Advanced Macroeconomics 2024-2025

Euler 1744 optimization, a complement

Etienne Wasmer

NYUAD

March 10, 2025

1/47
Optimization

▶ Static case
▶ Dynamic case
▶ Transversality conditions
▶ Source: the excellent Further Mathematics for Economic Analysis, Knut Sydsaeter, Peter
Hammond, Atle Seierstad, Arne Strom, Pearson Ed., ISBN-13: 9780273713289

2/47
Optimization

▶ Static case
▶ Dynamic case
▶ Transversality conditions

3/47
Static optimization
▶ Suppose that you want to maximize a function of n variables

f (x1 , x2 , ...xn )

subject to m constraints

gj (x1 , x2 , ...xn ) = bj , j = 1, ...m < n

▶ Standard method is to generalize the problem: more
unknowns, then the solution of this more general problem has
conditions
▶ which are necessary for your original problem. Implies that you
know the shape of solutions.
▶ Then restrictions lead to necessary and sufficient conditions.
▶ Pose vector notations x = (x1 , x2 , ..., xn ), b = (b1 , ..., bm ) and
g(x) = (g1 (x), g2 (x), ..., gm (x))
▶ Define an admissible solution as a vector x satisfying the
constraints. 4/47
Static optimization, Lagrangian

▶ Introduce the Lagrangian

m
X
L(x) = f (x) − λj [gj (x) − bj ]
j=1

▶ It is a function of x but the multipliers λj are also inputs.

▶ The necessary conditions for optimality of the Lagrangian are:

X ∂gj (x) m
∂L(x) ∂f
=0⇔ = λj
∂xi ∂xi ∂xi
j=1

5/47
Necessary and sufficient conditions, static optimization
1. Part 1: Necessary condition
▶ Suppose that x∗ is an optimum of the initial problem and that
functions f , gj , j = 1, m are differentiable around x∗ .
▶ Suppose that the matrix m × n of

∂g1 (x∗ ) ∂g1 (x∗ )

 
...
 ∂x1 ∂xn 
 .. .. 
. .
 
 
 ∂g (x∗ ) ∗ 
∂gm (x )
m
...
∂xi ∂xn
has rank m
▶ Then there exists unique Lagrange multipliers satisfying the
FOC of the Lagrangian.
2. Part 2: Sufficiency condition
▶ If Lagrange multipliers exist and x∗∗ maximizing the
Lagrangian,
▶ If the Lagrangian L(x) is concave in x and the set over which
functions are defined is convex.
▶ Then x∗∗ solves the original problem.
6/47
Concavity of a function: definition
▶ Concave means in general for a real differentiable function
G (y)of n variables on an open convex set S that

G (y′ ) − G (y) ≤ ∇G (y).(y′ − y)

for all y, y′ in the support S of G
▶ where . is the scalar product
▶ and ∇ is the gradient vector, that is
 
∂G (y) ∂G (y) ∂G (y)
∇G (y) = , ..., , ...,
∂y1 ∂yi ∂yn
▶ so that
n
X ∂G (y)
G (y′ ) − G (y) ≤ (yi′ − yi )
∂yi
i=1
′
for all y, y .
▶ In short in one dimension: at point y , the function
subsequently grows less fast than if it followed its original
slope in y ′ 7/47
Necessary and sufficient conditions, static optimization,
proof of part 2- sufficiency

▶ A solution to the FOCs of the Lagrangian x ∗∗ is an interior

maximum if the Lagrangian is concave.
▶ This implies that L(x∗∗ ) ≥ L(x).
▶ If x is admissible for a solution, then the constraints are
satisfied, so that gj (x) = bj .
▶ Hence, f (x∗∗ ) ≥ f (x).
▶ Then x∗∗ solves the original problem.
▶ NB: If x∗ maximizes the original problem, it does not always
maximizes the Lagrangian.
▶ NB2: The Lagrangian multipliers are a shadow price of the
associated resource.

8/47
Necessary and sufficient conditions, static optimization,
proof of part 1

▶ The proof of the first part involves the rank condition. It uses
the theorem of implicit functions.
▶ Then one finds a set of m variables satisfying the constraints,
as a function of the n − m remaining variables.
▶ And one can express the remaining free variables to get the
optimum.

9/47
If constraints are inequality conditions
▶ Suppose now that you want to maximize a function of n
variables
f (x1 , x2 , ...xn )
subject to m constraints
gj (x1 , x2 , ...xn ) ≤ bj , j = 1, ...m < n
▶ Still introduce the Lagrangian
m
X
L(x) = f (x) − λj [gj (x) − bj ]
j=1

▶ Same FOC (called Kuhn-Tucker conditions), there are also

complementary slackness conditions λj ≥ 0
▶ Sufficiency condition is that the Lagrangian is concave when
multipliers satisfy the FOC.
▶ Another sufficiency condition is that f is concave and
gj , j = 1, m are quasi-convex. NB: f (x) = x 2 is convex, but f (x) = |x|
is only quasi-convex. 10/47
Optimization

▶ Static case
▶ Dynamic case
▶ Transversality conditions

11/47
A dynamic environment

▶ Country produces Y = f (K ) with f increasing concave.

▶ Country consumes and invest:

Y = f (K ) = C + K̇

▶ NB: all variables are function of time: K = K (t) as well as

C = C (t) and so the time derivatives K̇ , Ċ .
▶ Country has collective preference U(C ) increasing concave
and discounts the future exponentially e −σt .

12/47
A dynamic environment

▶ Let T be a final period.

▶ The objective is to find the best value for
Z T
U(C )e −σt dt
0

▶ Equivalent to maximizing
Z T
U[f (K ) − K̇ ]e −σt dt (1)
0

13/47
A dynamic environment

▶ A more general version of equation (1) is

Z t1
F (t, x, ẋ)dt (2)
t0

▶ with initial and final (landing) conditions

x(t0 ) = x0 , x(t1 ) = x1 .
▶ Problem: we know how to maximize a function with respect
to x scalar or even vector: FOC or Lagrangian.
▶ But maximize an operator of functions with respect to a
function itself: we do not really know. What is the derivative
with respect to a function?

14/47
Here comes Euler!

▶ In 1744, Euler says that a solution x(t) to (2) implies that

 
∂F d ∂F
− =0 (3)
∂x dt ∂ ẋ
▶ Note the total derivative of F above. In the most general case,

∂2F ∂2F ∂2F ˙

 
d ∂F (t, x, ẋ)
= + ẋ + ẋ (4)
dt ∂ ẋ ∂t∂ ẋ ∂x∂ ẋ ∂ ẋ∂ ẋ

15/47
Here comes Euler!

▶ The most complex Euler equation is therefore:

∂F ∂2F ∂2F ∂2F ˙

= + ẋ + ẋ (5)
∂x ∂t∂ ẋ ∂x∂ ẋ ∂ ẋ∂ ẋ
▶ This is a second-order differential equation in x(t) if
∂2F
̸= 0
∂ ẋ∂ ẋ

16/47
Proof 1/N: Necessary condition

▶ Very important logic here in the proof: calculus of variations.

▶ Leads to prove that Euler is a necessary condition for
optimality and identify sufficiency conditions.
▶ Denote by x ∗ = x ∗ (t) a solution to the original problem.
▶ Consider a variation function µ(t) twice differentiable such
that µ(t0 ) = µ(t1 ) = 0 and an arbitrary constant α.
▶ Given this, a function x = x ∗ + αµ is a candidate to the
problem since it has the same initial and final values.
▶ Idea: let’s use Euler to demonstrate that if not satisfied, the
maximization is not achieved around small values of α e.g. for
small variations.

17/47
Proof 2/N: Necessary condition

▶ Denote by I (α) the value

Z t1
I (α) = F (t, x ∗ + αµ, x˙∗ + αµ̇)dt (6)
t0
▶ By definition, I (0) is the highest value since x ∗ is a solution.
▶ Therefore, one knows that I ′ (α) = 0 in α = 0.

18/47
Proof 3/N: Necessary condition

▶ That’s almost it:

Z t1
′ ∂
I (α) = F (t, x ∗ + αµ, x˙∗ + αµ̇)dt
t0 ∂α

▶ First calculate that partial derivative under the integral.

∂ ∂F ∂F
F (t, x ∗ + αµ, x˙∗ + αµ̇) = µ+ µ̇
∂α ∂x ∂ ẋ

▶ Apply it to α = 0:
Z t1  
∂ ∂
I ′ (0) = F (t, x ∗ , x˙∗ ) × µ + F (t, x ∗ , x˙∗ ) × µ̇ dt
t0 ∂x ∂ ẋ
(7)

19/47
Proof 4/N: Necessary condition

▶ Now comes the boring part: integrate by part

▶ Formula is
Z t1 Z t1
′ t1
u.v dt = [u.v ]t0 − u ′ .vdt
t0 t0

▶ Apply it to the second part of equation (7) above, with here

v (t) = µ(t).
Z t1
∂
F (t, x ∗ , x˙∗ ) × µ̇dt
t0 ∂ ẋ
 t1 Z t1  
∂ ∗ ˙∗ d ∂ ∗ ˙∗
= F (t, x , x ) × µ − F (t, x , x ) × µdt
∂ ẋ t0 t0 dt ∂ ẋ

20/47
Proof 5/N: Necessary condition

▶ Reintegrate in I ′ (0)

Z t1 Z t1  
′ ∂ d ∂
I (0) = F (t, x ∗ , x˙∗ ) × µ − ∗ ˙∗
F (t, x , x ) × µdt
t0 ∂x t0 dt ∂ ẋ
 t1
∂ ∗ ˙∗
+ F (t, x , x ) × µ
∂ ẋ t0

▶ that is,
Z t1   
′ ∂ d ∂
I (0) = F (t, x ∗ , x˙∗ ) − F (t, x ∗ , x˙∗ ) × µdt
t0 ∂x dt ∂ ẋ
 t1
∂
+ F (t, x ∗ , x˙∗ ) × µ
∂ ẋ t0

21/47
Proof 6/N: Necessary condition

▶ First use that µ is 0 in the beginning t0 and end t1 of the time

interval. So the last term disappears. One is left with only
Z t1   
′ ∂ ∗ ˙∗ d ∂ ∗ ˙∗
I (0) = F (t, x , x ) − F (t, x , x ) × µdt
t0 ∂x dt ∂ ẋ
(8)

22/47
Proof 7/N: Necessary condition

▶ Second: this has to be true for all functions µ(t) twice

differentiable (no slope discontinuity), therefore it must be
that the integrand of (8) is zero:
 
∂ ∗ ˙∗ d ∂ ∗ ˙∗
F (t, x , x ) − F (t, x , x ) = 0
∂x dt ∂ ẋ

▶ That is, x ∗ (t) satisfies the Euler equation.

23/47
Proof 8/N: Sufficiency

▶ Sufficiency is another matter. We need to show that the

objective function applied to an x ∗∗ that would satisfy Euler,
is at least higher than any other function.
▶ For that, we need to assume that F (t, x, ẋ) is concave in
(x, ẋ).
▶ Why? Suppose a function x ∗∗ (t) that satisfies the Euler
equationand the two limit conditions.
▶ And compare it to all possible functions x ( t) that only satisfy
the two limit conditions, so that x ∗∗ (t0 ) = x(t0 ), same in t1 .

24/47
Proof 9/N: Sufficiency
▶ Concave means in general for a real differentiable function
G (y)of n variables on an open convex set S that

G (y′ ) − G (y) ≤ ∇G (y).(y′ − y)

for all y, y′ in the support S of G

▶ where . is the scalar product
▶ and ∇ is the gradient vector, that is
 
∂G (y) ∂G (y) ∂G (y)
∇G (y) = , ..., , ...,
∂y1 ∂yi ∂yn
▶ so that
n
′
X ∂G (y)
G (y ) − G (y) ≤ (yi′ − yi )
∂yi
i=1
′
for all y, y .
25/47
Proof 10/N: Sufficiency
▶ Get back to the proof. The concavity of F in (x, ẋ) implies
that:

∂F (t, x ∗∗ , x ˙∗∗ )
F (t, x, ẋ) − F (t, x ∗∗ , x ˙∗∗ ) ≤ (x − x ∗∗ )
∂x
∂F (t, x ∗∗ , x ˙∗∗ )
+ (ẋ − x ˙∗∗ )
∂ ẋ
▶ Now insert the Euler equation:
 
∂ ∗∗ ˙∗∗ d ∂ ∗∗ ˙∗∗
F (t, x , x ) = F (t, x , x )
∂x dt ∂ ẋ

▶ therefore
 
∗∗ d ∂
F (t, x, ẋ) − F (t, x , x ˙∗∗ ) ≤ F (t, x , x ) (x − x ∗∗ )
∗∗ ˙∗∗
dt ∂ ẋ
∂F (t, x ∗∗ , x ˙∗∗ )
+ (ẋ − x ˙∗∗ )
∂ ẋ
26/47
Proof 11/N: Sufficiency
▶ Invert the inequality: leads to:
 
∗∗ d ∂F
˙
∗∗
F (t, x , x ) − F (t, x, ẋ) ≥ (t, x , x ) .(x ∗∗ − x)
∗∗ ˙∗∗
dt ∂ ẋ
∂F (t, x ∗∗ , x ˙∗∗ ) ˙∗∗
+ (x − ẋ)
∂ ẋ
▶ Now the miracle: the right-hand side can be rewritten as:
 
∗∗ ˙∗∗ d ∂F ∗∗ ˙∗∗ ∗∗
F (t, x , x ) − F (t, x, ẋ) ≥ (t, x , x ).(x − x)
dt ∂ ẋ
(9)
▶ Why? Just redevelop.
 
d ∂F ∗∗ ˙∗∗ ∗∗
(t, x , x ).(x − x)
dt ∂ ẋ
 
d ∂F
= (t, x , x ). (x ∗∗ − x)
∗∗ ˙∗∗
dt ∂ ẋ
d ∂F
+ (x ∗∗ − x) (t, x ∗∗ , x ˙∗∗ )
dt ∂ ẋ 27/47
Proof 12/N: Sufficiency

▶ Statr from the inequality (9)

 
∗∗ d ∂F
F (t, x , x ˙∗∗ ) − F (t, x, ẋ) ≥ (t, x ∗∗ , x ˙∗∗ ).(x ∗∗ − x)
dt ∂ ẋ

and integrate it between t0 and t1 :

Z t1
F (t, x ∗∗ , x ˙∗∗ ) − F (t, x, ẋ) dt ≥


t
Z0 t1  
d ∂F ∗∗ ˙∗∗ ∗∗
(t, x , x ).(x − x)
t0 dt ∂ ẋ
 t1
∂F ∗∗ ˙∗∗ ∗∗
= (t, x , x ).(x − x)
∂ ẋ t0

28/47
Proof 13/N: Sufficiency

▶ And from the limit conditions (boundary conditions) of both

x, x ∗∗ , the right hand side is zero.
▶ So we just proved that
Z t1
F (t, x ∗∗ , x ˙∗∗ ) − F (t, x, ẋ) dt ≥0
 
t0

for all admissible functions.

▶ which implies that x ∗∗ (t) is indeed a solution to the original
problem. QED (quod erat demonstrandum).

29/47
Optimization

▶ Static case
▶ Dynamic case (continued)
▶ Transversality conditions

30/47
Euler is back!

▶ Euler found that an optimal solution must satisfy:

 
∂F d ∂F
− =0 (10)
∂x dt ∂ ẋ

with
∂2F ∂2F ∂2F ˙
 
d ∂F
(t, x, ẋ) = + ẋ + ẋ (11)
dt ∂ ẋ ∂t∂ ẋ ∂x∂ ẋ ∂ ẋ∂ ẋ

31/47
Apply it to our ”nation’s problem”

▶ Problem was to find the best capital accumulation policy -

and then deduce consumption.
Z T
max U[f (K ) − K̇ ]e −σt dt (12)
K 0

▶ That is
F (t, K , K̇ ) = U[f (K ) − K̇ ]e −σt
▶ Euler says

∂F ∂2F ∂2F ∂2F ˙

= + K̇ + K̇ (13)
∂K ∂t∂ K̇ ∂K ∂ K̇ ∂ K̇ ∂ K̇

32/47
Apply it to our ”nation’s problem”

▶ That is, with

F (t, K , K̇ ) = U[f (K ) − K̇ ]e −σt

∂F
= f ′ (K )U ′ [f (K ) − K̇ ]e −σt
∂K
∂F
= −U ′ ([f (K ) − K̇ ])e −σt
∂ K̇

▶ And so the RHS is

∂2F ∂2F ∂2F ˙

+ K̇ + K̇ =
∂t∂ K̇ ∂K ∂ K̇ ∂ K̇ ∂ K̇
σU ′ (C )e −σt − f ′ (K )U”(C )e −σt K̇ + U”(C )e −σt K̇˙

33/47
Apply it to our ”nation’s problem”

▶ Combining:

U”(C ) h ′ ˙
i
f ′ (K ) − σ = − f (K )K̇ − K̇
U ′ (C )
▶ How do we continue?

34/47
Apply it to our ”nation’s problem”
▶ Write it again:
U”(C ) h ′ ˙
i
f ′ (K ) − σ = − f (K )K̇ − K̇
U ′ (C )
▶ Now use:
C + K̇ = f (K )
▶ That is:
Ċ + K̇˙ = f ′ (K )K̇
▶ So we can replace and get:
U”(C )
f ′ (K ) − σ = − Ċ
U ′ (C )

CU”(C ) Ċ
f ′ (K ) − σ = −
U ′ (C ) C

Ċ f ′ (K ) − σ
=
C θ
our regular Euler equation. 35/47
Apply it to our ”nation’s problem”

▶ Why? We had earlier this semester:

c˙t log (βR)

=
ct θ

▶ log (β) = −log (1 + σ) = −σ when the time period goes to

zero.
▶ log (R) = log (1 + f ′ (K )) = f ′ (K ) when again the time period
to express the returns goes to zero.
▶ Implication: if return to capital too large, consumption will
increase; reverse if too small.
▶ Consumption stop progressing at the point where return to
investment will be equal to discounting.

36/47
Optimization

▶ Static case
▶ Dynamic case
▶ Transversality conditions

37/47
Return to the original problem: transversality

▶ One want to maximize the following objective:

Z t1
F (t, x, ẋ)dt (14)
t0

▶ With initial condition x(t0 ) = x0

▶ Final (landing) one : either i) x(t1 ) free or ii) x(t1 ) ≥ x1 .
▶ Example: capital stock cannot become negative, or
no-Ponzi-game, or debt cannot be too large (limit cannot be
too negative).

38/47
Return to the original problem: transversality
Theorem
1. Necessary condition:
▶ If x ∗ is an optimal solution to the problem, it must solve the
Euler equation.
▶ In case i) of a free value x ∗ (t1 ), one has

∂F
(t = t1 ) = 0
∂ ẋ

▶ In the case of case ii) this is only

∂F
(t = t1 ) ≤ 0
∂ ẋ
with inequality if x ∗ (t1 ) = x1 and equality if x ∗ (t1 ) > x1 .
2. Sufficient condition
▶ Reciprocally, if F (t, x, ẋ) is concave in (x, ẋ), then an
admissible solution x ∗∗ satisfying Euler and the transversality
condition will solve the original problem.
39/47
Return to Proof 1/N: Necessary condition

▶ Calculus of variations.
▶ Denote by x ∗ = x ∗ (t) a solution to the original problem.
▶ Consider a variation function µ(t) twice differentiable such
that µ(t0 ) = 0 (but not necessarily in t1 ) and an arbitrary
constant α.
▶ Given this, a function x = x ∗ + αµ is a candidate to the
problem since it has the same initial value.

40/47
Return to Proof 2/N: Necessary condition

▶ Denote by I (α) the value

Z t1
I (α) = F (t, x ∗ + αµ, x˙∗ + αµ̇)dt (15)
t0
▶ By definition, I (0) is the highest value since x ∗ is a solution.
▶ Therefore, one knows that I ′ (α) = 0 in α = 0.

41/47
Return to Proof 3/N: Necessary condition

▶ That’s almost it:

Z t1
′ ∂
I (α) = F (t, x ∗ + αµ, x˙∗ + αµ̇)dt
t0 ∂α

▶ First calculate that partial derivative under the integral.

∂ ∂F ∂F
F (t, x ∗ + αµ, x˙∗ + αµ̇) = µ+ µ̇
∂α ∂x ∂ ẋ

▶ Apply it to α = 0:
Z t1  
′ ∂ ∗ ˙∗ ∂ ∗ ˙∗
I (0) = F (t, x , x )µ + F (t, x , x )µ̇ dt (16)
t0 ∂x ∂ ẋ

42/47
Return to Proof 4/N: Necessary condition

▶ Now comes the boring part: integrate by part

▶ Formula is
Z t1 Z t1
u.v ′ dt = [u.v ]tt10 − u ′ .vdt
t0 t0

▶ Apply it to equation (7)(in fact here 15) above:

Z t1
′ ∂
I (part2) = F (t, x ∗ , x˙∗ )µ̇dt
t ∂ ẋ
0  Z t1  
∂ d ∂
= F (t, x ∗ , x˙∗ )µ tt10 − F (t, x ∗ , x˙∗ ) µdt
∂ ẋ t0 dt ∂ ẋ

43/47
Return to Proof 5/N: Necessary condition

▶ Reintegrate in I ′ (0)

Z t1 Z t1  
′ ∂ d ∂
I (0) = F (t, x ∗ , x˙∗ )µ − ∗ ˙∗
F (t, x , x ) µdt
t ∂x t0 dt ∂ ẋ
 0 
∂ ∗ ˙∗
+ F (t, x , x )µ tt10
∂ ẋ
▶ that is,
Z t1   
′ ∂ d ∂
I (0) = F− F µdt
t0 ∂x dt ∂ ẋ
 
∂
+ F (t, x ∗ , x˙∗ )µ t1
t0
∂ ẋ

44/47
Return to Proof 6/N: Necessary condition

▶ First use that µ is 0 in the beginning and possibly at the end

of the time interval. So the last term half disappears in 0 and
all depends on t1 .
Z t1     
′ ∂ d ∂ ∂F (t)
I (0) = F− F µdt + µ(t) (t1 )
t0 ∂x dt ∂ ẋ ∂ ẋ

45/47
Return to Proof 6/N, continued: Necessary condition

▶
Z t1     
′ ∂ d ∂ ∂F (t)
I (0) = F− F µdt + µ(t) (t1 )
t0 ∂x dt ∂ ẋ ∂ ẋ

▶ Therefore: two cases.

▶ First. If we have a free value for t1 , any µ(t1 ) is admissible so
∂F (t)
one needs to land in (t1 ) = 0
∂ ẋ

46/47
Return to Proof 6/N, continued: Necessary condition

▶ Second case:
Z t1     
′ ∂ d ∂ ∂F (t)
I (0) = F− F µdt + µ(t) (t1 )
t0 ∂x dt ∂ ẋ ∂ ẋ

▶ If one imposes instead that x ∗ (t1 ) is above x1 : two sub-cases,

x ∗ (t1 ) can be strictly above or exactly at the limit.
▶ It the candidate solution x ∗ is strictly above the limit, again
the partial derivative of F with respect to ẋ is zero for I ′ (0) to
be zero.
▶ It the candidate solution x ∗ is exactly at the limit, then it
means that there are many admissible functions such that
x(t1 ) = x ∗ (t1 ) + αµ(t1 ) ≥ x1 which implies µ(t1 ) ≥ 0 for all
positive α. Therefore, 0 = I ′ (α) ≥ I ′ (0) so I ′ (0) ≤ 0. This
implies that the partial derivative of F with respect to ẋ has to
be negative.

47/47