Life-Cycle Taxation with Persistent Shocks

A Recursive Formulation*

Job Boerma, with a few edits by Jonathan Heathcote

November 7, 2021

1 Introduction

The mechanism design approach to dynamic taxation is also used to study optimal taxation of earnings

and capital over the life-cycle. In this setup, the stochastic process for labor productivity is assumed

persistent. Persistent shocks to labor productivity are used to capture the observed persistence of wage

rates and labor earnings as emphasized by the empirical labor literature. The planner problems in this

literature (e.g. Farhi and Werning (2013), Golosov, Troshkin and Tsyvinski (2016), Stantcheva (2018),

N’Diaye (2018)) are formulated and solved recursively.

In this note we extend the example on Golosov, Kocherlakota and Tsyvinski (2003) and the inverse

Euler equation to derive the recursive formulation of the insurer’s maximization problem using techniques

due to Fernandes and Phelan (2000). Specifically, we use a two period example to show the use of promise-

keeping conditions and threat-keeping conditions to write the problem recursively.

2 Life-Cycle Problem

Skill Process. The model features two periods, period 0 and 1. Labor productivity in the initial period

is denoted θ0 ∈ {θL , θH }, with θL < θH . We use πj to denote the probability of being type j ∈ {H, L}

in the initial period. Labor productivity evolves between the initial and final period, with final period

productivity θ1 ∈ {θL , θH }. Let πi|j describe the conditional probability of realizing type i in the final
*
I thank Ellen McGrattan and Monica Tran Xuan for helpful comments.

1
period after being type j in the initial period. The ordering of the subscripts on the allocations is

chronological: the first index refers to the initial period, the second script to the final period.

Preferences. To illustrate another useful technique for this literature, I assume preferences are separable

in consumption c and hours worked ` = y/θ and given by:

1+ 1
u c, y; θ = u(c) − ` η ,

where u is increasing and strictly concave, and where η > 0 is the Frisch elasticity of labor supply.

Insurer Problem. The profit-maximizing insurer has access to a linear savings technology with gross

return rate R and chooses (cH , yH , cL , yL , cHH , yHH , cHL , yHL , cLH , yLH , cLL , yLL ) to solve:

1h
i
max πH (yH − cH ) + πH πH|H (yHH − cHH ) + πL|H (yHL − cHL )
R
1h
i
+πL (yL − cL ) + πL πH|L (yLH − cLH ) + πL|L (yLL − cLL )
R

subject to an ex-ante welfare constraint:

 
πH u(cH , yH /θH ) + βπH πH|H u(cHH , yHH /θH ) + πL|H u(cHL , yHL /θL )
 
+πL u(cL , yL /θL ) + βπL πH|L u(cLH , yLH /θH ) + πL|L u(cLL , yLL /θL ) ≥ V ,

where V is the ex-ante welfare promise. The insurer’s maximization problem is constrained by four

incentive constraints. First, the insurer dissuades high productivity types from misreporting in the initial

period (assuming truthful reporting in the final period):

h i
u(cH , yH /θH ) + β πH|H u(cHH , yHH /θH ) + πL|H u(cHL , yHL /θL )
h i
≥ u(cL , yL /θH ) + β πH|H u(cLH , yLH /θH ) + πL|H u(cLL , yLL /θL ) .

where it is should be noted that final period utilities are evaluated at the true probabilities. Similarly,

the low productivity type is discouraged to misreport in the initial period (in practice we might not need

to worry about this constraint, but we will include it for now):

h i
u(cL , yL /θL ) + β πH|L u(cLH , yLH /θH ) + πL|L u(cLL , yLL /θL )
h i
≥ u(cH , yH /θL ) + β πH|L u(cHH , yHH /θH ) + πL|L u(cHL , yHL /θL ) .

Furthermore, the solution to the insurer problem is restricted to respect the incentive constraints for

high types in the final period, given their report in the initial period (we won’t worry about low types

2
misreporting in the second period).

u(cHH , yHH /θH ) ≥ u(cHL , yHL /θH ) ;

u(cLH , yLH /θH ) ≥ u(cLL , yLL /θH ) .

Note the following about the formulation of the insurer problem:

1. We assume that the IC constraints for the high productivity types are binding in the final period.

Question: Is this without loss of generality? If yes, prove it. If not, construct a counterexample.

2. We assume that after misreporting in the initial period, the agent reports truthfully in the final

period (in the initial period IC constraints). We could consider misreporting in both periods.

Question: Is our formulation without loss? If yes, prove it. If not, construct a counterexample.

3
Reformulated Insurer Problem. Before writing out the optimality conditions to the insurer problem,

we examine whether the insurer problem has a unique solution. To verify this, reformulate the problem
 1+ 1
in utility space.1 To do so, define ūi ≡ u(ci ) and h̄i ≡ θyji
η
, where i ∈ {H, L, HH, HL, LH, LL}, and
1
1
1+ η
j = L for i ∈ {L, HL, LL} and j = H otherwise. These definitions imply ci = u−1 (ūi ) and yi = θj h̄i .

The insurer chooses (ūH , h̄H , ūL , h̄L , ūHH , h̄HH , ūHL , h̄HL , ūLH , h̄LH , ūLL , h̄LL ) to solve:
1 1 1
! " ! !!#
1+ η1 1 1+ 1 1+ 1
max πH θH h̄H − u−1 (ūH ) + πH πH|H θH h̄HH − u−1 (ūHH ) + πL|H θL h̄HL − u−1 (ūHL )
η η

R
1 1 1
! " ! !!#
1+ η1 1 1+ 1 1+ 1
+πL θL h̄L − u−1 (ūL ) + πL πH|L θH h̄LHη − u−1 (ūLH ) + πL|L θL h̄LLη − u−1 (ūLL )
R

subject to:




πH ūH − h̄H + βπH πH|H ūHH − h̄HH + πL|H ūHL − h̄HL




+πL ūL − h̄L + βπL πH|L ūLH − h̄LH + πL|L ūLL − h̄LL ≥ V ,

and four incentive compatibility constraints:

h

i
ūH − h̄H + β πH|H ūHH − h̄HH + πL|H ūHL − h̄HL
 1+ 1
θL η h

i
≥ ūL − h̄L + β πH|H ūLH − h̄LH + πL|H ūLL − h̄LL ,
θH

the incentive constraint for the low type in the initial period:
h

i
ūL − h̄L + β πH|L ūLH − h̄LH + πL|L ūLL − h̄LL
 1+ 1
θH η h

i
≥ ūH − h̄H + β πH|L ūHH − h̄HH + πL|L ūHL − h̄HL ,
θL

and the incentive constraints for high types in the final period:
 1+ 1
θL η
ūHH − h̄HH ≥ ūHL − h̄HL ;
θH
 1+ 1
θL η
ūLH − h̄LH ≥ ūLL − h̄LL .
θH

The insurer problem has a strictly concave objective and a linear constraint set. The objective function
1
1
1+ η
is strictly concave because the sum of two strictly concave functions is strictly concave: h̄i is strictly

concave for η > 0 and u−1 is strictly convex given that u is increasing and strictly concave. If the insurer

problem has a unique solution, it is unique.

1
Also see Rustichini and Phelan (2017).

4
 Return to our original problem: The profit-maximizing insurer chooses (cH , yH , cL , yL , cHH , yHH , cHL , yHL , cLH

to solve:

1h
i
max πH (yH − cH ) + πH πH|H (yHH − cHH ) + πL|H (yHL − cHL )
R
1h
i
+πL (yL − cL ) + πL πH|L (yLH − cLH ) + πL|L (yLL − cLL )
R

subject to (with multplier labels attached to each constraint)

 
η :πH u(cH , yH /θH ) + βπH πH|H u(cHH , yHH /θH ) + πL|H u(cHL , yHL /θL )
 
+πL u(cL , yL /θL ) + βπL πH|L u(cLH , yLH /θH ) + πL|L u(cLL , yLL /θL ) ≥ V ,

h i
µH : u(cH , yH /θH ) + β πH|H u(cHH , yHH /θH ) + πL|H u(cHL , yHL /θL )
h i
≥ u(cL , yL /θH ) + β πH|H u(cLH , yLH /θH ) + πL|H u(cLL , yLL /θL ) .

h i
µL : u(cL , yL /θL ) + β πH|L u(cLH , yLH /θH ) + πL|L u(cLL , yLL /θL )
h i
≥ u(cH , yH /θL ) + β πH|L u(cHH , yHH /θH ) + πL|L u(cHL , yHL /θL ) .

βπH µHH : u(cHH , yHH /θH ) ≥ u(cHL , yHL /θH ) ;

βπL µLH : u(cLH , yLH /θH ) ≥ u(cLL , yLL /θH ) .

First-order Conditions. The optimality conditions with respect to (cH , cL , yH , yL ) are respectively

given by:
 
0 µH µL
1 = u (cH ) η + −
πH πH
 
0 µH µL
1 = u (cL ) η − +
πL πL
  1      1
1 yH η 1 µH µL 1 yH η 1
1= 1+ η+ − 1+
η θH θH πH πH η θL θL
   1      1
1 yL η 1 µL µH 1 yL η 1
1= 1+ η+ − 1+
η θL θL πL πL η θH θH

For consumption in the final period, (cHH , cHL , cLH , cLL ), the corresponding optimality conditions

5
are:

µL πH|L
 
1 0 µHH µH
= u (cHH ) η + + −
βR πH|H πH πH πH|H
µL πL|L
 
1 0 µHH µH
= u (cHL ) η − + −
βR πL|H πH πH πL|H
µH πH|H
 
1 µLH µL
= u0 (cLH ) η + − +
βR πH|L πL πH|L πL
µH πL|H
 
1 0 µLH µL
= u (cLL ) η − − + .
βR πL|L πL πL|L πL

The optimality conditions with respect to consumption can be used to establish that cHH ≥ cHL and that

cLH ≥ cLL . Given that the incentive constraints for high types in the second period hold with equality,

this implies yHH ≥ yHL as well as yLH ≥ yLL .

The optimality conditions with respect to labor in the final period (yHH , yHL , yLH , yLL ) are:
1
µL πH|L
   
1 1 yHH µHH
η 1 µH
= 1+ η+ + −
βR η θH πH|HθH πH πH πH|H
1 1
µL πL|L
     
1 1 yHL η 1 µH 1 yHL η 1 µHH
= 1+ η+ − − 1+
βR η θL θL πH πH πL|H η θH θH πL|H
1
µH πH|H
   
1 1 yLH η 1 µLH µL
= 1+ η+ − +
βR η θH θH πH|L πL πH|L πL
1 1
µL µH πL|H
     
1 1 yLL η 1 1 yLL η 1 µLH
= 1+ η+ − − 1+ .
βR η θL θL πL πL πL|L η θH θH πL|L

By using the optimality conditions with respect to consumption and labor, we establish that the marginal

decision for the high types in the second period is undistorted.

3. Practice Question: Derive the inverse Euler equation/equations for this setup.

This was Job’s question. We will answer it now. Take these three FOCs for cH , cHH and cHL :
 
0 µH µL
1 = u (cH ) η + −
πH πH

µL πH|L
 
1 0 µHH µH
= u (cHH ) η + + −
βR πH|H πH πH πH|H
µL πL|L
 
1 0 µHH µH
= u (cHL ) η − + −
βR πL|H πH πH πL|H

6
  
1 µH µL
= η+ −
u0 (cH ) πH πH
πH|H πL|H µL πH|L µL πL|L
     
1 µHH µH µHH µH
+ = πH|H η + + − + πL|H η − + −
βR u0 (cHH ) u0 (cHL ) πH|H πH πH πH|H πL|H πH πH πL|H
πH|H µH µL π µ
L|H H µL
= πH|H η + µHH + − π + πL|H η − µHH + − π
πH πH H|L πH πH L|L
 
µH µL
= η+ −
πH πH

So

πH|H πL|H
 
1
0
= βR 0
+ 0
u (cH ) u (cHH ) u (cHL )

This is the famous inverse Euler equation. Suppose utility is logarithmic. Then we have

cH = βR πH|H cHH + πL|H cHL

Suppose we want to define the implicit savings wedge for this allocation as the tax rate that would

satisfy a standard optimal savings condition. Thus the wedge is the solution 1 − τ to

u0 (cH ) = βR(1 − τ ) πH|H u0 (cHH ) + πL|H u0 (cHL )

 
1 1 1
= βR(1 − τ ) πH|H + πL|H
cH cHH cHL

Suppose βR = 1. If this decentralization decentralizes the optimal allocation, then combining the

inverse Euler and the decentralization FOC we have

 
1 1
0
= (1 − τ )E 0
E[c ] c

Jensen’s inequality states that if f is convex, E[f (x)] > f (E[x]). Now x1 is a convex function, so

E c10 > E[c1

 
0 ] , which implies τ > 0. Thus, the optimum implicitly taxes saving.

Recursive Problem: Profit Maximization. This two period example was fine, except that even with

only two periods we ended up with many first-order conditions. Imagine trying to compute the optimal

consumption and labor allocations for any possible history with multiple periods and multiple values for

θ. The problem will explode unless we can formulate it recursively.

So let’s write this problem in a recursive fashion, using the techniques of Fernandes and Phelan

(2000). To start, let’s stick with the assumption that there are only two periods. We solve the planner

problem after invoking the standard argument that the incentive constraints for the second period high

types has to bind.

7
 In the initial period, the planner chooses allocation (cH , yH , cL , yL ), promised utilities (VH , VL ),

and threat utilities (VHL , VLH ). Threat utility VHL is the promised utility to innate type H that reports

L. We need this object because we need to make sure that misreporting does not pay off in terms of

lifetime utility. And just as the planner can manipulate promised future values to make truthtelling more

attractive, so can she manipulate promised future values conditional on lying today in order to make

lying less attractive.

The insurer thus solves:

1h i
max πH (yH − cH ) + πL (yL − cL ) + πH Π (VH , VLH , θH ) + πL Π (VL , VHL , θL ) ,
R

where Π is the value function for the final period which we discuss below, subject to the ex-ante welfare

constraint:
 
πH U (cH , yH /θH ) + πL U (cL , yL /θL ) + β πH VH + πL VL ≥ V ,

where V is the endowment of ex-ante welfare. Furthermore, maximization of profits is initially constrained

by the two incentive compatibility constraints for the initial period:

 
0 ≤ U (cH , yH /θH ) − U (cL , yL /θH ) + β VH − VHL
 
0 ≤ U (cL , yL /θL ) − U (cH , yH /θL ) + β VL − VLH .

Comparing these expressions to the initial period incentive constraints gives an interpretation for the

promised utilities and threat utilities. The promised utility VH is the expected utility delivered in the

final period when the individual reports truthfully today. When the high productivity type reports L

instead, the expected utility is VHL . The planner chooses the threat value to prevent misreporting in the

initial period.

The constraints are respectively given multipliers η, µH , µL . As a result, the optimality conditions

to the initial period allocation (cH , yH , cL , yL ) are given by:

 
µH µL
1 = u0 (cH ) η + −
πH πH
 
0 µH µL
1 = u (cL ) η − +
πL πL
  1      1
1 yH η 1 µH µL 1 yH η 1
1= 1+ η+ − 1+
η θH θH πH πH η θL θL
   1      1
1 yL η 1 µL µH 1 yL η 1
1= 1+ η+ − 1+ ,
η θL θL πL πL η θH θH

8
(which are exactly the same conditions we had for the previous formulation) while the optimality condi-

tions with respect to future values (VH , VL , VHL , VLH ) are:

 
µH
βR η + = −Π1 (VH , VLH , θH )
πH
 
µL
βR η + = −Π1 (VL , VHL , θL )
πL
µH
βR = Π2 (VL , VHL , θL )
πL
µL
βR = Π2 (VH , VLH , θH )
πH

For the final period, we solve two separate component problems given the state variables promised

utility, threat utility, and the past realization of labor productivity type. When the initial productivity

level is θH , the final period planner problem chooses (cHH , yHH , cHL , yHL ) to solve:

Π (VH , VLH , θH ) ≡ max πH|H (yHH − cHH ) + πL|H (yHL − cHL ) ,

subject to promise keeping, threat keeping, and the incentive compatibility condition:

πH|H U (cHH , yHH /θH ) + πL|H U (cHL , yHL /θL ) ≥ VH

πH|L U (cHH , yHH /θH ) + πL|L U (cHL , yHL /θL ) ≤ VLH

U (cHH , yHH /θH ) ≥ U (cHL , yHL /θH )

The multipliers on the constraints are given by φH , φLH and µ̃HH ≡ βRµHH . Evaluating the promise

keeping and threat keeping condition, we observe the key complication faced when assuming persistent

shocks: privately observed histories of productivity shocks influence the way in which agents evaluate

continuation contracts. The evaluation of contracts designed for the final period (cHH , yHH , cHL , yHL )

depends on the individual’s type in the initial period. When the shocks are instead time independent,

the evaluation of future contracts is identical, and the threat keeping constraint is redundant.

The final period insurer problem is constrained in two ways. The promise keeping condition restricts

the planner to deliver an allocation delivering utility VH to agents reporting truthfully. At the same time,

the threat keeping conditions ensures that no gains are made by an individual misreported in the initial

period, which is ensured by delivering VLH .

9
 The optimality conditions with respect to (cHH , yHH , cHL , yHL ) are given by:

πH|L
 
0 µ̃HH
1 = u (cHH ) φH + φLH +
πH|H πH|H
1
πH|L
   
1 yHH η 1 µ̃HH
1= 1+ φH + φLH +
η θH θH πH|H πH|H
πL|L
 
µ̃HH
1 = u0 (cHL ) φH + φLH −
πL|H πL|H
1 1
πL|L
     
1 yHL η 1 1 yHL η 1 µ̃HH
1= 1+ φH + φLH − 1 +
η θL θL πL|H η θH θH πL|H

Combining the first-order conditions for type HH, we note that the optimality conditions imply that the

marginal decisions for the high type are undistorted (so the zero tax at the top result survives in the last

period). In addition, this problem gives rise to the following envelope conditions:

Π1 (VH , VLH , θH ) = −φH

Π2 (VH , VLH , θH ) = −φLH

Similarly, we solve the social planner problem given previously realized productivity level is θL .

The final period insurer chooses (cLH , yLH , cLL , yLL ) to solve:

Π (VL , VHL , θL ) ≡ max πH|L (yLH − cLH ) + πL|L (yLL − cLL ) ,

subject to promise keeping, threat keeping, and the incentive compatibility condition:

πH|L U (cLH , yLH /θH ) + πL|L U (cLL , yLL /θL ) ≥ VL

πH|H U (cLH , yLH /θH ) + πL|H U (cLL , yLL /θL ) ≤ VHL

U (cLH , yLH /θH ) ≥ U (cLL , yLL /θH )

The multipliers on the constraints are respectively given by φL , φHL , and µ̃LH ≡ βRµLH .

The optimality conditions with respect to (cLH , yLH , cLL , yLL ) are given by:

πH|H
 
0 µ̃LH
1 = u (cLH ) φL + φHL +
πH|L πH|L
1
πH|H
   
1 yLH η 1 µ̃LH
1= 1+ φL + φHL +
η θH θH πH|L πH|L
πL|H
 
µ̃LH
1 = u0 (cLL ) φL + φHL −
πL|L πL|L
1 1
πL|H
     
1 yLL η 1 1 yLL η 1 µ̃LH
1= 1+ φL + φHL − 1 +
η θL θL πL|L η θH θH πL|L

10
Note that the optimality conditions imply that the marginal decisions for the high type are undistorted.

In addition, this problem gives rise to the following envelope conditions:

Π1 (VL , VHL , θL ) = −φL

Π2 (VL , VHL , θL ) = −φHL

We fully characterized the solution to the recursive profit maximization problem. Next, we establish that

the solution to the recursive profit maximization problem aligns with the sequential profit maximization

problem.

Proposition 1. Equivalence Sequential and Recursive Profit Maximization Problem.

(cH , yH , cL , yL , cHH , yHH , cHL , yHL , cLH , yLH , cLL , yLL , η, µH , µL , µHH , µLH ) solves the sequential cost min-

imization problem given V if and only if the recursive cost minimization problem given V is solved by

(cH , yH , cL , yL , cHH , yHH , cHL , yHL , cLH , yLH , cLL , yLL , η, µH , µL , µHH , µLH ).

Proof. For the initial period, we directly observe that the optimality conditions are identical. For the final

period, an identical observation is made after using the optimality conditions and envelope conditions

with respect to promised utilities and threat utilities. Specifically, we rewrite the optimality conditions

with respect to future utilities (VH , VL , VHL , VLH ) as:

 
µH
βR η + = φH
πH
 
µL
βR η + = φL
πL
µH
βR = −φHL
πL
µL
βR = −φLH
πH

Substituting these expressions into the optimality conditions on the previou page shows the optimality

conditions are indeed identical to those from the sequence problem. 

4. We wrote the planner problem as a cost minimization problem subject to a participation constraint

for a risk-neutral insurer.

Question: Is this the dual problem of the welfare maximization problem subject to a resource

constraint?

11
5. Formulate the welfare maximization problem in recursive form.

Question: Prove equivalence between the sequential welfare maximization problem and the recursive

welfare maximization problem.

12
3 A Formulation with Multiple Values for the Shock and Multiple

Periods

The invididual states are: (i) age a, (ii) last period productivity θ−1 , (iii) expected promised value from

today onwards V

There is also (iv) a threat value Ṽ which applies to someone one type more productive than θ−1
+
yesterday – call this type θ−1 . (We could imagine a threat value for any possible false report in the

previous period, but we will assume that the only misreport we need to worry about is from a type

pretending to one type less productive than they truly are)

If this type misreported yesterday they will be pooled with the θ−1 type today.
 
Call the individual state vector s = a, θ−1 , V, Ṽ

Prior to the realization of θ, for each point in the individual state space s, the planner chooses

c(θ, s), y(θ, s), V (θ, s), Ṽ (θ, s) in order to solve

X 
1

Π(s) = max π(θ|θ−1 ) y(θ, s) − c(θ, s) + Π(a + 1, θ, V (θ, s), Ṽ (θ, s))
c(θ,s),y(θ,s),V (θ,s),Ṽ (θ,s) R
θ

subject to promise-keeping and threatkeeping

X
π(θ|θ−1 ) {u(c(θ, s), y(θ, s)/θ) + βV (θ, s)} ≥ V
θ
X
+
π(θ|θ−1 ) {u(c(θ, s), u(y(θ, s)/θ) + βV (θ, s)} ≤ Ṽ
θ

and to local downward incentive constraints

u(c(θ, s), y(θ, s)/θ) + βV (θ, s) ≥ u(c(θ− , s), y(θ− , s)/θ) + β Ṽ (θ− , s) θ = 2, ..., N

where θ− is one type less productive than θ.

At each age this problem can be solved, one value for s at at time.

Then we can move back one age.

13
References

14