Viscosity Solutions of Hamilton-Jacobi Equations

and Optimal Control Problems

(an illustrated tutorial)

Alberto Bressan

Department of Mathematics, Penn State University

Contents

1 - Preliminaries: the method of characteristics 1

2 - One-sided differentials 4
3 - Viscosity solutions 9
4 - Stability properties 10
5 - Comparison theorems 12
6 - Control theory 19
7 - The Pontryagin maximum principle 21
8 - Extensions of the P.M.P. 29
9 - Dynamic programming 35
10 - The Hamilton-Jacobi-Bellman equation 38
References 43

0
1 - Preliminaries: the method of characteristics
A first order, scalar P.D.E. has the form

F (x, u, ∇u) = 0 x ∈ Ω ⊆ IRn . (1.1)

.
It is convenient to introduce the variable p = ∇u, so that (p1 , . . . , pn ) = (ux1 , . . . , uxn ). We assume
that the F = F (x, u, p) is a continuous function, mapping IRn × IR × IRn into IR.

Given the boundary data

u(x) = ū(x) x ∈ ∂Ω, (1.2)
a solution can be constructed (at least locally, in a neighborhood of the boundary) by the classical
method of characteristics. The idea is to obtain the values u(x) along a curve s 7→ x(s) starting
from the boundary of Ω, solving a suitable O.D.E. (figure 1.1).

x(s)

y
figure 1.1

Fix a point y ∈ ∂Ω and consider a curve s 7→ x(s) with x(0) = y. Call

. ¡ ¢ . ¡ ¢ ¡ ¢
u(s) = u x(s) , p(s) = p x(s) = ∇u x(s) .

We seek an O.D.E. describing the evolution of u and p = ∇u along the curve. Denoting by a dot
the derivative w.r.t. the parameter s, we clearly have
X X
u̇ = uxi ẋi = pi ẋi , (1.3)
i i

X
ṗj = uxj xi ẋi . (1.4)
i

In general, ṗj thus depends on the second derivatives of u. Differentiating the basic equation (1.1)
w.r.t. xj we obtain
∂F ∂F X ∂F
+ uxj + uxi xj = 0 . (1.5)
∂xj ∂u i
∂p i

1
Hence
X ∂F ∂F ∂F
uxj xi = − − pj . (1.6)
i
∂p i ∂x j ∂u

If we now make the choice ẋi = ∂F/∂pi , the right hand side of (1.4) is computed by (1.6). We
thus obtain a system where the second order derivatives do not appear:

 ∂F

 ẋi = i = 1, . . . , n ,

 ∂pi


 X ∂F
u̇ = pi , (1.7)
 ∂pi

 i



 ∂F ∂F
 ṗj = − − pj j = 1, . . . , n .
∂xj ∂u

This leads to a family of Cauchy problems, which in vector notation take the form
 
 ∂F  x(0) = y

 ẋ = 


 ∂p 


 

∂F
u̇ = p · u(0) = u(y) y ∈ ∂Ω . (1.8)

 ∂p 


 

 
 ṗ = − ∂F − ∂F · p
 

∂x ∂u p(0) = ∇u(y)

The resolution of the first order boundary value problem (1.1)-(1.2) is thus reduced to the solution
of a family of O.D.E’s, depending on the initial point y. As y varies along the boundary of Ω, we
expect that the union of the above curves x(·) will cover a neighborhood of ∂Ω, where our solution
u will be defined.

Remark 1.1. If F is linear w.r.t. p, then the derivatives ∂F/∂pi do not depend on p. Therefore,
the first two equations in (1.7) can be solved independently, without computing p from the third
equation.

Example 1.2. The equation

|∇u|2 − 1 = 0 x∈Ω (1.9)
on IR2 corresponds to (1.1) with F (x, u, p) = p21 + p22 − 1. Assigning the boundary data

u=0 x ∈ ∂Ω ,

a solution is clearly given by the distance function

u(x) = dist (x, ∂Ω) .

The corresponding equations (1.8) are

ẋ = 2p , u̇ = p · ẋ = 2 , ṗ = 0 .

Choosing the initial data at a point y we have

x(0) = y, u(0) = 0, p(0) = n ,

2
 y2

γ
−
x
x(s) Ω
y
y
1

figure 1.2

where n is the interior unit normal to the set Ω at the point y. In this case, the solution is
constructed along the ray x(s) = y + 2sn, and along this ray one has u(x) = |x − y|. Assuming that
the boundary ∂Ω is smooth, in general the distance function will be smooth only on a neighborhood
of this boundary. If Ω is bounded, there will certainly be a set γ of interior points x̄ where the
distance function is not differentiable (fig. 1.2). These are indeed the points such that
dist (x̄, ∂Ω) = |x̄ − y1 | = |x̄ − y2 |
for two distinct points y1 , y2 ∈ ∂Ω.

The previous example shows that, in general, the boundary value problem for a first order
P.D.E. does not admit a global C 1 solution. This suggests that we should relax our requirements,
and consider solutions in a generalized sense. We recall that, by Rademacher’s theorem, every
Lipschitz continuous function u : Ω 7→ IR is differentiable almost everywhere. It thus seems
natural to introduce

Definition 1.3. A function u is a generalized solution of (1.1)-(1.2) if u is Lipschitz continuous on

the closure Ω, takes the prescribed boundary values and satisfies the first order equation (1.1) at
almost every point x ∈ Ω.

Unfortunately, this concept of solution is far too weak, and does not lead to any useful unique-
ness result.

Example 1.4. The boundary value problem on the unit interval

|ux | − 1 = 0 x ∈ [0, 1] , x(0) = x(1) = 0, (1.10)
admits infinitely many generalized solutions (fig. 1.3a).

In view of the previous example, one seeks a new concept of solution for the first order equation
(1.1), having the following properties:

1. For every boundary data (1.2), a unique solution exists, depending continuously on the boundary
values and on the function F .

2. This solution u coincides with the limit of vanishing viscosity approximations. Namely, u =
limε→0+ uε , where the uε are solutions of
F (x, uε , ∇uε ) = ε ∆uε .

3
 u0

u1 u
uε
xε x
0 1 0 1
u
2

figure 1.3a figure 1.3b

3. In the case where (1.1) is the Hamilton-Jacobi equation for the value function of some opti-
mization problem, our concept of solution should single out precisely this value function.

In connection with Example 1.4, we see that the distance function

½
x if x ∈ [0, 1/2],
u0 (x) =
1−x if x ∈ [1/2, 1],
is the only one, among those shown in fig. 1.3a, that can be obtained as a vanishing viscosity
limit. Indeed, any other generalized solution u with polygonal graph has at least one strict local
minimum in the interior of the interval [0, 1], say at a point x. If uε → u uniformly on [0, 1], for
some sequence of smooth solutions to
|uεx | − 1 = ε uεxx ,
then each uε will have a local minimum at a nearby point xε (fig. 1.3b). But this is impossible,
because ¯ ε ¯
¯ux (xε )¯ − 1 = −1 6= ε uεxx (xε ) ≥ 0.

In the following sections we shall introduce the definition of viscosity solution and see how it
fulfils the above requirements.

2 - One-sided differentials
Let u : Ω 7→ IR be a scalar function, defined on an open set Ω ⊆ IRn . The set of super-
differentials of u at a point x is defined as
½ ¾
+ . n u(y) − u(x) − p · (y − x)
D u(x) = p ∈ IR ; lim sup ≤0 . (2.1)
y→x |y − x|
In other words, a vector p ∈ IRn is a super-differential iff the plane y 7→ u(x) + p · (y − x) is tangent
from above to the graph of u at the point x (fig. 2.1a). Similarly, the set of sub-differentials of u
at a point x is defined as
½ ¾
− . n u(y) − u(x) − p · (y − x)
D u(x) = p ∈ IR ; lim inf ≥0 , (2.2)
y→x |y − x|

4
 u
u

x x
figure 2.1a figure 2.1b

so that a vector p ∈ IRn is a sub-differential iff the plane y 7→ u(x) + p · (y − x) is tangent from
below to the graph of u at the point x (fig. 2.1b).

Example 2.1. Consider the function (fig. 2.2)

(
0√ if x < 0,
.
u(x) = x if x ∈ [0, 1],
1 if x > 1.

In this case we have

D+ u(0) = ∅, D− u(0) = [0, ∞[ ,
© √ ª
D+ u(x) = D− u(x) = 1/2 x x ∈ ]0, 1[ ,
£ ¤
D+ u(1) = 0, 1/2 , D− u(1) = ∅.

1 x
0
figure 2.2

If ϕ ∈ C 1 , its differential at a point x is written as ∇ϕ(x). The following characterization of

super- and sub-differential is very useful.

Lemma 2.2. Let u ∈ C(Ω). Then

(i) p ∈ D+ u(x) if and only if there exists a function ϕ ∈ C 1 (Ω) such that ∇ϕ(x) = p and u − ϕ
has a local maximum at x.

5
(ii) p ∈ D− u(x) if and only if there exists a function ϕ ∈ C 1 (Ω) such that ∇ϕ(x) = p and u − ϕ
has a local minimum at x.

By adding a constant, it is not restrictive to assume that ϕ(x) = u(x). In this case, we are
saying that p ∈ D+ u(x) iff there exists a smooth function ϕ ≥ u with ∇ϕ(x) = p, ϕ(x) = u(x).
In other words, the graph of ϕ touches the graph of u from above at the point x (fig. 2.3a). A
similar property holds for subdifferentials: p ∈ D− u(x) iff there exists a smooth function ϕ ≤ u,
with ∇ϕ(x) = p, whose graph touches from below the graph of u at the point x. (fig. 2.3b).

ϕ u
ϕ
u

x x
figure 2.3a figure 2.3b

Proof of Lemma 2.2. Assume that p ∈ D+ u(x). Then we can find δ > 0 and a continuous,
increasing function σ : [0, ∞[ 7→ IR, with σ(0) = 0, such that
¡ ¢
u(y) ≤ u(x) + p · (y − x) + σ |y − x| |y − x|
for |y − x| < δ. Define Z r
.
ρ(r) = σ(t) dt
0
and observe that
ρ(0) = ρ0 (0) = 0, ρ(2r) ≥ σ(r) r .
By the above properties, the function
. ¡ ¢
ϕ(y) = u(x) + p · (y − x) + ρ 2|y − x|
is in C 1 (Ω) and satisfies
ϕ(x) = u(x), ∇ϕ(x) = p.
Moreover, for |y − x| < δ we have
¡ ¢ ¡ ¢
u(y) − ϕ(y) ≤ σ |y − x| |y − x| − ρ 2|y − x| ≤ 0.
Hence, the difference u − ϕ attains a local maximum at the point x.

To prove the opposite implication, assume that Dϕ(x) = p and u − ϕ has a local maximum
at x. Then
u(y) − u(x) − p · (y − x) ϕ(y) − ϕ(x) − p · (y − x)
lim sup ≤ lim sup = 0. (2.3)
y→x |y − x| y→x |y − x|
This completes the proof of (i). The proof of (ii) is entirely similar.

6
Remark 2.3. By possibly replacing the function ϕ with ϕ̃(y) = ϕ(y) ± |y − x|2 , it is clear that in
the above lemma we can require that u − ϕ attains a strict local maximum or local minimum at
the point x. This is particularly important in view of the following stability result.

Lemma 2.4. Let u : Ω 7→ IR be continuous. Assume that, for some φ ∈ C 1 , the function u − φ
has a strict local minimum (a strict local maximum) at a point x ∈ Ω. If um → u uniformly, then
there exists a sequence of points xm → x with um (xm ) → u(x) and such that um − φ has a local
minimum (a local maximum) at xm .

Proof. Assume that u − φ has a strict local minimum at x. For every ρ > 0 sufficiently small,
there exists ερ > 0 such that

u(y) − φ(y) > u(x) − φ(x) + ερ whenever |y − x| = ρ .

By the uniform convergence um → u, for all m ≥ Nρ sufficiently large one has um (y) − u(y) < ερ /4
for |y − x| ≤ ρ. Hence

ερ
um (y) − φ(y) > um (x) − φ(x) + |y − x| = ρ ,
2

This shows that um − φ has a local minimum at some point xm , with |xm − x| < ρ. Letting
ρ, ερ → 0, we construct the desired sequence {xm }.

This situation is illustrated in fig. 2.4a. On the other hand, if x is a point of non-strict local
minimum for u − φ, the slightly perturbed function um − φ may not have any local minimum xm
close to x (fig. 2.4b).

um− φ u− φ

xmx xm x

figure 2.4a figure 2.4b

Some simple properties of super- and sub-differential are collected in the next lemma.

Lemma 2.5. Let u ∈ C(Ω). Then

7
 (i) If u is differentiable at x, then
© ª
D+ u(x) = D− u(x) = ∇u(x) . (2.4)

(ii) If the sets D+ u(x) and D− u(x) are both non-empty, then u is differentiable at x, hence (2.4)
holds.

(iii) The sets of points where a one-sided differential exists:

. © ª . © ª
Ω+ = x ∈ Ω; D+ u(x) 6= ∅ , Ω− = x ∈ Ω; D− u(x) 6= ∅ (2.5)
are both non-empty. Indeed, they are dense in Ω.

Proof. Concerning (i), assume u is differentiable at x. Trivially, ∇u(x) ∈ D± u(x). On the other
hand, if ϕ ∈ C 1 (Ω) is such that u − ϕ has a local maximum at x, then ∇ϕ(x) = ∇u(x). Hence
D+ u(x) cannot contain any vector other than ∇u(x).

To prove (ii), assume that the sets D+ u(x) and D− u(x) are both non-empty. Then there we
can find δ > 0 and ϕ1 , ϕ2 ∈ C 1 (Ω) such that (fig. 2.5)
ϕ1 (x) = u(x) = ϕ2 (x), ϕ1 (y) ≤ u(y) ≤ ϕ2 (y) |y − x| < δ.
By a standard comparison argument, this implies that u is differentiable at x and ∇u(x) =
∇ϕ1 (x) = ∇ϕ2 (x).

Concerning (iii), fix any ball B(x0 , ρ) ⊆ Ω. By choosing ε > 0 sufficiently small, the smooth
function (fig. 2.6)
. |x − x0 |2
ϕ(x) = u(x0 ) −
2ε
is strictly negative on the boundary of the ball, where |x − x0 | = ρ. Since u(x0 ) = ϕ(x0 ), the
function u − ϕ attains a local minimum at an interior point x ∈ B(x0 , ρ). By Lemma 2.2, the
sub-differential of u at x is non-empty. Indeed, ∇ϕ(x) = (x − x0 )/ε ∈ D− u(x). The previous
argument shows that, for every x0 ∈ Ω and ρ > 0, the set Ω− has non-empty intersection with the
ball B(x0 , ρ). Therefore Ω− is dense in Ω. The case of super-differentials is entirely similar.

u u

ϕ
2

ϕ
ϕ
1

x x0 x
figure 2.5 figure 2.6

8
3 - Viscosity solutions
In the following, we consider the first order, partial differential equation
¡ ¢
F x, u(x), ∇u(x) = 0 (3.1)

defined on an open set Ω ∈ IRn . Here F : Ω × IR × IRn 7→ IR is a continuous (nonlinear) function.

Definition 3.1. A function u ∈ C(Ω) is a viscosity subsolution of (3.1) if

¡
F x, u(x), p) ≤ 0 for every x ∈ Ω, p ∈ D+ u(x). (3.2)

Similarly, u ∈ C(Ω) is a viscosity supersolution of (3.1) if

¡
F x, u(x), p) ≥ 0 for every x ∈ Ω, p ∈ D− u(x). (3.3)

We say that u is a viscosity solution of (3.1) if it is both a supersolution and a subsolution

in the viscosity sense.

Similar definitions also apply to evolution equations of the form

¡
ut + H t, x, u, ∇u) = 0, (3.4)

where ∇u denotes the gradient of u w.r.t. x. Recalling Lemma 1, we can reformulate these defini-
tions in an equivalent form:

Definition 3.2. A function u ∈ C(Ω) is a viscosity subsolution of (3.4) if, for every C 1 function
ϕ = ϕ(t, x) such that u − ϕ has a local maximum at (t, x), there holds

ϕt (t, x) + H(t, x, u, ∇ϕ) ≤ 0. (3.5)

Similarly, u ∈ C(Ω) is a viscosity supersolution of (3.4) if, for every C 1 function ϕ = ϕ(t, x)
such that u − ϕ has a local minimum at (t, x), there holds

ϕt (t, x) + H(t, x, u, ∇ϕ) ≥ 0. (3.6)

Remark 3.3. In the definition of subsolution, we are imposing conditions on u only at points x
where the super-differential is non-empty. Even if u is merely continuous, say nowhere differen-
tiable, there are many of these points. Indeed, by Lemma 2.5, the set of points x where D+ u(x) 6= ∅
is dense on Ω. Similarly, for supersolutions we impose conditions only at points where D− u(x) 6= ∅.

Remark 3.4. If u is a C 1 function that satisfies (3.1) at every x ∈ Ω, then u is also a solution in
the viscosity sense. Viceversa, if u is a viscosity solution, then the equality (3.1) must hold at every
point x where u is differentiable. In particular, if u is Lipschitz continuous, then by Rademacher’s
theorem it is a.e. differentiable. Hence (3.1) holds a.e. in Ω.
.
Example 3.5. Set F (x, u, ux ) = 1 − |ux |. Then the function u(x) = |x| is a viscosity solution of

1 − |ux | = 0 (3.7)

9
defined on the whole real line. Indeed, u is differentiable and satisfies the equation (3.7) at all
points x 6= 0. Moreover, we have

D+ u(0) = ∅, D− u(0) = [−1, 1]. (3.8)

To show that u is a subsolution, there is nothing else to check. To show that u is a supersolution,
take any p ∈ [−1, 1]. Then 1 − |p| ≥ 0, as required.
It is interesting to observe that the same function u(x) = |x| is NOT a viscosity solution of
the equation
|ux | − 1 = 0 . (3.9)
Indeed, at x = 0, taking p = 0 ∈ D− u(0) we find |0| − 1 < 0. In conclusion, the function u(x) = |x|
is a viscosity subsolution of (3.9), but not a supersolution.

4 - Stability properties
For nonlinear P.D.E’s, the set of solutions may not be closed w.r.t. the topology of uniform
convergence. In general, if un → u uniformly on a domain Ω, to conclude that u is itself a solution
of the P.D.E. one should know, in addition, that all the derivatives Dα un that appear in the
equation converge to the corresponding derivatives of u. This may not be the case in general.

Example 4.1. A sequence of solutions to the equation

|ux | − 1 = 0 , u(0) = u(1) = 0 (4.1)

is provided by the saw-tooth functions (fig. 4.1)

 · ¸
 k−1 k−1 k−1 1

x − m if x ∈ , +
. m m 2m
um (x) = · ¸ k = 1, . . . , m . (4.2)

 k k 1 k
 −x if x ∈ − ,
m m m m

Clearly um → 0 uniformly on [0, 1], but the zero function is not a solution of (4.1). In this case,
the convergence of the functions un is not accompanied by the convergence of their derivatives.

u4

0 1
figure 4.1

The next lemma shows that, in the case of viscosity solutions, a general stability theorem
holds, without any requirement about the convergence of derivatives.

10
Lemma 4.2. Consider a sequence of continuous functions um , which provide viscosity sub-
solutions (super-solutions) to

Fm (x, um , ∇um ) = 0 x ∈ Ω. (4.3)

As m → ∞, assume that Fm → F uniformly on compact subsets of Ω × IR × IRn and um → u in

C(Ω). Then u is a subsolution (a supersolution) of (3.1)

Proof. To prove that u is a subsolution, let φ ∈ C 1 be such that u − φ has a strict local maximum
at a point x. We need to show that
¡ ¢
F x, φ(x), ∇φ(x) ≤ 0. (4.4)

By Lemma 2.4, there exists a sequence xm → x such that um − φ has a local maximum at xm , and
um (xm ) → u(x) as m → ∞. Since um is a subsolution,
¡ ¢
Fm xm , um (xm ), ∇φ(xm ) ≤ 0 . (4.5)

Taking the limit in (4.5) as m → ∞, we obtain (4.4).

The above result should be compared with Example 4.1. Clearly, the functions un in (4.2) are
not viscosity solutions.

The definition of viscosity solution is naturally motivated by the properties of vanishing vis-
cosity limits.

Theorem 4.3. Let uε be a family of smooth solutions to the viscous equation

¡ ¢
F x, uε (x), ∇uε (x) = ε ∆uε . (4.6)

Assume that, as ε → 0+, we have the convergence uε → u uniformly on an open set Ω ⊆ IRn .
Then u is a viscosity solution of (3.1).

Proof. Fix x ∈ Ω and assume p ∈ D+ u(x). To prove that u is a subsolution we need to show that
F (x, u(x), p) ≤ 0.

1. By Lemma 2.2 and Remark 2.3, there exists ϕ ∈ C 1 with ∇ϕ(x) = p, such that u − ϕ has a
strict local maximum at x. For any δ > 0 we can then find 0 < ρ ≤ δ and a function ψ ∈ C 2 such
that ¯ ¯
¯∇ϕ(y) − ∇ϕ(x)¯ ≤ δ if |y − x| ≤ ρ , (4.7)
kψ − ϕkC 1 ≤ δ (4.8)
and such that each function uε − ψ has a local maximum inside the ball B(x; ρ), for ε > 0 small
enough.

2. Let xε be the location of this local maximum of uε − ψ. Since uε is smooth, this implies

∇ψ(xε ) = ∇u(xε ), ∆u(xε ) ≤ ∆ψ(xε ), (4.9)

hence from (4.6) it follows ¡ ¢

F x, uε (xε ), ∇ψ(xε ) ≤ ε ∆ψ(xε ). (4.10)

11
3. Extract a convergent subsequence xε → x̃. Clearly |x̃ − x| ≤ ρ. Since ψ ∈ C 2 , we can pass to
the limit in (4.10) and conclude
¡ ¢
F x, u(x̃), ∇ψ(x̃) ≤ 0 (4.11)

By (4.7)-(4.8) we have
¯ ¯ ¯ ¯ ¯ ¯
¯∇ψ(x̃) − p¯ ≤ ¯∇ψ(x̃) − ∇ϕ(x̃)¯ + ¯∇ϕ(x̃) − ∇ϕ(x)¯
(4.12)
≤ δ+δ.

Since δ > 0 can be taken arbitrarily small, (4.11) and the continuity of F imply F (x, u(x), p) ≤ 0,
showing that u is a subsolution. The fact that u is a supersolution is proved in an entirely similar
way.

5 - Comparison theorems
A remarkable feature of the notion of viscosity solutions is that on one hand it requires a
minimum amount of regularity (just continuity), and on the other hand it is stringent enough to
yield general comparison and uniqueness theorems.
The uniqueness proofs are based on a technique of doubling of variables, which reminds of
Kruzhkov’s uniqueness theorem for conservation laws [K]. We now illustrate this basic technique
in a simple setting.

Theorem 5.1 (Comparison). Let Ω ⊂ IRn be a bounded open set. Let u1 , u2 ∈ C(Ω) be,
respectively, viscosity sub- and supersolutions of

u + H(x, ∇u) = 0 x ∈ Ω. (5.1)

Assume that
u1 (x) ≤ u2 (x) for all x ∈ ∂Ω. (5.2)

Moreover, assume that H : Ω × IRn 7→ IR is uniformly continuous in the x-variable:

¯ ¯ ³ ¡ ¢´
¯H(x, p) − H(y, p)¯ ≤ ω |x − y| 1 + |p| , (5.3)

for some continuous and non-decreasing function ω : [0, ∞[ 7→ [0, ∞[ with ω(0) = 0. Then

u1 (x) ≤ u2 (x) for all x ∈ Ω. (5.4)

Proof. To appreciate the main idea of the proof, consider first the case where u1 , u2 are smooth.
If the conclusion (5.4) fails, then the difference u1 − u2 attains a positive maximum at a point
.
x0 ∈ Ω. This implies p = ∇u1 (x0 ) = ∇u2 (x0 ). By definition of sub- and supersolution, we now
have
u1 (x0 ) + H(x0 , p) ≤ 0,
(5.5)
u2 (x0 ) + H(x0 , p) ≥ 0.

12
 u1
u1
u
2 u2

Ω Ω
x x0
0

figure 5.1a figure 5.1b

Subtracting the second from the first inequality in (5.5) we conclude u1 (x0 ) − u2 (x0 ) ≤ 0, reaching
a contradiction.

Next, consider the non-smooth case. We can repeat the above argument and reach again a
contradiction provided that we can find a point x0 such that (fig. 5.1a)

(i) u1 (x0 ) > u2 (x0 ),

(ii) some vector p lies at the same time in the upper differential D+ u1 (x0 ) and in the lower
differential D− u2 (x0 ).

A natural candidate for x0 is a point where u1 − u2 attains a global maximum. Unfortunately,

at such point one of the sets D+ u1 (x0 ) or D− u2 (x0 ) may be empty, and the argument breaks down
(fig. 5.1b). To proceed further, the key observation is that we don’t need to compare values of u1
and u2 at exactly the same point. Indeed, to reach a contradiction, it suffices to find nearby points
xε and yε such that (fig. 5.2)

(i’) u1 (xε ) > u2 (yε ),

(ii’) some vector p lies at the same time in the upper differential D+ u1 (xε ) and in the lower
differential D− u2 (yε ).

Can we always find such points? It is here that the variable-doubling technique comes in. The
trick is to look at the function of two variables
. |x − y|2
Φε (x, y) = u1 (x) − u2 (y) − . (5.6)
2ε
This clearly admits a global maximum over the compact set Ω × Ω. If u1 > u2 at some point x0 ,
this maximum will be strictly positive. Moreover, taking ε > 0 sufficiently small, the boundary
conditions imply that the maximum is attained at some interior point (xε , yε ) ∈ Ω × Ω. Notice
that the points xε , yε must be close to each other, otherwise the penalization term in (5.6) will be
very large and negative.
We now observe that the function of a single variable
µ ¶
|x − yε |2
x 7→ u1 (x) − u2 (yε ) + = u1 (x) − ϕ1 (x) (5.7)
2ε

13
 u1

u2

Ω
yε x
ε
figure 5.2

attains its maximum at the point xε . Hence by Lemma 2.2

xε − yε
= ∇ϕ1 (xε ) ∈ D+ u1 (xε ).
ε
Moreover, the function of a single variable
µ ¶
|xε − y|2
y 7→ u2 (y) − u1 (xε ) − = u2 (y) − ϕ2 (y) (5.8)
2ε
attains its minimum at the point yε . Hence
xε − yε
= ∇ϕ2 (yε ) ∈ D− u2 (yε ).
ε
We have thus discovered two points xε , yε and a vector p = (xε − yε )/ε which satisfy the conditions
(i’)-(ii’).

We now work out the details of the proof, in several steps.

1. If the conclusion fails, then there exists x0 ∈ Ω such that

© ª .
u1 (x0 ) − u2 (x0 ) = max u1 (x) − u2 (x) = δ > 0. (5.9)
x∈Ω

For ε > 0, call (xε , yε ) a point where the function Φε in (5.6) attains its global maximum on the
compact set Ω × Ω. By (5.9) one has

Φε (xε , yε ) ≥ δ > 0. (5.10)

¯ ¯ ¯ ¯
2. Call M an upper bound for all values ¯u1 (x)¯, ¯u2 (x)¯, as x ∈ Ω. Then

|x − y|2
Φε (x, y) ≤ 2M − ,
2ε
14
 Φε (x, y) ≤ 0 if |x − y|2 ≥ M ε.
Hence (5.10) implies √
|xε − yε | ≤ Mε . (5.11)

3. By the uniform continuity of the functions u2 on the compact set Ω, for ε0 > 0 sufficiently small
we have
¯ ¯ √
¯u2 (x) − u2 (y)¯ < δ whenever |x − y| ≤ M ε0 . (5.12)
2
We now show that, choosing ε < ε0 , the points xε , yε cannot lie on the boundary of Ω. For example,
if xε ∈ ∂Ω, then by (5.11) and (5.12)
¡ ¢ ¯ ¯ |xε − yε |2
Φε (xε , yε ) ≤ u1 (xε ) − u2 (xε ) + ¯u2 (xε ) − u2 (yε )¯ −
2ε
≤ 0 + δ/2 + 0,
against (5.10).

4. Having shown that xε , yε are interior points, we consider the functions of one single variable
ϕ1 , ϕ2 defined at (5.7)-(5.8). Since xε provides a local maximum for u1 − ϕ1 and yε provides a
local minimum for u2 − ϕ2 , we conclude that
. xε − yε
pε = ∈ D+ u1 (xε ) ∩ D− u2 (yε ). (5.13)
ε
From the definition of viscosity sub- and supersolution we now obtain
u1 (xε ) + H(xε , pε ) ≤ 0,
(5.14)
u2 (yε ) + H(yε , pε ) ≥ 0.

5. Observing that
¯ ¯ |xε − yε |2
u1 (xε ) − u2 (xε ) = Φ(xε , xε ) ≤ Φε (xε , yε ) ≤ u1 (xε ) − u2 (xε ) + ¯u2 (xε ) − u2 (yε )¯ − ,
2ε
by (5.9) we see that
¯ ¯ 2
¯u2 (xε ) − u2 (yε )¯ − |xε − yε | ≥ 0 .
2ε
Hence, by the uniform continuity of u2 ,
|xε − yε |2
→0 as ε → 0. (5.15)
2ε

6. Recalling (5.10) and (5.13), and subtracting the second from the first inequality in (5.14), we
obtain
δ ≤ Φε (xε , yε )
≤ u1 (xε ) − u2 (yε )
¯ ¯ (5.16)
≤ ¯H(xε , pε ) − H(yε , pε )¯
³ ¡ ¢´
≤ ω (|xε − yε | · 1 + |xε − yε |ε−1 .
This yields a contradiction, Indeed, by (5.3) and (5.15) the right hand side of (5.16) becomes
arbitrarily small as ε → 0.

15
 An easy consequence of the above result is the following uniqueness result for the boundary
value problem
u + H(x, ∇u) = 0 x ∈ Ω, (5.17)
u=ψ x ∈ ∂Ω. (5.18)

Corollary 5.2 (Uniqueness). Let Ω ⊂ IRn be a bounded open set. Let the Hamiltonian function
H satisfy the equicontinuity assumption (5.3). Then the boundary value problem (5.17)-(5.18)
admits at most one viscosity solution.

Proof. Let u1 , u2 be viscosity solutions. Since u1 is a subsolution and u2 is a supersolution, and

u1 = u2 on ∂Ω, by Theorem 1 we conclude u1 ≤ u2 on Ω. Reversing the roles of u1 and u2 , we
deduce u2 ≤ u1 , completing the proof.

By similar techniques, comparison and uniqueness results can be proved also for Hamilton-
Jacobi equations of evolutionary type. Consider the Cauchy problem
ut + H(t, x, ∇u) = 0 (t, x) ∈ ]0, T [ ×IRn , (5.19)
u(0, x) = ū(x) x ∈ IRn . (5.20)
.
Here and in the sequel, it is understood that ∇u = (ux1 , . . . , uxn ) always refers to the gradient of
u w.r.t. the space variables.

Theorem 5.3 (Comparison). Let the function H : [0, T ] × IRn × IRn satisfy the Lipschitz
continuity assumptions
¯ ¯ ¡ ¢¡ ¢
¯H(t, x, p) − H(s, y, p)¯ ≤ C |t − s| + |x − y| 1 + |p| , (5.21)
¯ ¯
¯H(t, x, p) − H(t, x, q)¯ ≤ C |p − q| . (5.22)
Let u, v be bounded, uniformly continuous sub- and super-solutions of (5.19) respectively. If
u(0, x) ≤ v(0, x) for all x ∈ IRn , then
u(t, x) ≤ v(t, x) for all (t, x) ∈ [0, T ] × IRn . (5.23)

Toward this result, as a preliminary we prove

Lemma 5.4. Let u be a continuous function on [0, T ] × IRn , which provides a subsolution of (5.19)
for t ∈ ]0, T [. If φ ∈ C 1 is such that u − φ attains a local maximum at a point (T, x0 ), then
¡ ¢
φt (T, x0 ) + H T, x0 , ∇φ(T, x0 ) ≤ 0. (5.24)

Proof. We can assume that (T, x0 ) is a point of strict local maximum for u − φ. For each ε > 0
consider the function
ε
φε (t, x) = φ(t, x) + .
T −t
Each function u − φε will then have a local maximum at a point (tε , xε ), with
tε < T, (tε , xε ) → (T, x0 ) as ε → 0 + .
Since u is a subsolution, one has
¡ ¢ ε
φε,t (tε , xε ) + H tε , xε , ∇φε (tε , xε ) ≤ − . (5.25)
(T − tε )2
Letting ε → 0+, from (5.25) we obtain (5.24).

16
Proof of Theorem 5.3.

1. If (5.23) fails, then we can find λ > 0 such that

n o
.
sup u(t, x) − v(t, x) − 2λt = σ > 0. (5.26)
t,x

Assume that the supremum in (5.26) is actually attained at a point (t0 , x0 ), possibly with t0 = T .
If both u and u are differentiable at such point, we easily obtain a contradiction, because

ut (t0 , x0 ) + H(t0 , x0 , ∇u) ≤ 0 ,

vt (t0 , x0 ) + H(t0 , x0 , ∇v) ≥ 0 ,

∇u(t0 , x0 ) = ∇v(t0 , x0 ), ut (t0 , x0 ) − vt (t0 , x0 ) − 2λ ≥ 0 .

2. To extend the above argument to the general case, we face two technical difficulties. First,
the function in (5.26) may not attain its global maximum over the unbounded set [0, T ] × IRn .
Moreover, at this point of maximum the functions u, v may not be differentiable. These problems
are overcome by inserting a penalization term, and doubling the variables. As in the proof of
Theorem 5.1, we introduce the function
¡ ¢ 1¡ ¢
Φε (t, x, s, y) = u(t, x) − v(s, y) − λ(t + s) − ε |x|2 + |y|2 − 2 |t − s|2 + |x − y|2 .
ε
Thanks to the penalization terms, the function Φε clearly admits a global maximum at a point
¡ ¢2
(tε , xε , sε , yε ) ∈ ]0, T ] × IRn . Choosing ε > 0 sufficiently small, one has

Φε (tε , xε , sε , yε ) ≥ max Φε (t, x, t, x) ≥ σ/2 .

t,x

3. We now observe that the function

h ¡ ¢ 1¡ ¢i .
(t, x) 7→ u(t, x) − v(sε , yε ) + λ(t + sε ) + ε |x|2 + |yε |2 + 2 |t − sε |2 + |x − yε |2 = u(t, x) − φ(t, x)
ε

takes a maximum at the point (tε , xε ). Since u is a subsolution and φ is smooth, this implies
µ ¶
2(tε − sε ) 2(xε − yε )
λ+ + H tε , xε , + 2εxε ≤ 0 . (5.27)
ε2 ε2

Notice that, in the case where tε = T , (5.27) follows from Lemma 5.5.
Similarly, the function
h ¡ ¢ 1¡ ¢i .
(s, y) 7→ v(s, y) − u(tε , xε ) − λ(tε + s) − ε |xε |2 + |y|2 − 2 |tε − s|2 + |xε − y|2 = v(s, y) − ψ(s, y)
ε

takes a maximum at the point (tε , xε ). Since v is a supersolution and ψ is smooth, this implies
µ ¶
2(tε − sε ) 2(xε − yε )
−λ + + H sε , y ε , − 2εyε ≥ 0 . (5.28)
ε2 ε2

17
4. Subtracting (5.28) from (5.27) and using (5.21)-(5.22) we obtain
µ ¶ µ ¶
2(xε − yε ) 2(xε − yε )
2λ ≤ H sε , yε , − 2εyε − H tε , xε , + 2εxε
ε2 ε2
µ ¶ (5.29)
¡ ¢ ¡ ¢ |xε − yε | ¡ ¢
≤ Cε |xε | + |yε | + C |tε − sε | + |xε − yε | 1 + + ε |xε | + |yε | .
ε2

To reach a contradiction we need to show that the right hand side of (5.29) approaches zero as
ε → 0.

5. Since u, v are globally bounded, the penalization terms must satisfy uniform bounds, indepen-
dent of ε. Hence
C0
|xε |, |yε | ≤ √ , |tε − sε |, |xε − yε | ≤ C 0 ε (5.30)
ε
for some constant C 0 . This implies
¡ ¢ √
ε |xε | + |yε | ≤ 2C 0 ε . (5.31)

To obtain a sharper estimate, we now observe that Φε (tε , xε , sε , yε ) ≥ Φε (tε , xε , tε , xε ), hence

¡ ¢ 1¡ ¢
u(tε , xε ) − v(sε , yε )−λ(tε + sε ) − ε |xε |2 + |yε |2 − 2 |tε − sε |2 + |xε − yε |2
ε
≥ u(tε , xε ) − v(tε , xε ) − 2λtε − 2ε|xε |2 ,

1¡ 2 2
¢ ¡ 2 2
¢
|t ε − s ε | + |xε − y ε | ≤ v(t ε , x ε ) − v(sε , y ε ) + λ(tε − s ε ) + ε |xε | − |yε | . (5.32)
ε2
By the uniform continuity of v, the right hand side of (5.32) tends to zero as ε → 0, therefore

|tε − sε |2 + |xε − yε |2
→0 as ε → 0. (5.33)
ε2

By (5.30), (5.31) and (5.33), the right hand side of (5.29) also approaches zero, This yields the
desired contradiction.

Corollary 5.5 (Uniqueness). Let the function H satisfy the assumptions (5.21)-(5.22). Then the
Cauchy problem (5.19)-(5.20) admits at most one bounded, uniformly continuous viscosity solution
u : [0, T ] × IRn 7→ IR.

18
6 - Control theory
The evolution of a deterministic system described by a finite number of parameter can be
modelled by an O.D.E.
ẋ = f (x) x ∈ IRn .
In some cases, the evolution can be influenced also by the external input of a controller. An
appropriate model is then provided by a control system, having the form
ẋ = f (x, u). (6.1)
Here x ∈ IR , while the control u : [0, T ] 7→ U is required to take values inside a given set U ⊆ IRm .
n

We denote by n o
. m
U = u : IR 7→ IR measurable, u(t) ∈ U for a.e. t
the set of admissible control functions. To guarantee local existence and uniqueness of solutions,
it is natural to assume that the map f : IRn × IRm 7→ IRn is Lipschitz continuous w.r.t. x and
continuous w.r.t. u. The solution of the Cauchy problem (6.1) with initial condition
x(t0 ) = x0 (6.2)
will be denoted as t 7→ x(t; t0 , x0 , u). It is clear that, as u ranges over the whole set of control
functions, one obtains a family of possible trajectories for the system. These are precisely the
solutions of the differential inclusion
n o
.
ẋ ∈ F (x) F (x) = f (x, ω) ; ω ∈ U . (6.3)

Example 6.1. Call x(t) ∈ IR2 the position of a boat on a river, and let v(x) be the velocity of the
water at the point x. If the boat simply drifts along with the currrent, its position is described by
the differential equation
ẋ = v(x).
If we assume that the boat is powered by an engine, and can move in any direction with speed ≤ ρ
(relative to the water), the evolution can be modelled by the control system
ẋ = f (x, u) = v(x) + ρu |u| ≤ ρ .
This is equivalent to a differential inclusion where the sets of velocities are balls with radius ρ
(fig. 6.1): ¡ ¢
ẋ ∈ F (x) = B v(x); ρ

figure 6.1

19
Example 6.2. An important class of control systems have the form

ẋ = f (x) + g(x) u u ∈ [−1, 1]

where f, g are vector fields on IRn . This is equivalent to a differential inclusion

© ª
ẋ ∈ F (x) = f (x) + g(x) u ; u ∈ [−1, 1]

where each set F (x) of possible velocities is a segment (fig. 6.2).

x
g 2

0 x1

figure 6.2

For the Cauchy problem (6.1)-(6.2), the reachable set at time T starting from x0 at time t0
(fig. 6.3) will be denoted by
n o
.
R(T ) = x(T ; t0 , x0 , u) , u∈U . (6.4)

R(T)

x
0

figure 6.3

20
 The control u can be assigned as an open loop control, as a function of time: t 7→ u(t), or
as a feedback control, as a function of the state: x 7→ u(x).

Among the major issues that one can study in connection with the control system (6.1) are
the following.

1 - Dynamics. Starting from a point x0 , describe the set of all possible trajectories. Study the
properties of the reachable set R(T ).

2 - Stabilization. For each initial state x0 , find a control u(·) that steers the system toward the
origin, so that
x(t; 0, x0 , u) → 0 as t → ∞ .
Preferably, the stabilizing control should be found in feedback form. One thus looks for a function
u = u(x) such that all trajectories of the system
¡ ¢
ẋ = f x, u(x)

approach the origin asymptotically as t → ∞.

3 - Optimal Control. Find a control u(·) ∈ U which is optimal w.r.t. a given cost criterion. For
example, given the initial condition (6.2), one may seek to minimize the cost
Z T
. ¡ ¢ ¡ ¢
J(u) = h x(t), u(t) dt + ϕ x(T )
t0

over all control functions u ∈ U. Here it is understood that x(t) = x(t; t0 , x0 , u), while

h : IRn × U 7→ IR, ϕ : IRn 7→ IR

are continuous functions. We call h the running cost and ϕ the terminal cost.

7 - The Pontryagin Maximum Principle

In connection with the system

ẋ = f (x, u), u(t) ∈ U, t ∈ [0, T ], x(0) = x0 , (7.1)

we consider the Mayer problem: ¡ ¢

max ψ x(T, u) . (7.2)
u∈U

Here there is no running cost, and only a terminal payoff to be maximized over all admissible
controls. Let t 7→ u∗ (t) be an optimal control function, and let t 7→ x∗ (t) = x(t; 0, x0 , u∗ ) be the
corresponding optimal trajectory (fig. 7.1). We seek necessary conditions that will be satisfied by
u∗ .

As a preliminary, we recall some basic facts from O.D.E. theory. Let t 7→ x(t) be a solution
of the O.D.E.
ẋ = g(t, x) . (7.3)

21
 ψ=ψ
max

x*

x
0
R(T)

figure 7.1

Assume that g : [0, T ] × IRn 7→ IRn is measurable w.r.t. t and continuously differentiable w.r.t. x.
Consider a family of nearby solutions (fig. 7.2), say t 7→ xε (t). Assume that at a given time s one
has
xε (s) − x(s)
lim = v(s) .
ε→0 ε
Then the first order tangent vector

. xε (t) − x(t)
v(t) = lim
ε→0 ε
is well defined for every t ∈ [0, T ], and satisfies the linearized evolution equation

v̇(t) = A(t) v(t) , (7.4)

with ¡ ¢
.
A(t) = Dx g t, x(t) . (7.5)

v(t)
x ε(t)
v(s)
xε(s)
x(t)
x(s)
figure 7.2

22
 Using the Landau notation, we can write xε (t) = x(t) + εv(t) + o(ε), where o(ε) denotes an
infinitesimal of higher order w.r.t. ε.

Together with (7.4), it is useful to consider the adjoint system

ṗ(t) = −p(t)A(t) (7.6)

Here A is an n × n matrix, with entries Aij = ∂gi /∂xj , p ∈ IRn is a row vector and v ∈ IRn is
a column vector. If t 7→ p(t) and t 7→ v(t) satisfy (7.6) and (7.4) respectively, then the product
t 7→ p(t)v(t) is constant in time. Indeed
d¡ ¢ £ ¤ £ ¤
p(t) v(t) = ṗ(t)v(t) + p(t)v̇(t) = − p(t)A(t) v(t) + p(t) A(t)v(t) = 0. (7.7)
dt

After these preliminaries,

¡ we ¢can now derive some necessary conditions for optimality. Since u∗
is optimal, the payoff ψ x(T, u∗ ) cannot be further increased by any perturbation of the control
u∗ (·). Fix a time τ ∈ ]0, T ] and a control value ω ∈ U . For ε > 0 small, consider the needle
variation uε ∈ U (fig. 7.3):

½
ω if t ∈ [τ − ε, τ ] ,
uε (t) = (7.8)
u∗ (t) if t∈/ [τ − ε, τ ] .

ω uε

u*

0 τ−ε τ T
figure 7.3

Call t 7→ xε (t) = x(t, uε ) the perturbed trajectory. We shall compute the terminal point
xε (T ) = x(T, uε ) and check that the value of ψ is not increased by this perturbation.
Assuming that the optimal control u∗ is continuous at time t = τ , we have
. xε (τ ) − x∗ (τ ) ¡ ¢ ¡ ¢
v(τ ) = lim = f x∗ (τ ), ω − f x∗ (τ ), u∗ (τ ) . (7.9)
ε→0 ε
Indeed, xε (τ − ε) = x∗ (τ − ε) and on the small interval [τ − ε, τ ] we have
¡ ¢ ¡ ¢
ẋε ≈ f x∗ (τ ), ω , ẋ∗ ≈ f x∗ (τ ), u∗ (τ ) .

23
Since uε = u∗ on the remaining interval t ∈ [τ, T ], as in (7.4) the evolution of the tangent vector

. xε (t) − x∗ (t)
v(t) = lim t ∈ [τ, T ]
ε→0 ε
is governed by the linear equation
v̇(t) = A(t) v(t) (7.10)
. ¡ ∗ ¢ ¡ ¢ ¡ ¢
with A(t) = Dx f x (t), u∗ (t) . By maximality, ψ xε (T ) ≤ ψ x∗ (T ) , therefore (fig. 7.4)
¡ ¢
∇ψ x∗ (T ) · v(T ) ≤ 0 . (7.11)

v(T)

p(T)= ψ
∆
v(τ ) xε

x*(T)

x*( τ)
p( τ)
x
0
ψ = const.
figure 7.4

Summing up, the previous analysis has established the following:

For every time τ (where u∗ is continuous) and every admissible control value ω ∈ U , we can
generate the vector
. ¡ ¢ ¡ ¢
v(τ ) = f x∗ (τ ), ω − f x∗ (τ ), u∗ (τ )
and propagate it forward in time, by solving the linearized equation (7.10). The inequality (7.11)
is then a necessary condition for optimality.

Instead of propagating the (infinitely many) vectors v(τ ) forward in time, it is more convenient
to propagate the single vector ∇ψ backward. We thus define the row vector t 7→ p(t) as the solution
of ¡ ¢
ṗ(t) = −p(t) A(t), p(T ) = ∇ψ x∗ (T ) . (7.12)
This yields p(t)v(t) = p(T )v(T ) for every t. In particular, (7.11) implies
h ¡ ¢ ¡ ¢i ¡ ¢
p(τ ) · f x∗ (τ ), ω − f x∗ (τ ), u∗ (τ ) = ∇ψ x∗ (T ) · v(T ) ≤ 0 ,

¡ ¢ n ¡ ¢o
p(τ ) · ẋ∗ (τ ) = p(τ ) · f x∗ (τ ), u∗ (τ ) = max p(τ ) · f x∗ (τ ), ω . (7.13)
ω∈U

24
 f(x*, u)
x*(T)

p(T)= ψ
∆
x*( τ)
p( τ)
x
0
figure 7.5

According to (7.13), for every time τ ∈ ]0, T ], the speed ẋ∗ (τ ) corresponding to the optimal control
u∗ (τ ) is the one having has inner product with p(τ ) as large as possible (fig. 7.5).
With some additional care, one can show that the maximality condition (7.13) holds at every
τ which is a Lebesgue point of u∗ , hence almost everywhere. The above result can be restated as

Theorem 7.1: Pontryagin Maximum Principle (Mayer Problem, free terminal point).
Consider the control system

ẋ = f (x, u) u(t) ∈ U, x(0) = x0 .

Let t 7→ u∗ (t) be an optimal control and t 7→ x∗ (t) = x(t, u∗ ) be the corresponding optimal trajectory
for the maximization problem ¡ ¢
max ψ x(T, u) .
u∈U

Define the vector t 7→ p(t) as the solution to the linear adjoint system
. ¡ ¢
ṗ(t) = −p(t) A(t), A(t) = Dx f x∗ (t), u∗ (t) ,

with terminal condition ¡ ¢

p(T ) = ∇ψ x∗ (T ) .
Then, for almost every τ ∈ [0, T ] the following maximality condition holds:
¡ ¢ n ¡ ¢o
p(τ ) · f x∗ (τ ), u∗ (τ ) = max p(τ ) · f x∗ (τ ), ω .
ω∈U

Relying on the Maximum Principle, the computation of the optimal control requires two steps:

STEP 1: solve the pointwise maximixation problem (7.13), obtaining the optimal control u∗ as
a function of p, x, i.e. © ª
u∗ (x, p) = argmax p · f (x, ω) . (7.14)
ω∈U

STEP 2: solve the two-point boundary value problem

( ¡ ¢ (
ẋ = f x, u∗ (x, p) x(0) = x0
¡ ¢ ¡ ¢ (7.15)
ṗ = −p · Dx f x, u∗ (x, p) p(T ) = ∇ψ x(T )

25
 • In general, the function u∗ = u∗ (p, x) in (7.14) is highly nonlinear. It may be multivalued or
discontinuous.

• The two-point boundary value problem (7.15) can be solved by a shooting method: guess
an initial value p(0) = p0 and solve the corresponding Cauchy problem. Try to adjust the
value of p0 so that the terminal values x(T ), p(T ) satisfy the given conditions.

Example 7.2 (Linear pendulum). Let q(t) = be the position of a linearized pendulum, con-
trolled by an external force with magnitude u(t) ∈ [−1, 1].

q̈(t) + q(t) = u(t), q(0) = q̇(0) = 0, u(t) ∈ [−1, 1] .

We wish to maximize the terminal displacement q(T ).

An equivalent control system is obtained by introducing the variables x1 = q, x2 = q̇:
( (
ẋ1 = x2 x1 (0) = 0
ẋ2 = u − x1 x2 (0) = 0

We thus seek
max x1 (T, u).
u∈U
∗ ∗
Let t 7→ x (t) = x(t, u ) be an optimal trajectory. The linearized equation for a tangent vector
is µ ¶ µ ¶µ ¶
v̇1 0 1 v1
=
v̇2 −1 0 v2
The corresponding adjoint vector p = (p1 , p2 ) satisfies
µ ¶
0 1 ¡ ¢
(ṗ1 , ṗ2 ) = −(p1 , p2 ) , (p1 , p2 )(T ) = ∇ψ x∗ (T ) = (1, 0) (7.16)
−1 0
.
because ψ(x) = x1 . In this special linear case, we can explicitly solve (7.16) without needing to
know x∗ , u∗ . An easy computation yields
¡ ¢
(p1 , p2 )(t) = cos(T − t), sin(T − t) . (7.17)

For each t, we must now choose the value u∗ (t) ∈ [−1, 1] so that

p1 x2 + p2 (−x1 + u∗ ) = max p1 x2 + p2 (−x1 + ω) .

ω∈[−1,1]

By (7.17), the optimal control is

¡ ¢ ¡ ¢
u∗ (t) = sign p2 (t) = sign sin(T − t) .

Example 7.3. Consider the problem on IR3

maximize x3 (T ) over all controls u : [0, T ] 7→ [−1, 1]

26
 .
x=q
2
u=1
p
u= −1

−1 0 1 2 x1= q

figure 7.6

for the system  

 ẋ1 = u
  x1 (0) = 0

ẋ2 = −x1 x2 (0) = 0 (7.18)

 

ẋ3 = x2 − x21 x3 (0) = 0
The adjoint equations take the form

(ṗ1 , ṗ2 , ṗ3 ) = (p2 + 2x1 p3 , −p3 , 0) (p1 , p2 , p3 )(T ) = (0, 0, 1) . (7.19)

Maximizing the inner product p · ẋ we obtain the optimality conditions for the control u∗

p1 u∗ + p2 (−x1 ) + p3 (x2 − x21 ) = max p1 ω + p2 (−x1 ) + p3 (x2 − x21 ) , (7.20)

ω∈[−1,1]
(
u∗ = 1 if p1 > 0 ,
u∗ ∈ [−1, 1] if p1 = 0 ,
u∗ = −1 if p1 < 0 .
Solving the terminal value problem (7.19) for p2 , p3 we find

p3 (t) ≡ 1, p2 (t) = T − t .

The function p1 can now be found from the equations

p̈1 = −1 + 2u∗ = −1 + 2 sign(p1 ), p1 (T ) = 0, ṗ1 (0) = p2 (0) = T ,

27
with the convention: sign(0) = [−1, 1]. The only solution is found to be
 µ ¶2

 3 T

 −2 3 − t if 0 ≤ t ≤ T /3 ,
p1 (t) =




0 if T /3 ≤ t ≤ T .

The optimal control is ½

−1 if 0 ≤ t ≤ T /3 ,
u∗ (t) =
1/2 if T /3 ≤ t ≤ T .
Observe that on the interval [T /3, T ] the optimal control is derived not from the maximality
condition (7.20) but from the equation p̈1 = (−1 + 2u) ≡ 0. An optimal control with this property
is called singular.

p
1
..
p=1
1

0 T/3 T

..
p = −3
1

figure 7.7

28
8 - Extensions of the P.M.P.
In connection with the control system
ẋ = f (t, x, u) u(t) ∈ U, x(0) = x0 ,
the more general optimization problem with terminal payoff and running cost
( Z T )
¡ ¢ ¡ ¢
max ψ x(T, u) − h t, x(t), u(t) dt
u∈U 0

can be easily reduced to a Mayer problem with only terminal payoff. Indeed, it suffices to introduce
an additional variable xn+1 which evolves according to
¡ ¢
ẋn+1 = h t, x(t), u(t) , xn+1 (0) = 0 ,
and consider the maximization problem
n ¡ ¢ o
max ψ x(T, u) − xn+1 (T, u) .
u∈U

Another important extension deals with the case where terminal constraints are given, say
x(T ) ∈ S, where the set S is defined as
. © ª
S = x ∈ IRn ; φi (x) = 0, i = 1, . . . , m .
Assume that, at a given point x∗ ∈ S, the m + 1 gradients ∇ψ, ∇φ1 , . . . , ∇φm are linearly
independent. Then the tangent space to S at x∗ is
© ª
TS = v ∈ IRn ; ∇φi (x∗ ) · v = 0 i = 1, . . . , m , (8.1)
while the tangent cone to the set
© ª
S + = x ∈ S ; ψ(x) ≥ ψ(x∗ )
is © ª
TS + = v ∈ IRn ; ∇ψ(x∗ ) · v ≥ 0, ∇φi (x∗ ) · v = 0 i = 1, . . . , m . (8.2)

When x∗ = x∗ (T ) is the terminal point of an admissible trajectory, we think of TS + as the

cone of profitable directions, i.e. those directions in which we should like to move the terminal
point, in order to increase the value of ψ and still satisfy the constraint x(T ) ∈ S (fig. 8.1).

T +
S
ψ (x*)
∆
S

x*

TS

figure 8.1

29
 This cone has a useful characterization:

Lemma 8.1. A vector p ∈ IRn satisfies

p·v ≥0 for all v ∈ TS + (8.3)

if and only if it can be written as a linear combination

m
X
p = λ0 ∇ψ(x∗ ) + λi ∇φi (x∗ ) (8.4)
i=1

with λ0 ≥ 0.

Proof.
Define the vectors

. .
w0 = ∇ψ(x∗ ) , wi = ∇φi (x∗ ) i = 1, . . . , m .

By our previous assumption, these vectors are linearly independent. We can thus add vectors wj ,
j = m + 1, . . . , n − 1 so that
© ª
w0 , w1 , · · · , wN , wm+1 , . . . , wn−1

is a basis of IRn . Let

© ª
v0 , v1 , · · · , vN , vm+1 , . . . , vn−1

be the dual basis, so that

½
1 if i = j,
vi · wj =
0 if i 6= j.

We observe that
n−1
X
v ∈ TS + if and only if v = c0 v0 + ci vi
i=m+1

for some c0 ≥ 0, ci ∈ IR. An arbitrary vector p ∈ IRn can now be written as

m
X n−1
X
p = λ0 w0 + λi wi + λi wj .
i=1 i=m+1

If v ∈ TS + then
n−1
X
p · v = λ0 c0 + λi ci .
i=m+1

It is now clear that (8.3) holds if and only if λ0 ≥ 0 and λi = 0 for all i = m + 1, . . . , n − 1.

30
 x *(T)

x *( τ) Γ
xε

R(T)
x0
figure 8.2

Let now t 7→ x∗ (t) = x(t, u∗ ) be a reference trajectory. As in the previous section, given
τ ∈ ]0, T ] and ω ∈ U , consider the family of needle variations
½
ω if t ∈ [τ − ε, τ ] ,
uε (t) = ∗ (8.5)
u (t) if t∈/ [τ − ε, τ ] .

Call
. x(T, uε ) − x(T, u∗ )
v τ,ω (T ) = lim
ε→0 ε
the first order variation of the terminal point of the corresponding trajectory. Define Γ as the
smallest convex cone containing all vectors v τ,ω . This is a cone of feasible directions, i.e. di-
rections in which we can move the terminal point x(T, u∗ ) by suitably perturbing the control u∗
(fig. 8.2).
We can now state necessary conditions for optimality for the

Mayer Problem with terminal constraints:

¡ ¢
max ψ x(T, u) , (8.6)
u∈U

for the control system

ẋ = f (t, x, u), u(t) ∈ U, t ∈ [0, T ], (8.7)

with initial and terminal constraints

¡ ¢
x(0) = x0 , φi x(T ) = 0, i = 1, . . . , m . (8.8)

Theorem 8.2 (PMP, geometric version). Let t 7→ x∗ (t) = x(t, u∗ ) be an optimal trajectory for
the problem (8.6)–(8.8), corresponding to the control u∗ (·). Then the cones Γ and TS + are weakly
separated, i.e. there exists a non-zero vector p(T ) such that

p(T ) · v ≥ 0 for all v ∈ TS + , (8.9)

31
 S T +
S

p(T)

x* Γ

R(T)
x0
S
figure 8.3

p(T ) · v ≤ 0 for all v ∈ Γ . (8.10)

This separation property is illustrated in fig. 8.3. An equivalemt statement is:

Theorem 8.3 (PMP, analytic version). Let t 7→ x∗ (t) = x(t, u∗ ) be an optimal trajectory,
corresponding to the control u∗ (·). Then there exists a non-zero vector function t 7→ p(t) such that
m
¡ ¢ X ¡ ¢
p(T ) = λ0 ∇ψ x∗ (T ) + λi ∇φi x∗ (T ) with λ0 ≥ 0 , (8.11)
i=1
¡ ¢
ṗ(t) = −p(t) Dx f t, x∗ (t), u∗ (t) t ∈ [0, T ] , (8.12)
¡ ¢ n ¡ ¢ o
p(τ ) · f τ, x∗ (τ ), u∗ (τ ) = max p(τ ) · f τ, x∗ (τ ), ω for a.e. τ ∈ [0, T ] . (8.13)
ω∈U

We show here the equivalence of the two formulations.

By Lemma 8.1, (8.9) is equivalent to (8.11).
Recalling that every tangent vector v τ,ω satisfies the linear evolution equation
¡ ¢
v̇ τ,ω (t) = Dx f t, x∗ (t), u∗ (t) v τ,ω (t),

we see that, if t 7→ p(t) satisfies (8.12), then the product p(t) · v τ,ω (t) is constant. Hence

p(T ) · v τ,ω (T ) ≤ 0

32
if and only if
h ¡ ¢ ¡ ¢i
p(τ ) · v τ,ω (τ ) = p(τ ) · f τ, x∗ (τ ), ω − f τ, x∗ (τ ), u∗ (τ ) ≤ 0

if and only if (8.13) holds.

As a special case, consider the

Lagrange Minimization Problem with fixed terminal point:

Z T
min L(t, x, u) dt , (8.14)
u∈U 0

for the control system on IRn

ẋ = f (t, x, u) u(t) ∈ U , (8.15)

with initial and terminal constraints

x(0) = x[ , x(T ) = x] . (8.16)

An adaptaition of the previous analysis yields

Theorem 8.4 (PMP, Lagrange problem). Let t 7→ x∗ (t) = x(t, u∗ ) be an optimal trajectory,
corresponding to the optimal control u∗ (·). Then there exist a constant λ ≥ 0 and a row vector
t 7→ p(t) (not both = 0) such that
¡ ¢ ¡ ¢
ṗ(t) = −p(t) Dx f t, x∗ (t), u∗ (t) − λ Dx L t, x∗ (t), u∗ (t) , (8.17)
¡ ¢ ¡ ¢
p(t) · f t, x∗ (t), u∗ (t) +λ L t, x∗ (t), u∗ (t)
n ¡ ¢ ¡ ¢o (8.18)
= min p(t) · f t, x∗ (t), ω + λ L t, x∗ (t), ω .
ω∈U

This follows by applying the previous results to the Mayer problem

min xn+1 (T, u)

u∈U

with
ẋ = f (t, x, u), ẋn+1 = L(t, x, u), xn+1 (0) = 0 .
Observe that the evolution of the adjoint vector (p, pn+1 ) = (p1 , . . . , pn , pn+1 ) is governed by the
linear system
 ∂f /∂x ··· ∂f1 /∂xn 0
1 1
 .. .. .. .. 
(ṗ1 , . . . , ṗn , ṗn+1 ) = −(p1 , . . . , pn , pn+1 )  . . . ..
 
∂fn /∂x1 ··· ∂fn /∂xn 0
∂L/∂x1 ··· ∂L/∂xn 0

33
Because of the terminal constraints (x1 , . . . , xn )(T ) = (x]1 , . . . , x]n ), the only requirement on the
terminal value (p1 , . . . , pn , pn+1 )(T ) is
pn+1 (T ) ≥ 0.
Since ṗn+1 = 0, we have pn+1 (t) ≡ λ for some constant λ ≥ 0.

Theorem 8.4 can be further specialized to the

Standard Problem of the Calculus of Variations:

Z T
¡ ¢
minimize L t, x(t), ẋ(t) dt (8.19)
0

over all absolutely continuous functions x : [0, T ] 7→ IRn such that

x(0) = x[ , x(T ) = x] . (8.20)

This corresponds to the optimal control problem (8.14), for the trivial control system
.
ẋ = u, u(t) ∈ U = IRn . (8.21)

We assume that L is smooth, and that x∗ (·) is an optimal solution. By Theorem 8.4 there
exist a constant λ ≥ 0 and a row vector t 7→ p(t) (not both = 0) such that
∂ ¡ ¢
ṗ(t) = −λ L t, x∗ (t), ẋ∗ (t) , (8.22)
∂x
¡ ¢ n ¡ ¢o
p(t) · ẋ∗ (t) + λ L t, x∗ (t), ẋ∗ (t) = minn p(t) · ω + λ L t, x∗ (t), ω . (8.23)
ω∈IR

If λ = 0, then p(t) 6= 0. But in this case ẋ∗ cannot provide a minimum over the whole space IRn .
This contradiction shows that we must have λ > 0.
Since λ, p are determined up to a positive scalar multiple, we can assume λ = 1. With this
choice (8.22) implies
∂ ¡ ¢
p(t) = − L t, x∗ (t), ẋ∗ (t) . (8.23)
∂ ẋ
The evolution equation
∂ ¡ ¢
ṗ(t) = − L t, x∗ (t), ẋ∗ (t)
∂x
now yields the famous Euler-Lagrange equations
· ¸
d ∂ ¡ ∗ ∗
¢ ∂ ¡ ¢
L t, x (t), ẋ (t) = L t, x∗ (t), ẋ∗ (t) . (8.24)
dt ∂ ẋ ∂x
Moreover, the minimality condition
¡ ¢ n ¡ ¢o
p(t) · ẋ∗ (t) + L t, x∗ (t), ẋ∗ (t) = minn p(t) · ω + L t, x∗ (t), ω
ω∈IR

yields the Weierstrass necessary conditions

¡ ¢
∗
¡ ∗ ∗
¢ ∂L t, x∗ (t), ẋ∗ (t) ¡ ¢
L(t, x (t), ω) ≥ L t, x (t), ẋ (t) + · ω − ẋ∗ (t) , (8.25)
∂ ẋ
valid for every ω ∈ IRn . In other words¡(fig. 8.4), for ¢every time t, the graph of ω 7→ L(t, x∗ (t), ω)
lies entirely above its tangent plane at t, x∗ (t), ẋ∗ (t)

34
 L(t, x *,ω )

a b .
x*(t) ω
figure 8.4

9 - Dynamic programming
Consider again a control system of the form
¡ ¢
ẋ(t) = f x(t), u(t) , u(t) ∈ U. (9.1)

We now assume that the set U ⊂ IRm of admissible control values is compact, while f : IRn × U 7→
IRn is a continuous function such that
¯ ¯ ¯ ¯
¯f (x, u)¯ ≤ C, ¯f (x, u) − f (y, u)¯ ≤ C |x − y| for all x, y ∈ IRn , u ∈ U , (9.2)

for some constant C. Given an initial data

x(s) = y ∈ IRn , (9.3)

under the assumptions (9.2), for every choice of the measurable control function u(·) ∈ U the
Cauchy problem (9.1)-(9.2) has a unique solution, which we denote as t 7→ x(t; s, y, u) or sometimes
simply as t 7→ x(t). We seek an admissible control function u∗ : [s, T ] 7→ U , which minimizes the
sum of a running and a terminal cost
Z T
. ¡ ¢ ¡ ¢
J(s, y, u) = h x(t), u(t) dt + g x(T ) . (9.4)
s

Here it is understood that x(t) = x(t; s, y, u), while

h : IRn × U 7→ IR, g : IRn 7→ IR

are continuous functions. We shall assume that the functions h, g satisfy the bounds
¯ ¯ ¯ ¯
¯h(x, u)¯ ≤ C, ¯g(x)¯ ≤ C, (9.5)

35
 ¯ ¯ ¯ ¯
¯h(x, u) − h(y, u)¯ ≤ C |x − y|, ¯g(x) − g(y)¯ ≤ C |x − y|, (9.6)
for all x, y ∈ IRn , u ∈ U . As in the previous sections, we call
n o
.
U = u : IR 7→ IRm measurable, u(t) ∈ U for a.e. t (9.7)

the family of admissible control functions. According to the method of dynamic programming,
an optimal control problem can be studied by looking at the value function:
.
V (s, y) = inf J(s, y, u). (9.8)
u(·)∈U

We consider here a whole family of optimal control problem, all with the same dynamics (9.1) and
cost functional (9.4). We are interested in how the minimum cost varies, as a function of the initial
conditions (9.3). As a preliminary, we state

Lemma 9.1. Let the functions f, g, h satisfy the assumptions (9.2), (9.5) and (9.6). Then the
value function V in (9.8) is bounded and Lipschitz continuous. Namely, there exists a constant C 0
such that ¯ ¯
¯V (s, y)¯ ≤ C 0 , (9.9)
¯ ¯ ¡ ¢
¯V (s, y) − V (s0 , y 0 )¯ ≤ C 0 |s − s0 | + |y − y 0 | . (9.10)

For a proof, see [E].

We want to show that the value function V can be characterized as the unique viscosity
solution to a Hamilton-Jacobi equation. Toward this goal, a basic step is provided by Bellman’s
principle of dynamic programming.

x( T ; τ, y ’, u)

y y’= x( τ ;s, y, u)

s τ T
figure 9.1

Theorem 9.2 (Dynamic Programming Principle). For every τ ∈ [s, T ] and y ∈ IRn , one has
½Z τ ¾
¡ ¢ ¡ ¢
V (s, y) = inf h x(t; s, y, u), u(t) dt + V τ, x(τ ; s, y, u) . (9.11)
u(·) s

36
 In other words (fig. 9.1), the optimization problem on the time interval [s, T ] can be split into
two separate problems:

• As a first step, we solve the optimization problem on the sub-interval [τ, T ], with running cost
h and terminal cost g. In this way, we determine the value function V (τ, ·), at time τ .

• As a second step, we solve the optimization problem on the sub-interval [s, τ ], with running
cost h and terminal cost V (τ, ·), determined by the first step.

At the initial time s, by (9.11) we are saying that the value function V (s, ·) obtained in step 2 is
the same as the value function corresponding to the global optimization problem over the whole
interval [s, T ].

Proof. Call J τ the right hand side of (9.11).

1. To prove that J τ ≤ V (s, y), fix ε > 0 and choose a control u : [s, T ] 7→ U such that

J(s, y, u) ≤ u(s, y) + ε.

Observing that
Z T
¡ ¢ ¡ ¢ ¡ ¢
V τ, x(τ ; s, y, u) ≤ h x(t; s, y, u), u(t) dt + g x(T ; s, y, u) ,
τ

we conclude Z τ
τ
¡ ¢ ¡ ¢
J ≤ h x(t; s, y, u), u(t) dt + V τ, x(τ ; s, y, u)
s
≤ J(s, y, u) ≤ V (s, y) + ε.
Since ε > 0 is arbitrary, this first inequality is proved.

2. To prove that V (s, y) ≤ J τ , fix ε > 0. Then there exists a control u0 : [s, τ ] 7→ U such that
Z τ ¡ ¢ ¡ ¢
h x(t; s, y, u0 ), u(t) dt + V τ, x(τ ; s, y, u0 ) ≤ J τ + ε. (9.12)
s

Moreover, there exists a control u00 : [τ, T ] 7→ A such that

¡ ¡ ¢
J τ, x(τ ; s, y, u0 ), u00 ) ≤ V τ, x(τ ; s, y, u0 ) + ε. (9.13)

One can now define a new control u : [s, T ] 7→ A as the concatenation of u0 , u00 :
½
. u0 (t) if t ∈ [s, τ ],
u(t) =
u00 (t) if t ∈ ]τ, T ].

By (9.12) and (9.13) it is now easy to check that

V (s, y) ≤ J(s, y, u) ≤ J τ + 2ε.

Since ε > 0 can be arbitrarily small, this second inequality is also proved.

37
 10 - The Hamilton-Jacobi-Bellman Equation
The main goal of this section is to characterize the value function as the unique solution of
a first order P.D.E., in the viscosity sense. In turn, this will provide a sufficient condition for the
global optimality of a control function u(·). As in the previous section, we assume here that the
set U is compact and that the functions f, g, h satisfy the bounds (9.2), (9.5) and (9.6).

Theorem 10.1. In connection with the control system (9.1), consider the value function V =
V (s, y) defined by (9.8) and (9.4). Then V is the unique viscosity solution of the Hamilton-Jacobi-
Bellman equation h i
− Vt + H(x, ∇V ) = 0 (t, x) ∈ ]0, T [ ×IRn , (10.1)

with terminal condition

V (T, x) = g(x) x ∈ IRn , (10.2)
and Hamiltonian function
. © ª
H(x, p) = min f (x, ω) · p + h(x, ω) . (10.3)
ω∈U

Proof. By Lemma 9.1, the value function is bounded and uniformly Lipschitz continuous on
[0, T ] ס IRn . The terminal
¢ condition (10.2) is obvious. To show that V is a viscosity solution, let
1 n
ϕ ∈ C ]0, T [ ×IR . Two separate statements need to be proved:

(P1) If V − ϕ attains a local maximum at a point (t0 , x0 ) ∈]0, T [ ×IRn , then

© ª
ϕt (t0 , x0 ) + min f (x0 , ω) · ∇ϕ(t0 , x0 ) + h(x0 , ω) ≥ 0. (10.4)
ω∈U

(P2) If V − ϕ attains a local minimum at a point (t0 , x0 ) ∈]0, T [ ×IRn , then

©
ϕt (t0 , x0 ) + min f (x0 , ω) · ∇ϕ(t0 , x0 ) + h(x0 , ω) ≤ 0. (10.5)
ω∈U

1. To prove (P1), we can assume that

V (t0 , x0 ) = ϕ(t0 , x0 ), V (t, x) ≤ ϕ(t, x) for all t, x .

If (10.4) does not hold, then there exists ω ∈ U and θ > 0 such that

ϕt (t0 , x0 ) + f (x0 , ω) · ∇ϕ(t0 , x0 ) + h(x0 , ω) < −θ. (10.6)

We shall derive a contradiction by showing that this control value ω is “too good to be true”.
Namely, by choosing a control function u(·) with u(t) ≡ ω for t ∈ [t0 , t0 + δ] and such that u
is nearly optimal on the remaining interval [t0 + δ, T ], we obtain a total cost J(t0 , x0 , u) strictly
smaller than V (t0 , x0 ). Indeed, by continuity (10.6) implies

ϕt (t, x) + f (x, ω) · ∇ϕ(t, x) < −h(x, ω) − θ. (10.7)

38
whenever
|t − t0 | < δ, |x − x0 | ≤ Cδ, (10.8)
.
for some δ > 0 small enough and C the constant in (9.2). Let x(t) = x(t; t0 , x0 , ω) be the solution
of ¡ ¢
ẋ(t) = f x(t), ω , x(t0 ) = x0 ,
i.e. the trajectory corresponding to the constant control u(t) ≡ ω. We then have
¡ ¢ ¡ ¢
V t0 + δ, x(t0 + δ) − V (t0 , x0 ) ≤ ϕ t0 + δ, x(t0 + δ) − ϕ(t0 , x0 )
Z t0 +δ
d ¡ ¢
= ϕ t, x(t) dt
t0 dt
Z t0 +δ n
¡ ¢ ¡ ¢ ¡ ¢o (10.9)
= ϕt t, x(t) + f x(t), ω · ∇ϕ t, x(t) dt
t0
Z t0 +δ ¡ ¢
≤− h x(t), ω dt − δθ,
t0

because of (10.7). On the other hand, the Dynamic Programming Principle (9.11) yields
Z t0 +δ
¡ ¢ ¡ ¢
V t0 + δ, x(t0 + δ) − V (t0 , x0 ) ≥ − h t, x(t) dt. (10.10)
t0

Together, (10.9) and (10.10) yield a contradiction, hence (P1) must hold.

2. To prove (P2), we can assume that

V (t0 , x0 ) = ϕ(t0 , x0 ), V (t, x) ≥ ϕ(t, x) for all t, x .

If (P2) fails, then there exists θ > 0 such that

ϕt (t0 , x0 ) + f (x0 , ω) · ∇ϕ(t0 , x0 ) + h(x0 , ω) > θ for all ω ∈ U. (10.11)

In this case, we shall reach a contradiction by showing that no control function u(·) is good enough.
Namely, whatever control function u(·) we choose on the initial interval [t0 , t0 + δ], even if during
the remaining time [t0 + δ, T ] our control is optimal, the total cost will still be considerably larger
than V (t0 , x0 ). Indeed, by continuity, (10.11) implies

ϕt (t, x) + f (x, ω) · ∇ϕ(t, x) > θ − h(x, ω) for all ω ∈ U, (10.12)

for all t, x close to t0 , x0 , i.e. such that (10.8) holds. Choose an arbitrary control function u :
[t0 , t0 + δ] 7→ A, and call t 7→ x(t) = x(t; t0 , x0 , u) the corresponding trajectory. We now have
¡ ¢ ¡ ¢
V t0 + δ, x(t0 + δ) − V (t0 , x0 ) ≥ ϕ t0 + δ, x(t0 + δ) − ϕ(t0 , x0 )
Z t0 +δ
d ¡ ¢
= ϕ t, x(t) dt
t0 dt
Z t0 +δ
¡ ¢ ¡ ¢ ¡ ¢ (10.13)
= ϕt t, x(t) + f x(t), u(t) · ∇ϕ t, x(t) dt
t0
Z t0 +δ ¡ ¢
≥ θ − h x(t), u(t) dt,
t0

39
because of (10.12). Therefore, for every control function u(·) we have
Z t0 +δ
¡ ¢ ¡ ¢
V t0 + δ, x(t0 + δ) + h x(t), u(t) dt ≥ V (t0 , x0 ) + δθ. (10.14)
t0

Taking the infimum of the left hand side of (10.14) over all control functions u, we see that this
infimum is still ≥ V (t0 , x0 ) + δθ. On the other hand, by the Dynamic Programming principle
(9.11), this infimum should be exactly V (t0 , x0 ). This contradiction shows that (P2) must hold,
completing the proof.

One can combine Theorems 5.3 and 10.1, and obtain sufficient conditions for the optimality
of a control function. The usual setting is the following. Consider the problem of minimizing the
cost functional (9.4). Assume that, for each initial condition (s, y), we can guess a “candidate”
optimal control us,y : [s, T ] 7→ U . We then call

.
Ve (s, y) = J(s, y, us,y ) (10.15)

the corresponding cost. Typically, these control functions us,y are found by applying the Pontryagin
Maximum Principle, which provides a necessary condition for optimality. On the other hand,
consider the true value function V , defined at (9.8) as the infimum of the cost over all admissible
control functions u(·) ∈ U. By Theorem 10.1, this function V provides a viscosity solution to
the Hamilton-Jacobi equation (10.1) with terminal condition V (T, y) = g(y). If our function Ve at
(10.15) also provides a viscosity solution to the same equations (10.1)-(10.2), then by the uniqueness
of the viscosity solution stated in Theorem 5.3, we can conclude that Ve = V . Therefore, all controls
us,y are optimal.

We conclude this section by exhibiting a basic relation between the O.D.E. satisfied by extremal
trajectories according to Theorem 7.1, and the P.D.E. of dynamic programming (10.1). Namely:

The trajectories which satisfy the Pontryagin Maximum Principle provide characteristic curves
for the Hamilton-Jacobi equation of Dynamic Programming.

We shall justify the above claim, assuming that all functions involved are sufficiently smooth.
As a first step, we derive the equations of characteristics, in connection with the evolution equation

Vt + H(x, ∇V ) = 0. (10.16)
.
Call p = ∇V the spatial gradient of V , so that p = (p1 , . . . , pn ) = (Vx1 , . . . , Vxn ). Observe that

∂2V ∂pi ∂pj

= = .
∂xi ∂xj ∂xj ∂xi

Differentiating (10.16) w.r.t. xi one obtains

∂pi ∂2V ∂H X ∂H ∂pi

= =− − . (10.17)
∂t ∂xi ∂t ∂xi j
∂pj ∂xj

40
If now t 7→ x(t) is any smooth curve, the total derivative of pi along x is computed by

d ¡ ¢ ∂pi X ∂pi
pi t, x(t) = + ẋj
dt ∂t j
∂x j
µ ¶ (10.18)
∂H X ∂H ∂pi
=− + ẋj −
∂xi j
∂pj ∂xj

In general, the right hand side of (10.18) contains the partial derivatives ∂pi /∂xj . However, if we
choose the curve t 7→ x(t) so that ẋ = ∂H/∂p, the last term will disappear. This observation lies at
the heart of the classical method of characteristics. To construct a smooth solution of the equation
(10.16) with terminal data
V (T, x) = g(x), (10.19)
we proceed as follows. For each point x̄, we find the solution to the Hamiltonian system of O.D.E’s

 ∂H 

 ẋi = (x, p) ,  xi (T ) = x̄i ,
∂pi
∂ḡ (10.20)

 ∂H  pi (T ) = (x̄) .
 ṗi = − (x, p) , ∂xi
∂xi

This solution will be denoted as

t 7→ x(t, x̄), t 7→ p(t, x̄) . (10.21)

¡ ¢
For every t we have ∇V t, x(t, x̄) = p(t, x̄). To recover the function V , we observe that along
each solution of (10.20) one has

d ¡ ¢ ∂H
V t, x(t, x̄) = Vt + ẋ · ∇V = −H(x, p) + p · . (10.22)
dt ∂p

Therefore Z µ ¶
¡ ¢ T
∂H
V t, x(t, x̄) = g(x̄) + H(x, p) − p · ds , (10.23)
t ∂p
where the integral is computed along the solution (10.21).

Next, assume that the hamiltonian function H comes from a minimization problem, and is
thus given by (10.3). For simplicity, we only consider the easier case where U = IRm and h ≡ 0,
so that no running cost is present. We thus have
¡ ¢ © ª
H(x, p) = p · f x, u∗ (x, p) = min p · f (x, ω) , (10.24)
ω

where © ª
u∗ (x, p) = arg min p · f (x, ω) . (10.25)
ω

At the point u∗ where the minimum is attained, one has

∂f ¡ ¢
p· x, u∗ (x, p) = 0 .
∂u

41
Hence
∂H ¡ ¢ ¡ ¢
x, u∗ (x, p) = f x, u∗ (x, p) ,
∂p
∂H ¡ ¢ ∂f ¡ ¢
x, u∗ (x, p) = p · x, u∗ (x, p) .
∂x ∂x
The Hamiltonian system (10.20) thus takes the form
 ¡ ¢ 
 ẋ = f x, u∗ (x, p) ,  x(T ) = x̄ ,

¡ ¢ (10.26)
 ṗ = −p · ∂f x, u∗ (x, p) , 

∂x p(T ) = ∇g(x̄) .

We now recognize that the evolution equations in (10.26) and the optimality conditions (10.25) are
precisely those given in the Pontryagin Maximum Principle. In other words, let u∗ (·) be a control for
which the Pontryagin Maximum Principle is satisfied. Then the corresponding trajectory x(·) and
the adjoint vector p(·) provide a solution to the equations of characteristics for the corresponding
hamiltonian system (10.16).

42
 References

• General theory of P.D.E.

[E] L. C. Evans, Partial Differential Equations, Amer. Math. Soc. Graduate Studies in Mathe-
matics, Vol. 19, Providence 1998.

• Control systems: basic theory and the Pontryagin Maximum Principle

[Ce] L. Cesari, Optimization - Theory and Applications, Springer, 1983.

[FR] W. H. Fleming and R. W. Rishel, Deterministic and Stochastic Optimal Control, Springer,
1975.

[LM] E. B. Lee and L. Markus, Foundations of Optimal Control Theory, Wiley, 1967.

[PBGM] L. S. Pontryagin, V. Boltyanskii, R. V. Gamkrelidze and E. F. Mishenko, The Mathematical

Theory of Optimal Processes, Wiley, 1962.

• Viscosity solutions to Hamilton-Jacobi equations

[Ba] G. Barles, Solutions de viscosité des équations de Hamilton-Jacobi, Springer, 1994.

[BC] M. Bardi and I. Capuzzo Dolcetta, Optimal Control and Viscosity Solutions of Hamilton-
Jacobi-Bellman Equations, Birkhäuser, 1997.

[C-L] M. Crandall and P. L. Lions, Viscosity solutions of Hamilton-Jacobi equations, Trans. Amer.
Math. Soc. 277 (1983), 1-42.

[CILL] M. Crandall, H. Ishii and P. L. Lions, User’s guide to viscosity solutions of second order partial
differential equations, Bull. Amer. Math. Soc. 27 (1992), 1-67.

[TTN] D. V. Tran, M. Tsuji and D. T. S. Nguyen, The Characteristic Method and its Generalizations
for First Order Nonlinear Partial Differential Equations, Chapman&Hall/CRC, 1999.

• Scalar conservation laws

[K] S. Kruzhkov, First order quasilinear equations with several space variables, Math. USSR
Sbornik 10 (1970), 217-243.

43