Professional Documents
Culture Documents
Ali Shourideh
April 3, 2012
• Contracts?
• Applications:
– Optimal Taxation
– Financial Contracting
1
1.2 Example
• Consider an economy with two homogenous goods: apples, c a , and bananas, cb
θu(c a ) + u(cb )
• Allocations:
cia (θi , θ−i ), cib (θi , θ−i )
• Feasibility:
c1a (θ1 , θ2 ) + c2a (θ1 , θ2 ) = 2ea
c1b (θ1 , θ2 ) + c2b (θ1 , θ2 ) = 2eb
• Consumption of bananas is equated; whoever has a higher taste for apples should
consume more.
• Competitive equilibrium?
2
• Assume for now they announce their type to the planner and he/she gives them
allocations
1.3 Mechanism
• Need to figure out what is the best way to communicate
• Let’s generalize:
• What is a mechanism: ( M, g)
– M: messaging space
– g: allocation rule
• Agent i send signal mi (can only communicate with the planner); m = (m1 , · · · , m N ) ∈
M
such that
3
h i h i
E u gi mi∗ (θi ), m−
∗
i
( θ −i ) , θ i | θ i ≥ E u g i
m̂ i
, m −i
∗ ( θ −i ) , θi |θi ∀m̂i , i
• Most common outcome concept; could also consider dominant strategy equilib-
rium
In particular
h i h i
E u gi mi∗ (θi ), m−
∗
i
( θ −i ) , θ i | θ i ≥ E u g i
m i
(
∗ i θ̂ ) , m −i
∗ ( θ −i ) , θi |θi , ∀θ̂i , i
New mechanism
M = Θ, c(θ ) = g(m∗ (θ ))
Truth-telling is ‘an’ equilibrium.
• Most applied work; people usually don’t announce their ’types’; decentralization
with real world contracts – will see an example soon.
4
subject to
∑ cij (θ ) = e j , ∀θ ∈ Θ, j = 1, · · · , K
i
h i h i
E u ci mi∗ (θi ), m−
∗
i
( θ −i ) , θ i | θ i ≥ E u c i
m i
(
∗ i θ̂ ) , m −i
∗ ( θ −i ) , θi |θi , ∀θ̂i , i
Continuum of agents
• Suppose there is a continuum of agents and shocks are i.i.d. across agents: no
aggregate shocks
• feasible allocations:
∑ π ( θi ) c j ( θi ) = e j
i
u(c(θ ), θ ) ≥ u(c(θ̂ ), θ )
• Last week: two types example; marginal tax rate zero at the top, positive at the
bottom
• How general?
5
• Here continuous types; Recall problem
Z
y (θ )
max u (c (θ )) − v dH (θ )
θ
subject to
Z Z
c (θ ) dF (θ ) + G ≤ y (θ ) dF (θ )
u (c (θ )) − v (y (θ ) /θ ) ≥ u c θ̂ − v y θ̂ /θ
• One way to simplify the constraint: replace it with its first and second order con-
ditions!!
FOC:
0
0
1 0 0
u c θ̂ c θ̂ − y θ̂ v y θ̂ /θ =0
θ θ̂ =θ
or
1
u0 (c (θ )) c0 (θ ) − y0 (θ ) v0 (y (θ ) /θ ) = 0
θ
SOC: After cumbersome algebra:
y0 (θ ) ≥ 0
6
• Can write this as
Z 1/θ̂ Z 1/θ̂
h x ( x, y (θ )) dx ≥ h x x, y θ̂ dx
1/θ 1/θ
or
Z 1/θ̂ Z 1/θ̂
0
v0 xy θ̂
y (θ ) v ( xy (θ )) dx ≥ y θ̂ dx
1/θ 1/θ
Now because v is convex, it must be that y ( θ ) ≥ y θ̂ . To see this, suppose not,
y (θ ) < y θ̂ .Then
v0 ( xy (θ )) < v0 xy θ̂
• So relaxed problem
Z
y (θ )
max u (c (θ )) − v dH (θ )
θ
subject to
Z Z
c (θ ) dF (θ ) + G ≤ y (θ ) dF (θ )
1
u0 (c (θ )) c0 (θ ) − y0 (θ ) v0 (y (θ ) /θ ) = 0
θ
y0 (θ ) ≥ 0
• We call the second constraint local incentive compatibility. You can show that when
local IC holds and y0 (θ ) ≥ 0, then IC holds. For proof see Fudenberg and Tirole,
chapter 7.
• Now this problem is still ugly. A trick a la Mirrlees: this FOC is equivalent to an
Envelop condition. How? Let
U (θ ) = u (c (θ )) − v (y (θ ) /θ )
Then
U (θ ) = max u c θ̂ − v y θ̂ /θ
θ̂
7
• Envelope:
∂
U 0 (θ ) =
u c θ̂ − v y θ̂ /θ θ̂ =θ
∂θ
y θ̂ 0
= v y θ̂ /θ
θ2
θ̂ =θ
y (θ ) 0
= v (y (θ ) /θ )
θ2
Rewrite the problem with this
Z
max U (θ ) dH (θ )
subject to
y (θ )
u (c (θ )) − v = U (θ )
θ
Z Z
c (θ ) dF (θ ) + G ≤ y (θ ) dF (θ )
y (θ ) 0
U 0 (θ ) = v (y (θ ) /θ )
θ2
y0 (θ ) ≥ 0
• Common technique in Mechanism Design problem: Let’s for now ignore the last
constraint. Later we have to check if it is satisfied. If violated, it means that the
original allocation has bunching of types; people have the same allocations. There
is a procedure called ironing; see Myerson’s ’Optimal Auction Design’.
• Looks like a standard Hamiltonian. We can use the standard techniques. If you
wanna learn more about calculus of variation, look at Luenberger, ’Optimization
by Vector Space Methods’. Let me quickly go over it; we can write the Lagrangian
y (θ )
Z Z
L = U (θ ) h (θ ) dθ + γ (θ ) u (c (θ )) − v − U (θ ) dθ
θ
Z Z
+λ y (θ ) f (θ ) dθ − c (θ ) f (θ ) dθ − G
y (θ ) 0
Z
0
+ µ (θ ) U (θ ) − 2 v (y (θ ) /θ ) dθ
θ
If U 0 wasn’t there, we could simply take ‘derivative’!! So let’s try and get rid of it!!
We can use integration by part to do just that
8
Z θ̄ Z
0
µ(θ )U (θ )dθ = µ (θ ) dU (θ )
θ
Z
µ0 (θ ) U (θ ) dθ
= µ θ̄ U θ̄ − µ (θ ) U (θ ) −
h − γ − µ0 = 0
γu0 (c) − λ f = 0
1 0 y 1 0 y y 00 y
−γ v + λf − µ 2v + 3v = 0
θ θ θ θ θ θ
µ θ̄ = µ (θ ) = 0
• Then
Z θ̄
λf 1 0 y 1 0 y y 00 y 1
λf − 0 v = 2v + 3v λ f − h dθ̂
u (c) θ θ θ θ θ θ θ u0 (c)
divide by 1θ v0 (y/θ ),
"Z θ̄ #
y/θv00 (y/θ )
dF θ̂ 1−H
1 1 1 1
− = 1+ −
1/θv0 (y/θ ) u0 (c) v0 (y/θ ) θ u0 c θ̂ 1−F λ (1 − F )
θf
"Z θ̄ #
y/θv00 (y/θ )
1−F dF θ̂ 1−H
1 1 1
− = 1+ −
u0 (c) (1 − τ (θ )) u0 (c) v0 (y/θ ) θ u0 c θ̂ 1−F λ (1 − F )
θf
" #
y/θv00 (y/θ )
Z θ̄ 0
(1 − H ) u 0 ( c )
1−F u (c (θ )) dF θ̂
τ
= 1+ −
1−τ v0 (y/θ ) θ u0 c θ̂ 1−F λ (1 − F )
θf
Also
So called Mirrlees-Diamond-Saez formula. Can be useful and misleading – A func-
tion of a bunch of endogenous stuff!
9
Loosely speaking, it says that marginal tax rates are a function of labor supply
elasticity, tail of the distribution and how redistributive the objective is. Let’s go
through them and see why
– The first term 1θ−fF is the tail of the productivity distribution. If the produc-
tivity distribution looks like a lognormal distribution or bounded above this
term converges to zero; taxes should be zero at the top. If productivity distri-
bution looks like a Pareto distribution at the top – has a fat tail, then this term
is positive.
– The second term looks a lot like an elasticity. In fact it is one plus the Frisch
elasticity of labor supply. So the more elastic is labor supply, the lower the
taxes.
– The last term is a complicated term. Somehow it measures how unequal is
consumption allocated as well as how redistribitive is the planner’s motive.
Suppose that H (θ ) goes up and nothing else changes, then it implies that the
tax rates should go up; a planner that would like to do more redistribution
would like to increase taxes.
• What happens to the top tax rate if the distribution of θ is unbounded. Suppose
that H converges to 1 really fast!!!, then
τ 1−F
lim = lim lim (1 + 1/ε)
1−τ θf
• Before providing some numbers. An exactly solved case. suppose that u (c) = c.
Then:
G (θ ) − F (θ )
τ (θ ) 1
= 1+
1 − τ (θ ) ε θ f (θ )
• We can calculate taxes at the top from micro data. What’s the shape of the income
distribution. Can approximate it with Pareto. See Saez’s pictures.
Suppose that
1 − F = Cθ −a
Then
f = aCθ −a−1
So
1−F 1
=
θf a
10
• From data we know the pareto tail for income ( dont really see θ). a from income
distribution is not the same as a above. But we know current taxes and can back
out θ from that
intratemporal Euler equation
(1 − τauctual ) θ = ψl 1/ε
So aθ = (1 + ε) aincome .So
τ 1 1+ε
=
1−τ (1 + ε) aincome ε
1
=
εaincome
• Then 1−τ τ should be between 2.5 and 1. So τ should be somewhere between .5 and
.71. Very high numbers!!!
• With income effect, the numbers would be higher. Now let’s look at Saez’s num-
bers.
11
3 Second Application - Diamond and Dybvig
• A dynamic model but we can analyze the problem using our tools.
• Maturity transformation:
– People need to eat everyday, i.e., they are hit by liquidity shocks
(1 − θ )u(c1 ) + θu(c1 + c2 )
• Utility function
cu00 (c)
− >1
u0 (c)
• Technology:
R>1
12
• Allocations: ct (θ ), t = 1, 2, x:investment in liquid asset,y: investment in illiquid
asset
• Feasibility:
x+y = e
πc1 (0) + (1 − π )c1 (1) = x
πc2 (0) + (1 − π )c2 (1) = Ry
or
1 1
π c1 (0) + c2 (0) + (1 − π ) c1 (1) + c2 (1) ≤ 1
R R
• Incentive constraint:
u(c1 (0)) ≥ u(c1 (1))
u(c1 (1) + c2 (1)) ≥ u(c1 (0) + c2 (0))
c2 (0) = c1 (1) = 0
0
u (c1 (0)) = Ru0 (c2 (1))
• Since risk aversion is higher than 1, 1 < c1 (0) < c2 (1) < R. There is some
insurance going on. Boring mechanism design problem!!! i.e., incentive constraints
are non-binding. What happens if we have other types?
• If they happen to become impatient, they can sell their claims to long asset to the
patient guys in exchange for claims to the short asset.
c1 (0) = xi + Pyi
13
Then the consumption of patient guys is
c2 (1) = xi /P + yi R
subject to
x+y = 1
c1 (0) = x + Py
R
c2 (1) = (y + x/P) R = ( x + Py)
P
• Market clearing:
Z Z
at t = 1, π c1i (0) di = xi di
Z Z
at t = 0, (1 − π ) c2i (1) = R yi di
• In equilibrium P = 1. Why?
14
• Banks offer a deposit contract, deposit your money at t = 0; if withdraw at t = 1,
then return is r1 if the fraction of people withdrawing is less then 1/r1 . If this
fraction is higher, then return is zero.
r1 f j ≤ 1/r1
V1 f j , r1 =
0 f j > 1/r1
V2 ( f , r1 ) = max { R (1 − f r1 ) / (1 − f ) , 0}
• A game between patient and impatient guys. Now, what happens in the equilib-
rium of this game: two equilibria
• No Run: impatient guys are boring. if patient guys think other patient guys will
not run
R (1 − πr1 )
No Run : u
1−π
Run : u (r1 )
• Run equilibrium: if patient guys think other patient guys will run:
No Run : u (0)
Run : 1/r1 u (r1 ) + (1 − 1/r1 ) u (0)
• Suppose we just ask them and dont care about sequential service. Also suppose
that the number of households is finite, I; Number of patient guys: stochastic
• Planner’s problem
15
subject to
1
( I − α (ω )) c1 (0, ω ) + α (ω ) c2 (1, ω ) = I
R
where ω = (θ1 , · · · , θ I ).
• FOCs
u0 (c1 (0, ω )) = Ru0 (c2 (1, ω ))
Can represent the state of the world by α: c1 (0, α), c2 (1, α).
I
c1 (0, α) =
α + R1/σ−1 ( I − α)
I
=
α + R1/σ−1 ( I − α)
IR1/σ
c2 (1, α) =
α 1 − R1/σ + IR1/σ−1
Since R > 1, c2 (1, α) > c1 (0, α). Also c1 (0, α) > c1 (0, α − 1).
– Intuition: The last patient trader does not want to lie(Given the remaining
resources with the bank, the planner always gives more to the patient guy).
Using a backward induction argument the second to last trader doesn’t want
to lie either. Nobody wants to lie! No Bank Run.
• Ennis and Kiester(2009, JET): if you allow for correlated types the truthfull revela-
tion game might have other equilibria.
• Bank Runs??!!
• DD: Bank run one issue that can be a result of maturity mismatch.
16
3.3 Non-Exclusive Contracts and The Role of Banks: Jacklin/Farhi-
Golosov-Tsyvinski
• Another issue pointed by Jacklin elaborated by Farhi, Golosov, and Tsyvinski: The
above analysis assumes exclusivity; households cannot contract with each other
and with other banks. Can potentially undo the insurance provided by the inter-
mediaries.
– There is a lot of banks; all of them are large, i.e., can service the whole market.
– They offer contracts at t = 0 and commit to them
– Households choose the contract with highest utility
– Before talking about non-exclusivity, let’s focus on exclusive contracts: As-
sume that banks can control how much people consume.
– Equilibrium is given by {ct (θ )} , Ū, where Ū is the value of equilibrium con-
tract to households and ct (θ ) solves the following maximization problem by
the intermediary(they take Ū as given)
1
max d1 + d2
R
subject to
x+y = 1
d1 + πc1 (0) + (1 − π ) c1 (1) = x
d2 + πc2 (0) + (1 − π ) c2 (1) = Ry
u (c1 (0)) ≥ u (c1 (1))
u (c1 (1) + c2 (1)) ≥ u (c1 (0) + c2 (0))
πu (c1 (0)) + (1 − π ) u (c1 (1) + c2 (1)) ≥ Ū
• Let’s focus on the truth-telling equilibrium so do not have to deal with bank runs.
• Now, we want to show that the equilibrium outcome is efficient: when contracts
are exclusive competitive markets are providing the efficient amount of insurance
that a planner offers.
17
• To show this first notice that as usual c2 (0) = c1 (1) = 0. Then the IC’s can be
written as
If we didn’t have the incentive constraints in there, we could use concavity proper-
ties to rewrite the bank’s problem as its dual and then bank’s problem would coin-
cide with planner’s problem. So how do we deal with incentive constraints(potentially
they can make the problem non-covex because a concave function is appearing on
the RHS of the IC’s). One way to deal with this in this environment is to rename
stuff: Let
1
max 1 − πC (û1 (0)) − (1 − π ) C (û2 (1))
R
subject to
where C (u) is the inverse of the utility function u (c). Now the constraint set
is linear and the objective is strictly concave – C is strictly convex and −C (u) is
strictly concave.
• So we can write down the dual of the above problem and use the fact that profits
in equilibrium are zero
subject to
18
Townsend(1984, Econometrica). They show that if contracts are signed before any
uncertainty is realized and people have access to lotteries (This is a trick to convex-
ify the constraint set – in the above proof, we didn’t really need it given the special
structure of this problem), then Competitive Equilibrium is constrained efficient.
• Let’s assume that household’s have access to trading market where they can trade
with each other but the intermediaries cannot see what they do in those markets,
i.e., hidden trading markets. There is an interest rate R̂ in this market. So if they
face a contract {ct (θ )} from an intermediary, this is the maximization problem
they solve
max (1 − θ ) u ( x1 ) + θu ( x2 )
θ̂,s,xt
subject to
x1 + s = c1 θ̂
x2 = R̂s + c2 θ̂
As it can be seen, they can decide to announce their type to the bank, take the
allocations from the bank, and trade in the hidden trade market. Call the value
associated with the above V {ct (θ )} , θ, R̂ . We also call the optimal choice of
reporting by θ̂ {ct (θ )} , R̂, θ and consumption by xt {ct (θ )} , R̂, θ .
1
max d0 + d1
R
such that
x+y = 1
d1 + πc1 (0) + (1 − π ) c1 (1) = x
d2 + πc2 (0) + (1 − π ) c2 (1) = Ry
θ = θ̂ {ct (θ )} , R̂, θ
πV {ct (θ )} , R̂, 0 + (1 − π ) V {ct (θ )} , R̂, 1 ≥ Ū
19
– Hidden Trade Market has to clear:
πxt {ct (θ )} , R̂, 0 + (1 − π ) xt {ct (θ )} , R̂, 1 ≤
πct θ̂ {ct (θ )} , R̂, 0 + (1 − π ) ct θ̂ {ct (θ )} , R̂, 1
1 1
c1 (0) + c2 (0) = c1 (1) + c2 (1)
R̂ R̂
Note that since consumers can do smoothing on their own in the hidden trad-
ing market, all they care about is the present value of consumption bundles of-
fered(where PV is calculted using hidden trade market prices, R̂). So they always
choose the bundle with the highest present value. IC means that the present values
have to be equal.
• Now, given the fact that banks make zero profits in equilibrium and a similar du-
ality result as before, competitive equilibrium is a solution to the following maxi-
mization problem
max πV {ct (θ )} , R̂, 0 + (1 − π ) V {ct (θ )} , R̂, 1
subject to
1 1
c1 (0) + c2 (0) = c1 (1) + c2 (1)
R̂ R̂
1 1
π c1 (0) + c2 (0) + (1 − π ) c1 (1) + c2 (1) ≤ 1
R R
• Second Result: R̂ = R. Suppose that R̂ < R, then clearly in the solution to the
above problem c1 (0) = c1 (1) = 0.Why? because if not you can decrease
c1 (0) by ε,
and increase c2 (0) by πR + (1 − π ) R̂ ε and c2 (1) by πε R − R̂ . This causes the
present values to increase equally(hence V increases for both types) and feasibility
to be satisfied. In other words, it is optimal for the intermediary to invest in the
long asset only. However, this cannot be an equilibrium because, when c1 (0) =
c1 (1) = 0, impatient types cannot consume anything given market clearing in the
hidden trading market. Similarly we can show that R < R̂ cannot be an equilibrium
either.
20
• So is the above allocation constraint efficient? Clearly it is not efficient with respect
to the original planning problem. But we have to define what constrainted efficient
means with hidden trade.
• We will assume that a planner is subject to the same hidden trading constraint –
people can trade in the hidden trade market behind planner’s back. However, by
choosing allocations, the planner can affect the interest rate in the hidden trade
market indirectly. So, here is the planning problem
subject to
1 1
π c1 (0) + c2 (0) + (1 − π ) c1 (1) + c2 (1) ≤ 1
R R
U (c1 (θ ) , c2 (θ ) ; θ ) ≥ V {ct (θ )} , R̂, θ
R̂ = Eqilibiurm Int. Rate in HT Mkt
subject to
1
x1 +
x2 = I
R̂
Call the policy functions x1 I, R̂, θ , x2 I, R̂, θ . Then it is easy to show that the
planner’s problem is equivalent to
max π V̂ I, R̂, 0 + (1 − π ) V̂ I, R̂, 1
I, R̂
subject to
1 1
π x1 I, R̂, 0 + x2 I, R̂, 0 + (1 − π ) x1 I, R̂, 1 + x2 I, R̂, 1 ≤1
R R
21
• Note that if there exists I and R̂ such that
where c1∗ (0) and c2∗ (1) is the solution to the original planning problem, then this
must be optimal because the above problem is obtained from adding extra con-
straints on the original planning problem.
c2∗ (1)
• Obviously I = c1∗ (0) and R̂ = c1∗ (0)
. To see that this is the case, we just have to show
that R̂ > 1, which is clear from the properties of the allocation described above.
Moreover, we can see that R̂ < R. This is because risk aversion is bigger than 1.
u0 ( I ) = Ru0 R̂I
Note that
d u00 ( x )
log u0 ( x ) = 0
dx u (x)
So
R̂I u00 ( x )
Z
log u0 R̂I − log u0 ( I ) =
dx
I u0 ( x )
Since risk aversion is bigger than 1,
u00 ( x ) x
− >1
u0 ( x )
Hence,
u00 ( x ) 1
< −
u0 ( x ) x
So, we must have
Hence, 1 < R̂ < R. So the planner would achieve the original efficient allocation.
• Note: this is not in general true. The Alternative2 Constrained Efficient Allocation
does not usually coincide with the Constrained Efficient Allocation.
• Why this inefficiency: As the above proof shows, the planner essentially affects
22
the interest rate in the hidden trading market. This interest rate affect agents’
incentives to lie. In particular, loweing interest rate R̂ benefits impatient agents
since it increases their income.
• So how can we implement this desired outcome? FGT show that a simple policy
when the government can see banks’ portfolios – or the banking sector’s protfolio
as a whole; there is no difference.
• Prefs:
θu (c a ) + u (cb )
As we saw in the previous example, we can just call consumption across at each
time a different good, i.e., we can call consumption in the first period apples and
consumption in the second period bananas.
T
∑ βt θt u (ct )
t =0
• Allocations: ct θ t . Feasbility
Z
ct θ t dµ θ t ≤ e
Θt
23
• Optimal risk sharing with full info
βt θt u0 (ct ) = λt
T Z T Z
∑β t t t
∑β t
θt u ct σt θ t dµt θ t
θt u ct θ dµt θ ≥
t =0 Θt t =0 Θt
• Planning problem
T Z
max ∑ βt θt u ct θ t dµt θ t
t =0 Θt
subject to
Z
ct θ t dµt θ t = e
Θt
T Z T Z
∑ βt θt u ct θ t dµt θ t ∑ βt θt u ct σt θ t dµt θ t , ∀σt
≥
t =0 Θt t =0 Θt
θ −1 : given
• One-shot deviation principle: Suppose that β < 1 and u is bounded or T < ∞. Then
an allocation is incentive compatible if and only if it satisfies one-shot incentive
t
compatibility: σt θ = θt except for a unique period t̂ and for a positive measure
of histories θ̂ t . With discrete processes,
∃!θ t σt θ t 6= θt
• With T = ∞, do not have to worry about deviations far away in time. same idea
works. See Fudenberg and Tirole, Section 5.
• This simplifies the set of incentive constraints. How do we write them down?
24
Define the following utility value:
vT θ T = θT u cT θ T
Z
t
= θt u ct θ + β vt+1 θ t , θt+1 π (dθt+1 |θt )
t
vt θ
One-shot deviation: lie today and tell the truth from then on.
Value of telling the truth from now on:
Z
t
vt+1 θ t , θt+1 π (dθt+1 |θt )
θt u ct θ +β
Value of lying today and telling the truth from tomorrow on:
Z
θt u ct θ t−1 , θ̂t + β vt+1 θ t−1 , θ̂t , θt+1 π (dθt+1 |θt )
• Planning problem Z
∗
w = max v0 (θ0 ) π (dθ0 |θ−1 )
subject to
Z
θt u ct θ t vt+1 θ t , θt+1 π (dθt+1 |θt ) = vt θ t
+β
Z
ct θ t dµt θ t = e
Θt Z
θt u ct θ t + β vt+1 θ t , θt+1 π (dθt+1 |θt ) ≥
Z
θt u ct θ t−1 , θ̂t + β vt+1 θ t−1 , θ̂t , θt+1 π (dθt+1 |θt )
• Would like to rewrite it recursively. For dynamic problems, very practical and
useful to deal with the dual. How do we know it is equivalent. For now, let’s just
assume!! Most problem have no way of knowing, in this case we do!!
25
• Then the above is equivalent to
T Z
max ∑ λt e − t t
ct θ dµt θ
t =0 Θt
subject to
Z
t
vt+1 θ t , θt+1 π (dθt+1 |θt ) = vt θ t
θt u ct θ +β
Z
v0 (θ0 ) π (dθ0 |θ−1 ) ≥ w̄
Z
t
+ β vt+1 θ t , θt+1 π (dθt+1 |θt ) ≥
θt u ct θ
Z
t −1 t −1
θt u ct θ , θ̂t + β vt+1 θ , θ̂t , θt+1 π (dθt+1 |θt )
• Suppose that |Θ| = N. State variable: θt−1 , vt θ t−2 , θt−1 , θt π dθt |θ̂t−1 = wt θt−1 , θ̂t−1 :
R
the value a true type of θ̂t−1 gets if pretends to be θt−1 at t − 1. N + 1 state variables
Pt θ− , w (θ− ) , w θ− , θ̂− θ̂− 6=θ− =
Z
λ t +1 0
0
max e − c (θ ) + Pt+1 θ, w (θ ) , w θ, θ̂ π (dθ |θ− )
Θ λt
subject to
Z
θu (c (θ )) + βw0 (θ ) π (dθ |θ− ) = w (θ− )
Z
θu (c (θ )) + βw0 (θ ) π dθ |θ̂− = w θ− , θ̂−
θu (c (θ )) + βw0 (θ ) ≥ θu c θ̂ + βw0 θ, θ̂
w θ, θ̂ : feasible
• Period 0 problem
Z
λ1 0
0
max e − c (θ ) + P1 θ, w (θ ) , w θ, θ̂ π (dθ |θ−1 )
w(θ̂−1 ) Θ λ0
26
subject to
Z
θ0 u (c (θ0 )) + βw0 (θ0 ) π (dθ0 |θ−1 )
w̄ =
Z
θ0 u (c (θ0 )) + βw0 (θ0 ) π dθ0 |θ̂−1
w θ̂−1 =
w θ̂−1 : feasible
• How do we find the set of feasible w θ, θ̂ ’s; A procedure similar to Abreu, Pearce
and Stachetti (To read more see Fernandes and Phelan, JET, 2000). The general
problem with persistence is very complicated. Later we will see a special case(θt is
a random walk) that we can actually solve by hand!
subject to
Z
θu (c (θ )) + βw0 (θ ) dπ (θ ) = w
θu (c (θ )) + βw0 (θ ) ≥ θu c θ̂
+ βw0 θ̂
Have to find λt ’s together with w0 such that P0 (w0 ) = 0 and that ct θ t dµt θ t =
R
e.
• How do we know that solution to the above problem is unique? IC’s make the
problem potentially non-covex.
27
subject to
Z
θu (θ ) + βw0 (θ ) dπ (θ ) = w
θu (θ ) + βw0 (θ ) ≥ θu θ̂ + βw0 θ̂
Constraint set is linear now and the period objective is strictly concave. The usual
method from SLP can be used to show that the value function is (strictly) concave.
• It turns out for many practical utility functions we have closed form solutions. For
now, let’s assume that qt+1 = λt+t 1 is constant over time, q, and also that T = ∞. It
λ
is fairly easy to generalize the following analysis to the case where qt+1 is changing
over time and finite horizon. Given this, we can drop the subscript t:
Z
−c (θ ) + qP w0 (θ )
P (w) = max dπ (θ )
subject to
Z
θu (c (θ )) + βw0 (θ ) dπ (θ ) = w
θu (c (θ )) + βw0 (θ ) ≥ θu c θ̂ + βw0 θ̂
subject to
Z
θu (θ ) + βw0 (θ ) dπ (θ ) = w
θu (θ ) + βw0 (θ ) ≥ θu θ̂ + βw0 θ̂
Now, we can see that the constraint set is linear of degree one in all the variables.
That suggests that the policy functions should be homogenous too, as long as the
objective is homothetic. We can consider three cases:
28
functions and value functions are homogeneous too. Guess:
1−β
u (θ, w) = û (θ ) + w
Eθ
w0 (θ, w) = ŵ (θ ) + w
1− β
P (w) = − Ae Eθ w
Problem becomes
1− β
Z h 1− β
i
max e Eθ w −eû(θ ) − qAe Eθ ŵ ( θ ) dπ (θ )
subject to
Z
[θ û (θ ) + βŵ (θ )] dπ (θ ) = 0
θ û (θ ) + βŵ0 (θ ) ≥ θ û θ̂ + βŵ θ̂
1− σ 1
– u (c) = c1−σ , σ 6= 1; In this case, C (u) = ((1 − σ) u) 1−σ . Multiplying the
constraint set by a positive factor doesn’t change anything, so we can make
the following guess
u (θ, w) = û (θ ) · (1 − σ ) w
w0 (θ, w) = ŵ (θ ) · (1 − σ) w
1
P (w) = A [(1 − σ ) w] 1−σ
29
Putting the 1 − σ there makes sure that the factor is positive. To see why it
works, under the above guess the problem becomes
Z h
1 1 1
i
max [(1 − σ ) w] 1−σ − [(1 − σ) û (θ )] 1−σ − qA [(1 − σ) ŵ (θ )] 1−σ dπ (θ )
subject to
1
Z
[θ û (θ ) + βŵ (θ )] dπ (θ ) =
1−σ
θ û (θ ) + βŵ (θ ) ≥ θ û θ̂ + βŵ θ̂
and
1 1
t −1 1
t
∏ [(1 − σ) ŵ (θs ; q)] 1−σ
ct θ , w0 ; q = [(1 − σ ) w0 ] 1− σ [(1 − σ) û (θt ; q)] 1− σ
s =0
– u (c) = − ψ1 e−ψc ; In this case C (u) = − ψ1 log (−ψu). Again multiplying the
constraint set by a positive factor does not change anything, so we have the
following guess
u (θ, w) = û (θ ) · (−w)
w0 (θ, w) = ŵ (θ ) · (−w)
1
P (w) = log (−w) + B
(1 − q ) ψ
subject to
Z
[θ û (θ ) + βŵ (θ )] dπ (θ ) = −1
θ û (θ ) + βŵ (θ ) ≥ θ û θ̂ + βŵ θ̂
30
This confirms the guess. So we have
1
c (θ, w; q) = − log (−ψû (θ; q) · (−w))
ψ
and
1 1 1 t −1
ct θ t , w0 ; q = − log (−ψû (θ; q)) − log (−w0 ) − ∑ log (−ŵ (θs ; q))
ψ ψ ψ s =0
A note about the CARA case: Notice that utility function is given by
1 T t −ψct
− ∑ β θt e
ψ t =0
1 T t −ψ(ct + ϕt )
ψ t∑
− βe
=0
• To completely solve the original planning problem, we need to find q and w0 . For
this, we will use feasibility:
Z
ct θ t , w0 ; q dµt θ t = e, ∀t ≥ 0
Θt
– Log case Z
ct θ t , w0 ; q dµt θ t = e
Θt
or
1− β
Z
e û(θt ;q)
e Eθ [w0 +∑ts−=10 ŵ(θs ;q)] dµt θ t = e
Θt
31
This will come down to the followig equations(why?)
Z 1− β
e Eθ ŵ ( θ;q ) dπ (θ ) = 1
e
P ( w0 ) + = 0
1−q
Similarly, we have
Z
1
[(1 − σ) ŵ (θ; q)] 1−σ dπ (θ ) = 1
e
P ( w0 ) + = 0
1−q
– CARA case: Z
ct θ t , w0 ; q dµt θ t = e
Θt
or
" #
1 1 1 t −1
Z
− log (−ψû (θ; q)) − log (−w0 ) − ∑ log (−ŵ (θs ; q)) dµt θ t = e
Θt ψ ψ ψ s =0
Hence,
Z
log (−ŵ (θ; q)) dπ (θ ) = 0
e
P ( w0 ) + = 0
1−q
For more general formulations, it is possible to write the above but wouldn’t be
able to bring them down to single equations.
Question: What happens when the problem is non-stationary, i.e, when T < ∞ or
when qt ’s are not constant.
• Now, we have fully characterized the solution to original planning problem. One
way to think about the model: what is the amount of optimal inequality? How
much inequality in consumption and wealth should societies tolerate given hetero-
32
geneity in taste or income? Let’s see what the model implies about optimal amount
of inequality in consumption:
1−β
E [ŵ (θs ; q)]
E [θ ]
Take logs Z 1− β
log e Eθ ŵ ( θ;q ) dπ (θ ) = 0
or
1−β
Z
ŵ (θ; q) dπ (θ ) < 0
Eθ
Hence the drift is negative. So consumption is a random walk with a negative drift.
So the variance of consumption converges to infinity and consumption converges
to 0 almost surely (can show this for other cases too). Immiseration. Disappointing
theory of optimal inequality!! This property is very robust in dynamic contracting
models– general preferences, finite horizon(increasing variance) or even shock pro-
cess (how persistent it is, we will see an example for this soon)
• Alternative theories to address this issue: can we modify the model to think about
optimal inequality? (if we have time, we’ll talk about them)
33
– Unequal discounting: Interpret each period as a generation; The planner puts
more weight on future generations than households: Phelan(2006, Restud),
Farhi and Werning(2007, JPE)
– Lack of Commitment: Planner cannot committ to future contracts; it realizes
that inequality is really high; will reneg. Can be shown to be equivalent to a
version of the above; Sleet and Yeltekin(RED, 2006)
– Fertility: Generational interpretation; planner can use both number of children
and promised utility to provide incentives; will use fertility more heavily and
hence there is a stationary distribution; Hosseini, Jones, and Shourideh(2010)
• Another way of doing this is to see that truth-telling values are related to each other
through the future stream of consumption so may be we can use some properties
of the shock process to relate these two values to each other. It turns out that there
is a special case where we can do this,the case where θt is a geometric random walk
Now, just think about a two period example. Moreover, assume that we represent
1
the allocations in terms of innovation shocks: ε t , c0 (ε 0 ) , c1 ε
The future utility of telling the truth, type ε 0
Z
w (ε 0 ) = θ−1 ε 0 ε 1 u (c1 (ε 0 , ε 1 )) dG (ε 1 )
34
It is easy to see that
ε0
w (ε̂ 0 , ε 0 ) = w (ε̂ 0 )
ε̂ 0
and hence the incentive constraint is given by
ε0
θ−1 ε 0 u (c0 (ε 0 )) + βw (ε 0 ) ≥ θ−1 ε 0 u (c0 (ε̂ 0 )) + β w (ε̂ 0 )
ε̂ 0
Notice that in the above two period example, if we let persistence to be ρ < 1, or
then ρ
ε0
w (ε̂ 0 , ε 0 ) = w (ε̂ 0 )
ε̂ 0
and the incentive constraint can be written as
ρ
ρ ρ ε0
θ−1 ε 0 u (c0 (ε 0 )) + βw (ε 0 ) ≥ θ−1 ε 0 u (c0 (ε̂ 0 )) + β w (ε̂ 0 ) (1)
ε̂ 0
T Z Z
w ε t −1
, ε̂ t = ∑ β s − t −1
··· θ−1 ε 0 · · · ε t−1 ε̂ t ε t+1 · · · ε s u cs εt−1 , ε̂ t , ε t+1 , · · · , ε s dG (ε t+1 ) · · ·
s = t +1
T Z Z
t
∑ s − t −1
θ−1 ε 0 · · · ε t−1 ε t ε t+1 · · · ε s u cs εt−1 , ε̂ t , ε t+1 , · · · , ε s dG (ε t+1 ) · · ·
ŵ ε ; ε̂ t = β ···
s = t +1
t
ε t t −1
w ε ; ε̂ t = w ε , ε̂ t
ε̂ t
• Note that the above formulation with persistence ρ < 1 does not work any more.
This is because
T Z Z
ρ s +1 ρ s ρ s − t +1 ρ s − t ρ s − t −1
w ε t −1
, ε̂ t = ∑ β s − t −1
··· θ−1 ε 0 · · · ε t−1 ε̂ t ε t +1 · · · ε s u cs εt−1 , ε̂ t , ε t+1 , · · · , ε s dG (ε
s = t +1
T Z Z
ρ s +1 ρ s ρ s − t +1 ρ s − t ρ s − t −1
t
∑ s − t −1 t −1
ŵ ε ; ε̂ t = β ··· θ −1 ε 0 · · · ε t −1 ε t ε t +1 · · · ε s u cs ε , ε̂ t , ε t+1 , · · · , ε s dG (ε
s = t +1
35
• Now, let’s see what happens in the infinite horizon case – now only a two dimen-
sional state variable θ− and w.
Z
−C (u (θ )) + qP w0 (θ ) , θ
P (w, θ− ) = max dF (θ |θ− )
subject to
Z
θu (θ ) + βw0 (θ ) dF (θ |θ− ) = w
θ
θu (θ ) + βw0 (θ ) ≥ θu θ̂ + β w0 θ̂
θ̂
w
• Further characterization: conjecture P (w, θ− ) = P̂ θ− . Why?
w
Let v = θ− . We can rewrite the problem as
Z 1
0 θ
max −C (u (θ )) + q P̂ v (θ ) g dθ
θ− θ−
subject to
Z
0
1 θ
θu (θ ) + βθv (θ ) g dθ = θ− v
θ− θ−
θu (θ ) + βθv0 (θ ) ≥ θu θ̂ + βθv0 θ̂
or Z
−C (û (ε)) + q P̂ v̂0 (ε)
max g (ε) dε
subject to
Z
θ− εû (ε) + βεv̂0 (ε) g (ε) dε = θ− v
Notice that the incentive constraint implies that û (ε) + βv̂0 (ε) is constant. So we
can combine promise promise keeping and IC’s and rewrite them as
v
û (ε) + βv̂0 (ε) =
E [ε]
36
subject to
v
û (ε) + βv̂0 (ε) =
E [ε]
It is clear that in the solution to the above problem
û (ε) = û ε0
No INSURANCE.
If the steady state of the above problem is given by v∗ and q is such that û (ε, v∗ ) = e,
then
ct θ t
= e
t
= θt v∗
wt θ
• Why? loose intuition: in this problem the planner uses differences in marginal rate
of substitution to provide incentives. With geometric random walk, the marginal
rate of substituition is the same across all types. No incentive for risk sharing can
be provided.
• This can be more clear by looking at the incentive constraint (1). Suppose that
−ρ
θ−1 = 1 and let v = ε 0 w (ε 0 ). Given this the incentive constraint becomes
ρ ρ
ε 0 u0 (ε 0 ) + βε 0 v (ε 0 ) ≥ ε 0 u0 (ε̂ 0 ) + βε 0 v (ε̂ 0 )
Now think about the trade-off between current utility, u0 , and future ’utility’, v (ε 0 ).
The slope of the indifference curve for a person of type ε 0 is given by
ρ −1
− βε 0
When ρ = 1, the slope of these indifference curves are all equal. Hence the planner
cannot exploit these differences to provide incentives. So the planner must treat all
agents similarly. The lower ρ, the higher it is the difference between these slopes.
Hence, for lower ρ’s, it is easier to provide incentives for truth telling and easier to
separate types.
• In the income shock interpretation of this model – with CARA, constant consump-
tion means autarky. So the variance of consumption still goes off to infinity in
the long run – geometric random walk. See Williams(2011, Econometrica) for a
continuous time version of this; Hopefully we will discuss continuous time soon.
37
5 Application: Optimal Unemployment Insurance
• We have learned the basic way to deal with dynamic contracting/mechanism de-
sign problems. The first application of those methods, we will consider is optimal
unemployment insurance.
• How do we model this tention? job seeking effort: private information with effort
being costly
• If provide full insurance against unemployment, nobody would ever look for a job.
• Basic environment: one agent and one unemployment insurance agency.(Can also
think of a continuum; doesn’t really make a difference in the way we write down
the contracting problem)
• Agent’s preference:
∞
∑ βt [u (ct ) − at ]
t =0
38
• Full information
ct ( ht ) = c
consumption is constant. a∗ solves
∞
βu0 (c) w
a = arg max ∑ β (1 − p ( a))
∗ t t
−a
a
t =0 1−β
• Private Information: revelation principle still applies; the planner suggest an action
to the agent; that action has to be optimal, i.e.,
∞ Z
∑ βt [u (ct (ht )) − (1 − st (ht )) at (ht )] dπ (ht ; at ) ≥
t =0
∞ Z
∑β t
[u (ct (ht )) − (1 − st (ht )) ât (ht )] dπ (ht ; ât )
t =0
• As before, we can focus on one-shot deviations. They only matter after histories
for which the worker has been unemployed. We can also define promise utility
ve ( ht ) , vu ( ht )
where ve (ht ) is the promised utility of a person who finds a job at the begining of
t + 1; vu (ht ) is the promised utility for a person who does not find a job; one-shot
IC is given by
• We are also gonna look at the dual as usual; so the unemployment insurance
agency minimizes the cost of insuring the agent subject to providing a least level
of utility.
• We are gonna jump right into the recursive formulation: Suppose Pu (v) and Pe (v)
are two value functions associated with employment and unemployment, then we
have
Pu (v) = min c + β [ p ( a) Pe (ve ) + (1 − p ( a)) Pu (vu )]
subject to
u (c) − a + β [ p ( a) ve + (1 − p ( a)) vu ] = v
− a + β [ p ( a) ve + (1 − p ( a)) vu ] ≥ − â + β [ p ( â) ve + (1 − p ( â)) vu ]
39
• As long as ve > vu , − â + β [ p ( â) ve + (1 − p ( â)) vu ] is a concave function. We can
replace by its first order condition
1 = βp0 ( a) (ve − vu )
•
Pu (v) = min c + β [ p ( a) Pe (ve ) + (1 − p ( a)) Pu (vu )]
subject to
u (c) − a + β [ p ( a) ve + (1 − p ( a)) vu ] = v
βp0 ( a) (ve − vu ) = 1
• Solving for Pe (v) is really easy; when people get a job their consumption is con-
stant. Also they do not have to look for jobs:
subject to
u (c) + βv0 = v
By Euler equation c has to be constant. Hence
u (c)
=v
1−β
and therefore
−w + u−1 ((1 − β) v)
Pe (v) =
1−β
• Now, let’s turn to the unemployed state’s problem. In general proving convex-
ity of the value function is almost impossible in dynamic contracting problems.
This is because of the non-convexity that incentive constraints introduce. So what
Hopenhayn and Nicolini do is to assume that the value function is convex and
differnetiable. Hence, we will be able to take first order conditions. They are given
by – let’s call consumption cu :
1 − λu0 (cu ) = 0
βp ( a) Pe0 (ve ) − λβp ( a) − ηβp0 ( a) = 0
β (1 − p ( a)) Pe0 (vu ) − λβ [1 − p ( a)] + ηβp0 ( a) = 0
βp0 ( a) [ Pe (ve ) − Pu (vu )] − λ −1 + βp0 ( a) [ve − vu ] − ηβp00 ( a) [ve − vu ] = 0
40
We can rewrite the above
1 p0 ( a)
Pe0 (ve ) = + η
u0 (cu ) p ( a)
1 p0 ( a)
Pu0 (vu ) = − η
u0 (cu ) 1 − p ( a)
We can think about the above equation as an Euler Equation for the UI agency.
Consider an increase in v by ε. As a result, the UI agency can increase cu by ε or
can increase ve and vu by εβ−1 . Note that the latter, does not affect the incentive
constraint. The beneft of the former change to the insurance agency is given by
1
ε = Pu0 (v ) ε
u0 (cu )
At the optimal allocation, the UI agency should be indifferent between these two
perturbations, so we get the original equation. Notice that if we use Envelop con-
ditions, the above equation implies that
1 1
= Et
u0 (c t) u0 (c t +1 )
This is the so-called Inverse Euler Equation. It turns out that it is a very general
property of consumption in dynamic contracting problems, as long as the source
of private information, here cost of effort, is separable from consumption – think
about the above perturbation; what part of the argument fails if utility function
was of the form u (c, a)?
It also has a strong implication, that these agents should be saving constrained
when they are in the unemployed state. Why? This is because of Jensen’s inequality
1 1 1
= Et ≥
u0 (c t) u0 (c t +1 ) Et u0 (c t +1 )
41
and hence
u0 (ct ) ≤ Et u0 (ct+1 )
Note that the above inequality is strict in the unemployment state since future
consumption is not degenerate. So if the agent had access to saving technology
with return β−1 , he would like to save in order to decrease future promsied utility.
The reason is that in this model, saving tightens future incentive constraints. An
unemployed agent would rather save a little bit and put in less effort, lower than
what the UI agency would like him/her to do. This is because of concavity of
utlity function. If the agent increases his saving, his utility in the unemployed
state increases more than his utility in the employed state since utility function is
concave and therefore he has a lower incentive to exert effort.
• Decreasing Benefits.
Consider a decentralization in which UI agency doesn’t let the agents save. In this
decntralization, unemployment benefit is simply cu (v) when the state is unem-
ployed is at state v. Moreover, the agent pays unemployment taxes once he gets a
job. The tax level is given by w − ce (v) at state v. Given the strucutre of the model
we can define three objects:
1. vut : the promised utility for an agent who has been unemployed since t = 0
and is unemployed at the beginning of t.
2. vet+1 : the promised utility for an agent who has been unemployed since t = 0
and finds a job at the beginning of t + 1.
3. cut : the consumption of a person who is unemployed at t.
1 p0 ( a) 1
Pu0 vut+1 =
0 u − η = 0 u
u (ct ) 1 − p ( a) u c t +1
Lemma 2. η > 0.
Proof.
Suppose not. Then
Pu0 (vu ) ≥ Pu0 (v ) ≥ Pe0 (ve )
together with convexity of the value function and local incentive compatibility im-
plies that
v ≤ vu < ve
42
Note that cu ≥ ce = u−1 ((1 − β) ve ) since η ≤ 0. Now, we can write
or
Note that the RHS of the above is positive since cu ≥ ce and that a is incentive
compatible. However, the left hand side is negative since v ≤ vu < ve and β < 1.
A contradiction. Q.E.D.
Since η > 0, then
1 p0 ( a) 1
0 u − η = 0 u
u (ct ) 1 − p ( a) u c t +1
implies that
1 1
> 0 u ⇒ cut > cut+1
u0 (cut ) u c t +1
Moreover, by assumption, Pu is convex and therefore
Proof.
We have
1
p0 ( a) [ Pe (ve ) − Pu (vu )] = η p00 ( a)
βp0 ( a)
and
p ( a) (1 − p ( a))
η = Pe0 (ve ) − Pu0 (vu )
p0 ( a)
hence
p00 ( a) p ( a) (1 − p ( a))
Pe (ve ) − Pu (vu ) = Pe0 (ve ) − Pu0 (vu )
βp0 ( a)3
43
Now suppose to the contrary that vet+1 > vet , then since vut+1 < vut . This together
with IC and concavity of p ( a) implies that at+1 > at . Under assumption i.,
Moreover,
Pe0 vet+1 > Pe0 (vet ) > Pu0 (vet ) ≥ Pu0 vut+1
However, this is in contradiction with the fact that Pe and Pu are both increasing.
When assumption ii. holds, the argument is similar.
• Intuitively, absent incentive constraints and due to insurance reasons, the planner
would like to equate vut and vet . So a decrease in vut should be accompanied by a
decrease in vet . However, the planner might want to make sure that agents exert
the efficient level of effort. So provided that the effect of keeping ve and vu close is
not so strong on incentives, then it is efficient to have vut be decreasing over time.
• We make the following change in the environment. Assume that the timing of the
events in each period is as follows: upon entering date t the households consume,
ct , they then look for jobs by exerting effert, and will realize whether they find a
job or not, with probability pt . We assume that the utility function in each period
is given by
u ( c t , (1 − s t −1 ) p t −1 )
Notice that we have assumed that utility is a direct function of job finding rate as
opposed to effort.
44
this case is consisted of
where τt (ht ) is the transfer to the agent, bt is a suggested level of saving by the
planner with b0 = 0 and ct is defined by the following budget constraint
subject to
ĉt (ht ) + b̂t (ht ) = τt (ht ) + wst (ht ) + β−1 b̂t−1 (ht−1 )
together with an appropriate borrowing limit that is low enough so that it is never
binding.
• So the UI agency now would like to minimize the cost of providing an allocation
subject to incentive compatibility and an appropriate promise keeping. The UI
agency also has access to the same borrowing and lending rate β−1 .
• First observation: without loss of generality, we can assume that in any incentive
compatible allocation, bt (ht ) = 0. Intuitively, as long as agents don’t want to
deviate, the UI agency can save for them. More technically, one can start from an
allocation with bt (ht ) 6= 0 and define transfers as
It can be shown that the contract {c (ht ) , τ̃ (ht ) , pt (ht ) , 0} is incentive compatible.
• Incentive constraint is very complicated now. One way to simplify, as always, is to
focus on the first order conditions. What are they? The first order condition with
respect to at is given by
u (cet (ht ) , pt−1 (ht−1 )) + βvet (ht ) − u (cut (ht ) , pt−1 (ht−1 )) − βvut (ht )
+ pu p (cet (ht ) , pt−1 (ht )) + (1 − p) u p (cet (ht ) , pt−1 (ht−1 )) = 0
45
An Euler equation, c.f., Inverse Euler Equation.
In order for this relaxed version of incentive compatibility to be equivalent to the
original definition, the objective in the original must be concave. Nothing guar-
antees that this is the case. In fact, here I show you a two period example where
the two constraints are not equivalent – for more detailed discussion see Kocher-
lakota(2004, Review of Economic Dynamics).
u ( c1 ) + u ( c2 ) − v ( p1 )
Suppose the optimal value of p is given. Then the partially planning problem is
given by
min c1 + pc E + (1 − p) cU
c1 ,c E ,cU
subject to
u (c1 ) + pu (c E ) + (1 − p) u (cU ) − v ( p) ≥ w
(0, p) ∈ arg max u (c1 − s) + p̂u (c E + s) + (1 − p̂) u (cU + s) − v ( p̂) (2)
s, p̂
min c1 + pc E + (1 − p) cU
c1 ,c E ,cU
subject to
u (c1 ) + pu (c E ) + (1 − p) u (cU ) − v ( p) ≥ w
u0 (c1 ) = pu0 (c E ) + (1 − p) u0 (cU )
v 0 ( p ) = u ( c E ) − u ( cU )
The three constraints pin down the level of consumption. For the allocation to be
incentive compatible, this has to be the solution to the maximization problem (2).
For this to happen, the Hessian matrix associated with that problem should be
negative semi-definite at that point. This hessian matrix is given by
p
− 1c−2 p
" #
− c12 − c2E
1
cE − 1
cU
1 U
1
cE − 1
cU −v00 ( p)
46
When v00 ( p) is small – very little curvature in the cost function, it is possible that
the determinant of this matrix is negative so there are potential second order gains
from deviations.
• Next we consider some cases where we have closed form solutions and we can
check whether the first order approach is valid.
• Let’s switch to the original problem. How can we rewrite the problem recursively
to incorporate the Euler equation? Werning(2000) shows that if we keep track of
uc = λ in addition to v, then we can rewrite the problem recursively. The only
thing we have to worry about is that not all pairs of (v, λ) are attainable. In order
to ensure this, he proposes a procedure similar to Abreu, Pearce, and Stachetti with
set operators and the fixed point of that operator leads to two sets, ∆u and ∆e such
that
∆s = {(v, λ) | (v, λ) is attainable by some contract}
I will side step this issue here since we will work with cases in which we can
characterize these sets analytically.
subject to
u (c, 0) + βv0 ≥ v
λ = uc (c, 0)
v0 , λ0 ∈ ∆e
subject to
47
• Let’s assume that the utility function satisfies the following
1
u (c, p) = − e−ψ(c−α( p))
ψ
• The first step is to characterize ∆s . Given the above utility function, we have
and therefore
ut = Et ut+1 = Et us , s > t
Therefore
∞
u t −1
v = Et−1 ∑ βs us =
s=t 1−β
1 uc,t−1 1
= − =− λ
ψ1−β ψ (1 − β )
When we assume that households can only save the above equalities become and
inequality. In the optimal contract, however, one can show that these inequalities
bind, so we ignore them here.
The above equality not only characterizes the sets ∆s , it also means that we can
drop λ as state variable.
subject to
u (c, 0) + βv0 = v
uc (c, 0) = −ψ (1 − β) v
48
subject to
Note that
subject to
Similar to the taste shock cases studied before, the constraint set is linear in terms
of u (c, p)’s and v’s. So we conjecture the following
1
ce (v) = ĉe − log (−v)
ψ
1
cu (v) = ĉu − log (−v)
ψ
e e
v (v) = v̂ · (−v)
vu (v) = v̂u · (−v)
p (v) = p̂
and
1
Pu (v) = Au − log (−v)
ψ (1 − β )
1
Pe (v) = Ae − log (−v)
ψ (1 − β )
Not gonna go through the algebra. Can check again that vu (v) < v.
49
• Are there simple policies that can implement the above optimal allocation? Let’s
try a simple policy: constant unemployment benefit b̂, constant tax τ̂. Now think
about a household’s problem
subject to
ce + ae ≤ w − τ̂ + β−1 a
cu + au ≤ b̂ + β−1 a
together with
V e ( a) = max u (c, 0) + βV e a0
subject to
c + a0 ≤ w − τ̂ + β−1 a
• We can show that there exists τ̂ and b̂ so that the solution to the above problem
coincides with the optimal allocations. To do so note that the functional forms for
the above value functions is given by
−1
V u ( a) = − Âu e−ψ( β −1) a
−1
V e ( a) = − Âe e−ψ( β −1) a
1
Pu (v) = Au − log (−v)
ψ (1 − β )
1
Pe (v) = Ae − log (−v)
ψ (1 − β )
or
u u (v)
v = − e ψ (1− β ) A e − ψ (1− β ) P
e e (v)
v = − e ψ (1− β ) A e − ψ (1− β ) P
u e
Now, if we set τ̂ and b̂ so that eψ(1− β) A = Âu and eψ(1− β) A = Âe , then Pu (v) and
Pe (v) are associated with β−1 a. Moreover, it can be checked that the solution for p
in (4) is identical to the solution in (3). Moreover, since the margin between c and
v0 in (3) is undistorted – the same as in (4), one can show that the solution to the
planner’s problem coincides with the solution to the agent’s problem given v0 , a0
is chosen so that Pu (v0 ) = β−1 a0 – we assume that the ange starts unemployed.
• To illustrate how the model changes with search explicitly modeled, we setup the
50
model with McCall search – Shimer and Werning (2008, AER): agents draw a wage
from a distribution F (w); decide whether to accept or reject the offer; The value of
the offer is private information. The model is very similar to the above.
u (w − τt )
1−β
u (bt ) + βvt+1
where vt+1 is the ex-ante utility of an unemployed agent at t + 1. For the agent
to follow the recommended reservation wage, we must have
u (w̄t − τt )
= u (bt ) + βvt+1
1−β
where
Z ∞
u (w − τt )
vt = (1 − F (w̄t )) (u (bt ) + βvt+1 ) + dF (w)
w̄t 1−β
subject to
Z
TdF (θ ) = g
V ( T; θ ) = max u (y − T (y) , y; θ )
y
51
Now, we need to turn this into a mechanism design problem. To do so, consider
a tax function and the allocation that is implied by it, i.e., c (θ; T ) = y (θ; T ) −
T (y (θ; T )) so that u (c (θ; T ) , y (θ; T ) ; θ ) = V ( T; θ ). This allocation has to be in-
centive compatible because
This means that a function T (y) must exist such that c (θ ) = y (θ ) − T (y (θ )). What
we have proved is the Taxation Principle.
In dynamic environments this becomes a bit trickier. It is very plausible to as-
sume that there are some insurance markets to insure households against some
of the productivity shocks they receive; health insurance for instance. This means
that when we right down the problem of an agent, we should take into account
what type of insurance markets are available to the household. Unfortunately, our
underlying friction – private information – does not pin down the limits of govern-
ment and markets. So for now, we assume that markets are incomplete and that
the only tool available to households for insurance is self-insurance a la Aiyagari.
More instead of writing the problem as a problem of finding best taxes, we start
directly from the mechanism design problem and then derive the properties of the
tax functions that can implement efficient allocation.
• Preferences:
T
yt
∑β t
u (ct ) − v
θt
t =0
• Allocations:
T
ct θ : ΘT → R
yt θ T : ΘT → R
52
such that ct θ T and yt θ T are both measurable with respect to θ t – they cannot
• Feasible allocations
Z Z
ct θ dπ θ + Kt+1 − (1 − δ) Kt = F Kt , yt θ dπ θ T
T T T
• Incentive compatibility:
( " !#)
T yt θ T
E0 ∑ βt u ct θ T −v ≥
t =0 θt
( " !#)
T yt σ θ T
E0 ∑ βt u ct σ θ T −v
θt
, ∀σ : Θ T → Θ T
t =0
• A key property that we have come across before is the Inverse Euler Equation. Any
solution of the mechanism design problem should satisfy
1 1 1
= Et 0
u0 (c T
t ( θ )) β [1 − δ + FK,t+1 ] u (ct+1 (θ T ))
• The idea behind this equation is that this is an Euler equation for the insurance
agency. Consider an insurance agency who wants to decide between whether to
deliver utility(from consumption) in period t and history θ t or to deliver utility in
period t + 1 for all histories θ t+1 following θ t . At the margin this insurance agency
should be indifferent to deliver utility at t vs. t + 1. To see what this implies,
consider a small decrease in utility at history θ t by ε and a small increase in utility
from consumption at t + 1 in all states θ t+1 following θ t , i.e.,
u ĉt θ T = u ct θ T − ε
u ĉt+1 θ T = u c t +1 θ T + β −1 ε
53
Now, the benefit of the above perturbation at t is given by
ε
u0 (c T
t ( θ ))
1 ε
Et 0
β u (ct+1 (θ T ))
To make the benefit in terms of period t goods, we have to discount them by the
gross interest rate given by 1 − δ + FK,t+1 and hence we have the above equation.
• To see more intuitively why the above equation holds, we consider the following
simple example:
• F (K, Y ) = RK + Y and δ = 1.
subject to
1 1
c0 +
R i∑ πi c1i ≤ y0 + ∑ πi y1i
R i
y1L y1H
u (c1L ) − v ≤ u (c1H ) − v
θH θH
n o
Now take the solution to the above problem c0 , y0 , {c1i , y1i }i= H,L . There are two
key properties that will provide intuition for us. First, c1H > c1L . Second, the IC
constraint binds with equality.
54
truth-telling in the future. The RHS of the IC constraint is increased by u0 (c1H ) Rε
and its LHS is increased by u0 (c1L ) Rε. Now since c1H > c1L and u is concave, the
RHS is increased by less than the LHS. Hence saving decreases the incentive for
truth-telling in the future. Hence, in addition to decreasing consumption today,
saving tightens future incentive constraints and therefore it has an additional cost.
Hence, on the margin
Since this incentive constraint binds in the second period, the solution to the mech-
anism design problem must satisfy
y1L y1H
c1L − v = c1H − v
θH θH
Now consider increasing saving by ε and increasing c1L and c1H by Rε. As it can
be seen from above, this perturbation changes the RHS and LHS of the above IC
by the same amount. Hence, the incentive cost of saving in this model is zero and
therefore Euler Equation must hold
• As the above analysis suggests, whether the houehold should be saving constrained
or borrowing constrained, depends on how marginal utility when telling the truth
is compared to marginal utility when lying for binding incentive constraint. In the
separable case, marginal utility when the telling the truth is lower than when lying
hence saving tightens the incentive constraints in the future. In Shourideh(2011),
I develop a model with capital income risk where truth-telling marginal utility is
higher than lying marginal utility and hence sometimes it is optimal to subsidize
saving.
Implementation
• So far we have focused on the mechanism design problem. Here we try to see, first
through the lens of the above example, how to implement the solution to the above
55
mechanism design problem. As mentioned before, we assume that the only asset
available to the households is a risk free bond. Now consider the above two period
example and suppose that there are no taxes in the first period and taxes in the
second period are a function of labor income and asset/capital income. Given this
market structure, the household budget constraint are given by
c0 + b1 ≤ y0
c1 ≤ y1 + (1 + r1 ) b1 − T (y1 , r1 b1 )
subject to
c0 + b1 ≤ y0
c1i ≤ y1i + (1 + r1 ) b1 − T (y1i , r1 b1 )
• The goal is to find a tax function T (·, ·) so that the solution to the above problem
coincides with the solution to the mechanism design problem c0∗ , y0∗ , c1i ∗ , y ∗ , b ∗ (=
1i 1
y0∗ − c0∗ ).
• Let’s start from a simple tax function T (y1 , r1 b1 ) = T y (y1 ) + τk r1 b1 . Given the
allocation, the candidate for τk is given by
1 u0 (c0∗ )
τk∗ = 1 + −
r1 rE0 u0 c1∗
56
utility. Consider the following allocation:
c0 = y0∗ − ε
b1 = b1∗ + ε
∗
y1i = y1L
∗ ∗
c1i = y1i − T y (y1i ) + (1 + r1 (1 − τk∗ )) (b1∗ + ε)
∗
= c1L + (1 + r1 (1 − τk∗ )) ε
The idea behind this deviation is that when facing a linear tax function on saving,
the household finds it optimal to save a bit more than what she is supposed to and
work less.
• One way to resolve this issue is to allow for non-linear asset/capital income tax.
For example, suppose that the capital tax function has a kink at r1 b1∗ . That is
with
b0 1 u0 (c0∗ )
lim T (r1 b1 ) = 1 + − 0 ∗
b1 &b1∗ r ru c1L
1 u0 (c0∗ )
lim∗ T b0 (r1 b1 ) = 1 + − ∗
r rEu0 c1L
b1 %b1
Then clearly the above double deivation doesn’t not deliver higher utility. However,
there might be some other deviations that deliver a higher level of utility. To find
the tax function T b (r1 b1 ) we do as follows:
57
Consider the following maximization problem
subject to
c1i = y1i − T y (y1i ) + x
Now, define the function T b (r1 b1 ):
u (y0 − b1 ) − v y0 /θ̄ + V (1 + r1 )b1 − T b (b1 r1 ) = u∗
where u∗ is the utility level from the solution to the mechanism design problem.
This is the lowest asset tax function that can implement the optimal allocation.
See Werning(2011, ‘Nonlinear Capital Taxation’) for a proof of differentiability of
T b (r1 b1 ) when the distribution of shocks is continuous as well as its extension of
multiple periods.
i.i.d. Shocks
• Here we study optimal capital and labor income taxes in a dynamic setting with
i.i.d. shocks. As before, the problem can be written in recursive form using promise
utility w as state variable. The mechanism design problem as before is given by
Z
c (θ ) − y (θ ) + qt+1 Pt+1 w0 (θ )
Pt (w) = min dF (θ )
subject to
Z
u (c (θ )) + h (y (θ ) , θ ) + βw0 (θ ) dF (θ ) = w
u (c (θ )) + h (y (θ ) , θ ) + βw0 (θ ) ≥ u c θ̂ + h y θ̂ , θ + βw0 θ̂
h (y, θ ) = θ ĥ (1 − y)
• Before trying to characterize the solution to the above problem, we first try and see
if we can show the convexity of the value function. Obviously at its current form
the constraint set of the above problem is not concave – concave functions showing
up on the RHS of the IC. This mean that the usual techniques from SLP cannot
58
be applied. However, there is a transformation of variables that might make the
constraint set convex. Consider the following variables
υ (θ ) = u (c (θ ))
η (θ ) = h (y (θ ) , θ )
C (υ (θ )) − Y (η (θ ) , θ )
where
u (C (υ)) = υ
h (Y (η, θ ) , θ ) = η
υ (θ ) + η (θ ) + βw0 (θ ) ≥ υ θ̂ + h Y η θ̂ , θ̂ , θ + βw0 θ̂
For this constraint
set to be convex, we must have that h Y η θ̂ , θ̂ , θ is a convex
function of η θ̂ . In general it is hard to come up with conditions on h (·, ·) so that
this function is convex. There are some cases, however, that the above function
becomes linear in η θ̂ . These are cases where h (y, θ ) = ĥ (y) g (θ ) where ĥ (y) is
concave and decreasing. Under this assumption
−1 η
Y (η, θ ) = ĥ
g (θ )
and hence
! !
η θ̂
= h ĥ−1
h Y η θ̂ , θ̂ , θ ,θ
g θ̂
!!
η θ̂
= ĥ ĥ−1 g (θ )
g θ̂
η θ̂
= g (θ )
g θ̂
59
– Productivity shocks + constant Frisch Elasticity specification
1 y 1+ γ
h (y, θ ) = −
1+γ θ
h (y, θ ) = θh (1 − y)
• Another way of ensuring that the constraint set is convex is to assume that only
downward constraints bind and that hh11 is a decreasing function of θ – can be
1
interpreted as decreasing absolute risk aversion with respect to θ.
• Example: Here we focus on an example where we have closed form solutions and
see how taxes evolve over time and in the cross section. Assume that preferences
have the following form
1
log c + log (1 − y)
θ
These preferences imply that optimal static distortions are independent of how
much wealth is brought in by the household, i.e., MRS between c and l is the same
when telling the truth and when lying.
subject to
Z
U (θ ) dF (θ ) = w
60
subject to
Z
U (θ ) dF (θ ) = w
υ (θ ) + θ −1 η (θ ) + βw0 (θ ) = U (θ )
U 0 ( θ ) = − θ −2 η ( θ )
where {η̂T (θ )} and {υ̂T (θ )} are the solution to the planning problem
Z h i
min eυ(θ ) + eη (θ ) dF (θ )
{υ(θ ),η (θ )}
subject to
Z
[υ (θ ) + θη (θ )] dF (θ ) = 0
υ 0 ( θ ) + θ −1 η 0 ( θ ) = 0
h0 (1 − y (θ ))
1 − τl,T (θ ) =
θu0 (c (θ ))
1
= eυ̂T (θ )−η̂T (θ )
θ
and hence is independent of history so far.
61
subject to
Z
U (θ ) dF (θ ) = w
υ (θ ) + θ −1 η (θ ) + βw0 (θ ) = U (θ )
U 0 ( θ ) = − θ −2 η ( θ )
Note that since consumption is separable from labor supply, th margin between c
and w0 is undistorted, i.e.,
1 w0 (θ )
e υ ( θ ) = β −1 q T A T e 1+E[1/θ ]
1 + E [1/θ ]
subject to
Z
U (θ ) dF (θ ) = w
υ (θ ) + θ −1 η (θ ) + βw0 (θ ) = U (θ )
U 0 ( θ ) = − θ −2 η ( θ )
Policy functions:
w
υT −1 (w, θ ) = υ̂T −1 (θ ) +
(1 + β) (1 + E [1/θ ])
w
ηT −1 (w, θ ) = η̂T −1 (θ ) +
(1 + β) (1 + E [1/θ ])
w
wT −1 (w, θ ) = ŵT −1 (θ ) +
1+β
w
PT −1 (w) = A T −1 e (1+β)(1+E[1/θ ])
subject to
Z h i h i
υ (θ ) (1 + β (1 + E [1/θ ])) + θ −1 η (θ ) dF (θ ) = β (1 + E [1/θ ]) log β−1 q T A T / (1 + E [1/θ ])
υ0 (θ ) (1 + β (1 + E [1/θ ])) + θ −1 η 0 (θ ) = 0
62
• In general, we can show that
w
υt (w, θ ) = υ̂t (θ ) +
αt
w
ηt (w, θ ) = η̂t (θ ) +
αt
0
wt (w, θ ) = αt+1 υt (w, θ ) − ψt
w
Pt (w) = At e αt
subject to
Z h i
(1 + βαt+1 ) υ (θ ) + θ −1 η (θ ) dF (θ ) = βψt
(1 + βαt+1 ) υ0 (θ ) + θ −1 η 0 (θ ) = 0
h0 1 − lt θ t
t
1 − τl,t θ =
θt u0 (ct (θ t ))
ct θ t
1 υ̂t (θt )−η̂t (θt )
= = e
θt (1 − lt (θ t )) θt
Hence, labor wedges are indpendent of history. However, they depend on t. More-
over, taxes should increase over life cycle.
63