Notes on Intertemporal Optimization

Econ 204A - Henning Bohn *

Most of modern macroeconomics involves models of agents that optimize

over time. The basic ideas and tools are the same as in microeconomics,
notably the Lagrange multiplier technique. At a technical level, the main
differences are interpretational or notational: In macroeconomics, goods are
usually “dated commodities” (see Debreu, Theory of Value, ch.2) that are
consumed at different times and their relative prices can often be expressed
in terms of interest rates or discount factors. Part 1 of this note will
explain the linkage. In terms of economic intuition, time adds two interesting
dimensions. First, it is sometimes instructive to keep track of agents’ asset
positions as time passes. Second, with uncertainty, new information may arrive
that enables agents to make purchases of later-dated commodities contingent on
information on which early consumption cannot be conditioned. Part 2 of this
note will explain the relation between overall (“intertemporal”) budget
constraints and the period-by-period budget equations designed to track assets
positions. To avoid “math overload,” stochastic issues will be deferred until
later. Part 3 will survey alternative solution techniques and Part 4 examines
infinite horizon problems.

Micro and Macro Problems:

Our first objective is to see that the intertemporal decision problems in

macroeconomics are in principle nothing new. With a little re-labeling, they
fit nicely into the standard Lagrangian optimization approach familiar from
microeconomics courses.

Consider the following microeconomics problem. A consumer has

preferences over N commodities, u=u(c1,...,cN), increasing and concave, where
(c1,...,cN) is the consumption bundle. The consumer has endowments (y1,...,yN)
of these commodities. On an individual level, goods prices (p1,...,pN) are
given. Then the utility-maximization problem is

* Provided on the web for use by UCSB students.

(C) Copyright 2001 Henning Bohn

- 1 -
Problem#1:
Maximize u=u(c1,...,cN)

over choice variables c1,...,cN,

N N
(1) s.t. ∑ pi·ci = ∑ pi·yi.
i=1 i=1

The summation notation will be used frequently: For any variable x, read
N
∑ xi = x1 +...+ xN.
i=1

With λ as the Lagrange multiplier on the constraint, the first order

conditions (FOC for short) are
ui(c1,...,cN) = Λ·pi, i=1,..N

where subscripts denote partial derivatives (as in ui = ∂u/∂ci). Provided

p1≠0, this can be simplified to
ui p
(2) = i for i=2,..,N.
u1 p1

That is, consumption depends only on relative prices. The constraint (1) and
the FOC (2) form a system of N equations in the N choice variables c1,...,cN.
Often, the system cannot be solved in closed form. We can, however, say
something about the structure of the solution.
The resulting choice vector (c1,...,cN) is a function of the exogenous
variables of the problem, endowments and relative prices. Note that endowments
only enter through the r.h.s. of (1). It is therefore convenient to define
“wealth” W = ∑N
i=1 pi·yi. Then the problem yields a system of demand

functions:
c1=c1(W/p1,p2/p1,..pN/p1)

c2=c2(W/p1,p2/p1,..pN/p1)

.....
cN=cN(W/p1,p2/p1,..pN/p1).

Often, utility is only well-defined for non-negative consumption. To be

precise one should then add the non-negativity conditions ci≥0, i=1,...,n, to
Problem#1 and work with the relevant Kuhn-Tucker conditions. In macro, we will
usually skip these complications and implicitly assumed that u(·) is such that
we can focus on interior solutions.

For the special case of N=2, the consumer problem reduces to

MAX: u(c1,c2) s.t. p1·ci+p2·c2 = W

- 2 -
and can be graphed in the usual (c1,c2) indifference curve diagram.

So much for microeconomics. Please alert me or the TA if the above problem is

not an easy problem for you. To test your abilities, derive the demand
functions for the case of log-utility.

Turning to macro, consider an economy with a large number of agents. To

start, assume the economy operates for T+1 periods only, periods t=0,...,T.
Agents have preferences over consumption ct in each period, u=u(c0,...,cT),
and they have incomes yt, t=0,...,T. (Think of per-period consumption as a
fixed-weight basket of different physical commodities; one could keep track of
different commodities, but that would distract from the macro issues.) Let pt
be the period-0 price of a discount bond with a payoff in period t equal to
the cost of one unit of consumption ct; define p0=1. Then pt=pt/p0 is the
relative price of period-t over period-0 consumption and the consumer budget
constraint can be written as
T T
∑ pt·ct = ∑ pt·yt.
t=0 t=0

That is, the present value of consumption equals the present value of income.
This constraint is known as the intertemporal budget constraint (IBC). The
consumer maximization problem is then

Problem#2:
Maximize u=u(c0,...,cT)

over choice variables c0,...,cT,

T T
(3) s.t. ∑ pt·ct = ∑ pt·yt
t=0 t=0

Note that Problem#1 and Problem#2 are virtually identical, except for slightly
different notation. The lesson is that you should feel comfortable applying
all your microeconomics knowledge to macroeconomics. For example, for given
specifications of the utility function, you should be able to derive demand
functions expressing period-t consumption as function of the relative prices
and of the present value of income.

Sometimes, you will have to do a little algebra to express a macro

problem in the above format. Notably, it is often convenient in macro to
express relative prices in terms of interest rates and discount factors. Let
rt, t=1,..,T, be a sequence of exogenous market interest rates, where rt is
the real interest rate on a bond issued in period t that matures in period

- 3 -
t+1. That is, consumers obtain (1+rt) units of period-(t+1) consumption for
reducing period-t consumption by 1 unit. Then the relative cost of period-
(t+1) consumption ct+1 in terms of period-t consumption is
ρt = pt/pt-1 = 1/(1+rt),

the period-t discount factor. Going back from t to t-1, to t-2, and so on to
period-0, one finds that the cost of period-t consumption ct relative to
period-0 consumption c0 is
1
(4) pt = ρt·pt-1 = ρt·ρt-1·...·ρ1 =
∏tj=1 1+rj

The lesson is that all the relative prices pt in Problem#2 above can be
expressed in terms of interest rates and/or discount factors.

Asset Accumulation:

Thinking about an intertemporal optimization problem as a re-labeled static

problem is useful from a technical perspective, but it does not do justice to
the dynamic problem. Consumption and other activities happen sequentially; the
timing of consumption and income is usually not perfectly synchronized,
forcing consumers to borrow or lend on financial markets; and the observable
dynamics of asset and debt accumulation are often the economically most
interesting aspect of an intertemporal problem. Because of economists interest
in asset positions, intertemporal problems are often set up in a way that
looks quite different than Problem#2 above.

The alternative starting point is a series of period-by-period budget

equations that trace out the process of asset accumulation. Let bt be the bond
holdings (positive or negative; if negative, meaning loans) of an agent at the
end of period t, t=0,...,T. Given initial assets (1+rt)·bt-1 at the start of
period t, income yt, and consumption ct, we obtain a end-of-period assets as
sum of savings yt-ct and initial assets,

(5) bt = yt - ct + (1+rt)·bt-1.

If we define (1+r0)·b-1=0, this equation applies for all periods t=0,...,T.

The individual choice variables are then sequences of bt and ct.

Note that equations (5) are mere accounting identities but not really
constraints, because they do not impose a limit on the level of bt. To
constrain agents choices, an end-point restriction must be imposed. The
natural one is bT≥0, which prevents agents from borrowing without repayment.

- 4 -
To get rid of the inequality (which would require a Kuhn-Tucker approach),
note that rational consumers do not leave positive assets at the end. Hence,
one may treat the endpoint restriction as an equality: bT=0.

Then one can formulate the consumer optimization problem as

Problem#3:
Maximize u=u(c0,...,cT)

over choice variables c0,...,cT, b0,...,bT-1,

s.t. bt = yt-ct + (1+rt)·bt-1 for t=0,...,T

where bT = 0, b-1 = 0 are given.

This looks like a more complicated problem than Problem#2 because it involves
more variables and equations: Problem#2 has T+1 choice variables and 1
constraint. Problem#3 has 2·T+1 choice variables and T+1 constraints. On the
other hand, Problem#3 provides more insights about what the consumer actually
does over time, how savings and bond holdings evolve.

Our objective is to show that Problems #2 and #3 are actually

identical.1 To see the equivalence, multiply (5) by the relative price pt and
exploit (4),
pt·bt = pt·yt - pt·ct + pt·[(1+rt)·bt-1]

= pt·yt - pt·ct + pt·[pt-1/pt·bt-1]

= pt·yt - pt·ct + pt-1·bt-1

This can be interpreted as the budget equation discounted back to period-0.

Then sum over t,
T T T T-1
∑ [pt·bt] = ∑ [pt·yt] - ∑ [pt·ct] + ∑ [pt·bt]
t=0 t=0 t=0 t=0
keeping in mind that b-1=0. Canceling the offsetting part of the pt·bt sums,
one finds
T T
(6) pT·bT = ∑ [pt·yt] - ∑ [pt·ct]
t=0 t=0

So far, we have only worked with accounting identities, equations (4) and (5).
Equation (6) shows that the budget equations (5) are equivalent to the

1 If one treated b as choice variable, too, one more choice variable and one more
T
constraint, bT≥0, would have to be added; that would be equivalent, too. I leave it as an
exercise to show that the optimal solution will always satisfy bT=0, provided utility is
strictly increasing.

- 5 -
intertemporal budget constraint (3) if and only if pT·bT=0. Hence, the
endpoint restriction bT=0 is crucial (assuming pT≠0). With this conditions,
Problems #2 and #3 are equivalent because they have the equivalent
constraints.
Now suppose you are given an optimization problem like #3. Doing the
steps from eq.(5) to eq.(6) in reverse, you can always substitute away the
asset positions and solve the problem as a pure consumption problem of the
form #2, using your knowledge of static optimization. Given the optimal
consumer demand functions, you can read off the asset positions from eq.(5).

With regard to Problem#3, you may wonder why anyone would ever try to
solve this problem rather than transform it into Problem#2. One answer is that
the FOC of Problem#3 involves Lagrange multipliers that have a useful
interpretation as shadow values. The other answer is that the transformation
from a dynamic to a quasi-static problem does not always work as easily as in
case of Problem#3.

Regarding the FOC of Problem#3, it is sufficient for our purposes to

examine a more restricted class of utility functions, time-additive utility.
Time-additive utility is defined as
T
(7) u(c1,..,cT) = ∑ βt·U(ct)
t=0

where 0<β<1 is the individual time-discount factor. The discount factor can
also be written in terms of the rate of time preference τ = 1/β-1 as
β=1/(1+τ). [Warning: Don’t confuse r and τ! Time preference is a matter of
personal preferences, potentially heterogenous across agents. The interest
rate characterizes individuals trading opportunities on financial markets; it
affects individuals constraints, but not preferences.] The Lagrangian problem
can then be written as
T T
Maximize L = ∑ βt·U(ct) + ∑ λ t·[yt-ct+(1+rt)·bt-1-bt]
t=0 t=0
where λ t are the multipliers for each period t=0,..,T.

The first order conditions with respect to ct are

(8a) βt·U’(ct) - λ t = 0, t=0,..,T

since ct appears in the period-t objective and in the period-t constraint.

For bt, it is useful to write out the set of Lagrange terms as

- 6 -
 + λ 0·[y0-c0-b0]

+ λ 1 · [y1 - c1 + (1+r1)· b0 - b1]

+ ...
+ λ T-1·[yT-1-cT-1+(1+r1)·bT-2-bT-1]

+ λ T·[yT-cT+(1+rT)·bT-1],

to see that bt appears negatively in the period-t constraint and positively,

multiplied by (1+rt+1), in the period-t+1 constraint. Hence, the FOC with
respect to bt are

(8b) (1+rt+1)·λ t+1 - λ t = 0, t=0,..,T-1

The multiplier λ t can be interpreted as the shadow value of resources in

period t. According to (8a), individuals consume until marginal utility equals
the shadow value of resources. For concave utility, U’(·) is a decreasing
function of ct, so that a low shadow value implies high consumption, and vice
versa. According to (8b), it is optimal to save in period t (reduce ct) until
a real dollar in period t has the same utility value as (1+rt+1) real dollars
in period t+1.2

Regarding complications, two examples may suffice. First, consider

credit constraints: Suppose consumers cannot borrow more than the amount B, so
that the credit constraints bt≥-B apply for all t. Such constraints can easily
be added to Problem#3 as Kuhn-Tucker constraints; but substituting away the
bond holdings would be a mess--not worth doing. Similar problem arise if loan
rates and deposit interest rates differ. Second, consider a production
problem: To save in period t, individuals invest in capital kt+1. In period
t+1, capital is used to produce goods yt+1=f(kt+1), where f(·) is an increasing
and concave production function. The budget equation in period t would be kt+1
= f(kt) - ct, replacing (5) in Problem#3. We will examine this problem in more
detail later. The marginal return to capital f’(k) could still be interpreted
as an interest factor, but the interest rate would no longer be exogenous;
hence, one would not gain much by writing the problem in a present value form.

2 Equation (8a) reveals why this approach is most useful for time-additive utility. Here, λ
t
depends only on ct. For other utility functions, marginal utility depends on consumption in
all periods. Then the FOC ut(c0,...,cT)=λt are much less insightful.

- 7 -
Solution Techniques:

For now, we will stick to Problem#3 in unmodified form and examine different
ways in which it can be solved. For more complicated problems, one or more of
the same approaches will usually work. The principal techniques are:

(a) The Lagrange multiplier method (Applied to #2 or #3)

(b) Substituting the constraints into the objective function and then
solving an unconstrained problem.

(c) Dynamic Programming.

The Lagrange multiplier methods were discussed above. The dynamic

programming approach will be deferred until we deal with stochastic problems,
because it is most useful in that context. This leaves case (2), substituting
the constraints into the objective. The idea is to write (5) as
(9) ct = yt - bt + (1+rt)·bt-1

and to substitute into the objective. Then the problem reduces to

Problem#4:
Maximize u=u(y0-b0, y1-b1+(1+r1)b0,...

...,yT-1-bT-1+(1+rT-1)bT-2, yT+(1+rT)bT-1),

over choice variables b0,...,bT-1,

i.e., an unconstrained problem with T variables. This is again a problem that

looks easier than Problem#3. One a solution for b0,...,bT-1 is obtained, the
implied consumption path can easily be computed from (9).

To compare the different solution techniques, it is instructive to

compare the structure of the FOC for the case of time-additive utility. For
problem #3, this was done above. With time-additive utility, Problem#2 calls
for maximizing the Lagrangian
T T T
L = ∑ βt·U(ct) + Λ·[ ∑ pt·yt- ∑ pt·ct],
t=0 t=0 t=0

which yields first order conditions:

(10) βt·U’(ct) = pt·Λ.

Comparing (8a) and (10), one can see that the λt constraints can be
interpreted as the period-t values of the period-0 shadow value Λ, λ t = pt·Λ.

In Problem#4, each bt appears twice in the objective function so that

the FOC are
(11) U’(yt-bt+(1+rt)bt-1) = (1+rt+1)·β·U’(yt+1-bt+1+(1+rt)bt)

- 8 -
for t=0,...,T-1; they form a system of T equations in the T choice variables.

If one eliminates the multipliers from (8a-b) and (10) and inserts (9)
into (11), the FOC for all three problems (#2, #3, and #4) reduce to the same
optimality conditions for consumption:
β·U’(ct+1) p 1
(12) = t+1 = ρt+1 = , for t=0,..,T-1.
U’(ct) pt 1+rt+1

That is, the marginal rate of substitution of consumption between adjacent

periods must equal the discount factor (interest rate).

For small T, these FOC and the various constraints provide enough
conditions that one can compute the optimal solution either algebraically or
numerically.

For larger T, it is worth emphasizing that we are dealing with

difference equations and that the different problems lead to different types
of difference equations.

(a) Equation (12) represents a first-order non-linear difference equation for

consumption. For t=0,..,T-1, this provides T equations for the T+1 unknown
consumption values. For any “guess” of c0, a sequence
U’(ct)
ct+1 = (U’)-1((1+r )
t+1)·β
that pins down the “trajectory” of the consumption path. The starting value c 0
(the boundary condition) can be obtained by inserting the alternative
sequences into the IBC. For example, consider U(ct) = ln(ct), rt=r, yt=y for
all t. Then the FOC 1/ct = (1+rt+1)·β·1/ct+1 imply the difference equation
ct+1 = (1+r)·β·ct. Inserting ct = (1+r)t·βt·c0 and pt = 1/(1+r)t into the IBC
implies
T T (1+r)t·βt·c T
0 1-βT+1
∑ pt·ct = ∑ = ∑ βt-1·c0 = 1-β ·c0
t=0 t=1 (1+r)t
t=0
T T T+1
1-ρ
∑ pt·yt = ∑ ρt·y = 1-ρ ·y
t=0 t=1
1-β 1-ρT+1
=> c0 = · · y
1-βT+1 1-ρ

(b) Equation (11) represents a second-order difference equation for bond

holdings. This formulation will be more useful later when we deal with
capital, because (11) is often algebraically messy. The general point is that
the optimal solution can be characterized in terms of “state variables” that
characterize initial conditions in a period; here, (1+rt)·bt-1. In general, a

- 9 -
second-order difference equation needs two boundary conditions. Here, we have
bT = b-1 = 0, the initial and the endpoint conditions.

(c) Equations (8a), (8b), and (5) represent a system of first-order difference
equations for bond holdings, consumption and the Lagrange multipliers:
bt - bt-1 = yt - ct + rt·bt-1 for t=0,..,T
rt+1
λ t+1-λ t = - ·λ for t=0,..,T-1
1+rt+1 t

λ t = βt·U’(ct) for t=0,..,T

with boundary conditions b-1 = b T =0. This looks quite complicated here, but
turns out to be very convenient in the limit when the time interval becomes
small and is known as the “Maximum principle” (see Dixit, ch.10).

Infinite Horizon Problems

Actual economies do not end. In macroeconomics, we are therefore often

interested in the case T->∞. For preferences, time additivity is again
convenient to deal with the limit case. The objective function is assumed to
be
∞
u = ∑ βt·U(ct),
t=0

provided the sum converges. The main question is the how to write the
constraints. A very direct approach is to work with the IBC. Assuming that the
relevant sums in the IBC (3) converge, maximize utility subject to
∞ ∞
(13) ∑ pt·ct = ∑ pt·yt
t=0 t=0
The first order conditions are again βt·U'(ct) = pt·Λ, where Λ = U'(c0). This
is a system with an infinite number of equations, but sometime easy to solve,
e.g., in the log-utility case.

Alternatively, one may want work with the budget equations (5). Then the
main question is what to do about the terminal condition bT≥0. For any finite
T, recall that Problems #2 and #3 are equivalent, if and only if pT·bT = 0. In
the limit, the problems are therefore equivalent if and only if
(14) limT->∞ pT·bT = 0

is satisfied, the so-called the transversality condition. Note that for

rt=r>0, p T = (1+r)-T ->0 for T->∞. Hence, (14) allows asset position to grow
exponentially, provided the rate of growth is less than the interest rate. The

- 10 -
first order conditions

βt·U'(ct) = λ t,

and (1+rt+1)·λ t+1 = λ t,

or equivalently,
U'(ct) = (1+rt+1)·β·U'(ct+1)

again form a set of difference equations that are similar as in the case of
finite horizons, except that the boundary conditions are different: bT=0 is
replaced by (14), and/or (3) is replaced by (13).

- 11 -