Introduction to Optimal Control Theory

and Hamilton-Jacobi equations

Seung Yeal Ha
Department of Mathematical Sciences
Seoul National University

1
 A priori message from SYHA

”The main purpose of these series of lectures is to make

you understand ABC of OCT and help you prepare ad-
vanced course on stochastic optimal control theory that
you might have in future. ”

2
 Lecture Schedules

• Lecture 1: ABC of Optimal Control Theory

• Lecture 2: PMP v.s. Bellman’s dynamic programming

• Lecture 3: Hamilton-Jacobi equations (classical theory)

• Lecture 4: Hamilton-Jacobi equations (modern theory)

3
 References

• An introduction to mathematical optimal control theory by

Craig Evans. Available at www.math.berkeley.edu/ evans.

• Optimal Control Theory by Donald E. Kirk.

• Introduction to the mathematical theory of control by Alberto

Bressan and Benedetto Piccoli.

4
 What do you mean by control of system ?

Control of a system has a double meaning:

• (Weak sense): Checking or testing whether the system’s be-

havior is satisfactory.

• (Strong sense): Putting things in order to guarantee that

the system behaves as desired.

5
 Maxims

• ”Since the building of the universe is perfect and is created

by the wisdom creator, nothing arises in the universe in which
one cannot see the sense of some maximum or minimum.” by
Leonhard Euler

• ”The words control theory are, of course, of recent origin, but

the subject itself is much older, since it contains the classical
calculus of variations as a special case, and the first calculus of
variations problems go back to classical Greece.” by Hector J.
Sussmann.

6
Lecture 1: ABC of Optimal Control Theory

7
 Goal of OCT

The objective of OCT is to determine the control signals that

will cause a process to satisfy the physical constraints and at
the same time minimize (or maximize) some performance
criterion.

Minimize cost and maximize payoff, utility

”A problem well put is a problem half solved”

8
 Key words of Lecture 1

• Controlled system, state, control, Performance measure

• Controllability, reachable set, linear time-invariant system

• Bang-bang control (principle), Kalman’s rank theorem, ob-

servability,

9
 Three giants of modern control theory

• Richard Bellman (August 26, 1920 - March 19, 1984)

Bellman’s dynamic programming (1953)

10
• Lev Pontryagin (3 September 1908 – 3 May 1988)

Pontryagin’s Maximum principle (PMP) (1956)

11
• Rudolf E. Kalman (19 May 1930)

Kalman filter, Kalman’s rank theorem (1960)

12
 Controlled dynamical systems

• (Dynamical system)

ẋ = f (x, t), x = (x1, · · · , xn), x = x(t) ∈ Rn.

x: state (variable)

• (Controlled dynamical system)

ẋ = f (x, u, t), u = (u1, · · · , um), u = u(t) ∈ Rm.

u: control (parameter)

13
• Example 1: Merton’s optimal investment and consumption

 Step A: (Modeling of a problem)

Consider an individual whose wealth today is W 0, and who will

live exactly T years. His task is to plan the rate of consumption
of wealth C(s) for 0 < s < T . All wealth not yet consumed earns
interest at a fixed rate r. We have, for simplicity, assigned no
utility to final-time wealth (a bequest).

Let W (t) to be the amount of wealth at time t.

(
Ẇ = rW − C, 0 < t ≤ T,
W (0) = W 0,
 Step B: (Identification of physical constraints)

W (t) ≥ 0, W (T ) = 0, C(t) ≥ 0.

 Step C: (Performance measure)

Z T
P [C] = e−ρtU (C(s))ds, max P [C].
0
where ρ is a discounting rate, and U is the utility function of
consumption.

• Reformulation as calculus of variation problem

Z T
max e−ρth(rW − Ẇ )ds, subject to W (0) = W 0.
W (·) 0
• Example 2. (automobile problem): Minimum-time optimal
problem

 Step A: (Modeling of a problem)

The car is to be driven in a strainght line away from the point 0

to the point e. The distance of the car from 0 at time t is given
by d(t). For simplicity, we assume that the car is denoted by the
unit point mass that can be accelerate by using the throttle or
decelerated by using the brake.

We set
¨ = α(t) + β(t),
d(t)
where α and β stand for throttle accelerate and braking decel-
eration respectively.
Again we set

x1 := d, ˙
x2 := d, u1 := α, u2 := β.
Then our controlled dynamical system is given by
! !
0 1 0 1
ẋ(t) = x(t) + u(t),
0 0 1 1
and two point boundary conditions:
! !
0 e
x(t0) = , x(tf ) = .
0 0
 Step B: (Identification of physical constraints)

State constraints
0 ≤ x1 ≤ e, 0 ≤ x2 .

Control constraints
0 ≤ u1 ≤ M1, −M2 ≤ u2 ≤ 0.

Assume that the amount of gasoline at time t = t0 is G gal-

lons, and the rate of gas consumption is proportional to both
accelerate and speed, thus the amount of gasoline used
Z t  
f
k1u1(t) + k2x2(t) dt ≤ G.
t0
 Step C: (Performance measure)

Minimum-time control

minimize J := tf − t0.
In general, the performance measure takes the form of
Z t
f
J = g(x(tf ), tf ) + r(x(t), u(t), t)dt,
t0

g ≡ 0; Lagrange problem, r ≡ 0: Mayer problem

Admissible controls: controls satisfying physical constraints,

admissible trajectory
 Optimal control problem

Optimal control problem is to find an admissible control u∗ which

causes the system ẋ = f (x, u, t) to follow an admissible trajectory
x∗ that maximize the performance measure (payoff)
Z t
f
P = g(x(tf ), tf ) + r(x(t), u(t), t)} dt.
| {z } t0 | {z
terminal payoff running payoff

In this lecture, we assume that Uad denotes the set of all admis-
sible controls:

Uad := {u : R+ → U : u = u(·) is measurable

and satisfies constraints},
14
and U = [−1, 1]n : symmetric and convex.

• Basic problem

To find a control u∗ = u∗(t) ∈ Uad which maximize th payoff

P [u∗] ≥ P [u], u ∈ Uad.

 Main questions

1. Does an optimal control exist ? (Existence of optimal con-

trol)

2. How can we characterize an optimal control mathematically

? (characterization of optimal control)

3. How can we construct an optimal control ? (realization of

optimal control)
 Two examples

1. Control of production and consumption

x(t) := amount of output produced at time t ≥ 0

 Assumptions:

• We consume some fraction of output at each time

• We reinvest the remaining fraction

15
Let u = u(t) be the fraction of output reinvested at time t ≥ 0,
0 ≤ u ≤ 1. In this case
(
ẋ = ux, 0 < t ≤ T,
x(0) = x0,
where
x ≥ 0, 0 ≤ u ≤ 1, U = [0, 1].
and
Z T
P [u] = (1 − u(t))x(t)dt.
0
2. A pedulum problem
(
θ̈ + λθ̇ + ω 2 sin θ = u, 0 < t ≤ T,
θ(0) = θ0, θ̇(0) = ω0.
We use the approximation

sin θ ≈ θ, |θ|  1
to get the linear approximate equation
(
θ̈ + λθ̇ + ω 2θ = u, 0 < t ≤ T,
θ(0) = θ0, θ̇(0) = ω0.
So the main question is to determine the control u so that (θ, θ̇
approaches (0, 0) as soon as possible. (minimum-time control
problem)
We set
x1 = θ, x2 = θ̇,
then, we have
! ! ! !
d x1 0 1 x1 0
= + .
dt x2 −ω 2 −λ x2 u
We now set
! !
x1 0
τ = τ [α(·)] : first time such that = .
x2 0
We also define
Z τ
P [α] = − 1dt = −τ.
0
 Controllability

In order to be able to do whatever we want with the given dynam-

ical system under control input, the system must be controllable.

16
 Controllability

Consider a controlled dynamics

(
ẋ = f (x, u), t0 < t < ∞,
x(t0) = x0,
We set the solution of the above controlled system as x(t; t0, x0):

•Natural a prior question before optimal control

For a fixed desired state x0, xf ∈ Rn, can we find a control u ∈ Uad
and tf < ∞ such that

x(tf ; t0, x0) = xf .

17
• Controllability question

Given the initial point x0 ∈ Rn and a set S ⊂ Rn,

does there exist a control u steering the system to the set S in

finite time ?

For the case S = {xf }, controllability question asks us

∃
( T < ∞, u ∈ Uad such that
ẋ = f (x, u), 0 < t < ∞,
x(0) = x0, x(T ) = xf .
 Controllability for a linear system

From now on, consider a linear-control system and S = {0}, i.e.,

M ∈ M n×n, N ∈ M n×m,
(
ẋ = M x + N u,
(1)
x(0) = x0.

Definition:

1. A linear-control system (??) is completely controllable ⇐⇒

For any x0, xf ∈ Rn, there exists a control u : [0, tf ] → Rm
such that x(tf ) = xf .

18
2. Reachable set

C(t) := {x0 : ∃ u ∈ A such that x(t, u, x0) = 0},

C := ∪t≥0C(t) : reachable set

• Simple observation

0 ∈ C(t) and
x0 ∈ C(t), t̂ > t =⇒ x0 ∈ C(t̂).
Looking for sufficient and necessary
condition of controllability

19
 Snapshot of ODE theory

Consider a homogeneous linear ODE:

M ∈ M n×n : constant matrix
(
ẋ = M x, t > 0,
x(0) = x0.

Then we have
x(t) = Φ(t)x0,
where Φ is a fundamental matrix define by
∞ k k
t M

Φ(t) = etM
X
:= .
k=0
k!

20
Consider an inhomogeneous linear system:
(
ẋ = M x + f (t), t > 0,
x(0) = x0.

Then the solution x can be given by the variation of parameters

formula (Duhamel’s formula)
Z t
x(t) = Φ(t)x0 + Φ(t − s)f (s)ds,
0
where Φ(t − s) = Φ(t)Φ−1(s).
We now return to
(
ẋ = M x + N u, t > 0,
x(0) = x0.

Then by Duhamel’s formula, we have

Z t
x(t) = Φ(t)x0 + Φ(t) Φ−1(s)N u(s)ds.
0

Note that

x0 ∈ C(t) ⇐⇒ x(t) = 0Z
t
⇐⇒ x0 = − Φ−1(s)N u(s)ds, for some u ∈ Uad.
0
• Theorem (Geometry of reachable set)

Rechable set C is symmetric and convex, i.e.,

(i) x0 ∈ C =⇒ −x0 ∈ C.
(ii) x0, x̂0 ∈ C, λ ∈ [0, 1] =⇒ λx0 + (1 − λ)x̂0 ∈ C.
 Kalman’s rank theorem (1960)

Consider

ẋ = M x + N u, M ∈ M n×n, N ∈ M n×m. (2)

We define a controllability matrix

G(M, N ) := [N |M N | · · · |M n−1N ].

• Definition

The linear control system (??) is controllable ⇐⇒ C = Rn .

21
• Theorem

rankG(M, N ) = n ⇐⇒ 0 ∈ C 0.

Proof. • (⇐=)

Suppose that 0 ∈ C 0. Note that

rankG(M, N ) ≤ n.
If rankG(M, N ) < n, then there exists b 6= 0 such that

btG(M, N ) = 0.
This yields

btN = btM N = · · · = btM n−1N = O.

By Cayley-Hamilton’s theorem, we also have

btM k N = O, k ≥ 0, btΦ−1(t)N = O.
We now claim

b is perpendicular to C(t), i.e., C 0 = ∅.

If x0 ∈ C(t), then
Z t
x0 = − Φ−1(s)N u(s)ds, u ∈ Uad.
0
Therefore,
Z t
bt x 0 = − btΦ−1(s)N u(s)ds = 0.
0
• (=⇒) Suppose that 0 6∈ C 0, i.e.,

0 ∈ (C 0)c ⊂ ∩t≥0(C 0(t))c.

Then 0 6∈ C 0(t), ∀ t ≥ 0. Therefore

0 ∈ ∂C(t).
Since C(t) is convex, there exists b 6= 0 such that

btx0 ≤ 0, x0 ∈ C(t).
For x0 ∈ C(t),
Z t
x0 = − Φ−1(s)N u(s)ds, u ∈ Uad.
0
Thus
Z t
bt x 0 = − btΦ−1(s)N u(s)ds ≤ 0.
0
This yields
btΦ−1(s)N = 0.
By differentiating the above relation, we have

btN = btM N = · · · = btM n−1N = O, i.e., btG(M, N ) = 0.

Hence rankG(M, N ) < n.
• Theorem Let λ be the eigenvalue of M .

rankG(M, N ) = n and Re(λ) ≤ 0 =⇒ The system (??) is

controllable.
• Theorem Kalman (1960)

The system (??) is controllable ⇐⇒ rank(G) = n.

This rank indicates how many components of the system are sensitive to the
action of the control
• Examples

1.

n
( = 2, m = 1, A = [−1, 1]
ẋ1 = 0, t > 0,
ẋ2 = u(t).

2.
(
ẋ1 = x2, t > 0,
ẋ2 = u(t).

3.
(
ẋ1 = u(t), t > 0,
ẋ2 = u(t).
 Observability

In order to see what is going on inside the system under obser-

vation, the system must be observable

22
 Observability

• Consider uncontrolled system:

ẋ = M x, t > 0, M ∈ M n×n, x ∈ Rn ,
(
(3)
x(0) = x0 : unknown.
Note that
x(t) = etM x0. (4)
Once we know x0, then we know everything !!

Suppose that we have observed data y:

y(t) = N x(t), N ∈ M m×n, m  n.

23
• Observability question

”Given the observation y = y(·) which is low-dimensional, can

we in principle construct high-dimensional real dynamics x = x(·)
?

Thus problem is how to recover x0 ∈ Rn from the observed data

y.

Note that

y(0) = N x(0) = N x0,

ẏ(0) = N ẋ(0) = N M x0,
··· = ··· ,
y (n−1)(0) = N x(n−1)(0) = N M n−1x0.
This yields
   
y(0) N
y 0(0)
 
  
= NM 
 0
 x .
· ·
  
   
y (n−1)(0) N M n−1

Thus, we have
 
N
 NM 
(??) and (??) is observable ⇐⇒ rank   = n.
 
 · 
N M n−1
• Definition
(
ẋ = M x, t > 0,
is observable
x(0) = x0, y = N x
⇐⇒ For two solutions x1 and x2 such that
N x1 = N x2 on [0, t] x1(0) = x2(0).

Example: N = I, N = 0.

• Theorem (Observability is a dual concept of controllability)

(
ẋ = M x, t > 0,
is observable
y = N x,
⇐⇒ ż = M tz + N tu is controllable.
 La Salle’s Bang-bang control

Uad := {u : [0, ∞) → U = [−1, 1]m : u is measurable}.

• Definition Let u ∈ Uad.

u = (u1, · · · , um) is a bang-bang control
⇐⇒ |ui(t)| = 1, ∀ t > 0, i = 1, · · · , m.

• Theorem (Bang-bang principle)

(
ẋ = M x + N u, 0 ∈ C(t)
x
x(0) = x0,
=⇒ ∃Z u∗ = u∗(·) : bang-bang control such that
t
x0 = − Φ−1(s)N u∗(s)ds.
0
24
 Preliminaries for bang-bang principle

L∞ = L∞(0, t; Rm) = {u : [0, t] → Rm : sup |u(s)| < ∞},

0≤s≤t
||u||L∞ := sup |u(s)|, L1 ⊂ (L∞)∗.
0≤s≤t

• Definition Let un, u ∈ L∞.

un → u in Zweak-star topology
t Z t
⇐⇒ un(s)ϕ(s)ds → u(s)ϕ(s)ds,
0 0
⇐⇒ un converges to u in weak-star topology,
1 Rt
where ϕ is a L -test function with 0 |ϕ(s)|ds < ∞.

25
• Theorem (Banach-Alaoglu)

Any bounded set in L∞ is weak-star compact.

• Corollary

If un ∈ Uad := {u : [0, t] → [−1.1]m : u is measurable},

∃ {unk } : subsequence of un such that
unk → u weak-star topology.
• Definition Let K ⊂ Rn and z ∈ K.

1. K is convex ⇐⇒ ∀ x, y ∈ K, 0 ≤ λ ≤ 1, λx + (1 − λ)y ∈
K.

2. z ∈ K is an extreme point ⇐⇒
there does not exist x, x̂ ∈ K and λ ∈ (0, 1) such that z =
λx + (1 − λ)x̂.

• Theorem (Krein-Milman)

K 6= ∅ : convex and weak-star compact

=⇒ K has at least one extreme pint.
 The proof of bang-bang’s principle

Let x0 ∈ C(t). Then, we set

K := {u ∈ Uad : u steers x0 to 0 at time t}.

• Lemma

K 6= ∅ : convex and weak-star compact.

26