determines the firm’s elasticity of demand?

• The Elasticity of the market demand

– You may just be in an industry producing a product with many substitutes,
so consumers are pretty sensitive to overall price increases in the industry.
For example, the oil industry was successful in forming a cartel, but coffee
producers were not because there are substitutes to coffee.
– A firm may be in an industry with many others, but geography gives each
of these firms a strong local monopoly. This is particularly true if goods are
difficult to transport.
• The number of competitors a firm faces obviously determines how price-sensitive
its customers are. The determinants of the number of firms in an industry are the
barriers to entry these are: 1. Legal, 2. Scale, 3. Technological, 4. Strategic. 5
Strategic Behaviour of other firms.

1.2. Two Part Tariffs. This is a topic we will revisit later in this course but also should
appear in the section on Monopoly.
A two-part tariff is a common form of pricing where the price paid by the consumer for
buying Q units is P = A + pQ, where A is a fixed fee (connection charge, membership fee,
admission charge) and p is the price for units. Clearly, the average price = (A + pQ)/Q
declines in Q. Effectively this is a quantity discount, so high-usage customers pay less per
unit than low-usage consumers. On the other hand it seems to be a deterrent to consuming
small numbers of units. You see two-part pricing in: Theme Parks, Utilities, all you can
eat restaurants, razors. . .
We have talked about how one might choose a price optimally when there is no fixed
fee, but how does this decision change in the presence of the fee. To understand the best
way of setting the fee A it is easiest to think of the case where there is only one consumer
buying units of your good. In this case the area below the demand curve of the individual
and above the price represents the value the individual is getting from the good but is not
paying for. Here is the example of a demand function and a per unit price.
 3: MONOPOLY 5
Example:Telephone-service3
3
3
120
price
(cents per minute) Q=120-P Surplus = $50

charge
per minute 20

0 100 120

quantity (minutes/month)
33

The area above the price but below the demand is a triangle base 100 height 100 so it has
an area 5000 or $50. This represents the value the consumer is getting at this per unit
price that they are not paying for. This is the maximum fixed fee the seller could charge.
In all they would make 20 × 100 on per-unit sales and then 5000 in a fixed fee. The revenue
the firm gets is the entire area under the demand curve between q = 0 and q = 100. Thus
we can now address the optimal choice of a two part tariff by appreciating that in this case
the firm’s revenue is the area under the demand not just price times quantity.
Z q
Total Revenue = P (v)dv
0
Z q 
d
Marginal Revenue = P (v)dv = P (q)
dq 0
So to maximise profits you set M R = M C, which implies P (q) = M C or that price is set
to equal marginal cost—just like a perfectly competitive firm would.

1.3. Price Discrimination. In general setting only one price for a product is suboptimal
because there are usually customers who are willing to pay more. The practice of setting
several prices for the same product is called price discrimination.
Price Discrimination occurs when a firm sells the same good at different prices to different
customers. Firms do this to try to capture part of the consumer surplus; this is not a bad
thing as generally it increases supply. Here are some examples of it
• Airlines(Business/Vacation)
• Hotels(Business/Vacation)
• Drugs (Generic/Branded)
• Telephone Companies (Fixed Fee and Call charges)
• Service provides Fixed fee and per unit cost (IBM/Xerox)
• Manufacturers(Offer high initial prices and lower ones later)
• Businesses(offer student discounts)
• Manufactures (Retail/Wholesale)
6 MARTIN CRIPPS

For price discrimination to be possible it is necessary that it is possible to divide the

market up in some way and to get the two groups to pay different prices. Thus the necessary
conditions for price discrimination are:
• It is possible to identify at least two market segments with different elasticities of
demand. This can be done using personal characteristics such as age or gender as
well as geography and time. (Dry Cleaning/Haircuts)
• It is also necessary to think about how price discrimination will be enforced. In
particular what stops higher price buyers buying at the lower price? Must also
prevent low price types reselling to higher price types. They become the firm’s
competitors for the high price customers.Resale is easier to prevent in services
rather than goods so price discrimination is more common in the service sector.

1.3.1. First-Degree Price Discrimination. Here a customised price is set for each individual
customer and it may be observed in car dealers, dotcoms, telephone companies, college
tuition. In practice it is very unlikely, because firms cannot find out customers true value.
This theoretical extreme would and it maximise profit if it were feasible.
(If a firm is able to do this, then the revenue is the area under the demand curve. The
revenue obtained from selling one more unit = price. Thus, it is optimal for the firm to
produce the quantity at which M C = p.)

1.3.2. Second-Degree Price Discrimination. This describes a situation where a firm is able
to charge a price that depends on the quantity purchased (e.g. quantity discounts). Ex-
amples: Block Pricing by Power Generators, minutes on your phone, two-part tariffs.

1.3.3. Third-Degree Price Discrimination. This describes everything else. For example
where the firm divides a market into 2 broad segments and charges these segments different
prices. If a firm supplies two markets A and B with two independent plants, then to
maximise profits it wants to set the prices in the segments in the way that will maximize
its profit. That is
MCA MCB
PA = 1 , PB = 1
1+ 1+
Price ElasticityA Price ElasticityB
To see where this result comes from, suppose that the firm supplies output from one plant to
two separate markets, then there is a trade-off between supplying one market and another
because it increases costs.
Profits = Π = qA p(qA ) + p(qB )qB ?C(w, qA + qB )
Maximizing:
dΠ dΠ
0= , 0=
dqA dqB
Or
dC dC
= p(qA ) + qA p0 (qA ) = p(qB ) + qB p0 (qB )
dq dq
 3: MONOPOLY 7

Rearranging
dC dC
= pA [1 + (qA /pA )p0 (qA )] = pB [1 + (qB /pB )p0 (qB )]
dq dq
and so. . .
MC MC
= pA = pB .
C.#State#Control#of#Monopoly!"
1 + (ElasticityA)
−1 1 + (ElasticityB )−1
1.4. Regulation ofPrice#Regulation!of!a!monopoly:!"
a Monopoly. The most common form of regulation a firm faces is a
maximum price, or a•If!the!price!is!regulated!to!be!no!more!than!P(max)"
price cap. We will see that this will tend to increase the output of a
"
monopoly and decrease its price. Here is a picture of a monopoly and a potential regulated
price (P (max)). Note that this regulated price is less than the price it would charge if it
set M R = M C.
Demand
MC
P(max) AC

MR

State%Control%of%Monopoly!"
53

The regulated price means that the firm’s M R curve changes. If the firm is selling less
than the maximum it can sell at the regulated price, then it can always sell one more unit
Price%regulation%of%a%monopoly:%6
•Beyond!this!point!MR!is!as!it!was!before."
and get the regulated price for it. Thus the firms M R is flat and is equal to the regulated
•What!is!the!optimal!quantity!now?"
price, until the regulated price hits the demand curve. At this point the M R is exactly
"
what it was before. This is shown in the figure below.

Demand
MC
P(max) AC

MR
56

It should be clear from this figure that M C = M R at the vertical section of the red M R
curve. Thus the amount of output the firm has produced has increased and prices have
fallen as a result of this regulation.
 4: STATIC OLIGOPOLY

MARTIN CRIPPS

In this topic we take the tools of strategic form games (in particular Nash equilibrium
and dominance) to analyse models of industries where there are many firms who are in
competition with each other. There is no single “correct” model of these kinds of industry.
Here we will look at four models of Oligopoly, but there are others you should be aware of.

1. Cournot
To understand how Cournot’s model of oligopoly works it is best to work through an
example. Before doing this we imagine a world where firms are producing exactly the
same product (homogenous product) and the decision they must take is how much output
to produce. By producing more output they drive down the market price for themselves
and all other firms in the industry. (We say this negative effect of one firm on others in
the industry is a “negative externality”.) One interpretation of this model is it describes
competition in the long run. That is when firms are deciding what factories to build. It is
then, presumably, that they make their quantity decisions.

1.1. Example of Two-Firm Cournot Oligopoly. In the example we imagine there are
only two firms in the industry each producing exactly the same good (water or oil for
example). The demand curve for their product is P = 30 − Q, where Q is the total output
produced by all the firms and P is the market price. Thus we could write P = 30−(Q1 +Q2 ),
where Q1 and Q2 are the individual output of firm 1 and firm 2 respectively. Each of the two
firms has the same costs of production with a constant Marginal Cost of 6: C(Q1 ) = 6Q1
and C(Q2 ) = 6Q2 . (A quicker way of writing this is C(Qi ) = 6Qi for i = 1, 2.)
With this information it is possible to write down the firms’ profits (or payoffs) in this
game:
π1 (Q1 , Q2 ) = P Q1 − C(Q1 ) = (30 − (Q1 + Q2 ))Q1 − 6Q1 ,
π2 (Q1 , Q2 ) = P Q2 − C(Q2 ) = (30 − (Q1 + Q2 ))Q2 − 6Q2 .
What is different from Monopoly is that each firm’s profit depends on the output of another
firm.
To summarise the game: Each firm chooses and output level Qi ≥ 0. These choices are
made simultaneously. The payoffs these choices are given by the firms’ profits π1 and π2
in the above equation.
Our tool to analyse this situation will be to find a Nash equilibrium. And because of
the large number of actions the firms have the method of finding the Nash equilibrium will
1
2 MARTIN CRIPPS

be to find the firm’s best responses. We called this “Method 4 The Intersection of Best
Responses”.

1.1.1. Finding the Best Responses. Here are two ways of calculating Firm 1’s best response
to a given output of Firm 2. First write down firm 1’s profit.
π1 (Q1 , Q2 ) = (30 − (Q1 + Q2 ))Q1 − 6Q1
Observe that this is a quadratic function of Q1 os its maximum can be found by differen-
tiating and setting equal to zero.
∂π1
= 30 − (2Q1 + Q2 ) − 6 = 24 − Q2 − 2Q1 = 0.
∂Q1
Solving this for Q1 gives
1
24 − Q2 = 2Q1 , orQ1 = 12 − Q2 .
2
Thus firm 1’s optimal choice of output is a linear and decreasing function of firm 2’s output
choice.
A second and more Economist-y approach to this would be to say that Firm 1’s Revenue
= P Q1 = [30 − Q1 − Q2 ]Q1 . So we differentiate this to get its Marginal Revenue
dR
MR = = 30 − 2Q1 − Q2 .
dQ1
We also know that the firm’s Marginal Cost = 6 from above. So to find where the firm
maximises profit we would set MC equal to MR, that is,
1
MC = 6 = 30 − 2Q1 − Q2 = MR or Q1 = 12 − Q2 .
2
This tells us firm 1’s best output as a function of firm 2’s output. This relationship is
important and is often plotted by economists. Here’s a picture
In#pictures:-
Q2"

Q1"="12"–"0.5"Q2"

Q1"

The monopoly output is where Q =0

2
There are three points worth emphasizing about this picture
#-
(1) It usually slopes down, because as my competitor produces more, there is less
demand less for me and as a result of the9" reduction in demand it is optimal to
produce less.
 4: STATIC OLIGOPOLY 3

(2) The intercepts are informative. When Q2 = 0 and firm 1 produces its optimal
response it is alone in the market. Thus Firm 1’s optimal response to Q2 = 0 is
the monopoly level of output. In this example 12 is the output of a monopolist.
(3) The intercepts are informative (again). When Q2 is very high, Firm 1 wants to
produce zero output, because it makes zero profit. (In fact when Q = 24 this is
true.) Zero profit is actually a characteristic of Perfect Competition. Thus, in this
example Q = 24 is the perfectly competitive output.
Notice that everything is about Firm 1 is true for Firm 2. So when we try to calculate
Firm 2’s Best Response we will get a very similar kind of equation:
1
Q2 = 12 − Q1 .
2
Just the names of the outputs have been swapped.

1.1.2. Finding the Cournot/Nash Equilibrium. We now have two reaction functions of the
firms: Q1 = 12 − 12 Q2 and Q2 = 12 − 12 Q1 . If we are going to find a Nash equilibrium of
this game we are going to!Best
findresponse
where they intersect.
functions This
intersect is usually
at the Nash called a Cournot, or
Cournot-Nash equilibrium, because
equilibrium"Cournot found this solution to the model many many
years before Nash was even born. Here’s the picture of the two reaction functions:

Q2#

Cournot#(Nash)#Equilibrium#

Q2#=#12'0.5Q1#

Q1#

15#
To find the Cournot equilibrium we must substitute one reaction function into the other
1
Q1 = 12 − Q2
2
1 1
Q1 = 12 − [12 − Q1 ]
2 2
3
Q1 = 6
4
Q∗1 = 8
Performing a similar calculation for Firm 2 we get Q∗1 = Q∗2 = 8.
There is a short cut that we can use in this particular case because the firms have the
same costs. In this case we know the equilibrium will be symmetric Q∗1 = Q∗2 . So instead of
substituting firm 2’s reaction function into firm 1’s reaction function we can just substitute
4 MARTIN CRIPPS

Q2 = Q1 and get the right answer:

1
Q∗1 = 12 − Q∗2
2
1
Q∗1 = 12 − Q∗1
2
3 ∗
Q = 12
2 1
Q∗1 = 8
This method of find the Cournot equilibrium can make the algebra a lot easier, but you
must remember this only works when firms have the same costs and de-
mand!
1.2. Generalisation of this Approach to Many Firms. There is nothing particularly
difficult about extending the Cournot model to many firms. (Although drawing the pictures
gets quite hard.) Suppose now that we have an industry with N identical firms—we will give
them the names i = 1, 2, . . . , N . Firm i has to choose an output qi ≥ 0 for i = 1, 2, . . . , N .
These choices are made simultaneously.
Now we must write down the firms’ profits. Each firm has the marginal cost c, or total
costs cqi . The final ingredient we need is the demand function and we will assume that Q
units can be sold if the price satisfies Q = D − p (so the total amount demanded is always
less than D). Of course there are many firms so the price is a function of the total output
N
X X
p = D − (q1 + q2 + · · · + qN ) = D − qi = D − qi − qj .
i=1 j6=i

This writes the relationship between prices and output in three identical way. For our
purposes the last is the most useful. It says price equals D minus
P firm i’s output and the
output of all the other firms. It will be useful to write Qi = j6=i qj as the output of all
firms apart from firm i.
Now we are in a position to write down firm i’s profit.
πi (qi , Qi ) = pqi − cqi = (D − Qi − qi )qi − cqi .
To find Firm i’s best response we want to maximise this by choosing the best qi . So, as
above, we differentiate and set equal to zero.
∂πi
= D − Qi − 2qi − c = 0.
∂qi
Setting this equal to zero gives qi∗ = (N − Qi − c)/2. If we do this for each firm we will get
N equations in N unknowns:
q1∗ = (D − Q1 − c)/2, q2∗ = (D − Q2 − c)/2, ... ∗
qN = (D − QN − c)/2
This can be solved directly, but it is much easier to use the trick we talked about in the
earlier section. Here all the firms are the same and have the same costs. We would expect
that the Nash equilibrium had all the firms producing the same output q1∗ = q2∗ = · · · = qN
∗ .
 4: STATIC OLIGOPOLY 5

So let us substitute this into Firm 1’s reaction function. With this assumption, the output
of all firms apart from firm 1 can Q1 = N ∗ ∗
P
i=2 i = (N − 1)q1 can be calculated. Hence we
q
get
q1∗ = (D − Q1 − c)/2 = (D − (N − 1)q1∗ − c)/2.
This solves for q1∗ to give
D−c
qi∗ = .
N +1
Hence as N increases each firm’s output tends to 0, there are more and more firms and
each becomes vanishingly small relative to the size of the market. We would hope this
would look something like perfect competition, and we can check this by finding out how
the price behaves as N increases. The price will satisfy
X D−c Nc + D
p∗ = D − qi∗ = D − N = →c as N → ∞.
N +1 N +1
Here the price tends to Marginal Cost as the number of firms grows. Thus Cournot
competition starts to look a lot like perfect competition when there are a lot of firms.

2. Bertrand Competition
This model of competition is similar to Cournot’s in that it assumes that firms are
producing exactly the same product. However, it assumes that firms choose the price of
their product (not how much to produce). This model of price competition is a good
description of what goes on in financial markets or of the price wars that many industries
experience from time to time. One interpretation of this is that it describes what happens
in the short run, rather than the long run.
To be precise there are three assumptions in Bertrand competition:
(1) The lower-priced firm always claims the entire market. (Firms produce identical
products.)
(2) All competition is in prices.
(3) If the firms set equal prices they will share the market.
Again it is easiest to understand Bertrand competition by thinking about an example.
2.1. An Example of Bertrand Competition. There are two ferry companies (it is
more usual to consider airlines but I am quite find of boats). They serve the same route
offering identical service, the only difference between the two companies is the price they
charge. The cost per customer is 30 for both companies. There are 1000 customers who are
willing to pay up to 50 to make the crossing. Company 1 charges a price P1 and company
2 charges P2 .
First let us write down the profits of Firm 1. If it is the low price firm it gets all 1000
customers and has profit
π1 = 1000(P1 − 30), P1 < P2 .
If it has the same price as Firm 2 it gets half of 1000 customers and has profit
π1 = 500(P1 − 30), P1 = P2 .
6 MARTIN CRIPPS

If it has a higher price than Firm 2 it gets no customers and has profit
π1 = 0(P1 − 30) = 0, P1 > P2 .
B.#Nash#Equilibrium:#Bertrand#Competition8
Here is a picture of this function (remember M C = 30).

Firm#1;s#Profit8

Firm 1's Profit

Firm 1's Price

MC P2

The most important thing we notice from this is a problem. Firm 1’s profit is highest if it
just undercuts firm 2’s price. But it also wants to charge as high a price as possible. Thus
what Firm 1 would like to do is to charge the highest30# price just below firm 1’s price, but no
such price exists. Thus, in this game (with continuous prices) the best response function
of firm 1 does not exist!
The question then is: can we find a Nash equilibrium if we cannot find where the best
responses intersect? Well the answer is yes, by telling a story. Suppose, the firms set equal
prices of 50 and share the market (which is clearly a good situation). The profit of each
firm is 500(50 − 30) = 10, 000. If firm 1 cut its price to 49, then it can attract all the
customers and make profit 19, 000 = 1000(49 − 30). In response firm 2 can undercut firm 1
by asking the price 48 and attracting the whole market. This process continues until both
firms are charging the price 30 and making zero profit. At this point no firm benefits by
undercutting their rival and the firms do actually have a best response. Summary: The
Nash equilibrium is at (P1 , P2 ) = (30, 30), although there is no best response function.

2.2. General Properties of Bertrand Competition. Here is a list of points that should
be noted:
• Pure price competition drives oligopoly to look very much like perfect competition.
The Nash equilibrium of the game has the firms setting prices equal to marginal
costs. This is why we believe markets with price competition such as financial
markets may be quite efficient.
• Price competition does not expand the market in the above example—the demand
was always 1000 for low prices. As firms cut prices, one would usually expect more
customers to want to buy the good. This expansion of the size of the market will
increase the temptation of the firms to undercut their rival’s price.
• When prices are not allowed to vary continuously the problem of the non-existence
of a best response goes away (see the next section).

2.3. Extensions to the Basic Model of Bertrand Competition.

 4: STATIC OLIGOPOLY 7

2.3.1. Integer Prices. A lot of the problems with Bertrand competition came from the fact
that firms were allowed to continuously vary their prices. If we assume firms are only able
to charge whole number prices then there is no problem with finding the biggest price less
than some number. It also increases the number of Nash equilibria of the game. Here are
the firms profits for the prices 30, 31, 32, 33, in the model of ferry competition above.

P2 = 30 P2 = 31 P2 = 32 P2 = 33
P1 = 30 (0,0) (0,0) (0,0) (0,0) ...
P1 = 31 (0,0) (500,500) (1000,0) (1000,0) ...
P1 = 32 (0,0) (0,1000) (1000,1000) (2000,0) ...
P1 = 33 (0,0) (0,1000) (0,2000) (1500,1500) ...
.. .. .. .. ..
. . . . .

If you apply the underlining method to this game you will find that it has 3 Nash equilibria:
(P1 , P2 ) = (30, 30), (P1 , P2 ) = (31, 31), (P1 , P2 ) = (32, 32). This increase in the number of
Nash equilibria arises because now you really have to make a big change in your price if
you are going to undercut your rival. This big change in the price might hurt your profits
more than the increase in customers you experience from being the low-price firm.

2.3.2. Price Guarantees. If firms advertise deals like “if you find this good cheaper any-
where else we will refund twice the difference” then there is a strange effect. Suppose the
firms set prices (40, 41) where would the customers go? If they go to the firm with the
sticker price of 40, that is what they will pay. If they go to the firm with the sticker price of
41 they can claim a refund of twice the difference in prices and in fact pay only 39. Thus all
the customers would prefer to go to the higher price firm and claim a refund. The sensible
response of the firm setting the low price (40) is to raise its sticker price, so the customers
come to it and claim the refund. If such guarantees are in place we would expect prices to
increase rather than decrease. These guarantees are collusive, although they look like they
are good for the consumer.

2.3.3. Capacity Constraints. If firms are unable to fit all 1000 customers on their ferry it
never makes sense to cut prices down to Marginal Cost. Instead prices seem to go around
in circles. These circles are called Edgeworth cycles.
Consider the following slight change in our example. Suppose that only 800 customers
will fit on a ferry. So that when you are the low-price firm your profits are 800(P − 30)
and when you are the high-price firm your profits are 200(P − 30). The high price firm
always gets to serve those who cannot get on the low-price ferry. Suppose now that firm
1 sets the price of 35. Firm 2 can undercut this and set the price of (say) 34 making the
profit 800(34 − 30) = 3200. Or firm 2 can embrace being the high-price firm and charge
the price of 50 and make the profit 200(50 − 30) = 4000. Clearly the best response is to
set the price P2 = 50. Now (of course) firm 1 will respond by raising prices to P1 = 49 and
competition will drive prices back down again to 35 where the cycle begins again.
8 MARTIN CRIPPS

3. Differentiated Product Duopoly

If firms are not producing identical products, then we say their products are differen-
tiated. When products are differentiated it will not be the case that the low-price firm
gets all the market, because some consumers just prefer the product of the high-price firm.
However, if products are substitutes we would expect their prices to have an effect of the
demand for each other. Thus the firms will be playing a game in the choice of price they
make. This is what we study here. Again we will do this by working through an example.

3.1. Example of Price Competition with Differentiated Products. There are two
firms producing different goods. P1 is the price chosen by Firm 1 and P2 is the price chosen
by Firm 2. We will first describe the demand functions of each firm.
Firm 1 : Q1 = 12 − 2P1 + P2
Firm 2 : Q2 = 12 + P1 − 2P2
Notice that each firm’s demand is decreasing in its own price but increasing in its rival’s
price—they are producing substitute goods and are in competition with each other. We
will ignore costs here and suppose that all the costs are fixed costs. Fixed Costs = 20.
The first step in defining this game is to write down the Firms’ actions. These are the
prices they choose (P1 , P2 ). The next step is to write down the Firms’ payoffs or profits.
Firm 1’s Profit : = P1 Q1 − Cost = P1 (12 − 2P1 + P2 ) − 20
Firm 2’s Profit : = P2 Q2 − Cost = P2 (12 − 2P2 + P1 ) − 20

This completes the formal description of the game that is being played here.
Now we must find a Nash equilibrium of this game and again we will use the Reaction
Function method. We begin by finding the best P1 for firm 1 given it knows the price of
firm 2. To do this we maximize firm 1’s profit by differentiating and setting equal to zero.
dπ1
= 12?4P1 + P 2 = 0
dP1
This solves to give
1
P1 = 3 + P2 .
4
A similar process for firm 2 will also find its optimal choice of P2 as a function of P1 :
1
P2 = 3 + P1 .
4
Note here these reaction functions are upward sloping. As my rivals price goes up it is
optimal for me to increase my price too. (I can still undercut by raising my price and
an increase in my rivals price has a positive effect on my demand.) To find the Nash
equilibrium (P1∗ , P2∗ ) we must substitute one reaction function into the other.
1 1
P1 = 3 + (3 + P1 ).
4 4
 C.#Nash#Equilibrium:#Differentiated#Product#
4: STATIC OLIGOPOLY 9
Duopoly<
This solves to give (P1∗ , P2∗ ) = (4, 4). Here is a picture of what we have just done.
The#Firms’#Reaction#Curves#or#Optimal#Responses<
#<
#< P2#
P1 = 3 + 0.25P2
#<
#<
#<
#< P2 = 3 + 0.25P1
#<
#<
#<
#
<
< P1#
<
The#Nash#equilibrium#is#at#(P1,P2)=#(4,4).##<

47#
4. A Model of a Single-Unit Auction
All the previous models were games where a few sellers compete to sell to a set of buyers.
Now we will look at the reverse position where a few buyers compete to acquire a good
from a single seller.
First we will describe the buyers. The buyers? values for the good are written as
(v1 , v2 , . . . , vN ) where buyer I’s value, vi is in the interval 0 ≤ vi ≤ 1. We will suppose
these values are random and that the buyers only know their value but not the value of
the others. The seller does not know the values (v1 , v2 , . . . , vN ) and so has a random or
unknown demand curve.
How the (v1 , v2 , . . . , vN ) are determined has a big effect on the nature of competition
that among the buyers. Here are some different assumptions that might be made:
Independent Symmetric Private Values: vi is drawn independently from the den-
sity f (v) on [0, 1]. All buyers have the same distribution of values for the good.
Independent Private Values: vi is drawn independently from the density fi (v) on
[0, 1]. Buyers have distinct views about the good.
Here is the order of events in the auction
(1) The players observe their own values and no-one else?s.
(2) Then they submit a bid.
(3) The rules of the auction determines payoffs.
Here a strategy for a buyer is to describe how they should bid for each different value they
observe. Thus players’ strategies are bidding functions that takes the player’s value and
maps it to a bid.

bi : Observed Value → Bid

bi : [0, 1] → [0, ∞)
vi 7→ bi (vi )

Here is a picture of the bidding function

10
D:#Model#of#a#Single/Unit#Auction5
MARTIN CRIPPS

#5 #b#

bi#(.)#

bi#(vi)#

v##
0# vi## 1#

We have to find a whole function to describe a player’s equilibrium strategy. To make this
easier we will only look for certain kinds of equilibria. That is, equilibria where bids are
strictly increasing functions of values.

4.1. Equilibrium in Second-Price Auctions. The rules of the auction determine who
wins and who pays what. The easiest and simplest set of auction rules to analyse are
second-price auctions. These are auctions where the person submitting the highest bid
gets the object, but they pay a price equal to the second highest bid (not their own highest
bid). In an auction where the price paid is the second highest bid, the strategy bi (vi ) = vi
(that is submit a bid equal to your value), weakly dominates all other strategies. This is a
famous result due to Vickrey and is much used in Economics.
We now explain why this is true.
We first consider the possibility of bidding above your value. That is, overstating how
you feel about the good. Suppose you have a value vi and consider your payoffs for a bid
b0 > vi . Your payoff only depends on the highest bid from the other players call this B. We
will deal separately with the cases where: (1) the highest bid from everyone else is above
your new contemplated bid B > b0 . (In which case you always lose the auction whether you
bid truthfully or exaggerate to b0 ). (2) The highest bid from everyone else falls between
your value and your increased bid b0 > B > vi . (In which case bidding truthfully causes
you to lose the auction and get zero while bidding b0 causes you to win the auction and
pay B > vi so you get a negative payoff.) (3) The highest bid from everyone else is below
your value vi > B. (In which case you win the object whether you bid vi or b0 and the
price you pay B is independent of your bid. This gives the matrix of payoffs below. You
can see that the top row is weakly dominated by the bottom row.
B > b0 B ∈ (vi , b0 ) B < vi
0
Bid b Lose = 0 Win = vi − B < 0 Win = vi − B > 0
Bid vi Lose = 0 Lose = 0 Win = vi − B > 0
Now we consider the possibility of bidding below your value. That is, understating how
you feel about the good. Suppose you have a value vi and consider your payoffs for a bid
 4: STATIC OLIGOPOLY 11

b00 < vi . We will deal separately with the cases where: (1) the highest bid from everyone
else is above your value B > vi . (In which case you always lose the auction whether you
bid truthfully or understate your bid). (2) The highest bid from everyone else falls between
your value and your understated bid b00 < B < vi . (In which case bidding truthfully causes
you to win the auction and pay a price B < vi giving positive profit while bidding b00 causes
you to lose the auction and get nothing.) (3) The highest bid from everyone else is below
your understated bid b00 > B. (In which case you win the object whether you bid vi or
b00 and the price you pay B is independent of your bid. This gives the matrix of payoffs
below. Again you can see that the top row is weakly dominated by the bottom row.
B > vi B ∈ (b00 , vi , ) B < b00
Bid b00 Lose = 0 Lose = 0 Win = vi − B > 0
Bid vi Lose = 0 Win = vi − B > 0 Win = vi − B > 0
Hence we can conclude: it is a weakly dominating strategy to bid truthfully in a second
price auction.
 5. DYNAMIC OLIGOPOLY

MARTIN CRIPPS

1. Introduction
In the models of oligopoly we studied in the previous section we imagined that the firms
were moving simultaneously, either choosing quantities (Cournot) or prices (Bertrand) at
exactly the same time. Thus the firms were playing a game in strategic form. The right way
to study games in strategic form is to use the tools of Nash equilibrium and dominance.
So this is how we modelled competition in these games.
Now we are going to study games where there is an order to the moves. In the first
two sections we will study competition in the case where one firm moves first and then
the other firm moves. In the final section the firms will move at the same time but there
are many periods where they can make different moves. Thus we are considering games
in extensive form. In games like these we know Nash equilibrium is inadequate, as it may
have non-credible threats. So the way we model competition now will be to use either
backwards induction or subgame perfection.

2. Stackelberg Competition
Stackelberg criticised Cournot’s way of modelling quantity competition, because he be-
lieved industries always had one firm that got to dominate the competition in a particular
way. In particular he thought there was a specific order in which the firms got to make
their quantity choices, there was a leader and then other firms followed, and that this order
affected the quantity choices the firms in the industry made.
This can obviously be quite a complicated situation to describe as a game tree, if there
are many firms in an industry. But in its simplest form, with only two firms, we could
imagine that first firm 1 (the leader) chooses its quantity, then firm 2 (the follower) chooses
its quantity, then the price is determined, and finally the firms’ profits are realised. Here
is the simplest possible version of this in game-tree form.

Firm%1%

q1=3% q1=4%

Firm%2% Firm%2%

q2=3% q2=4% q2=3% q2=4%

(10,10)% (8,11)% (11,8)% (7,7)%

1
2 MARTIN CRIPPS

In this game Firm 1 chooses a quantity q1 = 3 or q1 = 4, then firm 2 sees the quantity
chosen by the leader and chooses a quantity q2 = 3 or q2 = 4 themselves. How does this
change the outcome of competition? Well it depends on what kind of equilibrium we look
for. Stackelberg said we should solve this game using backwards induction (he introduced
this idea long before game theorists took it up). Here is the backwards-induction solution
to the game.

Firm%1%

q1=3% q1=4%

Firm%2% Firm%2%

q2=3% q2=4% q2=3% q2=4%

(10,10)% (8,11)% (11,8)% (7,7)%

First look at what firm 2 (the follower) does. It takes Firm 1’s output as given and
chooses the profit-maximizing output given Firm 1’s choice. Firm 2 is choosing its best
response to every output choice of Firm 1. Thus, in the terminology of reaction functions,
we would say that whatever Firm 1 does Firm 2 is always making a choice on its reaction
function.
The consequence of this structure is the leader has an advantage, because it can produce
a lot and force the follower to cut back its own output. Thus, in contrast to the Cournot
equilibrium the firms now behave in different ways although they have the same costs and
demand.
Now we will generalise these ideas to the firms we studied in Cournot competition.

2.1. Stackelberg Competition with Continuous Quantity Choices. We will now

revisit our previous model of oligopoly where demand is P = 30 − Q1 − Q2 and firms have
the profits:
π1 = P Q1 − 6Q1 = (30 − Q1 − Q2 )Q1 − 6Q1 ;
π2 = P Q2 − 6Q2 = (30 − Q1 − Q2 )Q2 − 6Q2 .
Also recall the reaction functions of the two firms which are:
1
Q1 = 12 − Q2 ;
2
1
Q2 = 12 − Q1 .
2
To find the Stackelberg equilibrium we would like to solve this game by backwards induc-
tion, first solving for the follower’s (Firm 2’s) optimal output and then finding the leader’s
(Firm 1’s) optimal output. But drawing out the game tree will be impossible as there are
 5. DYNAMIC OLIGOPOLY 3

so many Q1 ’s to consider. How can we do backwards induction without the picture? Here’s
how. . .
2.1.1. Follower’s Behaviour. If we think about the previous example, what the follower
did was to choose the best Q2 given the observed Q1 . This is exactly what the reaction
function of player 2 does. It describes the best possible Q2 for each Q1 . So, the last stage
of the backwards induction is precisely described by the function.
1
Q2 = 12 − Q1 Final Stage of Backwards Induction
2
2.1.2. Leader’s Behaviour. The leader knows how the follower will respond to any Q1 they
choose. If they choose Q1 = 3 the follower will choose Q2 = 12 − 3/2 = 10.5, because this
is their optimal response as determined by their reaction function. How should the leader
choose Q1 then? Well it should choose Q1 to maximise its profit taking into account the
fact that Q2 is constrained to equal 12 − Q1 /2. That is, the leader solves the problem:
1
max(30 − Q1 − Q2 )Q1 − 6Q1 , subject to Q2 = 12 − Q1 .
Q1 2
This constrained optimization can be solved by substitution:
  
1
max 30 − Q1 − 12 − Q1 Q1 − 6Q1 .
Q1 2
Simplifying this gives:
  
1 1
max 30 − Q1 − 12 − Q1 Q1 − 6Q1 = max 12Q1 − Q21 .
Q1 2 Q1 2
Differentiating and setting equal to zero gives:
Q1 = 12.
We can substitute this back into the reaction function of Firm 2 to get Q2 = 6.
Let us compare this situation with that when the firms move simultaneously (under
Cournot competition). If there is Cournot competition the firms will both produce output
8 (Q1 = Q2 = 8), so the total supply is 16 and the price is 14. If there is Stackelberg
competition the outputs are Q1 = 12 and Q2 = 6, so the total supply is 18 and the price
is 12. This is quite generally true. Under Stackelberg competition the leader produces
more the follower produces less and the leader makes more profit than under Cournot
competition. Thus being a quantity leader confers considerable benefits on the leader and
harms the followers.

3. Price Leadership
Just as we can think of situations where the firms set quantities in a definite order so we
can also imagine situations where firms set prices in an order too. Let us also imagine that
Firm 1 is the price leader and Firm 2 is the price follower. Here the effects of the order are
entirely different—being a leader in pricing harms the leader and benefits the follower. In
its starkest form suppose the firms are producing identical products, then if Firm 1 goes
4 MARTIN CRIPPS

first and sets price, then Firm 2 can undercut this price and take all the customers. Here’s
a simple example of this.

Firm%1%

P1=3% P1=4%

Firm%2% Firm%2%

P2=3% P2=4% P2=3% P2=4%

(5,5)% (15,0)% (0,15)% (9,9)%

Here the second mover has and advantage and this is transparent when we consider the
backwards-induction solution.

Firm%1%

P1=3% P1=4%

Firm%2% Firm%2%

P2=3% P2=4% P2=3% P2=4%

(5,5)% (15,0)% (0,15)% (9,9)%

Now let us do price leadership in the case when the firms are not producing identical
products. Recall our model of price competition with differentiated products where as the
goods are not identical, demand does not jump to zero for the highest price firm
Firm1 Demand Q1 = 12 − 2P1 + P2 ,
Firm2 Demand Q2 = 12 + P1 − 2P2 .
This gave rise to the (Price) reaction functions
1
P1 = 3 + P2 ,
4
1
P2 = 3 + P1 .
4
and a Nash equilibrium in the simultaneous move game where P1 = P2 = 4.
3.0.3. Follower’s Behaviour. Just as in the quantity-setting game the price-follower (Firm
2) will choose their price to be an optimal response to what the leader does. Thus Firm 2
will see P1 and then choose
1
P2 = 3 + P1 .
4
 5. DYNAMIC OLIGOPOLY 5

3.0.4. Leader’s Behaviour. The leader wishes to maximise their profit by choosing the right
P1 . In the previous section we noted that the leader’s profit was equal to
π1 = P1 Q1 − 20 = 12P1 − 2P12 + P1 P2 − 20.
However,the leader is constrained because whatever P1 they chose will lead to the choice
of a P2 = 3 + 0.25P1 by the follower. Thus the leader solves the problem:
1
max 12P1 − 2P12 + P1 P2 − 20, subject to: P2 = 3 + P1 .
P1 4
This can be solved by substitution:
 
1
max 12P1 − 2P12 + P1 3 + P1 − 20 = 15P1 − 1.75P12 − 20.
P1 4
Differentiating this and setting this equal to zero to find the maximum gives
2
0 = 15 − 3.5P1 or P1 = 30/7 = 4 + .
7
Thus the price leader chooses to set a higher price than 4 (the equilibrium in the simulta-
neous move game). The price follower sets the price
1 30 1
P2 = 3 + =4+
4 7 14
and undercuts the leader. They do well at this lower price.Their demand is higher than at
4 because the leader has raised their price and they the follower are undercutting them.

4. Cooperation through Repetition

We have seen that competition on prices will typically drive an industry to low profits.
This is reflected in the strategic form of a (static/one-shot) pricing game below. In this
game it is strictly dominant for each firm to choose a low price although they would prefer
to be able to coordinate on high prices.

High Price Low Price

High Price (10,10) (5,12)
Low Price (12,5) (7,7)
Now we use this one shot game as a building block in a dynamic or repeated pricing game
that has no end at all. We suppose the pricing game above is played again and again over
infinitely many periods t=0,1,2,. . . . In each period the firms can choose a price and the
price they choose may depend upon what has happened in the past.
This repeated game structure presents a number of technical problems. First, it is
difficult to draw the game tree even if the players only have a couple of actions in each
period, because there are so many periods. Second, the players have so many strategies
it is difficult to find and check that we have a Nash Equilibrium or a Subgame Perfect
Equilibrium. Finally, we will also need to find a way of adding up the profits the firms
make over infinitely many periods of play to determine their profits for the entire game.
6 MARTIN CRIPPS

To add up the profits over many periods we will use the rate of interest r to calculate
the firm’s discounted present value of the profit it earns every period. For example, if the
firm earns 7 units of profit every period of the game we will say its payoff from playing the
entire repeated game is
7 7 7
=7+ + 2
+ + ...
(1 + r) (1 + r) (1 + r)3
(Recall that r > 0 is an interest rate.) We can use what we know about geometric series
to work out this sum.
7 7 7
Payoff = 7 + + + + ...
(1 + r) (1 + r)2 (1 + r)3
 
1 1 1
=7 1+ + + + ...
(1 + r) (1 + r)2 (1 + r)3
∞
X 1
=7
(1 + r)t
t=0
1
=7 1
1− (1+r)
1+r
=7
r
Tis repeated game has many equilibria. Now we will give an example of one equilibrium
which is particularly beneficial to the firms. When the play out the equilibrium the firms
coordinate on the high price and only threaten to play a low price if one player cheats on
this. We will assume they both play the strategy:
“Set a high price if my opponent has set the high price in every past period,
but play the low price in all other cases.”
Such a strategy promotes cooperation by making something very bad happen if a player
cheats on the cooperation. (The very bad thing is permanently low prices.) These strategies
are usually called “Grim Trigger” strategies by game theorists.
To see why it is an equilibrium first notice that if I stick to playing the high price (and
my opponent does) the previous calculation tells me I will have profits
10(1 + r)
Payoff of Not Cheating = .
r
If I cheat on high prices I get 12 for one period and then 7 forever, so in total I get
7 7 7
Payoff of Cheating = 12 + + 2
+ + ...
(1 + r) (1 + r) (1 + r)3
∞
7 X 1
= 12 +
(1 + r) (1 + r)t
t=0
7
= 12 +
r
 5. DYNAMIC OLIGOPOLY 7

For it not to be optimal to cheat we need

Payoff of Not Cheating > Payoff of Cheating
10(1 + r) 7
> 12 +
r r
10(1 + r) > 12r + 7
3 > 2r
This tells us that if r < 1.5 it is not optimal to cheat. That is, it is a Nash equilibrium
to play out the strategy. To check that this strategy is also a subgame perfect equilibrium
we also need to check that the players want to carry out the threat of always playing the
low price. This is easy to check because playing the low price is the best thing to do if you
believe your opponent will also play the low price today.
ECON2001 Lecture Notes

ECON2001 Microeconomics
Lecture Notes
Term 1: Production
Ian Preston

Technology
Suppose a firm produces a single output q with inputs z. The constraints
imposed by technology can be represented using a production function f(z) so
that it is feasible to produce q using z if q ≤ f (z). Production is technically
efficient if q = f (z) so the greatest possible output is being produced given the
inputs.
More general than this is a formulation representing a firm’s activities by
a production plan, production vector or net output vector, y, in which outputs
and inputs are not distinguished. Negative quantities in a particular production
plan correspond to goods consumed (inputs) and positive quantities to goods
produced (outputs). This allows for multiple outputs and also for certain goods
to be either inputs or outputs depending on the production plan. In this more
general setting we can represent technological feasibility with what is called a
transformation function, F (y), with feasible plans satisfying F (y) ≤ 0. For
the single output case, y = (q, −z) and F (q, −z) = q − f (z). The transforma-
tion frontier is the set of technically efficient production plans, which is to say
satisfying F (y) = 0.

Properties of production
• Possibility of shutdown: Shutdown is usually assumed possible which
means that 0 ≤ f (0) or, in the more general case, F (0) ≤ 0. If no output
can be produced without using some inputs then these are equalities.
• Monotonicity: If production is technically efficient then net output of
one good can be increased only by decreasing net output of another. Out-
put can only be increased by increasing some input. To produce the
same output after decreasing one input requires increasing another. If
there is only one output then the production functions must be increasing
∂f /∂zi ≥ 0. In the more general formulation ∂F/∂yi ≥ 0.

Input requirement sets

Suppose there is only one output. Pick a particular level q. The input
requirement set Z(q) is the set of all input vectors capable of producing q. In
other words, all z such that f (z) ≥ q (or equivalently, F (q, −z) ≤ 0). The
boundary of the input requirement set is known as an isoquant.

1 Ian Preston, 2010

ECON2001 Lecture Notes

Marginal rates of transformation

The slopes of outputs with respect to inputs, of inputs with respect to inputs
holding output constant and of outputs with respect to other outputs holding
inputs constant can all be thought of in terms of rates at which one net output
can be transformed into another while the firm remains on its transformation
frontier. These are known as marginal rates of transformation or MRT. If we
hold F (y) = 0 and differentiate with respect to any one output yi letting another
yj change to maintain technical efficiency then

∂F ∂F dyj
+ = 0
∂yi ∂yj dyi F (y)=0

dyj ∂F /∂yi
⇒ M RTij = = −
dyi F (y)=0 ∂F /∂yj

For example, if we assume only one output:

• The rate at which any input can be transformed into output is known as
its marginal product and is equal to the slope of the production function

dq ∂f
M Pi = =
dzi q=f (z) ∂zi

(It may seem pedantic to note it but it may help to avoid confusion in
later uses of the idea to note that since inputs are minus net outputs this
is minus the marginal rate of transformation between net outputs in this
case.)
• The rate at which any input has to be increased as we decrease another
holding output constant is known as the marginal rate of technical substi-
tution between the two and is equal to the slope of the isoquant

dzj ∂f /∂zi
M RT Sij = − =
dzi q=f (z) ∂f /∂zj

Convexity of input requirement sets

Convexity of the input requirement set Z(q) means that convex combinations
of input vectors z A and z B which produce the same output q will also produce
at least q. In terms of the slope of isoquants, convexity implies a decreasing
MRTS.
The rate at which the MRTS changes as you move around an isoquant is an
indication of its curvature and will prove to be important for the responsiveness
of input choices to economic incentives. This is usually measured, for any two

2 Ian Preston, 2010

ECON2001 Lecture Notes

inputs, by the elasticity of substitution, defined as the elasticity of the input

ratio zi /zj to the MRTS between the two
∂ ln (zi /zj )
σij = −
∂ ln (M RT Sij )

Homotheticity
Production is homothetic if, whenever input vectors z A and z B produce the
same output, f (z A ) = f (z B ), then scaling those input vectors up or down by
the same factor will also give input vectors which produce the same output.
In other words, for any λ > 0, f (λz A ) = f (λz B ). This means that isoquants
corresponding to higher levels of output are just magnified versions of isoquants
producing lower levels of output. It also means that marginal rates of technical
substitution are constant along rays through the origin.
Production is said to be homogeneous of degree α if f (λz) = λα f (z). Ho-
mogeneous production functions always have the property of homotheticity.

Returns to scale
Returns to scale are concerned with the feasibility of scaling up and down
production plans. (Remember that a production plan includes a specification
of outputs as well as inputs so we are now talking about scaling up inputs and
output.)
• There are decreasing returns to scale (DRS) if scaling up the input vector
results in a less than proportionate increase in output. Thus λf (z) >
f (λz) for λ > 1. It is therefore possible to scale production plans down
but not up.
If production is homogeneous then there are decreasing returns to scale if
α < 1.
If there is only one output and one input then concavity of the production
function, f 00 < 0, implies decreasing returns to scale given f (0) = 0.
• There are increasing returns to scale (IRS) if scaling up the input vector
results in a more than proportionate increase in output. Thus λf (z) <
f (λz) for λ > 1. It is therefore possible to scale production plans up but
not down.
If production is homogeneous then there are decreasing returns to scale if
α > 1.
If there is only one output and one input then convexity of the production
function, f 00 > 0, implies increasing returns to scale given f (0) = 0.
• There are constant returns to scale (CRS) if scaling up the input vector
results in exactly the same scaling up of production. Thus λf (z) = f (λz)
for any λ. This is equivalent to homogeneity of degree one of the produc-
tion function. If there is also only one input then production is linear,
q = az for some a.

3 Ian Preston, 2010

ECON2001 Lecture Notes

Some examples
• Perfect substitutes, CRS q = f (z1 , z2 ) = az1 + bz2 : MP of z1 is a
and MP of z2 is b and both are constant. The MRTS is the ratio of the
marginal products a/b and is constant. Isoquants are therefore parallel
straight lines. The elasticity of substitution is infinite. Production is
homothetic and there are constant returns to scale.
• Perfect substitutes, general q = f (z1 , z2 ) = G(az1 + bz2 ): The MRTS
is still constant at a/b, isoquants are therefore still parallel straight lines
and the elasticity of substitution is still infinite. Production is also homo-
thetic as the MRTS is constant along rays (because it is constant every-
where). Returns to scale depend however on the function G.
• Perfect complements, CRS q = f (z1 , z2 ) = min[az1 , bz2 ]: MP of z1
is a if az1 < bz2 and zero otherwise; MP of z2 is b if az1 > bz2 and zero
otherwise. Isoquants are L-shaped with the kinks lying on a ray through
the origin of slope a/b. Production is homothetic and there are constant
returns to scale.
• Cobb-Douglas: q = f (z1 , z2 ) = Az1a z2b : MP of z1 is aq/z1 and MP of
z2 is bq/z2 . The MRTS is therefore az2 /bz1 which is diminishing so input
requirement sets are convex. Since ln(z1 /z2 ) = ln(a/b) − ln M RT S the
elasticity of substitution is constant at 1. The MRTS depends only on
the input ratio z1 /z2 so production is homothetic. Since f (λz1 , λz2 ) =
α β
A (λz1 ) (λz2 ) = λa+b f (z1 , z2 ), the production function is homogeneous
of degree a + b and shows IRS if a + b > 1, DRS if a + b < 1 and CRS if
a + b = 1.
b
• Constant elasticity of substitution: q = f (z1 , z2 ) = [Az1a + Bz2a ]
with a < 1: MP of z1 is abAq 1−1/b z1a−1 and MP of z2 is abBq 1−1/b z2a−1 .
a−1
The MRTS is therefore (A/B) (z1 /z2 ) which is diminishing so input
requirement sets are convex. Since ln(z1 /z2 ) = [ln(a/b) + ln M RT S] /(a−
1) the elasticity of substitution is constant at 1/(1 − a). The MRTS
depends only on the input ratio z1 /z2 so production is homothetic. Since
f (λz1 , λz2 ) == λab f (z1 , z2 ), the production function is homogeneous of
degree ab and shows IRS if ab > 1, DRS if ab < 1 and CRS if ab = 1.
√
• A non-homothetic
√ example: q = f (z1 , z2 ) = z 1 + √ z2 The MP of z1
is 1/2 z 1 and the MP of z2 is 1. MRTS is therefore 1/2 z 1 which is not
constant along rays so production is not homothetic. Input requirement
sets are convex but the elasticity of √ substitution
√ is not constant. There is
DRS if z1 > 0since f (λz1 , λz2 ) = λ z 1 + λz2 < λf (z1 , z2 ) if λ > 1 (and
CRS if z1 = 0).
• A multiple product example: Suppose a firm uses a single input z
which it can allocate to production
√ of either √
of two goods. If it uses ζ
in producing good 1 then q1 = ζ and q2 = z − ζ. Thus q12 + q22 = z.

4 Ian Preston, 2010

ECON2001 Lecture Notes

The transformation function is F (q1 , q2 , z) = q12 + q22 − z. There is DRS

as F (λy) > 0 if F (y) = 0 and λ > 1.

Profit maximisation
Suppose there is a single output priced at p, input prices are w and the firm
regards these prices as given. This would be sensible for a small firm under
competitive conditions. Firm profits are pf (z) − w0 z and we assume that the
firm makes production decisions so as to maximise its profits.
First order conditions, assuming all inputs are used in positive but finite
quantities at the optimum, require
∂f
p − wi = 0
∂zi
for each input, or equivalently

M Pi = wi /p.

The firm uses the input until its marginal product equals its input price divided
by the price of the firm’s output. In other words, the price ratio between input
and output is equated to (minus) the marginal rate of transformation between
output and input.
Second order conditions include the requirement that ∂ 2 f /∂zi2 ≤ 0 so that
output is a concave function of each input.
If we take the conditions relating to any two inputs then we can divide
through to eliminate the product price

∂f /∂zi
= wi /wj
∂f /∂zj

which is another condition relating a marginal rate of transformation to a price

ratio
M RT Sij = wi /wj .

Supply functions and unconditional input

demand functions
The input demands which solve the profit maximisation problem can be
written as functions of output prices p and input prices w known as unconditional
input demand functions, which we denote by the vector-valued function D(p, w).
Putting these demands into the production function f (z) will give output as
a function of output and input prices which we call the firm’s supply function
and denote S(p, w).
Some basic properties of these functions follow by very simple arguments

5 Ian Preston, 2010

ECON2001 Lecture Notes

• The profit maximising input choice cannot be altered by scaling both p

and w by a common factor so both functions are homogeneous of degree
zero
D(λp, λw) = D(p, w) S(λp, λw) = S(p, w).

• Consider a comparison of two situations with different prices, say (pA , wA )

and (pB , wB ). Call the output and input decisions in the two cases (q A , z A )
and (q B , z B ). The fact that the firm is maximising profit in each case
shows that

pA q A − wA z A ≥ pA q B − wA z B
pB q A − w B z A ≤ pB q B − w B z B

By simple subtraction, letting ∆q = q A −q B , ∆p = pA −pB , ∆z = z A −z B

and ∆w = wA − wB

pA ∆q − wA ∆z ≥ 0 pB ∆q − wB ∆z ≤ 0

and therefore
∆p∆q − ∆w∆z ≥ 0.

Thus if output price goes up (∆p > 0) holding input prices constant
(∆w = 0) then output cannot fall (∆q ≥ 0). Likewise if one input price
rises (∆wi > 0) while output price stays the same (∆p = 0) as do all
other input prices (∆wj = 0, j 6= i) then demand for that input cannot
rise (∆zi ≤ 0).
Hence S(p, w) is increasing in p and Di (p, w) is decreasing in wi .

The profit function

If we evaluate the profit earned by the profit maximising choices then we
can also write maximum profit as a function of output and input prices

π(p, w) = max pf (z) − w0 z = pS(p, w) − w0 D(p, w).

z

This function is known as the profit function.

It also has some obvious properties.
• Since scaling up output and input prices by a common factor λ makes no
difference to the profit maximising choices, profit must also be scaled up
by λ. The profit function is therefore homogeneous of degree one in p and
w
π(λp, λw) = λπ(p, w).

• Increasing the price of output cannot decrease profit and increasing the
price of an input cannot increase profit.

6 Ian Preston, 2010

ECON2001 Lecture Notes

Hotelling’s Lemma
If we take π(p.w) and differentiate with respect to output price p then

∂π X  ∂f 
∂Dj (p, w)
= f (D(p, w)) + p − wj
∂p j
∂zj ∂p
= f (D(p, w)) = S(p, w)

using the first order condition that says p∂f /∂zj − wj = 0. Hence the derivative
is the supply function.
Similarly, if we differentiate with respect to an input price wi then
∂π X  ∂f 
∂Dj (p, w)
= −Di (p, w) + p − wj
∂wi j
∂zj ∂wi
= −Di (p, w)

so that this derivative is minus the conditional input demand function.

These are both examples of what is known as Hotelling’s Lemma.

Symmetry
Given these relationships

∂Di ∂2π ∂2π ∂Dj

=− = − = .
∂wj ∂wi ∂wj ∂wj ∂wi ∂wi
Thus the unconditional effect of a change in one input price wj on demand for
another input zi must be identical to the unconditional effect of a change in wi
on demand for input zj . If this is not true of the firm’s input decisions then the
firm cannot be maximising profit.
Also, by a similar argument,

∂Di ∂2π ∂2π ∂S

= − = − = −
∂p ∂wi ∂p ∂p∂wi ∂wi
so effects of output price on input demands must be of the same magnitude as
(but of opposite sign to) effects of input price on output supply.

Profit maximisation in the general case

In the more general formulation we can let r be the price vector correspond-
ing to the net output vector y, profits are simply r0 y and the firm’s problem is
to solve maxy r0 y subject to F (y) ≤ 0. The profit function π(r) defined by the
solution to this problem is homogeneous of degree one in p and ∂π/∂ri = yi
(this being the general form of Hotelling’s lemma). Hence π(r) is increasing in
the prices of goods for which net outputs are positive and decreasing in those
for which net outputs are negative.

7 Ian Preston, 2010

ECON2001 Lecture Notes

Cost minimisation
For firms producing a single output, it is useful to model firm decision making
as a two-stage process with inputs chosen to minimise costs given output q and
output q then chosen to maximise profit given what that implies for how costs
vary with q. An implication of profit maximisation is that the chosen output
level is always produced at least cost.
This is a useful perspective because firms may not be price takers in output
markets but nonetheless still act as input-price-taking cost minimisers in input
markets.
Consider, then, a firm choosing inputs z to solve

min w0 z s.t. q ≤ f (z).

z

We can solve by setting up a Lagrangean

w0 z + µ [q − f (z)]

where µ is a Lagrange multiplier associated with the constraint requiring pro-

duction of the necessary output q.
First order conditions require

µ∂f /∂zi − wi = 0

and therefore imply equality between the marginal rate of technical substitu-
tion and input price ratio, M RT Sij = wi /wj exactly as derived in the profit
maximisation problem above.
We can visualise the optimum choice as occuring at a tangency between an
isoquant and an isocost, at a point where the slope of the isocost equals the
ratio of input prices.

Conditional input demands

The cost minimising input quantities are known as the conditional input
demands, H(q, w), written as a function of output q and w rather than price p
and w since we are now taking q as given rather than chosen jointly with inputs
given p.
The properties of these functions follow from arguments somewhat similar
to those deployed above
• Since scaling up input prices by a common factor λ makes no difference to
the cost minimising choices, conditional input demands are homogeneous
of degree zero in w, H(q, λw) = H(q, w).
• Suppose that z A minimises the cost of producing q at prices wA and z B
minimises costs of producing the same output at prices wB . Then

wA z A ≤ wA z B wB z A ≥ wB z B

8 Ian Preston, 2010

ECON2001 Lecture Notes

so that, by repeated subtraction, letting ∆z = z A −z B and ∆w = wA −wB ,

∆w∆z ≤ 0.

If only one input price changes so that ∆wi > 0 and ∆wj = 0 for i 6= j,
then ∆zi ≤ 0. Conditional input demands must therefore be decreasing
in own price.

Cost function
The function giving minimum cost given q and z is known as the cost func-
tion, C(q, w) = w0 H(q, w). It is
• increasing in the price of every input which is used

• homogeneous of degree one in w.

Shephard’s Lemma
If we differentiate the cost function with respect to one of the input prices
then we get a relationship linking the derivative to the conditional input demand.

∂C X ∂Hj (p, w)
= Hi (q, w) + wj
∂wi j
∂wi
X ∂f ∂Hj (p, w)
= Hi (q, w) + µ
j
∂zj ∂wi
= Hi (q, w)

where substitution has been made

P from the first order condition µ∂f /∂zj = wj
and note has been taken that j ∂f /∂zj · ∂Hj (p, w)/∂wi = ∂q/∂wi = 0 since
q is being held fixed.
This result is known as Shephard’s lemma.
Because of this link between derivatives and demands, conditional input
demands have symmetry properties similar to unconditional input demands

∂Hi ∂2C ∂2C ∂Hj

= = = .
∂wj ∂wi ∂wj ∂wj ∂wi ∂wi

Relationship between conditional and

unconditional input demands
The two types of input demand are related by the identity

D(p, w) = H(S(p, w), w)

9 Ian Preston, 2010

ECON2001 Lecture Notes

where the unconditional demands are recognised as the conditional input de-
mands at profit maximising supply.
Differentiating, we see that
∂Di (p, w) ∂Hi (S(p, w), w) ∂Hi (S(p, w), w) ∂S(p, w)
= + ·
∂wi ∂wi ∂q ∂wi
so the unconditional effect of increased input price includes both an effect at
fixed output ∂Hi /∂wi and an effect coming through adjustment to chosen output
level ∂Hi /∂q · ∂S/∂wi .

Price responsiveness of input demands

Since cost minimisation requires M RT Sij = wi /wj , responsiveness of zi /zj
to M RT Sij around an isoquant is immediately informative about responsiveness
of zi /zj to the input price ratio wi /wj holding output constant. But this is
exactly what the elasticity of substitution measures, introduced earlier as a
measure of curvature of isoquants. Responsiveness of input choices to economic
incentives is therefore directly related to the curvature of input requirement
sets.

Homotheticity and input demands

What about responsiveness of cost-minimising input demands to changes in
output q? This is not about what happens as one moves around an isoquant
(holding q constant and changing w) but about what happens as one moves
between isoquants (changing q and holding w constant).
Homotheticity was a special property of technology under which MRTS re-
mains constant as inputs are scaled up or down in the same proportions. Hence,
under homotheticity, as we move between isoquants holding MRTS constant (be-
cause equated to unchanging input price ratios) then we keep input proportions
unchanged, expanding or contracting all input choices equiproportionately as
we move out or in along rays through the origin. Cost shares of different inputs
are accordingly independent of the scale of production.

Returns to scale and cost functions

If there are constant returns to scale (CRS) then the production function is
homogeneous, technology is homothetic and scaling up output by a factor λ will
be achieved by scaling up all inputs by the same factor λ, also therefore scaling
up costs by λ. The cost function is therefore proportional to output

C(q, w) = qκ(w).

If there are decreasing returns to scale (DRS) then it is impossible to scale

production plans up and multiplying output by λ > 1 will multiply costs by
more than λ
C(λq, w) > λC(q, w) for λ > 1.

10 Ian Preston, 2010

ECON2001 Lecture Notes

If there are increasing returns to scale (IRS) on the other hand then it is possible
to scale production up but not down and multiplying output by λ > 1 will
multiply costs by less than λ

C(λq, w) < λC(q, w) for λ > 1.

Marginal and average costs

Average cost is defined simply by C(q, w)/q. Straightforward calculus shows
that  
∂(C/q) 1 ∂C C
= −
∂q q ∂q q
so that average cost is falling if it exceeds marginal cost ∂C/∂q and rising if it is
below marginal cost. Any output at which marginal and average cost coincide
is a stationary point for average cost. It is common to draw a U-shaped average
cost curve cut from below by marginal cost at its minimum, though that is only
one of many possible configurations.
The lowest quantity q at which average cost reaches a minimum is known as
the minimum efficient scale.

Profit maximisation, again

Having defined the cost function we can use it to rewrite the problem of
output choice as
max pq − C(q, w).
q

The first order condition for this problem requires that price be equated to
marginal cost, p = ∂C/∂q. The second order condition requires ∂ 2 C/∂q 2 ≥ 0.
Because shutdown has been assumed possible, firms cannot make negative
profits. The condition that price exceed average cost p ≥ C/q can therefore be
imposed on firm decisions. Given that p = ∂C/∂q, this implies ∂C/∂q ≥ C/q
so that it is only output ranges over which average costs are rising that will be
chosen. The competitive firm’s supply curve is therefore that segment of the
marginal cost curve lying above the average cost curve.

Some examples, again

√
• Perfect substitutes, DRS q = f (z1 , z2 ) = az1 + bz2 : The firm pro-
duces using only z1 if a/w1 > b/w2 and using only z2 if a/w1 < b/w2 .
Conditional input demands are therefore
 2
q /a if a/w1 > b/w2
H1 (q, w) =
0 if a/w1 < b/w2

0 if a/w1 > b/w2
H2 (q, w) =
q 2 /b if a/w1 < b/w2

11 Ian Preston, 2010

ECON2001 Lecture Notes

The cost function is

C(q, w) = q 2 min[w1 /a, w2 /b].

Profits are therefore

π(p, w) = min pq − C(q, w) = p2 /4 min[w1 /a, w2 /b],

q

the supply function is

p
S(p, w) =
2 min[w1 /a, w2 /b]

and unconditional demands are

 2
p a/4w12 if a/w1 > b/w2
D1 (q, w) =
0 if a/w1 < b/w2

0 if a/w1 > b/w2
D2 (q, w) =
p2 b/4w22 if a/w1 < b/w2
p
• Perfect complements, DRS q = f (z1 , z2 ) = min[az1 , bz2 ]: The firm
uses z1 and z2 in the same ratio whatever input prices.

H1 (q, w) = q 2 /a H2 (q, w) = q 2 /b

The cost function is

C(q, w) = q 2 [w1 /a + w2 /b].

Profits are therefore

π(p, w) = min pq − C(q, w) = p2 /4[w1 /a + w2 /b],

q

the supply function is

p
S(p, w) =
2[w1 /a + w2 /b]

and unconditional demands are

D1 (q, w) = p2 /4a[w1 /a + w2 /b]2 D2 (q, w) = p2 /4b[w1 /a + w2 /b]2

• Cobb-Douglas: q = f (z1 , z2 ) = Az1a z2b : MRTS az2 /bz1 is equated to

w1 /w2 . Thus, by substituting into the production function
 q 1/(a+b)  bw −b/(a+b)  q 1/(a+b)  bw a/(a+b)
1 1
H1 (q, w) = H2 (q, w) =
A aw2 A aw2

12 Ian Preston, 2010

ECON2001 Lecture Notes

The cost function is

 q 1/(a+b)  w a/(a+b)  w b/(a+b)
1 2
C(q, w) = .
A a b
Assume a + b < 1 so that there is DRS. Then the profit function is
 1/(1−a−b)
Ap
π(p, w) =
(a + b)(w1 /a)a (w2 /b)b
√
• A non-homothetic example: q = f (z1 , z2 ) = z 1 + z2 : Equating the
√
MRTS to the input price ratio gives 1/2 z1 = w1 /w2 at interior solutions.
Cost-minimising input demands are therefore
2

(w2 /2w1 ) if q ≥ w2 /2w1
H1 (q, w) = 2
q if q ≤ w2 /2w1

q − w2 /2w1 if q ≥ w2 /2w1
H2 (q, w) =
0 if q ≤ w2 /2w1

The cost function is

w2 q − w22 /4w1

if q ≥ w2 /2w1
C(q, w) =
w1 q 2 if q ≤ w2 /2w1

Profits are (p − w2 ) q + w22 /4w1 if q ≥ w2 /2w1 and pq − w1 q 2 other-

wise. If p > w2 then profits are everywhere increasing in q; only if
p < w2 is there a well-defined maximum at finite q with z2 = 0 in
which case π(p, w) = p2 /4w1 . Unconditional input demands in that case
2
are D1 (p, w) = (p/2w1 ) and D2 (p, w) = 0 and the supply function is
S(p, w) = p/2w1 .
• A multiple product example: q12 +q22 = z: The firm maximises p1 q1 +
p2 q2 − w[q12 + q22 ]. Supply functions are S1 (p, w) = p1 /2w and S2 (p, w) =

2 2 2
p2 /2w and the unconditional input demand
function is D(p, w) = p1 + p2 /4w .
2 2
The profit function is π(p, w) = p1 + p2 /4w.

13 Ian Preston, 2010

 6. ADVERSE SELECTION

MARTIN CRIPPS

1. Introduction
We now study models of markets where there is uncertainty about the type of the buyers
or the type of the sellers. If a buyer knows what type she is but the seller doesn’t we are in
a situation of asymmetric information. (One person in the game/market knows more than
the other.) In models of trading that economists study, asymmetric information gives rise
to a phenomenon termed “adverse selection”. We will see that uncertainty (or asymmetric
information) has profound effects on markets. In particular uncertainty makes optimization
difficult, harms efficiency and introduces a role for signalling.

2. The Market for Lemons

We begin by, informally, studying a classic model of adverse selection due to George
Akerloff. Here we will just draw demand and supply curves and chat. In the later sections
of this topic we will make these ideas more rigourous. The message of this model is that
it may be impossible to trade some kinds of goods even though there are people who want
to buy the good and others who want to sell the good because of lack of information. In
this case the good that cannot be traded is good-quality cars. These are squeezed out of
the market place by the presence of bad cars or “lemons”.
The model assumes that there is a large number of small sellers of used cars and a large
number of small buyers (so the market is perfectly competitive). There are two types of
used cars for sale Good used cars or Bad used cars ”Lemons”. If it was clear whether a
used car was good or bad, then there would be two separate markets. There would be a
market for good used cars with an equilibrium price and quantity. And there would be a
market for bad used cars with a different equilibrium price and quantity. If there were lot
of buyers and sellers in each market they would be perfectly competitive and efficient.
Now we change this simple world, so there is asymmetric information. To be precise,
the sellers know whether their car is good or bad but the buyers don’t know the quality of
a particular car they are offered. This means there can be only one market price for used
cars. Sellers of good used cars might like to ask for a higher price for their car and claim
that their car was good. But there is no particular reason for buyers to believe this claim,
because sellers of bad cars could also make such a claim. Thus the sellers of cars (good or
bad) are forced to offer their cars for sale at the same price.
Buyers have to decide whether to buy a given car or not at the market price. This decision
is obviously dependent on what they think about the quality of cars being offered. One
can make very extreme assumptions: the buyers are very pessimistic and believe all cars
1
2 MARTIN CRIPPS

are bad, the buyers are very optimistic and believe all cars are good, or some intermediate
belief. We will assume that buyers are not completely ignorant and are not dogmatic, but
that their beliefs reflect the aggregate mix of good and bad cars being traded. That is,
in an equilibrium the buyers know the general shares of good and bad cars for sale at the
current price.
Now let us start analysing the model. First we plot the two supply curves for good and
bad used cars. The
Thesupply
supplyof good used cars
of good lies(green)
cars above theis
supply
aboveof bad
theones (less are
supply of bad cars (black),
supplied at each price).
because sellers of a good car require a higher price to be willing to part with it.

P"

Q"

If these two supply curves are summed, we get the total supply of cars at each price (red).
The total supply of good and bad cars is8" the horizontal sum of good and
Observe that bad
at low prices
cars supply. only the bad used cars are offered and as the price rises the
share of good used cars increases.

P"

Q"

Now we draw the demand for used cars9" from the buyers. This will obviously depend
upon the buyers’ beliefs about the quality of cars being offered for sale. This figure has 3
demand curves. The lowest (light pink) is the demand curve if the buyers believe all used
cars are bad. The highest demand curve is the demand if the buyers think all used cars
are good. And the intermediate demand is when the buyers think the mix is 50:50. We
cannot draw the demands for all possible beliefs, but we assume that they will smoothly
adjust downwards as the buyers become more pessimistic about quality.
 If buyers think all cars are good then demand is maximized.!
6. ADVERSE SELECTION 3

P#

Q#

Now we combine the demands and supply. Suppose first that the buyers were optimistic
12#
and believed that all the cars sold were good. Then demand equals supply at the rightmost
blue dot in the figure below. At this point it is clear that there are not all good cars being
sold but a mix of good and bad. Thus the buyers should become more pessimistic and the
demand should shift down. As the demand shifts down, less and less good cars are sold at
equilibrium. In some In the limit
cases only will
demand bad continue
cars are sold and buyers
to shift down know this.!bad cars a sold
until only
and buyers believe only bad cars are sold.

P’’’#

Notice that sellers of good cars would like to be able to convince the buyers that they
are selling a good car, but they cannot because anything
15# they say could also be said by
low-quality sellers. That is, there is no credible way of distinguishing your good car from
someone else’s bad car. As a result low quality squeezes out high quality. This is called
Adverse Selection. One interpretation of what is going on here is that buyers lower prices
to protect themselves against the risk of buying a bad used car. But by lowering prices
they encourage fewer good cars to be sold.
Adverse selection is important in financial markets, because sellers may be selling a stock
because the stock is bad and buyers may be buying because the stock is good. This leads
to Bid-Ask spreads: Sellers get offered lower prices to protect the market from those people
who are selling because the news is bad. Buyers get offered higher prices to protect the
market from those people who are buying because the stock is good.
4 MARTIN CRIPPS

2.1. Cursed Equilibrium. In the Lemons model, the buyers face a problem of adverse
selection. Another word for this problem is the winner’s curse. That is, when a buyer
succeeds in getting a car they may come to regret their success because the car is worse
than they thought it would be. This is a feature of many markets where there is less
than perfect information and competition among buyers—the winners curse is particularly
problematic in auctions.
In the above model the buyers are very smart and they make the correct guess about
the proportion of good and bad cars in the market, so they will not be disappointed on
average. One extension to the lemons model would be to allow some of the buyers to have
too optimistic beliefs about the quality of cars. That is, a share of the buyers are irrational
and believe the cars are good while the rest of the buyers correctly perceive the share of
good and bad cars. This allows us to incorporate irrationality into this world pretty easily.
This is studied in papers on cursed equilibrium

3. The Insurance Market and Adverse Selection

In this section we consider a second example of a perfectly competitive market with
asymmetric information, but in a much more rigourous model. In this model one side of
the market knows more about the product than the other: buyers know their riskiness
and sellers don’t. We will study an insurance market— not used cars—but the ideas are
the same. In this case the uniformed side of the market (sellers) raise prices to protect
themselves against the presence of bad types (high-risk buyers). But by raising prices they
reduce the share of low-risk buyers in the market and increase further the share of high-risk
buyers. The result we will get is that, although the market is competitive, the market is
inefficient, because not all trade that ought to occur does occur. This type of outcome is
called market failure.
Later we will talk about two kinds of remedies for this inefficiency: Signalling (where
the informed players are able to send credible signals about what they know). Screening
(where players’ choices indicate their type).

3.1. The Model. The model will derive a demand for insurance at each price. The price
will also determine the type of customers who demand insurance. To find an equilibrium
in this market we will then use the condition that perfectly competitive firms must make
zero profit. It is this zero-profit condition that will ultimately determine the equilibrium
in this market. The first step is to describe the buyers

3.1.1. The Buyers of Insurance. There is an infinity of buyers of insurance and the buyers
come in many different types. Each buyer’s type is described by π, which is their probability
of having an accident. We suppose their is a continuum of buyers. And the aggregate
distribution of buyers 0 ≤ π ≤ 1 is a uniform distribution on the interval [0, 1].
 B.#Market#Failure#and#Asymmetric#Info#8

Assumption#:#Suppose#that#there#is#a#continuum#of#
customers#with#types#that#are#distributed#uniformly#on#
6. ADVERSE SELECTION 5
the#interval#between#0#and#1.8
π#:=Pr(This#consumer#has#an#accident)8
8
88 1!
8
8

π!
0! 1!

So for example if we found that buyers with riskiness above π = 3/4 brought insurance,
we would have a demand for insurance that was equal to 1/4.
Now we describe the buyers’ utility for insurance. We will assume that each buyer has
a wealth w > 0, but if they have an accident they will incur some costs L. So as a result of
their accident their wealth will fall to w − L. We will also assume each buyer has a utility
of wealth U (w) = w1/2 .
Now let us consider the expected utility of a buyer that has a risk off accident π. With
probability π they will have an accident and end up with utility (w−L)1/2 . With probability
1 − π they will not have an accident and end up with utility w1/2 . Thus the buyer has
expected utility

Utility of No Insurance = π(w − L)1/2 + (1 − π)(w)1/2

Now suppose the buyer buys insurance. That is, she pays a price P > 0 but when an
accident occurs her loss gets compensated. This means when she has an accident she has
utility (w − P )1/2 , because she has to pay the price of the insurance. When she doesn’t
have an accident she also has utility (w − P )1/2 , because she brought the insurance before
she knew whether she was going to have an accident. Thus

Utility of Insurance = π(w − P )1/2 + (1 − π)(w − P )1/2 = (w − P )1/2

We can compare these utilities to figure out who is going to buy insurance. A buyer will
buy insurance if
Utility of Insurance > Utility of No Insurance
(w − P )1/2 > π (w − L)1/2 + (1 − π)(w)1/2
P 1/2 L 1/2
   
1− >π 1− + (1 − π)
w w
"  #
L 1/2 P 1/2
  
π 1− 1− >1− 1−
w w

1/2
1 − 1 − Pw
π>
L 1/2


1− 1− w
6 MARTIN CRIPPS

This inequality is very important to the subsequent analysis. First, notice that we do have
adverse selection, because this inequality says that only types with high riskiness will buy
insurance. That is, all the types with π satisfying:

1/2
1 − 1 − Pw
h(P ) = <π≤1
L 1/2


1− 1− w
strictly prefer to buy insurance. We will write the function on the left here as h(P ) so
all those people with h(P ) ≤ π ≤ 1 will buy insurance. How does the number of people
change who buy insurance when the price adjusts. Well it is not to hard to see that as P
increases so does h(P ) so fewer people buy insurance. It is also easy to see
1 − (1 − 0)1/2
h(0) = = 0,
L 1/2


1− 1− w
L 1/2


1− 1− w
h(L) = = 1.
L 1/2


1− 1− w
B.#Market#Failure#and#Asymmetric#Info#8
So if insurance were free everyone would buy it and if insurance cost as much as the loss
you were going to make no-one would buy it.
Once we Supply#is#determined#by#average#riskiness#of#the#buying#
know who is going to buy insurance we can work out the demand for insurance
at each price.customers.8
Below we have a picture of the mass of consumers who choose to buy
8
insurance at the price P .
8 h(P)!
88
Not!Buy! Buy!
8 1!
8

0! π!
1!

The total area in this figure is the size of the demand for insurance at the price P . This
area is equal to 1 × (1 − h(P )). So we now can write down the demand for insurance at
the price P

1/2 L 1/2
1 − Pw


− 1− w
Demand(P ) = 1 − h(P ) = .
L 1/2


1− 1− w
3.1.2. The Sellers of Insurance. There is one aspect of the buyers’ behaviour that will
particularly matter to the sellers and that is how risky are the customers when the current
price is P .
First let us investigate how sellers behave if they know the riskiness of their customer.
Consider a customer with known risk π. A seller will have a cost equal to zero for the
 6. ADVERSE SELECTION 7

customer if they don’t have an accident and a cost L for the customer if they do. This
makes the expected cost of the customer
Cost of Customer(π) = πL + (1 − π)0 = πL.
As the seller charges the price P , the profit from a customer of known riskiness is P − πL.
Thus, as long as P ≥ πL the firm will make a profit from selling to this customer if it
knows their riskiness.
Now suppose theB.#Market#Failure#and#Asymmetric#Info#8
price is P and the firm does know know the riskiness of their customers.
Of course, at an equilibrium they do know that only customers with riskiness above h(P )
will buy insurance. So what is the average riskiness of the customers that do buy insurance
Firms#costs#are#determined#by#average#riskiness#of#a#
at the price P ?customer.8
In our particular example this can be easily seen from the following picture.
8 Average!Riskiness!=0.5[1+h(P)]!
8
88
Not!Buy! Buy!
8 1!
8

π!
0! h(P)! 1!

The halfway point between h(P ) and 1 is (1 + h(P ))/2, so this is the average of the
customers in the blue region.1
Now we have the average riskiness of the customers that buy when the price is P , we
can work out the firm’s costs from selling insurance in this case. It is
Expected Costs = L(1 + h(P ))/2.
So now we can work out the firms expected profit from selling insurance as P − L(1 +
h(P ))/2.
We know that firms will enter the industry as long as the make strictly positive profit
from selling insurance so we could now write the supply curve in this market as:
(
∞ : P ≥ L(1 + h(P ))/2
Supply =
0 : P < L(1 + h(P ))/2

1BUT, this is the average riskiness here because the distribution is flat. In general it will not be possible
to just draw a line like this. The correct way to calculate the average riskiness for an arbitrary distribution
f (π) would be the following conditional expectation
R1
h(P )
πf (π)dπ
Average Riskiness of Buyers in[h(P ), 1] = R 1
h(P )
f (π)dπ

where f (π) = 1 is the density of the distribution of π. You can check this here by substituting in f (π) = 1
and doing the calculation.
8 MARTIN CRIPPS

This gives the supply curve as a horizontal line where P = L(1 + h(P ))/2. Where this
crosses the demand curve above gives the equilibrium in this market.
3.1.3. Equilibrium in a Perfectly Competitive Economy. To find the equilibrium price we
must find where supply equals demand. As supply will be zero or infinite if firms are not
making zero profits, to find equilibrium we find the place where firms make exactly zero
profit. That is, we must solve the equation
B.#Market#Failure#and#Asymmetric#Information8
P = L(1 + h(P ))/2.
Here is a picture ofJoining#dots#and#using#the#fact#that#h(P)#is#increasing8
the two sides of this equation
8 8 8#L[1+h(P)]/2#=#P8
P=revenue"
L"

L[1+h(0)]/2"
Expected"Costs"

B.#Market#Failure#and#Asymmetric#Information8
P"
L"

From this alone we can see that there must be an equilibrium where P = L and no-one
But#there#could#be#circumstances#in#which#some#(but#not#
buys insurance. Butall)#people#get#insurance#(L[1+h(P)]/2#=#P).8
maybe there are more equilibria and the picture looks like
P=revenue"
L"

L[1+h(0)]/2"
Expected"Costs"

P"
L"

It is quite hard to determine this. One way is to look at the slope of the function L(1 +
h(P ))/2 where P = L. this is because if the slope of this function is more than 1 when
P = L it means the picture looks like the above. (This has been set as an exercise.)
3.1.4. Equilibrium with Monopoly. We could also consider the problem of a monopolist
selling insurance. Then, if the monopolist set the price P they would sell to 1 − h(P )
customers. Each would pay the price P and on average the customers would cost the firm
L(1 + h(P ))/2. Thus the monopolist would make the profit
(1 − h(P ))(P − L(1 + h(P ))/2).
The monopolist would then choose P to maximise this expression.
 6. ADVERSE SELECTION 9

4. Two-Part Tariffs, Screening and Adverse Selection

We have now shown in two examples that adverse selection can result in extreme market
failure where only the highest possible insurance type and only the worst possible customer
gets insurance. This is a massive inefficiency, because almost nobody gets insurance (buys
a good car) although everybody wants one. There are two kinds of solution to this problem.
The first is to use signalling—this allows customers to tell insurance suppliers (credibly)
their type by showing they have good health if they are buying life insurance or a good
driving record if they are buying car insurance. The second is screening—here insurance
companies offer products that only low-risk types want to buy and get the different types
of consumer to choose different types of product. (This occurs when insurance companies
offer copays, excesses or deductibles because only low risk customers are willing to accept
this kind of policy.) Notice the difference in the order of moves. In screening, first the
firms select a list of a list of different products to offer for sale and then the different types
of customers choose one. In signalling, first the customers choose a signal then the firms
decide whether to sell to them.
We will now build a model of screening where a monopolist offers two deals call them
deal A and deal B. The deals are carefully constructed so the A-type customers prefer to
buy the A deal and the B-type customers prefer to buy the B deal. Thus the firm can get
the customers to tell it what type they are and, hopefully, increase their profits as a result.
4.1. No Adverse Selection. We will begin by describing a simple model of monopoly.
The demand the monopolist faces comes from (for now) identical customers. Each customer
has an income y. Their value from buying x units of a good is described by a function
ψ(x). We will assume ψ(.) has the following properties
dψ d2 ψ
ψ(0) = 0, > 0, < 0.
dx dx2
So ψ(.) is an increasing function (the consumer likes to have more x) but has a decreasing
slope (so the Marginal benefit of additional units of x decreases). When our consumer buys
x units of the good and pays F in total for these units they get a total value equal to
ψ(x) − F + y,
if they buy nothing they have value y.
If our monopolist chooses to set the price p for the good, each consumer will choose a
quantity to buy that maximizes their value. That is they solve the problem
Maximisex≥0 ψ(x) − px + y.
To solve this we differentiate with respect to x, set equal to zero and get the first order
condition
p = ψ 0 (x), or p = ψ 0 (x∗ (p)).
This can be interpreted as setting the marginal cost of extra units p equal to the marginal
benefit of extra units ψ 0 (x). As we assumed the function ψ 00 was decreasing this is the
picture of what’s going on here.
 C.#Two'Part#Tariffs#and#Adverse#Selection7
10 MARTIN CRIPPS

#7

ψ`(x)"=""Demand"Curve"for"good"x"

P"

x"
x*(p)"

We can also interpret ψ 0 (.) as this individual’s demand curve.

We can see from this picture that there is value that the customer is not paying for. The
monopolist could charge the customer a fixed fee F as well as a price per unit p and the
customer would still be willing to buy x∗ units. What is the biggest fixed fee the customer
is willing to pay? The customer is willing to pay any fee F that gives her greater utility
than not buying anything, that is:

ψ(x∗ ) − px∗ − F + y ≥ ψ(0) + y ⇔ ψ(x∗ ) − px∗ ≥ F.

So the customer will buy the deal x∗ at the price p and fee F if ψ(x∗ ) − px∗ ≥ F and will
not buy if ψ(x∗ ) − px∗ < F . That is, the customer either buys the quantity where marginal
benefit equals marginal cost (when the value it receives is greater than the fixed cost) or
nothing. Another way of saying this is that if the fixed fee is less than the total value the
C.#Two'Part#Tariffs#and#Adverse#Selection7
customer obtains then she is prepared to buy the good. The picture shows the optimal
fixed fee where ψ(x∗ ) − px∗ = F .
#7
The"fixed"fee"is"the"
ψ`(x)" area"below"the"
demand"curve"and"
above"the"per"unit"
F" sales."

P"

px*$

x"
x*(p)"

Now we can think about what is the price-fee combination that maximises the firm’s
profit. We suppose that the monopolist has a cost per unit of c > 0. As usual one of the
 6. ADVERSE SELECTION 11

most important things is writing down profit correctly:

Profit = Revenue − Cost
= px∗ + F − cx∗
Recall that the optimal fixed cost satisfies the condition ψ(x∗ ) − px∗ = F so if this is
substituted in we get
Profit = px∗ + ψ(x∗ ) − px∗ − cx∗
= ψ(x∗ ) − cx∗
Thus to maximise profit our monopolist wants to choose x∗ to make the above expression
as large as possible, that is, differentiating
0 = ψ 0 (x∗ ) − c
so the optimal thing for the monopolist to do is to set p = c (price equals marginal costs)
and all the profit comes from the fixed fee.
Summary: In this simple model the optimal thing for a monopolist to do is to sell the
quantity x∗ where Marginal Benefit equals Marginal Cost, ψ 0 (x∗ ) = c. Then it should make
the customer pay all their utility to them as a fixed fee and a price per unit ψ(x∗ ) = F +px∗ .
4.2. Introducing adverse Selection. Now we introduce adverse selection by introducing
two types of customers. Those who get high value from the product and those who get
a lower value from it. The way we do this is by allowing the two types to have different
value functions.
High Value = αψ(x) + y Population Share = φ;
Low Value = βψ(x) + y Population Share = 1 − φ.
C.#Two'Part#Tariffs#and#Adverse#Selection7
We assume α > β so the utility and marginal utility of the high value customers is above
that of the low value customers.
#7
The"fixed"fee"is"the"
ψ`(x)" area"below"the"
demand"curve"and"
above"the"per"unit"
F" sales."

P"

px*$

x"
x*(p)"

Adverse selection is now present in this model, because the monopolist would like to offer
to sell the good at marginal cost and take all the value. (But if it did this it would have to
offer a lower fixed fee to the low-value customers and a higher fixed fee to the high-value
12 MARTIN CRIPPS

customers.) This is impossible as all the high-value customers would then pretend to be
low-value customers.
The screening solution to the adverse selection problem. Now we will describe the menu
of deals the monopolist. The first deal (xa , Fa ) is a quantity and a payment (this payment
covers both the fixed fee and the per unit price). The monopolist would like this deal to
be bought by the high-value customers. The second deal (xb , Fb ) is a second quantity and
payment, which the monopolist would like to be bought by the low-value customers. In
total the monopolist will offer a menu of deals ((xa , Fa ); (xb , Fb )). Suppose the monopolist
were successful in getting the two types of customers to buy the appropriate deal, then it
would make the profit.
Profits = φ(Fa − cxa ) + (1 − φ)(Fb − cxb ).

4.2.1. Getting the Customers to Buy: Individual Rationality Constraints. The first thing
to notice is that the high-value customers are not willing to buy any deal. If the price was
very high they would rather not buy anything. If they are going to buy the deal (xa , Fa )
then they need to get positive from it. That is,
High-Value’s Utility from A-Deal ≥ Utility from not buying
αψ(xa ) − Fa + y ≥ y.
Or αψ(xa ) ≥ Fa . Similarly if the monopolist is going to get the low value types to buy the
B-Deal, then
Low-Value’s Utility from B-Deal ≥ Utility from not buying
βψ(xb ) − Fb + y ≥ y.
Or βψ(xb ) ≥ Fb . The result of these calculations are two constraints on the kinds of deals
the monopolist can choose to maximise its profit
αψ(xa ) ≥ Fa βψ(xb ) ≥ Fb .
These are called Individual Rationality Constraints which kind of means you cannot force
customers to buy things they don’t want to buy. They are sometimes also called partici-
pation constraints.

4.2.2. Getting the Customers to Buy the Right Deal: Incentive Compatibility Constraints.
The second thing the monopolist needs to worry about is that the high-value customers
actually choose to buy the low-value customers’ deal.(xb , Fb ) was offering the good at a
very low price the high-value types would rather buy this than (xa , Fa ). If they are going
to buy the deal (xa , Fa ) then they need to get higher utility from it than from (xb , Fb ).
That is,
High-Value’s Utility from A-Deal ≥ High-Value’s Utility from B-Deal
αψ(xa ) − Fa + y ≥ αψ(xb ) − Fb + y.
 6. ADVERSE SELECTION 13

Or αψ(xa ) − Fa ≥ αψ(xb ) − Fb . Similarly if the monopolist is going to get the low value
types to buy the B-Deal and not the A-Deal
Low-Value’s Utility from B-Deal ≥ Low-Value’s Utility from A-Deal
βψ(xb ) − Fb + y ≥ βψ(xa ) − Fa + y.
Or βψ(xb ) − Fb ≥ βψ(xa ) − Fa . The result of these calculations are two constraints on the
kinds of deals the monopolist can choose to maximise its profit
αψ(xa ) − Fa ≥ αψ(xb ) − Fb βψ(xb ) − Fb ≥ βψ(xa ) − Fa .
These are called Incentive Compatibility Constraints which kind of means you are giving
the customers the incentive to buy the things you want them to buy.

4.2.3. Reducing the Number of Constraints. Now we have found four different constraints
on the monopolist’s choice of a menu ((xa , Fa ); (xb , Fb )). Here is the list
(1) αψ(xa ) − Fa ≥ 0
(2) βψ(xb ) − Fb ≥ 0
(3) αψ(xa ) − Fa ≥ αψ(xb ) − Fb
(4) βψ(xb ) − Fb ≥ βψ(xa ) − Fa
Consider the individual rationality constraint for the high value types, (1), this is ensured
because the A’s could pretend to be low-value types and they can get positive value when
they do this. Here is the argument. First the high values don’t want the low-value deal,
this is the constraint (3)
(3)
z}|{
αψ(xa ) − Fa ≥ αψ(xb ) − Fb
Then we notice that α > β, so αψ(xb ) ≥ βψ(xb ) and
(3) α≥β
z}|{ z}|{
αψ(xa ) − Fa ≥ αψ(xb ) − Fb ≥ βψ(xb ) − Fb .
Finally we use the fact that this last expression is the low type’s utility from buying the
B-deal, which must be positive.
(3) α≥β (2)
z}|{ z}|{ z}|{
αψ(xa ) − Fa ≥ αψ(xb ) − Fb ≥ βψ(xb ) − Fb ≥ 0.
Thus by using (3) and α ≥ β and (2) we can conclude that αψ(xa ) − Fa ≥ 0 which is (1).
Thus this condition is implied by the remaining three constraints and we only need them.
Here is the list:
(5) βψ(xb ) − Fb ≥ 0,
(6) αψ(xa ) − Fa ≥ αψ(xb ) − Fb ,
(7) βψ(xb ) − Fb ≥ βψ(xa ) − Fa .
14 MARTIN CRIPPS

Now think about adding a small amount f to the payments Fa and Fb . If you did this the
new constraints would be:
βψ(xb ) − Fb − f ≥ 0,
αψ(xa ) − Fa − f ≥ αψ(xb ) − Fb − f,
βψ(xb ) − Fb − f ≥ βψ(xa ) − Fa − f.
This change would increase what the monopolist earned from all its customers. And so
would have to be good for the monopolist. The change would not be a problem for the
two last constraints, because the f affects both sides of these equally. However, it might
be a problem for the first constraint, because as f gets bigger eventually it would become
negative. Thus we can deduce that the first constraint has to be an equality when the
monopolist is maximising its profit.
βψ(xb ) − Fb = 0
+
z}|{
αψ(xa ) − Fa ≥ αψ(xb ) − Fb
+
z}|{
βψ(xb ) − Fb ≥ βψ(xa ) − Fa
Now the first constraint pins down Fb , but the monopolist would still like to make Fa as big
as possible. This will be easy in the last constraint, because it just makes the right smaller,
but will be harder for the middle constraint as the left gets smaller. We can conclude that
the monopolist will increase Fa until the middle constraint is also an equality:
βψ(xb ) − Fb = 0,
αψ(xa ) − Fa = αψ(xb ) − Fb ,
βψ(xb ) − Fb ≥ βψ(xa ) − Fa .
We have now done with manipulation of the constraints and we will consider the monop-
olist’s optimisation.

4.2.4. Writing Down the Monopolist’s Optimisation. Now we will consider how the monop-
olist will choose the menu ((xa , Fa ); (xb , Fb )) to maximise its profits. This is the problem
max φ(Fa − cxa ) + (1 − φ)(Fb − cxb ), subject to
((xa ,Fa );(xb ,Fb ))
βψ(xb ) − Fb = 0
αψ(xa ) − Fa = αψ(xb ) − Fb
βψ(xb ) − Fb ≥ βψ(xa ) − Fa
We have already simplified this problem by implementing the reduction in the constraints
we talked about in the previous section. Now let us use the 2 equality constraints to
 6. ADVERSE SELECTION 15

substitute out Fa and Fb . Substituting out Fb = βψ(xb ) we get

max φ(Fa − cxa ) + (1 − φ)(βψ(xb ) − cxb ), subject to
((xa ,Fa );(xb ,Fb ))
αψ(xa ) − Fa = αψ(xb ) − βψ(xb )
0 ≥ βψ(xa ) − Fa
Now if we take the remaining equality constraint we have Fa = αψ(xa ) + (β − α)ψ(xb ). If
this is substituted in to our optimisation we now have the problem
max φ(αψ(xa ) + (β − α)ψ(xb ) − cxa ) + (1 − φ)(βψ(xb ) − cxb ),
((xa ,Fa );(xb ,Fb ))
subject to 0 ≥ βψ(xa ) − αψ(xa ) − (β − α)ψ(xb )
Or tidying up this becomes,
max φαψ(xa ) + (β − φα)ψ(xb ) − φcxa − (1 − φ)cxb ,
((xa ,Fa );(xb ,Fb ))
subject to (α − β)ψ(xa ) ≥ (α − β)ψ(xb )
The constraint can be made even simpler, because dividing by (α − β) says ψ(xa ) ≥ ψ(xb ).
Then we know that ψ(.) is an increasing function, so the only way this constraint can hold
is if xa ≥ xb . So the problem we need to solve is
max φαψ(xa ) + (β − φα)ψ(xb ) − φcxa − (1 − φ)cxb , s.t. xa ≥ xb .
(xa ;xb )

We have got rid of all the constraints and we now have one very simple condition.

4.2.5. Solving the Monopolist’s Optimisation. We now have a simple inequality constrained
optimisation that tells us what the monopolist wants to do. Let us write down the La-
grangean for this problem
L(xa , xb ) = φαψ(xa ) + (β − φα)ψ(xb ) − φcxa − (1 − φ)cxb + λ(xa − xb ).
Finding the first order conditions.
∂L
= φαψ 0 (xa ) − φc + λ = 0,
∂xa
∂L
= (β − φα)ψ 0 (xb ) − (1 − φ)c − λ = 0.
∂xb
Now we need to worry about whether the constraint binds or not. Suppose the constraint
was binding (λ > 0) and xa = xb = x then if these two equations are added together we
get βψ 0 (xa ) = c. Then substitute this into the first equation to get
φ(α − β)ψ 0 (xa ) + λ = 0,
which is impossible, because everything on the left is positive. This contradiction tells us
the constraint cannot be binding and λ = 0.
16 MARTIN CRIPPS

Substituting the fact that the constraint is not binding into the first order conditions we
get:
β(1 − φ)
αψ 0 (xa ) = c, βψ 0 (xb ) = c > c.
β − φα
So we learn that the high value types get sold their optimal quantity and the low value
types get sold lessC.#Two'Part#Tariffs#and#Adverse#Selection7
than their optimal quantity.

#7
At"any"given"price"the"
αψ`(.)" type"α’s"will"buy"more"
and"get"more"value."
βψ`(.)"
c(1 φ)/[1 φ(α/β)]

c"

x"
xb " x a"

What fees do they pay? We substituted these out of the problem with the substitutions
Fb = βψ(xb ) and Fa = αψ(xa ) + (β − α)ψ(xb ). The first of these, Fb = βψ(xb ), says the
C.#Two'Part#Tariffs#and#Adverse#Selection7
low-value types get not extra value. The fee they pays takes all their utility.

#7
αψ`(.)"

βψ`(.)"
c(1 φ)/[1 φ(α/β)]

c"

x"
xb " x a"

But notice the high-value types don’t pay all their utility to the monopolist. They pay
less than their total utility: Fa = αψ(xa ) − (α − β)ψ(xb ). This is because they could
pretend to be low-value types. If they make this pretence they get strictly positive utility
αψ(xb ) − Fb = (α − β)ψ(xb ) > 0. So this is the amount of utility they get to keep.
 6. ADVERSE SELECTION 17

To conclude: The right way to solve the adverse selection problem for the monopolist is
to efficiently serve the high-value customers and treat the low-value types suboptimally.
 7. MORAL HAZARD AND INCENTIVES

MARTIN CRIPPS

1. Introduction
Economic relationships often have the form of a Principal who contracts with an Agent
to take certain actions that the principal cannot observe directly. In each of these cases the

Table 1. Examples of Principal-Agent Pairs

Principal Agent Hidden Action

Physician Patient exercise
Employer Employee effort
Stockholder Manager strategy
Insurer insuree caution
Bank Borrower prudence
Teacher Student study time

principal wants to control what the agent does, because this affects the principal and the
agent’s payoffs, but has limited ability to directly control the agent’s actions. These are
all examples of what economists call Moral Hazard. We think of moral hazard as arising
in situations where individuals have hidden or secret actions they can take.
For there to be an economic problem here it is necessary that three conditions hold: (1)
The Principal and Agent have different payoffs/objectives. (2) The Agent has actions it
can take. (3) The actions the Agent takes are not directly monitored by the Principal. If
any one of the three conditions fail, then there is not problem. If (1) is not true then the
Principal can rely on the Agent to take actions they both like. If (2) is not true then there
is nothing the Agent can do that will harm the Principal. And if (3) is not true then the
Principal can choose to only reward the Agent if they take the Principal’s most preferred
action.
If an incentive problem/moral hazard problem exists there will be a loss in efficiency.
The Principal will try to devise a reward/incentive scheme that will encourage the Agent
to take actions the Principal likes. These incentives impose an economic cost and harm
efficiency. We are all familiar with examples of bad incentive schemes. Here are two
examples:
• 1960’s: dentists in the UK health service had a government contract that paid
them for every cavity/filling they drilled independently of whether the tooth needed
drilling.
1
2 MARTIN CRIPPS

• Cost Plus Contracts: A Principal will reimburse a builder for allowable costs and a
provision for normal rates of profit (a cost plus contract) on a construction project.
However, even the very best incentive scheme will still impose economic costs and generate
inefficiency!

1.1. Adverse Selection or Moral Hazard? Under adverse selection there is hidden
information and under moral hazard there are hidden actions. Sometimes it is difficult
to tell the difference between moral hazard and adverse selection. Here is an example.
Suppose it was observed that Volvos go through stop signs more than other makes of car.
This has two potential explanations one is about moral hazard: “Volvo drivers believe
they?re safer and take less care when they drive as a result”. The other is about adverse
selection: “Bad drivers tend to buy Volvos.” Clearly both might be true but there are very
different implications for these two explanations of the same fact.

1.2. General Features of Incentive Schemes.

1.2.1. Proportionality. This is a common feature of incentive schemes. Managers get a

share in a company. Farmers get to keep a share of their crops. And students tend to
get grades that increase linearly with their performance on a course. These proportional
schemes tend to get used when good performance is easy to measure and closely correlated
with the Principal’s desired outcome.

1.2.2. High-Powered Incentives. This is when a small change in observable outcome has a
big change in the Principal’s decision. You see these when it is particularly important to
avoid a bad outcome: Pass-Fail Grading Schemes, Sentences for Serious Crimes, Certifica-
tion of Nuclear Power Plants are all examples of high-powered incentives.

1.2.3. Relative Performance Evaluation. When good performance by the agent is partic-
ularly difficult to measure, incentives will compare the performance of one agent with
another agent to judge good performance. Tournaments, Benchmarking and Competition
are all examples of relative performance evaluation.

1.2.4. Target and Penalty Schemes. These are schemes that award a bonus when a certain
level of performance is achieved. The bonus has the effect of discouraging further effort
and there are mild penalties for under-perfomance—these schemes provide low-powered
incentives.

1.2.5. Efficiency Wages. Firms sometimes choose to pay their workers more than the mar-
ket wage. If they paid exactly what a worker could get elsewhere, then the worker does
not care if she is fired or not. (She can immediately obtain the same wage elsewhere.) The
more a worker is paid in his current job relative to the market wage, the more costly it is
for her to get fired (and she doesn’t want to shirk). Companies want to raise wages and
perpetuate unemployment as this gives workers incentives to try hard in their current job.
 7. MORAL HAZARD AND INCENTIVES 3

1.3. Ownership: One Solution to the Incentive Problem. We’ve seen that incentive
problems arise if 3 conditions hold: 1. There is a divergence of interests. 2. There is a
need for the individuals to transact. 3. There are observation problems. One solution to
this problem is to sell the issue to the agent. Once the worker owns the company, they will
of course be willing to put in effort. Thus they should be willing to buy the company from
its owners at a price that correctly reflects its full value.
There are 2 problems with this argument. First the agent may not be able to borrow
sufficient money to acquire the company (we say the agent is credit-constrained). Second,
the company might be risky and the agent may care more about the risks of owning the
company than the current owners do.

1.4. Incentives and Risk. In a situation of moral hazard, the Principal observes an
outcome that is correlated with the action that she wants to encourage and rewards the
agent based on this outcome. Usually, this outcome has a random element in it. Sometimes
the agent puts in effort but is just unlucky and a bad outcome occurs. This randomness in
the outcomes makes the agent’s life risky because the rewards they get from the Principal
are random and usually individuals don?t like this risk.
Thus incentives introduce an inefficiency because they force Agents (workers, students
etc.) to bear risk. Although it is usually better for the Principal (firms, government etc.)
to bear the risk not the individual agents. It is actually essential for agents to bear some
risk, because if they experience no risk there are no incentives! So as economists we want
individuals to bear some risk (to give them incentives) but not too much (to impose too
great costs on them). Also, they might simply refuse to participate in schemes that are
too risky. (Of course, if the agent is risk-neutral and doesn’t care about risk this is not a
problem.)

1.5. Summary. In setting up incentives the general message is you must balance several
forces. They are (1) The increased benefit from better behaviour from agents. (2) The costs
of risk borne by agents (risk aversion). (3) How precisely you can measure performance.
(4) How much effort will increase in response to incentives anyway!
Incentive Intensity Principle: Incentives should be most intense when agents are able to
take actions to respond to them.

2. Moral Hazard in Lending: Loan Contracts and Bankruptcy

In this section we work through a very simple example of moral hazard. The Principal
will be a lender deciding whether to make a loan or not. The Agent will be a borrower who
can decide whether to put in effort or not into an investment project. We will see that the
optimal loan contract for the principal is not to make a loan unless the agent can put in
sufficient collateral. There will be no costs to risk here, both the Principal and the Agent
are risk neutral. However, the agent will be able to declare bankruptcy and this gives the
agent skewed incentives.
4 MARTIN CRIPPS

2.1. The Model. A company has £x in collateral, but needs a sum of money £1 > £x to
finance a risky investment project. The project has returns that are risky and also depends
on the effort put in by the company. If the company puts in no effort the project will pay
out
£3 probability = 1/3; £0 probability = 2/3.
If the company puts in high effort the project will pay

£3 probability = 2/3; £0 probability = 1/3.

High effort costs the company £0.8 in lost opportunities elsewhere but low effort is free.

2.2. Optimal Decision When Investment is Self Financed. If the investment is made
from the company’s own assets, the company’s profits from its two effort levels are:

Profit High Effort = (2/3)3 + (1/3)0 − 1 − 0.8 = £0.2,

Profit Low Effort = (1/3)3 + (2/3)0 − 1 = £0.

It is efficient in this case for the project to go ahead and for the company to put in high
effort.

2.3. Optimal Decision When Investment is Bank Financed. Now suppose the com-
pany borrows all the cost of the investment from the bank and must pay the £1 back from
any money it makes on the project. Then the above calculation becomes:

Profit High Effort = (2/3)(3 − 1) + (1/3)0 − 0.8 = £0.53,

Profit Low Effort = (1/3)(3 − 1) + (2/3)0 = £0.666.

In this case the company is better off putting in low effort (this save it the costs of effort).
Notice that when the project doesn’t succeed the loan is not repaid, so this doesn’t affect
the company’s profits. (This is the company declaring bankruptcy.) Essentially, what
happens when the investment is borrowed is that the company gets to gamble with the
bank’s money and would rather not put in any effort of its own. The borrower here bears
no real costs of risk.

2.4. Optimal Decision When Company Puts in Some Collateral. Suppose the
company puts in £x in collateral and borrows 1 − x from the bank. Then the profits from
high and low effort are

Profit High Effort = (2/3)(3 − (1 − x)) + (1/3)0 − 0.8,

Profit Low Effort = (1/3)(3 − (1 − x)) + (2/3)0.

Suppose he bank will only loan the funds if it knows the company will put in high effort,
because this maximises the probability of it getting its loan repaid. What value of collateral
 7. MORAL HAZARD AND INCENTIVES 5

ensures the company will put in high effort? This requires

Profit High Effort > Profit Low Effort
2/3[3 − (1 − x)] − 0.8 > 1/3[3 − (1 − x)]
6 − 2(1 − x) > 5.4 − (1 − x)
0.6 > 1 − x
x > 0.4
Thus if the bank insists that the borrower puts in collateral greater than 0.4 it can be
certain that the borrower will put in high effort into the project. The incentive for high
effort comes from the desire of the borrower not to lose its collateral. Both the bank and
the firm are bearing some of the risks.

3. A Labour Contracting Model

The previous section was a warm-up exercise. It didn’t really have anything for the
Principal to do—the bank either made a loan or it didn’t—the point was how the agent
responded to different levels of the collateral. Now we will study the optimal choice of an
incentive scheme by a principal, who takes into account how the agent responds to different
incentives. Here the principal will be a manager and the agent will be a worker.

3.1. The Model. The worker can choose either to put in low effort e = 0 or high effort
e = 1. The amount of effort the agent puts in affects how much output q gets produced.
Here is the relationship between effort and output:
e = 0, ⇒ q = 0;
(
q = 1 prob = π
e = 1, ⇒
q = 0 prob = 1 − π

Thus when output is produced (q = 1) the manager knows that the worker put in high
effort, but if no output is produced (q = 0) the manager does not know whether the worker
put in high effort but was unlucky or just put in low effort (shirked).
We will assume that e = 1 is costly for the worker and costs them c > 0, but e = 0 is
costless for the worker. We will also assume the worker can go elsewhere and get utility
U > 0, so it doesn’t have to work for the firm and cannot be forced to have arbitrarily bad
utility.
The manager is able to provide an incentive scheme for the worker. In this case she
can see what output the worker produces but not see the effort they put in. So we allow
the manager to choose an employment contract that pays two different wages (w0 , w1 ),
where w0 is the wage the worker gets paid if they produce nothing and w1 is the wage the
worker gets if they produce some output. Thus the worker has a payoff that depends on
its expected wage net of its effort costs. There are three different actions he can take and
6 MARTIN CRIPPS

three different expected wages:

(1) e=1 = πw1 + (1 − π)w0 − c

(2) e=0 = w0
(3) Outside Option =U

Now let us think about the profits of the firm (what the manager gets). These will
depend on how much output gets produced and how much wages are paid. We will assume
that output can be sold at the price of one, so profit equals 1q − w = q − w. Again the
manager’s expected profit will depend on the worker’s action:

(4) e=1 Expected Profit = π1 + (1 − π)0 − (πw1 + (1 − π)w0 )

(5) e=0 Expected Profit = −w0

This completes the description of the model, now we proceed to solve it. . .

3.2. The Worker’s Optimal Behaviour. As usual we work backwards. We first under-
stand what the worker will do for various different wage contracts (w0 , w1 ) and then use
this to understand what the right choice of contract is for the manager. We already have
figured out what the worker will get from each of his 3 actions. These are the expressions
(1),(2),(3) above. For the worker to be willing to come and work for the firm he must do
better at the firm than taking his outside option, that is, one of (1),(2) must be bigger
than (3).

max {πw1 + (1 − π)w0 − c, w0 } ≥ U Individual Rationality

So if one of: πw1 + (1 − π)w0 − c ≥ U or w0 ≥ U then the worker is better off working for
the firm than not. This is called the individual rationality constraint, or the participation
constraint, and is analogous to the IR constraint in adverse selection models.
If the worker works for the firm and prefers to provide high effort (e = 1) then his utility
from high effort (1) must be greater than his utility from low effort (2). This gives us a
second constraint

πw1 + (1 − π)w0 − c ≥ w0 Incentive Compatibility

which is called the incentive compatibility constraint. This says the wage contract gives
the worker the incentive to put in high effort.
If the manager of the firm wants the worker to work and to provide high effort then two
constraints must hold:

πw1 + (1 − π)w0 − c ≥ U IR
πw1 + (1 − π)w0 − c ≥ w0 IC

Here is a picture of these two constraints.

 7. MORAL HAZARD AND INCENTIVES 7
C.#A#Labour#Contracting#Model#4
w1#
#4 Incen.ve#compa.ble#
Contracts#

(U+c)/π
Individual#Ra.onal#
Contracts#

w1#=###π:1#(U+c)#:##π:1#(1:π)#w0##

c/π w1#=#π:1(c+w0)#:#π:1(1:π)#w0#

(U+c)/(1:π)
w0#

But if the manager wants the worker to provide low effort then the constraints are

w0 ≥ U IR
w0 ≥ πw1 + (1 − π)w0 − c IC

See if you can draw these two constraints.

3.3. The Manager’s Problem. The manager wants to maximise the profits made. We
have written profits down in the expressions (4) and (5). Suppose the manager wanted the
worker to provide high effort then she would choose a contract (w0 , w1 ) to solve the profit
maximisation problem

max π1 + (1 − π)0 − (πw1 + (1 − π)w0 )

w0 ,w1
subject to: πw1 + (1 − π)w0 − c ≥ U
πw1 + (1 − π)w0 − c ≥ w0

We have a picture of the constraints on this optimisation above. What about the manager’s
objectives. Well they want to maximise the profits which is equivalent to minimising the
expected wage they have to pay. Thus there profits increase as the expected wage decreases.
8 MARTIN CRIPPS
C.#A#Labour#Contracting#Model#4
w1#
Manager’s#indifference#curves#go#up#
#4 as#expected#wage#falls#

Constant =πw1#+(1-π)w0#

Higher#profit#and#
lower#costs#in#this#
direcAon#

w0#

If we now combine the objectives of the manager with the constraint set above we can see
that the manager would maximise her profit by choosing a wage contract on the yellow
segment of the figure below.C.#A#Labour#Contracting#Model#4
w1#
#4 There#are#a#range#of#
solu8ons#to#this#
problem.#
(U+c)/π

c/π

(U+c)/(1+π)
w0#

This line segment has the endpoints (w0 , w1 ) = (0, (U +c)/π) and (w0 , w1 ) = (U, U +(c/π)).
Any convex combination of these two points is an optimal contract. The expected wage
that is paid by the manager at the optimal contract is
U +c c
(1 − π)0 + π = (1 − π)U + π(U + = U + c.
π π
The manager thus gets a maximised profit equal to π − U − c.
If π > U + c the manager could make positive profit from this type of contract. Notice
that this is fully efficient. If the expected output π is greater than the social costs of high
effort c + U (where U is the opportunity cost of the worker coming to the firm), then it
is economically efficient for the worker to provide high effort. Thus there appear to be no
efficiency costs of incentives in this model.
 7. MORAL HAZARD AND INCENTIVES 9

The expected wage above barely satisfies the workers IR constraint. (The worker is just
indifferent between working for the firm or giving up and working elsewhere.) The situation
is different for the IC constraint. At the contract (w0 , w1 ) = (U, U +(c/π)) the worker is just
indifferent between supplying high effort and shirking. However, at all other contracts the
worker strictly prefers higher effort. For example, at the contract (w0 , w1 ) = (0, (U + c)/π)
the wage for no output is so low that the worker strictly prefers to provide high effort
rather than low effort. Notice that this wage contract is very risky. The worker only gets
paid if they give high effort. As the worker only cares about his expected wage, this is
not a problem here. But if the worker disliked risk, then they would prefer the contract
(w0 , w1 ) = (U, U + (c/π)) to the contract (w0 , w1 ) = (0, (U + c)/π).
Suppose the manager decided to have low effort from the worker, then they would get
zero output (from the low effort) and would want to minimise their wage bill. To do this
they could always pay zero (and encourage the worker to go elsewhere).

4. Contracts and Risk

The previous section found the full optimal contract for the manager, but missed a
crucial aspect of incentives—risk. Because this aspect was missing the optimal contract
was actually efficient. Now we include an agent who doesn’t like risk. This has the effect
of giving only one optimal contract—the one where the wage is least risky—and efficiency
costs.

4.1. The Model. The model is very similar to the one in the previous section again there
will be two effort levels and two output levels. Now the effort level will be called e− (low
effort) and e+ (high effort) and the output levels will be called y − (low output) and y +
(high output). When the worker puts in low effort, e− , the probability of high output
is π(e− ). And when the worker puts in high effort, e+ , the probability of high output is
π(e+ ). Of course we will assume π(e+ ) > π(e− ). In summary we have:
(
− y = y − prob = 1 − π(e− )
e=e ⇒ ;
y = y + prob = π(e− )
(
y = y − prob = 1 − π(e+ )
e = e+ ⇒ .
y = y + prob = π(e+ )
In contrast to the previous section, it is possible for the worker to produce high output
even when they put in low effort. Thus, the manager is never certain what the worker
did. The difference π(e+ ) − π(e− ) measures how effectively the manager can monitor the
worker. If π(e+ ) − π(e− ) is zero, it is impossible for the manager to ever know anything
about the worker’s behaviour or to provide incentives. If π(e+ ) − π(e− ) = 1 the manager
perfectly knows what the worker has done.
As before we will write a wage contract as a pair (w− , w+ ), where w− is the wage the
worker gets after low output is realised and w+ is the worker’s wage after high output is
achieved. The worker has a utility function u(w) and a cost of effort function c(e), so we
10 MARTIN CRIPPS

can write down the worker’s payoffs to the three different actions they can take.

e+ = π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) − c(e+ )

e− = π(e− )u(w+ ) + (1 − π(e− ))u(w− ) − c(e− )
Outside Option =U

4.2. The Constraints. We will focus on contracts where the worker provides high effort.
In this case the individual rationality constraint only requires that the worker prefers high
effort to quitting

π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) − c(e+ ) ≥ U. Individual Rationality

The IR constraint cannot be manipulated in a useful way. That is not true of the IC
constraint. If the worker is to provide high effort he must also prefer high effort to low
effort that is

π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) − c(e+ ) ≥ π(e− )u(w+ ) + (1 − π(e− ))u(w− ) − c(e− ). IC

Now we can do the following manipulations.

π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) − c(e+ ) ≥ π(e− )u(w+ ) + (1 − π(e− ))u(w− ) − c(e− )
π(e+ )[u(w+ ) − u(w− )] − c(e+ ) ≥ π(e− )[u(w+ ) − u(w− )] − c(e− )
π(e+ )[u(w+ ) − u(w− )] − c(e+ ) ≥ π(e− )[u(w+ ) − u(w− )] − c(e− )
[π(e+ ) − π(e− )][u(w+ ) − u(w− )] ≥ c(e+ ) − c(e− )

This says the increase in the probability of a high wage times in the increase in the utility
of the high wage must be greater than the increase in the cost of a high wage.

4.3. The Firm’s Objectives. If the worker puts in high effort, there is probability π(e+ )
of a high output and the firm then has to pay a high wage and there is probability 1−π(e+ )
of low output and the firm then has to pay a low wage. Thus the firm’s expected profits,
when the worker puts in high effort is

π(e+ )(y + − w+ ) + (1 − π(e+ ))(y − − w− )

So, if we assume that high effort is worth paying for, then the firm?s profit maximisation
problem boils down to maximising the above expression (by choosing wages) subject to the
two constraints in the previous section. That is

max π(e+ )(y + − w+ ) + (1 − π(e+ ))(y − − w− )

w+ ,w−

subject to: π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) ≥ U + c(e+ )

[π(e+ ) − π(e− )][u(w+ ) − u(w− )] ≥ c(e+ ) − c(e− )
 7. MORAL HAZARD AND INCENTIVES 11

4.4. Solving the Firm’s Problem. We now really need to use our Lagrangean technique
to solve the firm’s problem. It has two constraints and it wants to choose a wage contract to
maximise its expected profits. We are going to have to write down a rally big Lagrangean:
L(w+ , w− ) = π(e+ )(y + − w+ ) + (1 − π(e+ ))(y − − w− )
+ λ π(e+ )u(w+ ) + (1 − π(e+ ))u(w− ) − c(e+ ) − U


+ µ [π(e+ ) − π(e− )][u(w+ ) − u(w− )] − c(e+ ) + c(e− )



Then differentiate it with respect to w+ and w− . You can see that each of these two
variables appear in many places here. There are 3 different places where w− appears this
gives
∂L
= (1 − π(e+ ))(−1) + λ(1 − π(e+ ))u0 (w− ) + µ[π(e+ ) − π(e− )][−u0 (w− )] = 0
∂w−
If we divide this equation though by u0 (w− )(1 − π(e+ )) it becomes
−1 π(e+ ) − π(e− )
+ λ − µ =0
u0 (w− ) 1 − π(e+ )
Or
π(e+ ) − π(e− ) 1
λ−µ +
= 0 −
1 − π(e ) u (w )
Now let us differentiate the Lagrangean with respect to w+ . This gives
∂L
= π(e+ )(−1) + λπ(e+ )u0 (w+ ) + µ[π(e+ ) − π(e− )]u0 (w+ ) = 0
∂w−
Now we divide this through by π(e+ )u0 (w+ ) to get
−1 π(e+ ) − π(e− )
+ λ + µ =0
u0 (w+ ) π(e+ )
Or
π(e+ ) − π(e− ) 1
λ+µ +
= 0 +
π(e ) u (w )
We will now start to do some thinking about these conditions and what the tell us. Let
us start by writing them in a slightly different way:
1
u0 (w− ) = π(e+ )−π(e− )
λ − µ 1−π(e+ )
1
u0 (w+ ) = + −) .
λ + µ π(e π(e
)−π(e
+)

Suppose that λ > 0 and µ > 0 (we will address this issue in a little while). Then 1/(λ−. . . )
is bigger than 1/(λ+. . . ), so we know that the the marginal utility at the wage w− is greater
than the marginal utility at the wage w+ . This means that w+ > w− . The extent to which
this is true depends on the shape of the utility function and how quickly marginal utility
varies with the wages.
12 MARTIN CRIPPS

The gap between the two wages determines the strength of the incentives in this model.
What makes w+ − w− big? Making µ or π(e+ ) − π(e− ) big will strengthen incentives,
because it is this that makes the denominators in these two fractions far apart.
π(e+ ) − π(e− ) π(e+ ) − π(e− ) [π(e+ ) − π(e− )][1 − 2π(e+ )]
 
λ+µ − λ − µ = µ
π(e+ ) 1 − π(e+ ) (1 − π(e+ ))π(e+ )
But this makes perfect sense when µ is big the constraint on getting high effort is very
important so it makes sense for strong incentives to be provided. Also when π(e+ ) − π(e− )
is big the firm gets very accurate information from its output signals so strong incentives
are powerful.
Technical Issue: Now we might start to worry about the Lagrange multipliers being zero.
Suppose, λ = 0 then the first order condition for w− becomes
π(e+ ) − π(e− ) 1
−µ +
= 0 −
1 − π(e ) u (w )
but this is impossible because it says marginal utility is negative. So we can conclude the
λ > 0. Now suppose µ = 0. If this is substituted into the first order conditions we get
Then we have to have u0 (w1 1 + = w − , because marginal utility
− ) = u0 (w + ) = λ. This says w

increases in w. But it is impossible to satisfy the IC constraint if the same wage is paid
for both output levels.
 ECON2001 7. DESIGNING ECONOMIC SYSTEMS

MARTIN CRIPPS

1. An Example of a Mechanism Design Problem: Myerson-Satterthwaite

Result
We will begin by considering a simple problem. Is it possible for someone to organise a
buyer and a seller to trade efficiently? This problem is made difficult, because there are
some kinds of buyer and seller combinations that cannot beneficially trade (for example,
high-cost sellers and low-value buyers). So it is not always optimal for trade to occur.
What also makes it difficult is that the costs of the sellers and the values of the buyers are
private information and they may not be willing to reveal this to you—the organiser of the
trade. The final constraint on this problem is that agreeing to do what the organiser says
is voluntary—the organiser cannot force the buyer or the seller to take part. We will show
that it is impossible for the efficient trade to be organised in general. Thus no matter how
clever the organiser is, there has to be some gains from trade that cannot be realised in
this situation.
Here are the details of the example. There is a project under consideration. A seller
can build the project, a buyer can use it. The project gets built if the seller and buyer
can agree on terms (i.e. a price paid by the buyer to the seller). The project is either
“easy” or “hard” for the seller. If the project is easy the seller has costs C = 35 and if the
project is hard the seller has costs C = 60. These two types of seller are equally likely do
the high cost seller and the low cost seller each occur with probability 12 . (But of course
you as the organiser of trade don’t actually know which one has occurred.) The buyer also
comes in two types. The low-value and the high-value buyer. The low-value buyer places
value V = 40 on the project and the high-value buyer places value V = 65 on it. Again
the low and high-value buyer’s types are equally likely—each occurring with probability 12 .
Neither the buyer nor the seller can be forced to participate in the deal and they can get
zero by just walking away.
Suppose we look at the potential gains from trade in this world.

V=40 V=65
C=35 40-35=5 65-35=30
C=60 0 65-60=5

When C = 60 and V = 40 it is inefficient for the buyer and seller to trade, however, in all
other cases trade is efficient. The total expected increase in value that efficient trade can
1
2 MARTIN CRIPPS

generate is
1 1 1 1
(1) 10 = 0 + 5 + 5 + 30.
4 4 4 4
Thus if we could organise trade in all the cases where trade should occur, then welfare
could be increased by 10, however, we will no establish (in this example) the following
result:
Result 1 (Myerson Satterthwaite). There is no way of organising trade in this example
under which the project gets built if and only if it is efficient to do so.
1.1. Split the Difference Doesn’t Work. Suppose you tried to organise trade by getting
the buyer and seller to tell you their values and then choosing a price that would ensure
they shared the gains from trade equally. Then the prices you would choose are
V=40 V=65
C=35 P=37.5 P=50
C=60 P=62.5
(The missing entry in this table would be where no trade would occur.) Would the seller
want to tell the truth in this scheme? If the seller was low cost and truthfully announced her
type, she could expect to trade at the price P = 37.5 half the time and at the price P = 50
half the time, making an expected profit of (37.5 − 35)0.5 + (50 − 35)0.5 = 8.75. But if she
lied and said her costs were high she would trade half the time at the price 62.5 and not
trade the rest of the time. This would give her the expected profit 0.5(62.5 − 35) = 13.75.
Thus, she would be better off lying and not telling the truth and the split the difference
scheme does not work.
As an exercise check whether the high cost seller wants to tell the truth and whether
the buyer’s two types want to tell the truth here.

1.2. Myerson Satterthwaite in this Example. Now we will give an argument to show
that it is impossible to get the buyer and the seller to truthfully reveal their type in this
example. Thus it is impossible to organise trade efficiently. The reason this argument works
is that it is necessary to give the buyer and seller incentives to tell the truth, because you the
organiser needs to know when trade should occur and when it should not. These incentives
are costly. The total amount of value that can be created from trade is insufficient to
provide these incentives.
First notice that the total amount of surplus created from trade is 10, by the calculation
(1) above. Now suppose we begin by thinking about the high-cost seller. To get the high-
cost seller to trade with the high-value buyer we have to promise them a price of at least
60 half the time. But this means that the low-cost seller, by pretending to be the high-cost
seller, can get at least 25 = 60 − 35 half the time. So the very least the low-cost seller can
get from cheating on truthtelling is 0.5(60 − 35) = 12.5. But we want the low-cost seller
to tell the truth. So we must promise the low-cost seller at least 12.5 if they are to tell the
truth. Half of the time the seller is high cost (and we must give them some value) half of
 ECON2001 7. DESIGNING ECONOMIC SYSTEMS 3

the time the seller is low cost and we must give them at least 12.5. So we must expect to
give at least
1 1
0 + 12.5 = 6.25
2 2
of the surplus from trade to the seller.
Now suppose we begin by thinking about the low-value buyer. To get the low-value
buyer to trade with the low-cost seller we have to promise them a price of at most 40 half
the time. But this means that the high-value buyer, by pretending to be the low-value
buyer, can get at least 25 = 65 − 40 half the time. So the least the high-value buyer can
get from cheating on truthtelling is 0.5(65 − 40) = 12.5. But we want the high-value buyer
to tell the truth. So we must promise the high-value buyer at least 12.5 if they are to tell
the truth. Half of the time the buyer is low value (and we must give them some surplus)
half of the time the buyer is high value and we must give them at least 12.5. So we must
expect to give at least
1 1
0 + 12.5 = 6.25
2 2
of the surplus from trade to the buyer.
Combining these two calculations we have to give at least 6.25 to the seller to get
truthtelling and at least 6.25 to the buyer to get truthtelling and there is only 10 in surplus
to be shared. It is clearly impossible to get truthtelling because it requires resources of 12.5
and there is only 10 available. This result has profound implications for many economic
arguments. The idea that individual bargaining or negotiation leads to efficiency is not
true when the bargainers have private (payoff relevant) information. Arguments such as the
Coasian solution to the externality problem, or many Chicago-style analyses are vulnerable
to this issue.

2. Mechanisms and the Revelation Principle

A mechanism is a device for describing a possible economic processes where information
about individuals has to be collected and aggregated and then an economic decision must
then be taken.
A mechanism takes as its input a basic model of an economy:
(1) There are individuals in the economy: 1, 2, . . .
(2) A set of possible allocations, or social states, x ∈ X.
(3) Each individual has payoff/utility function ui (x, θi ), that depends on the social state
x chosen and on their private “type” which could be knowledge or information θi .
(4) Individual i’s type, θi , is selected from a set Θi by a known probability distribution.
In the example of the previous section there was: (1) Players: buyer and seller, (2) States:
trade or no trade and an exchange of money, (3) Types: high or low cost and high or low
value, (4) 50:50 probabilities. There are many other economic questions that fit this basic
structure.
• Should a bridge be built or not?
• Who should be sold a good?
4 MARTIN CRIPPS

• How much tax each person should pay?

• What firm should produce a good?
• How much pollution individuals should be allowed to produce and how much each
of them should be made to pay for the right to pollute?
A mechanism describes a process for deciding what allocation x to choose in this model.
It has two elements. The first is a set of signals Si for each person to send. (In the examples
above this could be: a vote, a bid, a tax return, a quote, a technology.) The second is a
policy rule γ(s1 , s2 , . . . , sN ) that takes the signals sent (one from each player) and maps
them to an allocation x ∈ X. (The the policy rule might take individuals tax returns and
map them into a tax bill and a decision on who to audit.) Other examples of mechanisms
would be: (1) A message the buyer and seller could send and a decision on whether trade
occurs and who pays what. (2) In the case of pollution the message could be how much
you are willing to reduce your pollution emissions and the rule determines what you have
to pay or be paid. (3) The message could also be how much you value the bridge and the
rule could be whether to build the bridge or not and who pays for it.
Once a mechanism (S1 , . . . , SN ; γ) is determined, the individuals in the world will find
themselves playing a game. First, they will observe their utility and their type. Then
they must decide what message si ∈ Si to send. And once all the messages are sent the
allocation γ(s1 , . . . , sN ) gets decided. Typically the messages that are sent will depend
on an individual’s type. The individuals are playing a game, because what message an
individual wishes to send will depend on the messages others send.
Summary: Choosing a mechanism is like choosing a game for people to play, the designer
want to choose the mechanism to produce an allocation that satisfies their objectives.

2.1. Equilibrium of the Mechanism Game and Implementation. The strategies in

the game determined by a given mechanism are functions: s∗i (θi ), i = 1, 2, . . . , N . These
functions tell individuals what message to send when they are type θi . This is just like the
bidding function in an auction, (in an auction a player must decide what bid to submit for
each value they might have).
First, we might think about different kinds of equilibrium of the mechanism. The pro-
file of equilibrium strategies, (s∗1 (θ1 ), . . . , s∗N (θN )), might be an equilibrium in dominant
strategies. This would clearly be very robust and give people strong incentives to send
the messages s∗i . It might just be an ordinary (Bayes) Nash equilibrium, where incentives
could be less robust. When we design a mechanism we ought to think about whether the
equilibrium is nice and robust (in dominant strategies) or perhaps less robust.
When the equilibrium of the game is played out the players send the messages (s∗1 (θ1 ), . . . , s∗N (θN ))
and the mechanism designer takes the decisions γ(s∗1 (θ1 ), . . . , s∗N (θN )). We can think of
this as a three step process. First the types of players are determined, next the signals are
sent, and finally a decision is taken

(θ1 , . . . , θN ) 7→ (s∗1 (θ1 ), . . . , s∗N (θN )) 7→ γ(s∗1 (θ1 ), . . . , s∗N (θN ))

 ECON2001 7. DESIGNING ECONOMIC SYSTEMS 5

If we look at the ends of this process, ignoring the intermediate step, we get a relationship
between the players’ types and the decision that gets taken
(θ1 , . . . , θN ) 7→ γ(s∗1 (θ1 ), . . . , s∗N (θN ))
(θ1 , . . . , θN ) 7→ Γ(θ1 , . . . , θN )
This tells us that if the players are the type (θ1 , . . . , θN ) it is possible to take the deci-
sion Γ(θ1 , . . . , θN ) ≡ γ(s∗1 (θ1 ), . . . , s∗N (θN )). This is called Implementation. That is, there
exists a mechanism that allows you to take the decision Γ(θ1 , . . . , θN ) when the types are
θ1 , . . . , θ N .
For example, one decision rule we would like to implement is the rule that obliges the
buyer and seller in to trade whenever it is efficient. We know this is impossible, by the
Myerson & Satterthwaite result. Another decision rule is to oblige the buyer and seller to
trade only if the buyer is high value and the seller is low cost. We know this is possible—just
choose an appropriate fixed price.
The main result we have on implementation is that in many situations we do not need
to give the agents complicated message spaces to get optimal behaviour. The Revelation
Principle states that if the strategies s∗i (θ) are an equilibrium in dominant strategies. It
is enough to allow the agents to just tell you their type Si = Θi . No more complicated
messages are necessary. The idea behind this is to take any arbitrary mechanism with an
equilibrium (s∗1 (θ1 ), . . . , s∗N (θN )) and add on to it an initial stage where players are asked
to report their type and if they report θi the initial stage then sends the message s∗i (θi ) to
the mechanism. This will not change the players’ incentives if (s∗1 (θ1 ), . . . , s∗N (θN )) was an
equilibrium of the original mechanism.

3. The Groves-Clark Mechanism

Suppose the set of social states is as follows. A project gets undertaken and the indi-
viduals in the society get values (v1 , v2 , . . . , vN ). Or a project does not get undertaken
and the individuals get values (0, 0, . . . , 0). The values vi are private information. The
Groves-Clark Mechanism describes a process to get individuals to truthfully reveal their
values and to get the project built if and only if it is socially efficient to do so.
The Groves-Clark Mechanism works as follows
(1) Individuals report their value for the project, say ṽi . This report does not neces-
sarily have to be truthful. P
(2) The reported values are summed i ṽi .
(3) Let C > 0 be the cost of the project. If the sum of individuals’ reports is greater
than the cost, then
X
ṽi ≥ C.
i
and the project is done. If the sum of individuals’ reports is less than the cost then
X
ṽi < C.
i
6 MARTIN CRIPPS

and the project is not done.

(4) In the final stage payments are determined. Suppose we sum the reported values
of all individuals apart
P from individual j and found that these reported values are
less than the cost: i6=j ṽi < C. But when we also add in individual j’s report the
P
values are more than the cost i ṽi > C. Such individual is called a decisive or
pivotal individual. Decisive or pivotal individuals must pay the sum
X
C− ṽi
i6=j

to the decision taker. All other individuals make no payment.

For this mechanism, the Nash equilibrium is for each individual to truthfully report their
value. In fact it is weakly dominant to truthfully report your value. Thus, the decision
taker will have the right information on whether to do the project and will take a socially
optimal decision. (However, the money raised by the mechanism will be different from the
cost of building the project.)
To see that it is optimal
P to for individual j tell the truth we will consider several cases.
A key variable U = i6=j ṽi the sum of everyone else’s reports. (You will notice that this
is very much like the proof that it is optimal to bid truthfully in a second-price auction.)

If U > C. Then the project will be built whatever individual j reports, they don’t care
what they report, so it is optimal for individual j to report truthfully.

Now suppose that U + vj > C > U . In this case reporting truthfully ṽj = vj , will result
in the project being built and individual j is pivotal. They will then have to pay C − U ,
thus reporting truthfully gets them the utility vj − (C − U ) > 0 (as vj + U > C). But if
they understate their value the project will not be built and they get value zero. Truthful
reporting is better than under reporting.

Finally C > U + vj . In this case reporting truthfully results in the project not being built
and the individual getting zero utility. Exaggerating the value and reporting ṽj > C − U
will get the project built, but because the individual is pivotal they must pay C − U which
is greater than their true value.

4. Auctions
Auctions are a very popular way of selling an object when the seller does not know
the demand function. In the classical model of monopoly a monopolist knows its demand
function and should just set an optimal price (this is called the Posted-Price Model).
When a monopolist doesn?t know demand it may make a big mistake by behaving in this
way (and not sell or sell too cheaply). Having an auction allows you to learn demand
(by observing the bids) and set a good price at the same time. Thus the monopolist is
finding a mechanism to reveal its demand and to determine its price. The idea here is that
competition amongst buyers gets them to tell you the true demand.
 ECON2001 7. DESIGNING ECONOMIC SYSTEMS 7

4.1. There are 5 different types of auction.

(1) First-Price Sealed Bid
• The buyers submit bids in secret.
• The seller awards the object to the highest bidder.
• The highest bidder pays their bid to the seller.
(2) Second-Price Sealed Bid
• The buyers submit bids in secret.
• The seller awards the object to the highest bidder.
• The highest bidder pays the second highest bid to the seller.
(3) Dutch Auction
• The auctioneer announces a very high price and slowly reduces it (using a
clock).
• The first buyer to press a button stops this clock.
• They get the object at the price on the clock.
•
(4) English Auction
• The auctioneer announces a very low price and slowly increases it in response
to bids.
• The price stops increasing when the last but one buyer drops out.
• The last buyer in gets the object at the price where the last but one buyer was
eliminated.
(5) Double Auction
• There are K sellers each with one unit to sell and many buyers each with one
unit to buy.
• Sellers and buyers simultaneously submit bids.
• The K highest bids get an object whether they come from a seller or a buyer.
• The price is determined by an average of the K and K + 1 highest bids.
• Sellers with bids above the price get to keep their objects, sellers with bids
below the price sell their objects at the price.
• Low buyers don’t get an object high buyers do.
4.2. A Model of a Single Unit Auction. We will study auctions where there is one indi-
visible unit of a good for sale, like a car, and N buyers with unknown values: (v1 , . . . , vN ).
It is these values, (v1 , . . . , vN ), that determine the demand curve. We will assume that
these values are drawn independently from a distribution on [0, 1].
The order of play in this game will be first individuals observe their own value vi and
no-one else’s. Then they submit a bid bi (vi ). The rules of the auction then determines who
gets the good and who pays what. Notice that a player’s strategy is a function that maps
their value to a bid.
There are many assumptions about the probability that generates the vector (v1 , . . . , vN )
which have technical names:
Independent Symmetric Private Values: vi is drawn independently from the den-
sity f (v) on [0, 1].
8 MARTIN CRIPPS

Independent Private Values: vi is drawn independently from the density f( v) on

[0, 1].
Correlated Private Values: The vector (v1 , . . . , vN ) is drawn from a full support
density on [0, 1] × [0, 1] × · · · × [0, 1].
Common Values: v1 = v2 = · · · = vN but players do not know what vi is. They do
have private signals on it.
4.3. Equilibrium in Second Price Auctions. In an auction where the price paid is
the second highest bid, the strategy bi (vi ) = vi (that is submit a bid equal to your value),
weakly dominates all other strategies.
To see this suppose you have a value vi and consider your payoffs for a bid b0 > vi .
Your payoff only depends on the highest bid from the other players call this B. WE have
already shown that whatever the value B it is weakly dominant to submit a bid equal to
your value.
4.4. Equilibrium in First-Price Auctions. We will now study an equilibrium of a
symmetric first-price auction. In a first-price auction a player wins when their bid is
higher than everyone else’s. This means the winner of the auction risks over bidding,
because there is a chance that others are going to bid a lot less and the price they end up
paying will not protect them against this.
It is a lot harder to find the equilibrium of a first-price auction than a second-price
auction. We will start by assuming that the players’ bids are increasing functions of their
values. In this case a player will have the highest bid if and only if they they also have the
highest value. Thus, they win the auction when they have the highest value. If I have the
value r the probability another player j has a value less than r equals F (r) = Pr(vj < r) this
is the cdf of the random variable r. As the players’ values are independent, the probability
that two players both have a value less than r is Pr(vj < r) × Pr(vj < r) = F (r)2 . The
player that all other players have a value less than r is therefore F (r)N −1 . Thus the
probability that a player with the value r wins a first price auction is
F (r)N −1 .
Now suppose that the equilibrium bidding rule of the players is b∗ (v). That is, each
player bids b∗ (v) when they have the value v. What is the player’s expected utility from
playing out the equilibrium? This can be written as
Probability wins auction × Utility from winning + Probability loses × 0
The utility from winning is the individual’s value less their bid. Putting symbols for these
expressions we get
Expected Utility from Playing the Equilibrium = F (v)N −1 [v − b∗ (v)] + 0.
The individual could also pretend to have another value, say v 0 , and make the bid, b∗ (v 0 ),
that this other type plays. This would change the probability that the individual wins and
change the price they would pay
Expected Utility from Playing v 0 = F (v 0 )N −1 [v − b∗ (v 0 )] + 0.
 ECON2001 7. DESIGNING ECONOMIC SYSTEMS 9

At an equilibrium the bid b∗ (v) is optimal for type v so pretending to be v 0 and bidding
according to this this has to be worse than bidding b∗ (v). That is
F (v)N −1 [v − b∗ (v)] ≥ F (v 0 )N −1 [v − b∗ (v 0 )] for all v 0 .
Another way of saying the same thing is to say the right hand side of the inequality above
is maximized at v 0 = v.
v0 = v Maximises F (v 0 )N −1 [v − b∗ (v 0 )].
If we differentiate this and set it equal to zero we will find the maximum
d  0 N −1 ∗ 0 0 0 N −2 ∗ 0 0 N −1 db∗ 0
F (v ) [v − b (v )] = (N − 1)f (v )F (v ) [v − b (v )] + F (v ) [− (v )]
dv 0 dv
Now setting the derivative equal to zero and v 0 = v we get the differential equation
db∗
0 = (N − 1)f (v)F (v)N −2 [v − b∗ (v)] − F (v)N −1
dv
A little re-arranging gives
db∗
F (v)N −1 + (N − 1)f (v)F (v)N −2 b∗ (v) = (N − 1)f (v)F (v)N −2 v.
dv
The left hand side of this is just the derivative of a product
d 
F (v)N −1 b∗ (v) = (N − 1)f (v)F (v)N −2 v
dv
So integrating both sides allows us to solve for the optimal bidding function
Z
N −1 ∗
F (v) b (v) = (N − 1)f (v)F (v)N −2 vdv
or
(N − 1)f (v)F (v)N −2 vdv
R
∗
b (v) = .
F (v)N −1
This expression tells us that the optimal bid in a first price auction is the bidders best
guess about the second highest bid.