See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224773123

Dynamic Programming and Optimal Control

Chapter · January 1995

CITATIONS READS

2,838 30,823

1 author:

Dimitri P. Bertsekas
Arizona State University
386 PUBLICATIONS 71,092 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

1) Approximate and abstract dynamic programming. 2) Proximal algorithms for large-scale linear systems of equations View project

All content following this page was uploaded by Dimitri P. Bertsekas on 21 December 2016.

The user has requested enhancement of the downloaded file.

 Dynamic Programming & Optimal Control

Adi Ben-Israel
Adi Ben-Israel, RUTCOR–Rutgers Center for Operations Research, Rut-
gers University, 640 Bartholomew Rd., Piscataway, NJ 08854-8003, USA
E-mail address: [email protected]
 LECTURE 1

Dynamic Programming

1.1. Recursion
A recursion is a rule for computing a value using previously computed values, for example,
the rule
fk+1 = fk + fk−1 , (1.1)
computes the Fibonnaci sequence, given the initial values f0 = f1 = 1.
Example 1.1. A projectile with mass 1 shoots up against earth gravity g. The initial
velocity of the projectile is v0 . What maximal altitude will it reach?
Solution. Let y(v) be the maximal altitude reachable with initial velocity v. After time ∆t,
the projectile advanced approximately v ∆t, and its velocity has decreased to approximately
v − g ∆t. Therefore the recursion
y(v) ≈ v δt + y(v − g ∆t) , (1.2)
gives
y(v) − y(v − g ∆t)
≈ v,
∆t
v
∴ y 0 (v) =
g
v2
∴ y(v) = , (1.3)
2g
and the maximal altitude reached by the projectile is v02 /2 g.
Example 1.2. Consider the partitioned matrix
µ ¶
B c
A= (1.4)
r α
where B is nonsingular, c is a column, r is a row, and α is a scalar. Then A is nonsingular
iff
α − rB −1 c 6= 0 , (1.5)
(verify!) in which case
µ ¶
−1 B −1 + βB −1 crB −1 −βB −1 c
A = , (1.6a)
−βrB −1 β
1
where β = . (1.6b)
α − rB −1 c
3
4 1. DYNAMIC PROGRAMMING

Can this result be used in a recursive computation of A−1 ? Try for example
   
1 2 3 4 −1 −2
A= 1  2 4 , with inverse A−1 =  0 −1 1  .
1 3 4 −1 1 0
Example 1.3. Consider the LP
max cT x
s.t. A x = b
x ≥ 0
and denote its optimal value by V (A, b). Let the matrix A be partitioned as A = (A1 , an ),
where an is the last column, and similarly partition the vector c as cT = (cT1 , cn ). Is the
recursion
V (A, b) = max {cn xn + V (A1 , b − an xn )} (1.7)
xn ≥0
valid? Is it useful for solving the problem?

1.2. Multi stage optimization problems and the optimality principle

Consider a process consisting of N sequential stages, and requiring some decision at each
stage. The information needed at the beginning of stage i is called its state, and denoted
by xi . The initial state x1 is assumed given. A decision ui is selected in stage i from the
feasible decision set Ui (xi ), that in general depends on the state xi . This results in a reward
ri (xi , ui ) and the next state becomes
xi+1 = Ti (xi , ui ) , (1.8)
where Ti (·, ·) is called the state transformation (or dynamics) in stage i. The terminal
state xN +1 results in a reward S(xN +1 ), called the salvage value of xN +1 . If the rewards
accumulate, the problem becomes
 
X N xi+1 = Ti (xi , ui ) , i = 1, N , 
max ri (xi , ui ) + S(xN +1 ) : ui ∈ Ui (xi ) , i = 1, N , . (1.9)
 1=1 
x1 given .
This problem can be solved, in principle, as an optimization problem in the variables
u1 , . . . , uN . However, this ignores the special, sequential, structure of the problem.
The Dynamic Programming (DP) solution is based on the following concept.
Definition 1.1. The optimal value function (OV function) at stage k, denoted Vk (·), is
 
X N xi+1 = Ti (xi , ui ) , i ∈ k, N , 
Vk (x) := max ri (xi , ui ) + S(xN +1 ) : ui ∈ Ui (xi ) , i ∈ k, N , . (1.10)
 
1=k xk = x .
The optimal value of the original problem is V1 (x1 ). An optimal policy is a sequence of
decisions {u1 , . . . , uN } resulting in the value V1 (x1 ).
Bellman’s Principle of Optimality (abbreviated PO) is often stated as follows:
 1.2. MULTI STAGE OPTIMIZATION PROBLEMS AND THE OPTIMALITY PRINCIPLE 5

An optimal policy has the property that whatever the initial state and the
initial decisions are, the remaining decisions must constitute an optimal
policy with regard to the state resulting from the first decision, [5, p. 15].
The PO can be used to recursively compute the OV functions
Vk (x) = max {rk (x, u) + Vk+1 (Tk (x, u), u)} , k ∈ 1, N , (1.11)
u∈Uk (x)

VN +1 (x) = S(x) . (1.12)

The last equation is called the boundary condition (BC).
P
N
Example 1.4. It is required to partition a positive number x in N parts, x = ui , such
i=1
P
N
that the sum of squares u2i is minimized.
i=1
For each k ∈ 1, N define the OV function
( k k
)
X X
Vk (x) := min u2i : ui = x .
i=1 i=1

The recursion (1.11) here becomes

© 2 ª
Vk (x) := min u + Vk−1 (x − u) ,
u∈[0,x]

with V1 (x) = x2 as BC.

For k = 2 we get
© 2 ª x2 x
V2 (x) := min u + (x − u)2 = , with optimal u = .
u∈[0,x] 2 2
x2 x
Claim: For genral N , the optimal value is VN (x) = , with optimal policy of equal ui = .
N N

Proof. (by induction)

© 2 ª
VN (x) = min u + VN −1 (x − u)
u∈[0,x]
½ ¾
2 (x − u)2
= min u + , by the induction hypothesis ,
u∈[0,x] N −1
x2 x
= , for u = .
N N
¤
The following example shows that the PO, as stated above, is incorrect. For details see
the book [29] by M. Sniedovich where the PO is given a rigorous treatment.
6 1. DYNAMIC PROGRAMMING

Example 1.5. ([29, pp. 294–295]) Consider the graph in Figure 1.1. It is required to
find the shortest path from node 1 to node 7, the length of a path is defined as the length
of its longest arc. For example,
length{1, 2, 5, 7} = 4 .
If the nodes are viewed as states, then the path {1, 2, 4, 6, 7} is optimal w.r.t. s = 1. However,
the path {2, 4, 6, 7} is not optimal w.r.t. the node s = 2, as its length is greater than the
length of {2, 5, 7}.

2
2 5

3 1
4 1

4 7
1
1 1
2
5
2
3 6

Figure 1.1. An illustration why the PO should be used carefully

1.3. Inverse Dynamic Programming

Consider a multi–stage decision process of § 1.2, where the state x is the project budget.
A reasonable question is to determine the minimal budget that will enable achieving a given
target v, i.e.
min {x : V1 (x) ≥ v} ,
denoted I1 (v) and called the optimal input for v. Similarly, at any stage we define the optimal
input as
½
min {x : Vk (x) ≥ v} ,
Ik (v) := k = 1, 2, · · · , N . (1.13)
∞ if the target v is unattainable ,
The terminal optimal input is defined as
IN +1 (v) := S −1 (v) (1.14)
assuming the salvage value function S(·) is monotonic.
A natural recursion for the optimal inputs is:
Ik (v) := min{x : ∃u ∈ Uk (x)  Tk (x, u) ≥ Ik+1 (v −rk (x, u))} , k = N, N −1, · · · , 1 , (1.15)
with (1.14) as BC.
 1.3. INVERSE DYNAMIC PROGRAMMING 7

Exercises.
Exercise 1.1. Use DP to maximize the entropy
( N N
)
X X
max − pi log pi : pi = 1 .
i=1 i=1

Exercise 1.2. Let Z+ denote the nonnegative integers. Use DP to write a recursion for
the knapsack problem
( N N
)
X X
max fi (xi ) : wi (xi ) ≤ W, , xi ∈ Z+ ,
i=1 i=1

where ci (·), wi (·) are given functions: Z+ → R and W > 0 is given. State any additional
properties of fi (·) and wi (·) that are needed in your analysis.
Exercise 1.3. ([7]) A cash amount of x cents can be represented by
x1 coins of 50 cents ,
x2 coins of 25 cents ,
x3 coins of 10 cents ,
x4 coins of 5 cents , and
x5 coins of 1 cent .
The representation is:
x = 50x1 + 25x2 + 10x3 + 5x4 + x5 .
(a) Use DP to find the representation with the minimal number of coins.
(b) Show that your solution agrees with the ”greedy” solution:
jxk ¹ º
x − 50x1
x1 = , x2 = , etc.
50 25
where bαc is the greatest integer ≤ α.
(c) Suppose a new coin of 20 cents is introduced. Will the DP solution still agree with the
greedy solution?
Exercise 1.4. The energy required to compress a gas from pressure p1 to pressure pN +1
in N stages is proportional to
µ ¶α µ ¶α µ ¶α
p2 p3 pN +1
+ + ··· +
p1 p2 pN
with α a positive constant. Show how to choose the intermediate pressures p2 , · · · , pN so as
to minimize the energy requirement.
Exercise 1.5. Consider the following variation of the game NIM, defined in terms of N
piles of matches containing x1 , x2 , · · · , xN matches. The rules are:
(i) two players make moves in alternating turns,
(ii) if matches remain, a move consists of removing any number of matches all from the same
pile,
8 1. DYNAMIC PROGRAMMING

(iii) the last player to move loses.

For example, consider a game with initial piles {x1 , x2 , x3 } = {1, 4, 7} where moves by players
I II
I, II are denoted by → and → resp.,
I II I II I
{1, 4, 7} → {1, 4, 5} → {0, 4, 5} → {0, 4, 4} → {0, 3, 4} → {0, 3, 3}
II I II I II
→ {0, 2, 3} → {0, 2, 2} → {0, 2, 1} → {0, 0, 1} → {0, 0, 0} , and II loses .
(a) Use DP to find an optimal move for an initial state {x1 , x2 , · · · , xN }.
(b) Find a simple rule to determine if an initial state is a winning position.
(c) Is {x1 , x2 , x3 , x4 } = {1, 3, 5, 7} a winning position?
Hint: Let V (x1 , x2 , · · · , xN ) be the optimal value of having the piles {x1 , x2 , · · · , xN } when
it is your turn to play, with
½
0, if {x1 , x2 , · · · , xN } is a losing position
V (x1 , x2 , · · · , xN ) =
1, otherwise.
Let yi be the number of matches removed from pile i. Then
½ ¾
V (x1 , x2 , · · · , xN ) = 1 − max max V (x1 , x2 , · · · , xi − ui , · · · , xN )
i=1,...,N 1≤yi ≤xi

where
V (x1 , x2 , · · · , xN ) = 0 if all xi = 0 except, say xj = 1
Exercise 1.6. We have a number of coins, all of the same weight except for one which
is of different weight, and a balance.
(a) Determine the weighing procedures which minimize the maximum time required to locate
the distinctive coin in the following cases:
• the coin is known to be heavier,
• it is not known whether the coin is heavier or lighter.
(b) Determine the weighing procedures which minimize the expected time required to locate
the coin.
(c) Consider the more general problem where there are two or more distinctive coins, under
various assumptions concerning the distinctive coins.
Exercise 1.7. A rocket consists of k stages carrying fuel and a nose cone carrying the
pay load. After the fuel carried in stage k is consumed, this stage drops off, leaving a k − 1
stage rocket. Let
W0 = weight of nose cone ,
wk = initial gross weight of stage k ,
Wk = Wk−1 + wk , initial gross weight of sub-rocket k ,
pk = initial propellant weight of stage k ,
vk = change in rocket velocity during burning of stage k .
Assume that the change in velocity vk is a known function of Wk and pk , so that .
vk = v(Wk , pk )
 1.3. INVERSE DYNAMIC PROGRAMMING 9

from which
pk = p(Wk , vk )
Since Wk = Wk−1 + wk , and the weight of the kth stage is a known function, g(pk ), of the
propellant carried in the stage, we have
wk = w(p(Wk−1 + wk , vk ))
whence, solving for wk , we have
wk = w(Wk−1 , vk )
(a) Use DP to design a k-stage rocket of minimum weight which will attain a final velocity
v.
(b) Describe an algorithm for finding the optimal number of stages k ∗ .
(c) Discuss the factors resulting in an increase of k ∗ i.e. in more stages of smaller size.
Exercise 1.8. Suppose that we are given the information that a ball is in one of N
boxes, and the a priori probability, pk , that it is in the kth box.
(a) Show that the procedure which minimizes the expected time required to find the ball
consists of looking in the most likely box first.
(b) Consider the more general problem where the time consumed in examining the kth box
is tk , and where there is a probability qk that the examination of the kth box will yield
no information about its contents. When this happens, we continue the search with the
information already available. Let F (p1 , p2 , · · · , pN ) be the expected time required to find
the ball under an optimal policy. Find the functional equation that F satisfies.
(c) Prove that if we wish to “obtain” the ball, the optimal policy consists of examining first
the box for which
pk (1 − qk )
tk
is a maximum. On the other hand, if we merely wish to “locate” the box containing the ball
in the minimum expected time, the box for which this quantity is maximum is examined
first, or not at all.
Exercise 1.9. A company has m jobs that are numbered 1 through m. Job i requires ki
employees. The “natural” monthly wage of job i is wi dollars, with wi ≤ wi+1 for all i. The
jobs are to be grouped into n labor grades, each grade consisting of several consecutive jobs.
All employees in a given labor grade receive the highest of the natural wages of the jobs in
that grade. A fraction rj of the employees in each jobs quit in each month. Vacancies must
be filled by promoting from the next lower grade. For instance, a vacancy in the highest of n
labor grades causes n − 1 promotions and a hire into the lowest labor grade. It costs t dollars
to train an employee to do any job. Write a functional equation whose solution determines
the number n of labor grades and the set of jobs in each labor grade that minimizes the sum
of the payroll and training costs.
Exercise 1.10. (The Jeep problem) Quoting from D. Gale, [18, p. 493],
· · · the problem concerns a jeep which is able to carry enough fuel to travel
a distance d, but is required to cross a desert whose distance is greater then
d (for example 2d). It is to do this by carrying fuel from its home base
10 1. DYNAMIC PROGRAMMING

and establishing fuel depots at various points along its route so that it can
refuel as it moves further out. It is then required to cross the desert on the
minimum possible amount of fuel.
Without loss of generality, we can assume d = 1.
A desert of length 4/3 can be crossed using 2 units of fuel, as follows:
Trip 1: The jeep travels a distance of 1/3, deposits 1/3 unit of fuel and returns to base.
Trip 2: The jeep travels a distance of 1/3, refuels, and travels 1 unit of distance.
The inverse problem is to determine the maximal desert that can be crossed, given the
quantity of fuel is x. The answer is:
1 1 1 x − dxe + 1
1 + + + ··· + + .
3 5 2dxe − 3 2dxe − 1
References: [1], [9], [11], [15], [17], [18], [21], [25].
 LECTURE 2

Dynamic Programming: Applications

2.1. Inventory Control

An inventory is a dynamic system described by three variables: state xt , decision u and
demand w, an exogeneous variable that may be deterministic or random (the interesting
case).
In the simplest case the state transformation is
xt+1 := xt + ut − wt , t = 0, 1, · · · (2.1)
where xt is the stock level at the beginning of day t, ut is the amount ordered from a supplier
that arrives at the beginning of day t, and wt is the demand during day t.
The initial stock x0 is given. Note that (2.1) allows for negative stocks.
The economic functions are:
• a holding cost A(x) for holding x units overnight (shortage cost if x < 0),
• an order cost B(u) of placing an order u, and
• revenue the R(w) from selling w units.
Let Vt (x) be the minimal cost of operations if day t begins with x units in stock. Then
Vt (x) := min{A(x) + B(u) − ER(w) + EVt+1 (x + u − w)} (2.2)
u

where E is expectation w.r.t. w. We assume the boundary condition

VN +1 (x) = S(x) (2.3)
where S(·) is the cost of terminal stock.
Typical constraints on u are: u ≥ 0 and u ≤ C − x where C is a storage capacity.
The order cost is typically of the form
½
0 if u = 0
B(u) = (2.4)
K + cu if u > 0
where K is a fixed cost.
It is convenient to use the normalized cost functions
V ∗ (x) := V (x) + cx (2.5)
in which case (2.2) becomes
Vt∗ (x) := min{A(x) + KH(u) − (ER(w) − cEw) + EVt+1
∗
(x + u − w)} (2.6)
u

where H(·) is the Heaviside function, H(u) = 1 if u > 0 and 0 otherwise.

11
12 2. DYNAMIC PROGRAMMING: APPLICATIONS

Theorem 2.1. Let A(x) and S(x) be convex with limit +∞ at x = ±∞, and let K = 0.
Then the optimal policy is (
0 if xt > St ,
ut = (2.7)
St − xt otherwise ,
for some St > 0 (the order to level on day t).
Proof. Let C be the class of convex functions with limit +∞ as the argument approaches
±∞. The function minimized in (2.6) is
∗
θ(x + u) = EVt+1 (x + u − w) (2.8)
If θ(x) ∈ C and has minimum at St then the rule (2.7) is optimal. The minimized function
is ½
θ(St ) if x ≤ St ,
min θ(x + u) = (2.9)
u θ(x) if x > St ,
that is constant for x ≤ St .
Writing (2.6) as Vt := LVt+1 , it follows that Vt ∈ C if Vt+1 ∈ C. Since VN +1 ∈ C by (2.3),
it follows from Vt = LN −t VN +1 is in C ¤
The determination of the optimal St is from
Vt∗ (x) = A(x) + EVt+1
∗
(S − w) = A(x) + θ(S) , x ≤ S ,
and the optimal S minimizes EA(S − w).
Theorem 2.2. If A(x) and S(x) are convex with the limit +∞ at x = ±∞ and if K 6= 0
then the optimal order policy is
(
0 if x > s ,
u= (2.10)
S − x otherwise ,
where S > s if K > 0.
Proof. The relevant optimality condition is (2.6). The expression to be minimized is
(
θ(x) if u = 0 ,
KH(u) + θ(x + u) = (2.11)
K + θ(x + u) if u > 0 ,
and θ is (2.8). If θ ∈ C with minimum at S then the rule (2.10) is optimal for s the smaller
root of
θ(s) = K + θ(S) .
∗
In general the functions V and θ are not convex. The proof resumes after Lemma 2.4
below. ¤
Definition 2.1. A scalar function φ(x) is K-convex if
φ(x + u) ≥ φ(x) + uφ0 (x) − K , ∀u , (2.12)
where φ0 is the derivative from the right,
φ(x + u) − φ(x)
φ0 (x) = lim .
u↓0 u
 2.1. INVENTORY CONTROL 13

The class of K-convex functions with limit +∞ at x = ±∞ is denoted by CK . ¤

Immediate consequences of the definition:
(a) K-convexity implies L-convexity for L ≥ K
(b) 0-convexity is ordinary convexity
(c) if φ(x) is K-convex, so is φ(x − w)
(d) if φ1 and φ2 are K1 and K2 convex respectively then α1 φ1 + α2 φ2 is α1 K1 + α2 K2
convex for α1 , α2 > 0
(e) if φ is K-convex so is Eφ(x − w)
The corresponding statements for CK are:
(a) CK increases with K
(b) C0 = C
(c) CK is closed under translations
P P
(d) φi ∈ CKi and αi ≥ 0 imply αi φi ∈ CK where K = αi Ki
(e) CK is closed under averaging.
Lemma 2.1. Let K > 0 and suppose θ is left-continuous. If one orders for all x in some
interval, then one orders up to a common level.
Proof. Let x0 < x00 be two points in the interval from where one orders up to S 0 and
S respectively. Then θ(S 0 ) ≤ θ(S 00 ), and S 0 < x00 (otherwise at x00 one should order up
00

to S 0 ). Therefore S 0 belongs to the interval, and in (x0 , S 0 ) one should order up to S 0 if at

all. For x near S 0 the inequality K + θ(S 0 ) < θ(x) is violated, and one does not order, a
contradiction. ¤
Lemma 2.2. Let θ(x) is left-continuous and K-convex, and let x1 < x2 < x3 < x4 , where
it is optimal not to order in (x1 , x2 ) and (x3 , x4 ). Then it is not optimal to order in (x2 , x3 ).
Proof. Follows from Lemma 2.1. ¤
Lemma 2.3. If θ(x) ∈ CK then an (s, S) policy is optimal.
Proof. By Lemma 2.2 the optimal policy is either
(
order if x ≤ s ,
do not order otherwise ,
or (
do not order if x ≤ s ,
order otherwise ,
for some s. By Lemma 2.1 the first case orders up to some S > s, and in the second case
one orders up to ∞. ¤
Lemma 2.4. If θ(x) is K-convex then so is
ψ(x) = min{KH(u) + θ(x + u)} .
u≥0

Lemma 2.4 says that Lφ is K-convex if φ is. In fact, Lφ ∈ CK if φ ∈ CK . Since Vt∗ = LN −t φ

it follows that Vt∗ is K-convex. Optimality of the (s, S)-policy follows from Lemma 2.3.
 LECTURE 3

Calculus of Variations

3.1. Necessary Conditions

Consider the problem
Z b
min J(y) := F (y(x), y 0 (x), x)dx (3.1)
a
subject to some conditions on y(x) at x = a and x = b. If y(x) is a local minimizer of (3.1)
then for any “nearby” function z(x)
J(z) ≥ J(y) . (3.2)
In particular, consider
z(x) := y(x) + δy(x) , (a test-function) , (3.3a)
with δy(x) = ²η(x) , (a variation) , (3.3b)
where ² is a parameter with small modulus. The test-function z satisfies the same boundary
conditions as y, imposing certain conditions on η, see e.g. (3.11). The variations of the
derivatives of y are
δy 0 (x) = ²η 0 (x) , δy 00 (x) = ²η 00 (x) , . . . (3.4)
provided η is differentiable.
Notation and terminology: Let Y denote the class of functions:[a, b] → R admissible for the above
problem. The variation δy is strong if δy = ²η → 0 as ² → 0, and weak if also δy 0 = ²η 0 → 0.
We use the following norms,
kyk0 := sup{|y(x)| : a ≤ x ≤ b} , (3.5a)
kyk1 := sup{|y(x)| + |y 0 (x)| : a ≤ x ≤ b} , (3.5b)
and associated neighborhoods of y ∈ Y,
U0 (y, δ) := {z : kz − yk0 ≤ δ} , (a strong neighborhood) , (3.6a)
U1 (y, δ) := {z : kz − yk1 ≤ δ} , (a weak neighborhood) . (3.6b)
Note that U1 (y, δ) ⊂ U0 (y, δ). The value J(y) is:
T
a strong local minimum if J(z) ≥ J(y) for all z ∈ Y U0 (y, δ),
T
a weak local minimum if J(z) ≥ J(y) for all z ∈ Y U1 (y, δ),
for some δ > 0.
Let J(y) be a weak local minimum. Then for all y + ²η in a weak neighborhood of y,
J(y + ²η) ≥ J(y) (3.7a)
Z b Z b
0 0
i.e. F (y + ²η, y + ²η , x)dx ≥ F (y, y 0 , x)dx (3.7b)
a a
15
16 3. CALCULUS OF VARIATIONS

Expanding the LHS of (3.7b) in ² we get

·Z b ¸
J(y) + ² (Fy η + Fy0 η ) dx + O(²2 ) ≥ J(y)
0
(3.8)
a
and as ² → 0 (since ²η is in a weak neighborhood of 0, ²η → 0 and ²η 0 → 0),
Z b
(Fy η + Fy0 η 0 ) dx = 0 . (3.9)
a
Integrating the second term by parts we get
Z b· ¸
d
η(x)Fy − η(x) (Fy0 ) dx + [η(x)Fy0 ]ba = 0 . (3.10)
a dx
The requirement that the test-function z of (3.3) satisfies the same boundary conditions as
y implies that
η(a) = η(b) = 0 , (3.11)
killing the last term in (3.10), leaving
Z b · ¸
d
η(x) Fy − (Fy0 ) dx = 0 (3.12)
a dx
which holds for all admissible η(x). It follows from Lemma 3.1 below that
d
Fy − Fy0 ≡ 0 , a ≤ x ≤ b , (3.13)
dx
the differential form of the Euler–Lagrange equation. This is a 2nd order differential
equation in y:
∂ 2F ∂ 2 F 0 ∂ 2 F 00
Fy − − y − 02 y = 0 . (3.14)
∂x∂y 0 ∂y∂y 0 ∂y
Lemma 3.1. If Φ is continuous in [a, b], and
Z b
Φ(x) η(x) dx = 0 ,
a
for every continuous function η(x) that vanishes at a and b, then Φ(x) ≡ 0 in [a, b].
Proof. Suppose Φ(ξ) > 0 for some ξ ∈ (a, b). Then Φ(x) > 0 in some neighborhood of
ξ, say α ≤ x ≤ β. Let
(
(x − α)2 (x − β)2 for α ≤ x ≤ β ,
η(x) :=
0 otherwise .
Rb
Then a Φ(x) η(x) dx > 0. Therefore Φ cannot be positive in [a, b]. It similarly cannot be
negative. ¤
Exercises.
Exercise 3.1. Derive (3.9) by the condition that ∂J/∂² must be zero at ² = 0.
d
Exercise 3.2. Does (3.13) follow from (3.12) by using η(x) := Fy − F0
dx y
in (3.12)?
 3.3. CALCULUS OF VARIATIONS AND DP 17

3.2. The second variation and sufficient conditions

Maximizing the value of the functional (3.1), instead of minimizing it as in § 3.1, would
give exactly the same necessary condition, the Euler–Lagrange equation (3.13). To distin-
guish between minima and maxima we need a second variation,
Let Z b
φ(²) := F (y + ²η, y 0 + ²η 0 , x)dx (3.15)
a
where y is an extremal solution, and η is fixed. Then
Z b
0
φ (0) = (Fy η(x) + Fy0 η 0 (x)) dx , compare with (3.9) , (3.16a)
a
Z b³ ´
2
φ00 (0) = Fyy η 2 (x) + 2Fyy0 η(x)η 0 (x) + Fy0 y0 η 0 (x) dx (3.16b)
a
Z b µ ¶µ ¶
0 Fyy Fyy0 η(x)
= (η(x), η (x)) dx (3.16c)
a
Fyy0 Fy0 y0 η 0 (x)
We call (3.16a) and (3.16b) the first and second variations of J(y), respectively.
Assuming the matrix in (3.16c) is positive definite, along the extremal y(x), for all
a ≤ x ≤ b, a sufficient condition for minimum is the Legendre condition
Fy 0 y 0 > 0 , (3.17)
the reverse inequality sufficient for maximum.
The Legendre condition (3.17) means that F is convex as a function of y 0 along the
extremal trajectory. Another aspect of convexity appears in the Weierstrass condition
F (y, Y 0 , x) − F (y, y 0 , x) + (Y 0 − y 0 )Fy0 ≥ 0 , (3.18)
for all admissible functions Y 0 = Y 0 (x, y). We recognize that (3.18) is the gradient inequality
for F , i.e. there is an implicit assumption that F is convex as a function of y 0 .
Both sufficient conditions (3.17) and (3.18) are strong, and difficult to check even if they
hold.

3.3. Calculus of Variations and DP

Consider the problem of minimizing the functional
Z b
J(y) := F (x, y, y 0 )dx (3.19)
a

subject to the initial condition y(a) = c. Introduce the function

f (a, c) := min J(y) (3.20)
over all feasible y. Here (a, c) are parameters with −∞ < a < b and −∞ < c < ∞. Breaking
the integral
Z b Z a+∆ Z b
= + (3.21)
a a a+∆
18 3. CALCULUS OF VARIATIONS

the Principle of Optimality gives

·Z a+∆ ¸
0
f (a, c) = min F (x, y, y )dx + f (a + ∆, c(y) (3.22)
y a
the minimization is over all y defined on [a, a + ∆] such that y(a) = c, and c(y) = y(a + ∆).
If ∆ is small, the function y(x) can be approximated in [a, a+∆] by y(x) ≈ y(a)+y 0 (a)x =
c + y 0 (a)x. The choice of y in [a, a + ∆] is therefore equivalent to the choice of
v := y 0 (a) . (3.23)
Therefore
Z a+∆
F (x, y, y 0 )dx = F (a, c, v)∆ + o(∆) , c(y) = c + v∆ + o(∆) . (3.24)
a
The optimality condition (3.22) then becomes
f (a, c) = min [F (a, , c, v)∆ + f (a + ∆, c + v∆] + o(∆) , (3.25)
v
with limit, as ∆ → 0, · ¸
∂f ∂f
− = min F (a, c, v) + v , (3.26)
∂a v ∂c
holding for all a < b. The initial condition here is f (b, c) = 0 for all c.
To derive the Euler–Lagrange equation, we rewrite (3.25), denoting the endpoint a by x
and the initial velocity v by y 0 ,
· ¸
0 ∂f ∂f 0
f (x, y) = min F (x, y, y )∆ + f (x, y) + ∆+ y ∆ + ··· (3.27)
y0 ∂x ∂y
with limit, as ∆ → 0, · ¸
0 ∂f 0 ∂f
0 = min F (x, y, y ) + +y (3.28)
y0 ∂x ∂y
a rewrite of (3.26). This equation is equivalent to the two equations
∂f
Fy0 + =0 (3.29)
∂y
obtained by partial differentiation w.r.t. y 0 , and
∂f ∂f
F+ + y0 =0 (3.30)
∂x ∂y
which holds form all x, y, y 0 satisfying (3.29). Differentiating (3.29) w.r.t. x, and (3.30) w.r.t.
y we get two equations
d ∂ 2f ∂ 2f
Fy0 + + 2 y0 = 0 , (3.31)
dx ∂x∂y ∂y
∂y 0 ∂ 2f ∂2f ∂f ∂y 0
Fy + Fy0 + + 2 y0 + = 0, (3.32)
∂y ∂x∂y ∂y ∂y ∂y
which, together with (3.29) give the Euler-Lagrange equation
d
(3.13) F0
dx y
− Fy = 0 .
 3.5. EXTENSIONS 19

3.4. Special cases

The Euler–Lagrance equation is simplified in the following cases:
(a) The function F (x, y, y 0 ) is independent of y 0 . Then (3.13) reduces to
∂F (x, y)
= 0
∂y
and y is implicitly defined (with no guarantee that the boundary conditions are satisfied;
these conditions cannot be arbitrary).
(b) F (x, y, y 0 ) is independent of y. Then (3.13) gives
d ∂F (x, y 0 )
Fy = 0 , or
0 = C (constant)
dx ∂y 0
which may be solved for y 0 ,
Z
0
y (x) = f (x, C) , ∴ y(x) = f (x, C) dx
∂F dy
(c) The autonomous case where ∂x
= 0. Multiply (3.13) by y 0 = dx
d
y0 Fy 0 − y 0 Fy = 0 .
µ ¶ dx
d 0 ∂F ∂F ∂F
y 0 − y 00 0 − y 0 = 0.
dx ∂y ∂y ∂y
µ ¶
d 0 ∂F 0
y − F (y, y ) = 0.
dx ∂y 0
∂F
∴ y 0 0 − F (y, y 0 ) is constant (3.33)
∂y
3.5. Extensions
The Euler–Lagrange equation is extended in three ways:
(a) Higher order derivatives. It is required a stationary value of
Z b µ ¶
dy dn y
J(y) := F y, , . . . , n , x dx (3.34)
a dx dx
Using (3.3), the first variation of J is
Z bµ ¶
dη dn η
δJ = η Fy + Fy0 + · · · + n Fy(n) dx
a dx dx
Successive integration by parts (assuming that y, y 0 , · · · , y (n) are given at x = a and x = b)
gives Z bµ ¶
d d2 n d
n
Fy − Fy0 + 2 Fy00 − · · · + (−1) F (n) η(x) dx = 0
a dx dx dxn y
and by Lemma 3.1,
d d2 dn
Fy − Fy0 + 2 Fy00 − · · · + (−1)n n Fy(n) ≡ 0 , a ≤ x ≤ b . (3.35)
dx dx dx
20 3. CALCULUS OF VARIATIONS

(b) Several dependent variables. Here

Z b
J(y) := F (y1 , y2 , · · · , yn , y10 , y20 , · · · , yn0 , x) dx
a

and a similar analysis gives the necessary conditions

d
Fy0 ≡ 0 , i ∈ 1, n
Fyi − (3.36)
dx i
(c) Several independent variables. Let y = y(x1 , · · · , xn ) and
Z Z Z µ ¶
∂y ∂y ∂y
J(y) = · · · F y, , ,··· , , x1 , x2 , · · · , xn dx1 dx2 · · · dxn
D ∂x1 ∂x2 ∂xn
The first variation is
Z Z Z Ã n
!
∂F X ∂η
² ··· η + Fk dx1 dx2 · · · dxn
D ∂y k=1
∂xk
∂y
where Fk is the derivative of F w.r.t. . The analog of integration by parts is Green’s
∂xk
Theorem. Let ∂D denote the boundary of D. Then
Z Z Z ÃXn
! Z Z Z X n
∂η
··· Fk dx1 dx2 · · · dxn = ··· (ηFk ) nk · ds −
D k=1
∂x k ∂D k=1
Z Z Z Xn µ ¶
∂Fk
− ··· η dx1 dx2 · · · dxn
D k=1
∂x k

where nk is unit normal to the surface ∂D. If y is given on ∂D then η = 0 there, and the
Euler–Lagrange equation becomes
Xn
∂Fk
Fy − = 0. (3.37)
k=1
∂x k

3.6. Hidden variational principles

Given a differential equation, is it the Euler–Lagrange equation of a variational problem?
We illustrate this for the 2nd order linear ODE
µ ¶
d dy
a(x) + b(x)y + λc(x)y = 0 (3.38a)
dx dx
y(x0 ) = y(x1 ) = 0 (3.38b)
The values of λ for which a solution exists are the eigenvalues of (3.38). Multiply (3.38a)
by the variation δy(x) and integrate
Z x1 µ µ ¶ ¶
d dy
dx δy a(x) + b(x)yδy + λc(x)yδy = 0
x0 dx dx
 3.6. HIDDEN VARIATIONAL PRINCIPLES 21

Using y δy = 12 δ(y 2 ) and integrating the first term by parts we get

Z x1 µ ¶
dy 0 b(x) 2 λc(x) 2
dx −a(x) (δy) + δy + δy = 0
x0 dx 2 2
Using the same notation for the variation of (y 0 )2 ,
dy 1 2
(δy)0 = y 0 (δy)0 = y 0 δy 0 = δy 0
dx 2
and therefore
Z x1 ³ ´
2
δ −a(x)y 0 (x) + [b(x) + λc(x)] y 2 (x) dx = 0
x0
Z x1 Z x1 ³ ´
2 02 2
∴ λδ c(x)y (x) dx − δ a(x)y (x) − b(x)y (x) dx = 0
x0 x0
or δ(J1 − λJ2 ) = 0
Z x1 ³ ´
2
where J1 (y) = a(x)y 0 (x) − b(x)y 2 (x) dx
Zx0x1
J2 (y) = c(x)y 2 (x) dx
x0

Therefore any solution of (3.38) is an extremal of the variational problem of finding the
stationary values of
Rx1 ¡ ¢
a(x)y 0 2 (x) − b(x)y 2 (x) dx
J1 x0
λ= = Rx1 (3.39)
J2 2
c(x)y (x) dx
x0

Example 3.1. The eigenvalues of

y 00 + λy = 0 (3.40a)
y(0) = y(1) = 0 (3.40b)
are π 2 , 4π 2 , 9π 2 , · · · . Using (3.39) these eigenvalues are the stationary values of the ratio
R1
y 0 2 dx
0
λ= (3.41)
R1
y2 dx
0

Exercises.
Exercise 3.3. Reverse the above argument, starting from the variational problem (3.39)
and deriving the Euler–Lagrange equation.
22 3. CALCULUS OF VARIATIONS

Exercise 3.4. Use a trial solution of (3.40)

y0 (x) = x(1 − x)(1 + αx)
where α is a changing parameter. Compute RHS(3.41), and minimize w.r.t α to obtain an
extimate of the smallest eigenvalue λ = π 2 .

3.7. Integral constraints

Consider the problem of minimizing (3.19) subject to the additional constraint
Z b
G(x, y, y 0 )dx = z , (3.42)
a

where z is given. We denote

Z b
f (x, y, z) := min F (x, y, y 0 )dx (3.43)
y a

subject to (3.42). Such a problem is called isoperimetric, for historic reasons (see Exam-
ple 3.2). The analog of (3.25) is
f (x, y, z) = min
0
[F (x, y, y 0 )∆ + f (x + ∆, y + y 0 ∆, z − G(x, y, y 0 )∆] (3.44)
y

with limit · ¸
0 ∂f 0 ∂f 0 ∂f
0 = min F (x, y, y ) + +y − G(x, y, y ) . (3.45)
y0 ∂x ∂y ∂z
Therefore,
∂f ∂f
0 = Fy0 + − Gy , (3.46)
∂y ∂z
and
∂f ∂f ∂f
0=F+ + y0 −G , (3.47)
∂x ∂y ∂z
holding for all x, y, z, y 0 satisfying (3.46). Differentiating (3.46) w.r.t. x and (3.47) w.r.t. y,
and combining, we get
µ ¶ µ ¶
∂ ∂f ∂ ∂f
F− G − F− G =0 (3.48)
∂y 0 ∂z ∂y ∂z
Partial differentiation of (3.47) w.r.t. z yields
∂ 2f ∂ 2f ∂ 2f
+ y0 −G 2 = 0. (3.49)
∂x∂z ∂y∂z ∂z
µ ¶
d ∂f
∴ = 0, (3.50)
dx ∂z
(3.51)
i.e. ∂f
∂z
is a constant, say λ, and (3.48) is the Euler-Lagrange equation for F − λG. The
parameter λ is the Lagrange multiplier of the constraint (3.42).
 3.7. INTEGRAL CONSTRAINTS 23

Example 3.2. ([10, p. 22]) The Dido problem is to find the smooth plane curve of
perimeter L which encloses the maximum area. We show that this curve is a circle. Fix any
point P of the curve C, and use polar coordinates with origin R π 1 at2 P , and θ = 0 along the
(half) tangent of C at P . The area enclosed by C is A = 0 2 r dθ, and the perimeter is
R π ³¡ dr ¢2 2
´1/2
L= 0 dθ
+ r dθ. Define F := A + λL, and maximize
 õ ¶ !1/2 
Z π 2 Z π
dr £1 2 ¤
F =  1 2
r +λ +r 2  dθ = r + λT dθ
2 2
0 dθ 0

subject to r(0) = r(π) = 0. The Euler-Lagrange equation gives, using (3.33),

dr λ dr/dθ
− λT − 21 r2 = constant
dθ T
λ(dr/dθ)2 − λT 2 1 2
∴ −2r = constant
T
λr2 1 2
∴ +2r = constant = 0 , since r = 0 is on C
T
õ ¶ !1/2 µ ¶2
2
dr dr
∴ 2λ + + r2 = 0. ∴ + r2 = 4 λ2
dθ dθ
dθ 1 ³ r ´
∴ = . ∴ θ = sin−1 −
dr (4 λ − r2 )1/2
2 2λ
∴ r = −2 λ sin θ
a circle of radius −λ. The radius is determined from
Z π Ã µ ¶2 !1/2 Z π
dr 2
L= +r dθ = − 2λdθ = −2 π λ ,
0 dθ 0

therefore the radius is L/2 π.

Example 3.3. ([10, p. 23])
Z 1 Z 1 Z 1
minimize ẋ2 dt s.t. x dt = 0 , xt dt = 1 and x(0) = x(1) = 0 .
0 0 0
Let F (x, ẋ, t) := ẋ2 + λx + µxt. The Euler–Lagrange equation is 2ẍ − λ − µt = 0
∴ x = 1
4
λt2 + 16 µt3 + At , since x(0) = 0
1
∴ + 16 µ + A = 0 , since x(1) = 0
4
λ
Z 1
1 1 1
∴ 12 λ + 24 µ + 2 A = 0 , since x dt = 0
0
Z 1
1 1 1
∴ 16 λ + 30 µ + 3 A = 1 , since xt dt = 1
0
three linear equations with solution λ = −µ = 360 , A = 30. Finally x = 90t2 − 60t3 + 30.
24 3. CALCULUS OF VARIATIONS

Exercises.
Exercise 3.5. A flexible fence of length L is used to enclose an area bounded on one
side by a straight wall. Find the maximum area that can be enclosed.
Exercise 3.6. (The catenary equation) A flexible uniform cable of length 2a hangs
between the fixed points (0, 0) and (2b, 0), where b < a. Find the curve y = y(x) minimizing
Z 2b
(1 + (y 0 )2 )1/2 y dx (potential energy)
0
Z 2b
s.t. (1 + (y 0 )2 )1/2 dx = 2a (given length)
0

Answer: y/k = cosh(b/k) − cosh{(x − b)/k} , where k is from a = k sinh b/k.

Exercise 3.7.
Z Z µ ¶
∂f 2 ∂f 2
Minimize + dxdy , where A := {(x, y) : x2 + y 2 ≤ 1}
A ∂x ∂y
RR
s.t. A
f dxdy = B with f given on ∂A. Show that f satisfies the partial differential
equation
∂ 2f ∂ 2f
+ + kf = 0
∂x2 ∂y 2
where k is a constant to be determined.

3.8. Natural boundary conditions

If y(a) is not specified then
∂f
|x=a = 0 (3.52)
∂y
for otherwise there is a better starting point. Then (3.29) gives
Fy0 |x=a = 0 , (3.53)
a natural boundary condition at the free endpoint x = a.
Another case is where even x = a is unspecified, and the y curve is only required to start
somewhere on a given curve y = g(x). Then the change in f as the initial point varies along
y = g(x) must be zero,
∂f ∂f
+ g0 =0 (3.54)
∂x ∂y
and by (3.30),
∂f ∂f
F + y0 − g0 =0 (3.55)
∂y ∂y
and finally, by (3.29),
F + (g 0 − y 0 )Fy0 = 0 , (3.56)
the transversality condition at the free initial point.
 3.9. CORNERS 25

x
3

1 2 3 4 t

Figure 3.1. At the corner, t = 2, the solution switches from x0 = 1 to x = 2

3.9. Corners
Consider the Calculus of Variations problem
Z T
opt F (t, x, x0 ) dt
o
s.t. x(0), x(T ) given
x piecewise smooth
An optimal x∗ satisfies the Euler-Lagrange equation
d
Fx = Fx0
dt
∗
in any subinterval of [0, T ] where x is continuously differentiable. The discontinuity points
of dtd x∗ are called its corners.
The Weierstrass-Erdmann corner conditions: The functions Fx0 , F − x0 Fx0 are contin-
uous in corners.
Example 3.4. Consider
Z 3
min (x − 2)2 (x0 − 1)2 dt
0
s.t. x(0) = 0, x(3) = 2
The lower bound, 0, is attained if
x = 2 or x0 = 1 , 0 ≤ t ≤ 3 .
This function has a corner at t = 2, but
Fx0 = 2(x − 2)2 (x0 − 1) and F − x0 Fx0 = (x − 2)2 (x0 − 1)(x0 + 1)
are continuous at t = 2.
 LECTURE 4

The Variational Principles of Mechanics

This lecture is based on [23] and [24]. Other useful refernces are [24], [20] and [32].

4.1. Mechanical Systems

A particle is a point mass, i.e. mass concentrated in a point. The position r, velocity
v and acceleration a of a particle in R3 are
   
x ẋ
r = r(t) = y  , v = ṙ = ẏ  , a = v̇ = r̈ .
z ż
The norm of the velocity vector v is called speed, and denoted by v = kvk.
Consider a system of N particles in R3 . Describing it requires 3N coordinates, less if
the N particles are not independent. The number of degrees of freedom of the system is the
minimal number of coordinates describing it. If the system has s degrees of freedom, let
q := (q1 , q2 , . . . , qs ) denote its generalized coordinates,
q̇ := (q̇1 , q̇2 , . . . , q̇s ) its generalized velocities.
The position & motion of the system are determined by the 2s numbers (q, q̇).
Example 4.1. A double pendulum in planar motion, see Fig. 4.1, has two degrees of
freedom. It is described in terms of the angles φ1 and φ2 }. The generalized coordinates and
velocities can be taken as {φ1 , φ2 }, and {φ̇1 , φ̇2 } respectively.

4.2. Hamilton’s Principle of Least Action

A mechanical system is characterized by a function
L = L(q, q̇, t) = L(q1 , . . . , qs , q̇1 , . . . , q̇s , t) (4.1)

x
φ
1 l1

m
1
l2
φ
2

m
2

Figure 4.1. A double pendulum.

27
28 4. THE VARIATIONAL PRINCIPLES OF MECHANICS

Hamilton’s Principle, [23, pp. 111-114]: The motion of a mechanical system between any
two points (t1 , q(t1 )) and (t2 , q(t2 )) occurs in such a way that the definite integral
Z t2
S := L(q, q̇, t) dt (4.2)
t1

becomes stationary for arbitrary feasible variations. R

L is called the Lagrangian of the system, and S = L dt its action.

4.3. The Euler-Lagrange Equation

Let s = 1, q = q(t) ∈ argmin S and consider a perturbed path q(t) + δq(t) where
δq(t1 ) = δq(t2 ) = 0 .
Z t2
δS := L(q + δq, q̇ + δ q̇, t) dt −tt21 L(q, q̇, t) dt
t1
Z t2 µ ¶
∂L ∂L
= δq + δ q̇ dt
t1 ∂q ∂ q̇
· ¸ t2 Z t2 µ ¶
∂L ∂L d ∂L d
= δq + − δq dt , since δ q̇ = δq
∂ q̇ t1 t1 ∂q dt ∂ q̇ dt
Z t2 µ ¶
∂L d ∂L
= − δq dt , since δq(t1 ) = δq(t2 ) = 0 .
t1 ∂q dt ∂ q̇
Since δq is arbitrary, we conclude that δS = 0 iff
∂L d ∂L
− =0, (4.3)
∂q dt ∂ q̇
the Euler-Lagrange equation (3.13), a necessary condition for minimal action.
The Euler–Lagrange equations for a system with s degrees of freedom are
∂L d ∂L
− = 0 , i ∈ 1, s . (4.4)
∂qi dt ∂ q̇i

4.4. Nonuniqueness of the Lagrangian

Consider
b q̇, t) := L(q, q̇, t) + d f (q, t) ,
L(q, (4.5)
dt
the last term does not depend on q̇. Then
Z t2
b
S = b q̇, t) dt = S + f (q(t2 ), t2 ) − f (q(t1 ), t1 ) .
L(q,
t1

∴ δ Sb = 0 ⇐⇒ δS = 0 .
d
The Lagrangian L(q, q̇, t) is therefore defined up to derivative dt
f (q, t).
 4.7. THE LAGRANGIAN OF A SYSTEM 29

4.5. The Galileo Relativity Principle

An inertial frame (of reference) is one in which time is homogeneous and space is
both homogeneous and isotropic. In such a system, a free particle at rest remains at rest.
Consider a particle moving freely in an inertial frame. Its Lagrangian must be indepen-
dent of t, r and the direction of v.
∴ L = L(kvk) = L(v 2 ) .
µ ¶
d ∂L
∴ = 0 , by the Euler–Lagrange equation .
dt ∂v
∂L
∴ = 0 ∴ v = constant .
∂v
Law of Inertia. In an inertial frame a free motion has a constant velocity.
Consider two frames, K and K,b where K is inertial, and K b moves relative to K with
velocity v. The coordinates r and b
r of a given point are related by
r=b
r+vt
(time is absolute, i.e. t = b
t).
b
Galileo’s Principle. The laws of mechanics are the same in K and K.

4.6. The Lagrangian of a Free Particle

Let an inertial frame K move with an infitesimal velocity ε relative to another inertial
b A free particle has velocities {v, v
frame K. b in these two frames.
b} and Lagrangians {L, L}
b = v+ε.
∴ v
¡ ¢
b
∴ L = L(b v 2 ) = L v 2 + 2v · ε + ε2
∂L
≈ L(v 2 ) + 2v · ε
∂(v 2 )
d
= L + f (r, t) , see (4.5) .
dt
∂L
∴ 2v · ε is linear in v .
∂(v 2 )
∂L
∴ is independent of v .
∂(v 2 )
∴ L = 12 m v 2 , (4.6)
where m is the mass of the particle.

4.7. The Lagrangian of a System

Consider a system with several particles. If they do not interact, the Lagrangian of the
system is the sum of (4.6) terms,
X
L= 1
2
mk vk2 . (4.7)
30 4. THE VARIATIONAL PRINCIPLES OF MECHANICS

If the particles interact with each other, but not with anything outside the system (in which
case the system is called closed), the Lagrangian is
X
L= 1
2
mk vk2 − U (r1 , r2 , . . .) (4.8)
where rk P is the position of the kth particle. Here
1
T := 2
mk vk2 the kinetic energy,
U := U (r1 , r2 , . . .) the potential energy.
If t is replaced by −t the Lagrangian does not change (time is isotropic).

4.8. Newton’s Equation

The equations of motion
µ ¶
d ∂L ∂L
= , k = 1, 2, . . .
dt ∂ q̇k ∂qk
can be written, using the generalized momentums,
∂L
pk := (4.9)
∂ q̇k
dpk ∂U
as = − , k = 1, 2, · · ·
dt ∂qk
In particular, the Lagrangian (4.8) gives
d ∂U
(mk vk ) = − , k = 1, 2, · · · (4.10)
dt ∂rk
If mk is constant, (4.10) becomes
dvk ∂U
mk =− (4.11)
dt ∂rk

4.9. Interacting Systems

Let A be a system interacting with another system B (A moves in a given external field
due to B), and let A + B be closed.
Using generalized coordinates qA , qB ,
LA+B = TA (qA , q̇A ) + TB (qB , q̇B ) − U (qA , qB ) .
∴ LA = TA (qA , q̇A ) − U (qA , qB (t)) .
The potential energy may depend on t.
Example 4.2. A single particle in an external field,
m v 2 − U (r, t) .
L = 1
2
∂U
∴ mv̇ = − .
∂r
A uniform field is where U = −F · r where F is a constant force.
 4.10. CONSERVATION OF ENERGY 31

Example 4.3. Consider the double pendulum of Fig. 4.1. For the first particle,
2
T1 = 12 m1 l12 φ˙1
U1 = −m1 g l1 cos φ1

For the second particle,

x2 = l1 sin φ1 + l2 sin φ2
y2 = l1 cos φ1 + l2 cos φ2
∴ T2 = 12 m2 (ẋ22 + ẏ22 )
³ ´
1 2 ˙ 2 2 ˙ 2 ˙ ˙
= 2 m2 l1 φ1 + l2 φ2 + 2l1 l2 cos(φ1 − φ2 )φ1 φ2
2 2
∴ L = 1
2
(m1 + m2 )l12 φ˙1 + 12 m2 l22 φ˙2 + m2 l1 l2 cos(φ1 − φ2 )φ˙1 φ˙2 +
+(m1 + m2 )gl1 cos φ1 + m2 gl2 cos φ2 .

4.10. Conservation of Energy

The homogeneity of time means that the Lagrangian of a closed system does not depend
explicitly on time,
dL X ∂L X ∂L
= q̇i + q̈i
dt i
∂q i i
∂ q̇ i

∂L ∂L d ∂L
with no term ∂t
. Using the Euler-Lagrange equations ∂qi
= dt ∂ q̇i
we write

dL X d ∂L X ∂L X d µ ∂L ¶
= q̇i + q̈i = q̇i
dt i
dt ∂ q̇i i
∂ q̇i i
dt ∂ q̇i
à !
d X ∂L
∴ q̇i −L = 0.
dt i
∂ q̇i

Therefore
X ∂L
E := q̇i −L (4.12)
i
∂ q̇i

remains constant during the motion of a closed system, see also (3.33).
If T (q, q̇) is a quadratic function of q̇, and L = T (q, q̇) − U (q), then (4.12) becomes
X ∂L
E = q̇i − L = 2T − L
i
∂ q̇i
= T +U (4.13)

the total energy.

32 4. THE VARIATIONAL PRINCIPLES OF MECHANICS

4.11. Conservation of Momentum

The homogeneity of space implies that the Lagrangian is unchanged under a translation
of the space,
b
r = r+ε
X ∂L X ∂L
∴ δL = · δri = ε · =0
i
∂ri i
∂r i

Since ε is arbitrary,
X ∂L
δL = 0 ⇐⇒ =0
i
∂ri
X d ∂L
∴ = 0 , by Euler–Lagrange .
i
dt ∂vi
d X ∂L
∴ = 0.
dt i
∂vi
Therefore
X ∂L
P := (4.14)
i
∂vi
remains constant during motion. For the Lagrangian (4.8),
X
P= mi vi (4.15)
i
the momentum of the system. P ∂L P ∂U
Newton’s 3rd Law. From i ∂r i
= 0 we conclude i Fi = 0 where Fi = − ∂ri , i.e. the
sum of forces on all particles in a closed system is zero.
4.12. Conservation of Angular Momentum
The isotropy of space implies that the Lagrangian is invariant under an infinitesimal
rotation δφ, see Fig. 4.2
δr = δφ × r ,
or kδrk = r sin θkδφk. Similarly,
δv = δφ × v .
Substituting these in
X µ ∂L ∂L
¶
δL = · δri + · δvi = 0 ,
i
∂r i ∂v i
∂L ∂L
and using pi = ∂vi
, ṗi = ∂ri
we get
X
(ṗi · δφ × ri + pi · δφ × vi ) = 0
i
or
X d X
δφ · (ri × ṗi + vi × pi ) = δφ · ri × pi = 0 .
i
dt i
 4.13. MECHANICAL SIMILARITY 33

δr
δφ

Figure 4.2. An infitesimal rotation.

Since δφ is arbitrary, we conclude

d X
ri × pi = 0
dt i
or, the angular momentum X
M= ri × pi (4.16)
i
is conserved in the motion of a closed system.

4.13. Mechanical Similarity

Let the potential energy be a homogeneous function of degree k, i.e.
U (αr1 , αr2 , . . . , αrn ) = αk U (r1 , r2 , . . . , rn ) (4.17)
for all α. Consider a change of units,
b
ri = αri ,
b
t = βt .
α
bi =
∴ v vi .
β
α2
∴ Tb = T .
β2
If
α2
= αk , i.e. if β = α1−k/2 ,
β2
then
b = Tb − U
L b = αk (T − U ) = αk L
and the equations of motion are unchanged. If a path of length l is travelled in time t, then
the corresponding path of length b
l is travelled in time b
t where
à !1−k/2
b
t b
l
= . (4.18)
t l
34 4. THE VARIATIONAL PRINCIPLES OF MECHANICS

Example 4.4. [Coulomb Force] If k = −1 then (4.18) gives

à !3/2
b
t b
l
= (4.19)
t l
that is Kepler’s 3rd Law.
Example 4.5. For small oscilations and k = 2,
b
t
=1
t
independet of amplitude.

4.14. Hamilton’s equations

Writing
X ∂L X ∂L
dL = dqi + dq̇i
i
∂qi i
∂ q̇i
X X
= ṗi dqi + pi dq̇i
i i
à !
X X X
= ṗi dqi + d pi q̇i − q̇i dpi
i i i
³X ´ X
∴ d pi q̇i − L = −sumi ṗi dqi + q̇i dpi
i

Defining the Hamiltonian ( Legendre transform of L),

X
H = H(p, q, t) := pi q̇i − L (4.20)
i

we have X X
dH = − ṗi dqi + q̇i dpi .
i i
Comparing with
X ∂H X ∂H
dH = dqi + dpi
i
∂qi i
∂pi
we get the Hamilton equations
∂H
q̇i = (4.21)
∂pi
∂H
ṗi = − (4.22)
∂qi
The total time derivative of H is
dH ∂H X ∂H X ∂H
= + q̇i + ṗi ,
dt ∂t i
∂q i i
∂p i
 4.16. THE HAMILTON–JACOBY EQUATIONS 35

and by (4.21)–(4.22),
dH ∂H
= .
dt ∂t
4.15. The Action as Function of Coordinates
R t2
The action S = t1 Ldt has the Lagrangian as its time derivative, dS dt
= L.
· ¸t2 Z t2 µ ¶
∂L ∂L d ∂L
δS = δq + − δqdt
∂ q̇ t1 t1 ∂q dt ∂ q̇
X ∂S
= pi δqi ∴ = pi
i
∂q i

dS ∂S X ∂S ∂S X
∴ L = = + q̇i = + pi q̇i
dt ∂t i
∂qi ∂t i
∂S X
∴ = L− pi q̇i = −H
∂t i
X
∴ dS = pi dqi − Hdt
i
Maupertuis’ Principle. The motion of a mechaincal system minimizes
Z ÃX !
pi dqi dt
i
True, if energy is conserved.

4.16. The Hamilton–Jacoby equations

∂S
+H =0 (4.23)
∂t
or
∂S
+ H(q1 , . . . , qs , p1 , . . . , ps , t) = 0 (4.24)
∂t
∂H
Solution. If ∂t
= 0 then
µ ¶
∂S ∂S ∂S
+ H q1 , . . . , q s , ,..., =0
∂t ∂q1 ∂qs
with solution
S = −ht + V (q1 , . . . , qs , α1 , . . . , αs , h)
where h, α1 , . . . , αs are arbitrary constants.
µ ¶
∂V ∂L
∴ H q1 , . . . , qs , ,..., =h.
∂q1 ∂qs
Separation of variables. If
H = G(f1 (q1 , p1 ), f2 (q2 , p2 ), . . . , fs (qs , ps ))
36 4. THE VARIATIONAL PRINCIPLES OF MECHANICS

then µ µ ¶ µ ¶ µ ¶¶
∂V ∂V ∂V
G f1 q1 , , f2 q2 , , . . . , fs ps , =h
∂q1 ∂q2 ∂qs
Let
µ ¶
∂V
αi = fi qi , , i ∈ 1, s
∂qi
Xs Z
∴ V = fi (qi , αi )dqi
i=1
s Z
X
∴ S = −G(α1 , . . . , αs )t + fi (qi , αi )dqi
i=1
 LECTURE 5

Optimal Control, Unconstrained Control

5.1. The problem

The simplest control problem
Z T
max f (t, x(t), u(t)) dt (5.1)
0
0
s.t. x (t) = g(t, x(t), u(t)) (5.2)
x(0) fixed , (5.3)

where f, g ∈ C 1 [0, T ] , u ∈ CS [0, T ]. This is a generalization of Calculus of Variations.

Indeed,
Z T
max f (t, x(t), x0 (t)) dt , x(0) given ,
0

can be written as (5.1) with x0 = u in (5.2).

5.2. Fixed Endpoint Problems: Necessary Conditions

Let the endpoint x(T ) be fixed (variable endpoints are studied in the next section). For
any x, u satisfying (5.2),(5.3) and any λ(t) ∈ C 1 [0, T ],
Z T Z T
f (t, x, u) dt = [f (t, x, u) + λ(g(t, x, u) − x0 )] dt .
0 0

Integrating by parts,
Z T Z T
0
− λ(t)x (t) dt = −λ(T )x(T ) + λ(0)x(0) + x(t)λ0 (t) dt
0 0
Z T Z T
∴ f (t, x, u)dt = [f (t, x, u) + λg(t, x, u) + xλ0 ] dt − λ(T )x(T ) + λ(0)x(0) .
0 0

Let
• u∗ (t) be an optimal control,
• u∗ (t) + ²h(t) a comparison control, with parameter ² and h fixed,
• y(t, ²) the resulting state, y(t, 0) = x∗ (t) and y(0, ²) = x(0) , y(T, ²) = x(T ) for all
².
37
38 5. OPTIMAL CONTROL, UNCONSTRAINED CONTROL

Define
Z T
J(²) := f (t, y(t, ²), u∗ (t) + ²h(t)) dt
0
Z T
∴ J(²) = [f (t, y(t, ²), u∗ (t) + ²h(t)) + λ(t)g(t, y(t, ²), u∗ (t) + ²h(t)) + y(t, ²)λ0 (t)] dt
0
−λ(T )y(T, ²) + λ(0)y(0, ²)
Since u∗ is the maximizing control, J(²) has a local maximum at ² = 0. Therefore J 0 (0) = 0.
Z T
0
J (0) = [(fx + λgx + λ0 ) y² + (fu + λgu ) h] dt = 0 , (5.4)
0
∂
where fx := ∂x f (t, x∗ (t), u∗ (t)), etc.
(5.4) holds for all y, h iff along (x∗ , u∗ ),
λ0 (t) = − [fx (t, x(t), u(t)) + λ(t)gx (t, x(t), u(t))] , (5.5)
fu (t, x(t), u(t)) + λ(t) gu (t, x(t), u(t)) = 0 . (5.6)
Define the Hamiltonian
H(t, x(t), u(t), λ(t)) := f (t, x, u) + λ g(t, x, u) . (5.7)
Then
∂H
x0 = ⇐⇒ (5.3) ,
∂λ
∂H
λ0 = − ⇐⇒ (5.5) ,
∂x
∂H
= 0 ⇐⇒ (5.6) .
∂u

5.3. Terminal Conditions

Consider the problem (with x ∈ Rn , u ∈ Rm )
Z T
max f (t, x, u) dt + φ(T, x(T )) (5.8)
0

subject to
x0i (t) = gi (t, x, u) , i ∈ 1, n (5.9)
xi (0) = fixed , i ∈ 1, n (5.10)
xi (T ) = fixed , i ∈ 1, q (5.11)
xi (T ) = free , i ∈ q + 1, r (5.12)
xi (T ) ≥ 0 , i ∈ r + 1, s (5.13)
K(xs+1 , . . . , xn , t) ≥ 0 at t = T (5.14)
 5.3. TERMINAL CONDITIONS 39

where 1 ≤ q ≤ r ≤ s ≤ n and K is continuously differentiable. Write

Z T
J= [f + λ · (g − x0 )] dt + φ(T, x(T )) (5.15)
0

and integrate by parts to get

Z T
J= [f + λ · g + λ0 · x] dt + φ(T, x(T )) + λ(0) · x(0) − λ(T ) · x(T ) (5.16)
0

Let x∗ , u∗ be optimal, and let x, u be nearby feasible trajectory satisfying (5.9)–(5.14) on

0 ≤ t ≤ T + δT . Let J ∗ , f ∗ , g∗ , φ∗ denote values along an (t, x∗ , u∗ ), and J, f, g, φ values
along (t, x, u). Then
Z T
J −J =∗
[f − f ∗ + λ · (g − g∗ ) + λ0 · (x − x∗ )] dt + φ − φ∗ (5.17)
0
Z T +δT
∗
−λ(T ) · [x(T ) − x (T )] + f (t, x, u) dt
T

The linear part is

Z T
δJ = [(fx + λgx + λ0 ) · h + (fu + λgu ) · δu] dt (5.18)
0
+φx · δx1 + φT δT − λ(T ) · h(T ) + f (T )δT
where
h(t) = x(t) − x∗ (t)
δu(t) = u(t) − u∗ (t) (5.19)
δx1 = x(T + δT ) − x∗ (T )
Approximating h(T ) by
h(T ) = x(T ) − x∗ (T ) ≈ δx1 − x∗ 0 (T )δT = δx1 − g∗ (T )δT (5.20)
and substituting in (5.18) gives
Z T
δJ = [(fx + λgx + λ0 ) · h + (fu + λgu ) · δu] dt (5.21)
0
+ [φx − λ(T )] · δx1 + (f + λ · g + φt ) |t=T δT ≤ 0
If the multipliers λ satisfy
λ0 = − (fx + λgx ) (5.22)
along (x∗ , u∗ ) then δJ ≤ 0 for all feasible h, δu only if (5.6) holds and at t = T
[φx − λ(T )] · δx1 + (f + λ · g + φt ) δT ≤ 0 (5.23)
40 5. OPTIMAL CONTROL, UNCONSTRAINED CONTROL

for all feasible δx , δT . This implies

∂φ
λi (T ) = , i ∈ q + 1, r (5.24)
∂xi
∂φ
λi (T ) ≥ , i ∈ r + 1, s (5.25)
∂xi
· ¸
∂φ
xi (T ) λi (T ) − = 0 , i ∈ r + 1, s (5.26)
∂xi
∂φ ∂K
λi (T ) = +p , i ∈ s + 1, n (5.27)
∂xi ∂xi
∂K
f + λ · g + φt + p = 0 at T (5.28)
∂t
p ≥ 0 , pK = 0 at T . (5.29)
The last two facts follow from the Farkas Lemma.
If the endpoint x(T ) is free, then the terminal conditions reduce to
λ(T ) = 0 . (5.30)
Theorem 5.1. If u∗ is optimal then u∗ (t) maximizes H(t, x∗ (t), u(t), λ(t)) , 0 ≤ t ≤ T ,
where λ satisfies (5.5) and the appropriate terminal condition.
Example 5.1 (Calculus of Variations). Consider the problem
½Z T ¾
0
max f (t, x(t), u(t)) dt : x (t) = u(t) , x(0) given , x(T ) free . (5.31)
0
and the Hamiltonian H(t, x, , , u, λ) = f (t, x, u) + λu .
∂H
λ0 = − = −fx (5.32)
∂x
λ(T ) = 0 (5.33)
∂H
= fu + λ = 0 . (5.34)
∂u
From (5.32) and (5.34) follows the Euler-Lagrange equation
d
fx = fu ,
dt
and from (5.33) and (5.34) the transversality condition
fu (T ) = 0 . (5.35)
 LECTURE 6

Optimal Control, Constrained Control

6.1. The problem

The simplest such problem is
Z T
max f (t, x(t), u(t)) dt (6.1)
0
0
s.t. x (t) = g(t, x(t), u(t)) , x(0) given , (6.2)
a≤u≤b (6.3)

6.2. Necessary Conditions

Let J be the value of (6.1), J ∗ the optimal value, and let δJ be the linear part of J − J ∗ .
Z T
∴ δJ = [(fx + λgx + λ0 ) + (fu + λgu ) δu] dt − λ(T )δx(T ) . (6.4)
o

If
λ0 = − (fx + λgx ) , λ(T ) = 0 , (6.5)
then
Z T
δJ = (fu + λgu ) δu dt
0

and the optimality condition is

δJ ≤ 0 for all feasible δu
meaning
δu ≥ 0 if u = a ,
δu ≤ 0 if u = b ,
δu unrestricted if a < u < b .
Therefore, for all t,
u(t) = a =⇒ fu + λgu ≤ 0 ,
a < u(t) < b =⇒ fu + λgu = 0 , (6.6)
u(t) = b =⇒ fu + λgu ≥ 0 .
41
42 6. OPTIMAL CONTROL, CONSTRAINED CONTROL

If (x∗ , u∗ ) is optimal for (6.1)–(6.2) then there is a function λ such that (x∗ , u∗ , λ) satisfying
(6.2),(6.3),(6.5) and (6.6).
(6.2) ⇐⇒ x0 = Hλ ,
(6.5) ⇐⇒ λ0 = −Hx ,
for the Hamiltonian
H(t, x, u, λ) = f (t, x, u) + λ g(t, x, u) . (6.7)
Then (6.6) is the solution of the NLP
max H = f + λg
s.t. a≤u≤b
whose Lagrangian is
L := f (t, x, u) + λg(t, x, u) + w1 (b − u) + w2 (u − a)
giving the necessary conditions
∂L
= fu + λgu − w1 + w2 = 0 (6.8)
∂u
w1 ≥ 0 , w1 (b − u) = 0 (6.9)
w2 ≥ 0 , w2 (u − a) = 0 (6.10)
which are equivalent to (6.6). For example, if u∗ (t) = b then
u∗ − a > 0 , w2 = 0 and fu + λgu = w1 ≥ 0 .
The control u is singular if its value does not change H, for example, the middle case in
(6.6). A singular control cannot be determined from H.

6.3. Application to Nonrenewable Resources

The model (Hotelling, 1931) is
Z ∞
max e−rt u(t)p(t) dt
0
s.t. ẋ = −u
x(0) given (positive)
x(t), u(t) ≥ 0
and 0 ≤ u(t) ≤ umax (finite) , ∀ t .
Let
H(t, x, u, λ) := e−rt pu − λu (6.11)
then
λ0 = −Hx = 0 =⇒ λ = constant .
 6.4. CORNERS 43

rt rt
S λ1 e λ2 e
rt
λ 3e

L1 L2 t
u
u max

0
t

Figure 6.1. Maximum extraction in two intervals of lenghts L1 + L2 = L

Assuming the resource will be exhausted by some time T ,

£ ¤
H(T, x(T ), u(T ), λ(T )) = e−rT p(T ) − λ(T ) u(T ) = 0
∴ λ = e−rT p(T )
½
0 if p(t) < λert = er(t−T ) p(T )
∴ u(t) =
umax if >
The duration L of the time extraction takes place (u = umax ) is
Lumax = x(0) (6.12)
so λ is adjusted to satisfy (6.12). For example, consider the price function p(t) plotted in
Figure 6.1, and three exponential curves λi ert , i = 1, 2, 3. The value λ1 is too high, allowing
for maximal extraction (p(t) > λ1 ert ) in a period too short (i.e. shorter than L). Similarly,
the value λ3 is too low. The value λ2 is just right, where the duration of p(t) > λ2 ert is
L1 + L2 = L.
Special case: p(t) is exponential.
  
 >   never
st
p(t) = p(0)e , s = = r =⇒ extract anytime
  
< immediately

6.4. Corners
The optimality of discontinuous controls raises questions about the continuity of λ, H
at points where u is discontinuous. Recall the Weierstrass-Erdmann corner conditions of
§ 3.9: The functions Fx0 , F − x0 Fx0 are continuous in corners. Returning to optimal control,
44 6. OPTIMAL CONTROL, CONSTRAINED CONTROL

consider
Z T
opt F (t, x, u) dt s.t. x0 = u
0
with H = F + λu
Hu = Fu + λ = 0
∴ λ = −Fx0
The Weierstrass-Erdmann corner conditions imply
λ = −Fx0 , H = F − x0 Fx0
are continuous in corners of u.

6.5. Control Problems Linear in u

Z T
max [F (t, x) + uf (t, x)] dt (6.13)
0
0
s.t. x = G(t, x) + ug(t, x)
x(0) given, and u satisfies (6.3)
The Hamiltonian is
H := (F + λG) + u(f + λg) (6.14)
and the necessary conditions include
λ0 = −Hx (6.15)
  
 a  < 
u = ? if f + λg = 0. (6.16)
 b  > 

6.6. A Minimum Time Problem

The position of a moving particle is given by
x00 (t) = u , x(0) = x0 given , x0 (0) = 0 .
The problem is to findnu bounded by
−1 ≤ u(t) ≤ 1
bringing the particle to rest (x0 = 0) at x = 0 in minimal time.
Solution (Pontryagin et al, 1962): Let x1 = x, x2 = x0 . The problem is
Z T
min dt
0
s.t. x01 = x2 , x1 (0) = x0 , x1 (T ) = 0
x02 = u , x2 (0) = 0 , x2 (T ) = 0
u − 1 ≤ 0 , −u − 1 ≤ 0
 6.6. A MINIMUM TIME PROBLEM 45

with Lagrangian
L = 1 + λ1 x2 + λ2 u + w1 (u − 1) − w2 (u + 1)
and necessary conditions
∂L
= λ2 + w1 − w2 = 0 , (6.17)
∂u
w1 ≥ 0 , w1 (u − 1) = 0 ,
w2 ≥ 0 , w2 (u + 1) = 0 ,
∂L
λ01 = − =0, (6.18)
∂x1
∂L
λ02 = − = −λ1 , (6.19)
∂x2
L(T ) = 0 , since T is free . (6.20)
Indeed, L(t) ≡ 0 , 0 ≤ t ≤ T .
(6.18) =⇒ λ1 = constant . ∴ (6.19) =⇒ λ2 = −λ1 t + C .
If λ2 = 0 in some interval then λ1 = 0 and L(T ) = 1 + 0 + 0 contradicting (6.20).
∴ λ2 changes sign at most once .
• If λ2 > 0 then w2 > 0 and u = −1.
• If λ2 < 0 then w1 > 0 and u = 1.
Consider an interval where u = 1. Then
x02 = u = 1
∴ x2 = t + C0
t2
∴ x1 = + C0 t + C − 1
2
(x2 − C0 )2
= + C0 (x2 − C0 ) + C1
2
x2 C2
= 2 + C2 , C2 = − 0 + C1 .
2 2
Similarly, in an interval where u = −1,
x22
x1 = −+ C3
2
The optimal path must end on one of the parabolas
x22 x2
x1 = or x1 = − 2
2 2
passing through x1 = x2 = 0. The optimal path begins on a parabola passing through
(x0 , 0), and ends up on the switching curve, see Figure 6.3.
46 6. OPTIMAL CONTROL, CONSTRAINED CONTROL

x2

x1

x22
Figure 6.2. The parabolas x1 = ± + C and the switching curve.
2

x2

x0
x1

Figure 6.3. The optimal trajectory starting at (x0 , 0).

6.7. An Optimal Investment Problem

The model is
Z ∞
max e−rt [p(t)f (K(t)) − c(t)I(t)] dt
I 0
s.t. K 0 = I − bK
K(0) = K0 given
I≥0
where: K = capital assets, I = investment, f (K) = output, p(t) = unit price of output, and
c(t) = unit price of investment.
The current value Hamiltonian is
H := pf (K) − cI + m(I − bK)
 6.8. FISHERIES MANAGEMENT 47

Necessary conditions include:

m(t) ≤ c(t), I(t) [c(t) − m(t)] = 0 , (6.21)
m0 = (r + b)m − pf 0 (K) (6.22)
At any interval with I > 0, m = c and m0 = c0 . Substituting in (6.22)
pf 0 (K) = (r + b)c − c0 , while I > 0 , (6.23)
a static equation giving the optimal K, with
LHS(6.23) = marginal benefit of investment ,
RHS(6.23) = marginal cost of investment ,

Differentiating (6.23)
p0 f 0 (K) + pf 00 (K)K 0 = (r + b)c0 − c00
and substituting K 0 = I − bK we get the singular solution I > 0.
Between investment periods we have I = 0. To see what this means collect m–terms in
(6.22) and integrate Z ∞
−(r+b)t
e m(t) = e−(r+b)s p(s)f 0 (K(s)) ds
t
Also
∞ Z
−(r+b)t d £ −(r+b)s ¤
e c(t) = − e c(s) ds
ds
Z ∞t
= e−(r+b)s [c(s)(r + b) − c0 (s)] ds
t
The condition m(t) ≤ c(t) gives
Z ∞
e−(r+b)s [p(s)f 0 (K(s)) − (r + b)c(s) + c0 (s)] ds ≤ 0
t
with equality if I > 0. Therefore at any interval [t1 , t2 ] with I = 0 (I > 0 for (t1 )− and (t2 )+ )
Z t2
e−(r+b)s [pf 0 − (r + b)c + c0 ] ds ≤ 0
t
with equality at t = t1 .

6.8. Fisheries Management

The model is
Z ∞
max e−rt (p(t) − c(x))u(t) dt
0
s.t. ẋ = F (x) − u (6.24)
x(0) = x0 given
x(t) ≥ 0
u(t) ≥ 0
48 6. OPTIMAL CONTROL, CONSTRAINED CONTROL

where: x(t) = stock at time t, F (x) = the growth function, u(t) = the harvest rate, p = unit
price, and c(x) = unit cost.
Substituting (6.24) in the objective
Z ∞
max e−rt (p − c(x))(F (x) − ẋ) dt
0
R∞
or max 0 φ(t, x, ẋ) dt with E-L equation
∂φ d ∂φ
=
∂x dt ∂ ẋ
where
∂φ
= e−rt {−c0 (x) [F (x) − ẋ] + [p = c(x)] F 0 (x)}
∂x
d ∂φ d © −rt ª
= −e [p − c(x)]
dt ∂ ẋ dt
= e−rt {r [p − c(x)] + c0 (x)ẋ}
c0 (x)F (x)
∴ F 0 (x) − = r (6.25)
p − c(x)
Economic interpretation: Write (6.25) as
F 0 (x) [p − c(x)] − c0 (x)F (x) = r [p − c(x)]
d
or {[p − c(x)] F (x)} = r [p − c(x)] (6.26)
dx
RHS = value, one instant later, of catching fish # x + 1
LHS = value, one instant later, of not catching fish # x + 1
 LECTURE 7

Stochastic Dynamic Programming

This lecture is based on [28].

7.1. Finite Stage Models

A system has finitely many, say N , states, denoted i ∈ 1, N . There are finitely many,
say n, stages. The state i of the system is observed at the beginning of each stage, and a
decision a ∈ A is made, resulting in a reward R(i, a). The state of the system then changes,
from i to j ∈ 1, N , according to the transition probabilities Pij (a). Taken together, a
finite–stage sequential decision process is the quintuple {n, N, A, R, P }.
Let Vk (i) denote the maximum expected return, with state i and k stages to go. The
optimality principle then states

( N
)
X
Vk (i) = max R(i, a) + Pij (a) Vk−1 (j) , k = n, n − 1, · · · , 2 , (7.1a)
a∈A
j=1

V1 (i) = max R(i, a) , (7.1b)

a∈A

a recursive computation of Vk in terms of Vk−1 , with boundary condition (7.1b). An equiva-

lent boundary condition is V0 (i) ≡ 0 for all i, with (7.1a) for k = n, n − 1, · · · , 1.

7.1.1. A Gambling Model. This section is based on [22]. A player is allowed to

gamble n times, in each he can bet any part of his current fortune. He either wins or loses
the amount of the bet with probabilities p or q = 1 − p, respectively.
Let Vk (x) be the maximal expected return with present fortune x and k gambles to go.
Assuming the gambler bets a fraction α of his fortune,

Vk (x) = max {p Vk−1 (x + αx) + q Vk−1 (x − αx)} , (7.2a)

0≤α≤1

V0 (x) = log x , (7.2b)

49
50 7. STOCHASTIC DYNAMIC PROGRAMMING

logarithmic utility used in (7.2b). If p ≤ 1/2 then it is optimal to bet zero (α = 0) and
Vn (x) = log x. Suppose p > 1/2. Then,
V1 (x) = max {p log(x + αx) + q log(x − αx)}
α
= max {p log(1 + α) + q log(1 − α)} + log x
α
∴ α = q−p (7.3)
∴ V1 (x) = C + log x
where C = log 2 + p log p + q log q (7.4)
Similarly, Vn (x) = n C + log x (7.5)
and the optimal policy is to bet the fraction (7.3) of the current fortune. See also Exercise 7.1.
7.1.2. A Stock–Option Model. The model considered here is a special case of [30].
Let Sk denote the price of a given stock on day k, and suppose
k+1
X
Sk+1 = Sk + Xk+1 = S0 + Xi
i=1
where X1 , X2 , · · · are i.i.d. RV’s with distribution F and finite mean µF . You have an option
to buy one share at a fixed price c, and N days in which to exercise this option. If the option
is exercised (on a day) when the stock price is s, the profit is s − c.
Let Vn (s) denote the maximal expected profit if the current stock price is s and n days
remain in the life of the option. Then
½ Z ¾
Vn (s) = max s − c , Vn−1 (s + x) dF (x) , n ≥ 1 (7.6a)

V0 (x) = max{s − c, 0} (7.6b)

Clearly Vn (s) is increasing in n for all s (having more time cannot hurt) and increasing in s
for all n (prove by induction).
Lemma 7.1. Vn (s) − s is decreasing in s.
Proof. Use induction on n. The claim is true for n = 0: V0 (s) − s is decreasing in s.
Assume that Vn−1 (s) − s is decreasing in s. Then
½ Z ¾
Vn (s) − s = max −c, [Vn−1 (s + x) − (s + x)] dF (x) + µF

is decreasing in s since Vn−1 (s + x) − (s + x) is decreasing in s for all x. ¤

Theorem 7.1. There are numbers s1 ≤ s2 ≤ · · · ≤ sn ≤ · · · such that if there are n
days to go and the present price is s, the option should be exercised iff s ≥ sn
Proof. If the price is s and n days remain, it is optimal to exercise the option if
Vn (s) ≤ s − c , by (7.6a)
Let
sn := min{s : Vn (s) − s = −c} ,
 7.1. FINITE STAGE MODELS 51

with sn := ∞ if the above set is empty. By Lemma 7.1,

Vn (s) − s ≤ Vn (sn ) − sn = −c , ∀ s ≥ sn
Therefore it is optimal to exercise the option if s ≥ sn . Since Vn (s) is increasing in n, it
follows that sn is increasing. ¤
Exercise 7.2 shows that it is never optimal to exercise the option if µF ≥ 0.
7.1.3. Modular Functions and Monotone Policies. This section is based on [31].
Let a function g(x, y) be maximized for fixed x, and define y(x) as the maximal value of y
where the maximum occurs,
max g(x, y) = g(x, y(x))
y
When is y(x) increasing in x? It can be shown that a sufficient condition is
g(x1 , y1 ) + g(x2 , y2 ) ≥ g(x1 , y2 ) + g(x2 , y1 ) , ∀ x1 > x2 , y1 > y2 . (7.7)
A function g(x, y) satisfying (7.7) is called supermodular, and is called submodular if
the inequality is reversed.
Lemma 7.2. If ∂ 2 g(x, y)/∂x∂y exists, then g is supermodular iff
∂ 2 g(x, y)
≥ 0, (7.8)
∂x∂y
and submodular iff ∂ 2 g(x, y)/∂x∂y ≤ 0.
Proof. If ∂ 2 g(x, y)/∂x∂y ≥ 0 then for x1 > x2 and y1 > y2 ,
Z y1 Z x1 2
∂ g(x, y)
dx dy ≥ 0 ,
y2 x2 ∂x∂y
Z y1
∴ [g(x1 , y) − g(x2 , y)] dy ≥ 0 ,
y2

and (7.7) follows. Conversely, suppose (7.7) holds. Then for all x1 > x2 and y1 > y,
g(x1 , y1 ) − g(x1 , y) g(x2 , y1 ) − g(x2 , y)
≥
y1 − y y1 − y
and as y1 ↓ y,
∂ ∂
g(x1 , y) ≥ g(x2 , y)
∂y ∂y
implying (7.8). ¤
Example 7.1. An optimal allocation problem with penalty costs. There are I identical
jobs to be done consecutively in N stages, at most one job per stage with uncertainty
about completing it. If in any stage a budget y is allocated, the job will be completed with
probability P (y), an increasing function with P (0) = 0. If a job is not completed, the amount
allocated is lost. At the end of the N th stage, if there are i unfinished jobs, a penalty C(i)
must be paid.
The problem is to determine the optimal allocation at each stage so as to minimize the
total expected cost.
52 7. STOCHASTIC DYNAMIC PROGRAMMING

Take the state of the system to be the number i of unfinished jobs, and let Vn (i) be the
minimal expected cost in state i with n stages to go. Then

Vn (i) = min {y + P (y) Vn−1 (i − 1) + [1 − P (y)] Vn−1 (i)} , (7.9a)

y≥0

V0 (i) = C(i) . (7.9b)

Clearly Vn (i) increases in i and decreases in n. Let yn (i) denote the minimizing y in (7.9a).
One may expect that
yn (i) increases in i and decreases in n . (7.10)
It follows that yn (i) increases in i if
∂2
{y + P (y) Vn−1 (i − 1) + [1 − P (y)] Vn−1 (i)} ≤ 0 ,
∂i∂y
∂
or P 0 (y) [Vn−1 (i − 1) − Vn−1 (i)] ≤ 0 ,
∂i
∂
where i is formally interpreted as continuous, so makes sense. Since P 0 (y) ≥ 0, it follows
∂i
that
yn (i) increases in i if Vn−1 (i − 1) − Vn−1 (i) decreases in i . (7.11)
Similarly,
yn (i) decreases in n if Vn−1 (i − 1) − Vn−1 (i) increases in n . (7.12)
It can be shown that if C(i) is convex in i, then (7.11)–(7.12) hold.
7.1.4. Accepting the best offer. This section is based on [19]. Given n offers in
sequence, which one to accept? Assume:
(a) if any offer is accepted, the process stops,
(b) if an offer is rejected, it is lost forever,
(c) the relative rank of an offer, relative to previous offers, is known, and
(d) information about future offers is unavailable.
The objective is to maximize the probability of accepting the best offer, assuming that
all n! arrangements of offers are equally likely.
Let P (i) denote the probability that the ith offer is the best,
1/n i
P (i) = Prob{offer is best of n| offer is best of i} = = ,
1/i n
and let H(i) denote the best outcome if offer i is rejected, i.e. the maximal probability of
accepting the best offer if the first i offers were rejected. Clearly H(i) is decreasing in i.
The maximal probability of accepting the best offer is
½ ¾
i
V (i) = max {P (i), H(i)} = max , H(i) , i = 1, · · · , n .
n
 7.2. DISCOUNTED DYNAMIC PROGRAMMING 53

It follows (since i/n is increasing and H(i) is decreasing) that for some j,
i
≤ H(i) , i ≤ j ,
n
i
> H(i) , i > j ,
n
and the optimal policy is: reject the first j offers, then accept the best offer.

Exercises.
Exercise 7.1. What happens in the gambling model in § 7.1.1 if the boundary condition
(7.2b) is replaced by V0 (x) = x ?
Exercise 7.2. For the stock option problem of § 7.1.2 show: if µF ≥ 0 then sn = ∞ for
n ≥ 1.

7.2. Discounted Dynamic Programming

Consider the model in the beginning of § 7.1 with two changes:
(a) the number of states is countable, and
(b) the number of stages is countable,
and assume that rewards are bounded, i.e. there exists B > 0 such that
|R(i, a)| < B , ∀ i, a . (7.13)
A policy is a rule a = f (i, t), assigning an action a :=∈ A to a state i at time t. It is:
randomized if it chooses a with some probability Pa , a ∈ A, and stationary if not
randomized, and if the rule is a = f (i), i.e. depends only on the state.
If a stationary policy is used, then the sequence of states {Xn : n = 0, 1, 2, · · · } is a
Markov chain with transition parobabilities Pij = Pij (f (i)).
Let 0 < α < 1 be the discount factor. The expected discounted return Vπ (i) from a
policy π and initial state i is the conditional expectation
(∞ )
X
Vπ (i) = Eπ αn R(Xn , an )| X0 = i (7.14)
n=0

and as consequence of (7.13),

B
Vπ (i) < , for all policies π . (7.15)
1−α
7.2.1. Optimal policies. The optimal value function is defined as
V (i) := sup Vπ (i) , ∀ i .
π

A policy π ∗ is optimal if
Vπ∗ (i) = V (i) , ∀ i .
54 7. STOCHASTIC DYNAMIC PROGRAMMING

Theorem 7.2. The optimal value function V satisfies

( )
X
V (i) = max R(i, a) + α Pij (a)V (j) , ∀ i (7.16)
a
j

Proof. Let π be any policy, choosing action a at time 0 with probability Pa , a ∈ A .

Then " #
X X
Vπ (i) = Pa R(i, a) + Pij (a) Wπ (j)
a∈A j

where Wπ (j) is the expected discounted return from time 1, under policy π. Since
Wπ (j) ≤ αV (j)
it follows that
" )
X X
Vπ (i) ≤ Pa R(i, a) + α Pij (a) V (j)
a∈A j
" #
X X
≤ Pa max R(i, a) + Pij (a) Wπ (j)
a
a∈A j
( )
X
= max R(i, a) + α Pij (a)V (j) (7.17)
a
j
( )
X
∴ V (i) ≤ max R(i, a) + α Pij (a)V (j) (7.18)
a
j

since π is arbitrary. To prove the reverse inequality, let a policy π begin with a decision a0
satisfying
( )
X X
R(i, a0 ) + α Pij (a0 )V (j) = max R(i, a) + α Pij (a)V (j) (7.19)
a
j j

and continue in such a way that Vπj (j) ≥ V (j) − ².

X
∴ Vπ (i) = R(i, a0 ) + α Pij (a0 )Vπj (j)
j
X
≥ R(i, a0 ) + α Pij (a0 )V (j) − α²
j
X
∴ V (i) ≥ R(i, a0 ) + α Pij (a0 )V (j) − α²
j
( )
X
∴ V (i) ≥ max R(i, a) + α Pij (a)V (j) − α²
a∈A
j

and the proof is completed since ² is arbitrary. ¤

 7.2. DISCOUNTED DYNAMIC PROGRAMMING 55

We prove now that a stationary policy satisfying the optimality equation (7.16) is optimal.
Theorem 7.3. Let f be a stationary policy, such that a = f (i) is an action maximizing
RHS(7.16)
( )
X X
R(i, f (i)) + α Pij (f (i))V (j) = max R(i, a) + α Pij (a)V (j) , ∀ i .
a∈A
j j

Then
Vf (i) = V (i) , ∀ i .
Proof. Using
( )
X
V (i) = max R(i, a) + α Pij (a)V (j)
a∈A
j
X
= R(i, f (i)) + α Pij (f (i))V (j) (7.20)
j

we interpret V as the expected return from a 2–stage process with policy f in stage 1, and
terminal reward V in stage 2. Repeating the argument
V (i) = E {n–stage return under f | X0 = i} + αn E(V (Xn |X0 = i)
→ V (i) , by (7.15) .
¤

7.2.2. A fixed point argument. We recall the following

Theorem 7.4. (The Banach Fixed Point Theorem) Let T : X → X be defined on a
complete metric space {X, d}, and let 0 < α < 1 satisfy
d(T (x), T (y)) ≤ αd(x, y) , ∀ x, y ∈ X . (7.21)
Then T has a unique fixed point x∗ , and the successive approximations xn+1 = T (xn )
converge uniformly to x∗ on any sphere d(x0 , x∗ ) ≤ δ.
Proof. For (7.21) it follows that
d(xn+1 , xn ) = d(T (xn ), T (xn−1 )) ≤ αd(xn , xn−1 )
∴ d(xn+1 , xn ) ≤ αn d(x1 , x0 )
n−1
X
m n
∴ d(x , x ) ≤ d(xi , xi+1 ) , for m < n ,
i=m
≤ α (1 + α + α2 + · · · + αn−1−m )d(x1 , x0 ) ,
m

αm
≤ d(x1 , x0 ) , (7.22)
1−α
∴ d(xm , xn ) → 0 as m, n → ∞ .
56 7. STOCHASTIC DYNAMIC PROGRAMMING

Since {X, d} is complete, the sequence {xn } converges, say to x∗ ∈ X. Then T (x∗ ) = x∗
since
d(xn+1 , T (x∗ )) = d(T (xn ), T (x∗ )) ≤ αd(xn , x∗ ) → 0 as n → ∞
If x∗∗ is another fixed point of T then
d(x∗ , x∗∗ ) = d(T (x∗ ), T (x∗∗ ) ≤ αd(x∗ , x∗∗ )
and x∗ = x∗∗ since α < 1. Since (7.22) is independent of n it follows that
αm
d(xm , x∗ ) ≤ d(x1 , x0 ) , m = 1, 2, · · · (7.23)
1−α
¤
A mapping T satisfying (7.21) is called contraction, and Theorem 7.4 is also called the
contraction mapping principle.
Theorem 7.4 applies here as follows: Let U be the set of bounded functions on the state
space, endowed with the sup norm
kuk = sup |u(i)| (7.24)
i
For any stationary policy f let Tf be an operator mapping U into itself,
X
(Tf u)(i) := R(i, f (i)) + α Pij (f (i))u(j)
j

Then Tf satisfies for all i,

|(Tf u)(i) − (Tf v)(i)| < α sup |u(j) − v(j)| , ∀ u, v ∈ U
j

and therefore,
sup |(Tf u)(i) − (Tf v)(i)| < α sup |u(j) − v(j)| , ∀ u, v ∈ U (7.25)
i j

i.e. Tf is a contraction in the sup norm (since 0 < α < 1). It follows that
Tfn u → Vf as n → ∞
It follows that the optimal value function V is the unique bounded solution of (7.16)
7.2.3. Successive approximations of the optimal value. For V0 (i) be any bounded
function, i ∈ 1, N , and define
X
Vn (i) := max {R(i, a) + α Pij (a)Vn−1 (j)} , n = 1, 2, · · · (7.26)
a∈A
j

Then Vn converges to the optimal value V .

Proposition 7.1.
(a) If V0 ≡ 0 then
B
|V (i) − Vn (i)| ≤ αn+1 .
1−α
(b) For any bounded Vo , Vn (i) → V (i) uniformly in i.
Note: (a) is the analog of (7.23).
 7.2. DISCOUNTED DYNAMIC PROGRAMMING 57

7.2.4. Policy improvement. Let h be a policy that improves on a given stationary

policy g in the sense of (7.27) below. The following proposition shows the seinse in which h
is closer to an optimal policy than g.

Proposition 7.2. Let g be a stationary policy with expected value Vg and define a policy
h by
X X
R(i, h(i)) + α Pij (h(i))Vg (i) = max {R(i, a) + α Pij (a)Vg (j)} . (7.27)
a
j j

Then

Vh (i) ≥ Vg (i) , ∀ i ,
and if Vh (i) = Vg (i) , ∀(i) , then Vg = Vh = V .

The policy improvement method starts with any policy g, and does (7.27) until no
improvement is possible. Finiteness follows if the state and action spaces are finite.

7.2.5. Solution by LP.

Proposition 7.3. If u is a bounded function on the state space such that

X
u(i) ≥ max{R(i, a) + α Pij (a)u(j)} , ∀ i , (7.28)
a
j

then u(i) ≥ V (i) for all i.

Proof. Let the mapping T on be defined by

X
(T u)(i) := max{R(i, a) + α Pij (a)u(j)}
a
j

for bounded functions u. Then (7.28) means u ≥ T u, and therefore u ≥ T n u for all u. ¤

It follows that the optimal value function V is the unique solution of the problem
P
min u(i)
i P
s.t. u(i) ≥ max{R(i, a) + α Pij (a)u(j)} , ∀ i ,
a j

or equivalently
P
min u(i)
i P (7.29)
s.t. u(i) ≥ R(i, a) + α Pij (a)u(j) , ∀ i , ∀ a ∈ A .
j
58 7. STOCHASTIC DYNAMIC PROGRAMMING

7.3. Negative DP
Assume countable state space and finite action space. “Negative” means here that the
rewards are negative, i.e. the objective is to minimize costs. Let a cost C(i, a) ≥ 0 result
from action a in state i, and for any policy π,
"∞ #
X
Vπ (i) = Eπ C(Xn , an ) : X0 = i
n=0
V (i) = inf Vπ (i)
π
∗
and call a policy π optimal if
Vπ∗ (i) = V (i) , ∀ i .
The optimality principle here is
X
V (i) = min{C(i, a) + Pij (a)V (j)}
a
j

Theorem 7.5. Let f be a stationary policy defined by

X X
C(i, f (i)) + Pij (f (i))V (j) = min{C(i, a) + Pij (a)V (j)} , ∀ i .
a
j j

Then f is optimal.
Proof. From the optimality principle it follows that
X
C(i, f (i)) + Pij (f (i))V (j) = V (i) , ∀ i . (7.30)
j

If V (j) is cost of stopping in state j, (7.30) expresses indifference between stopping and
continuing one more stage. Repeating the argument,
" n−1 #
X
Ef C(Xt , at ) : X0 = i + Ef [V (Xn ) : X0 = i] = V (i)
t=0
and, since costs are nonnegative,
" n−1 #
X
Ef C(Xt , at ) : X0 = i ≤ V (i) ,
t=0

and in the limit, Vf (i) ≤ V (i), proving optimality. ¤

 Bibliography

[1] G.C. Alway, Crossing the desert, Math. Gazette 41(1957), 209
[2] V.I. Arnold, Mathematical Methods of Classical Mechanics (Second Edition, English
Translation), Springer, New York, 1989
[3] R.E. Bellman, On the theory of dynamic programming Proc. Nat. Acad. Sci 38(1952),
716–719.
[4] R.E. Bellman, Dynamic Programming, Princeton University Press, 1957
[5] R.E. Bellman and S.E. Dreyfus, Applied Dynamic Programming, Princeton University
Press, 1962
[6] A. Ben-Israel and S.D. Flåm, Input optimization for infinite horizon discounted programs,
J. Optimiz. Th. Appl. 61(1989), 347–357
[7] A. Ben-Israel and W. Koepf, Making change with DERIVE: Different approaches, The
International DERIVE Journal 2(1995), 72–78
[8] D.P. Bertsekas, Dynamic Programming and Stochastic Control, Academic Press, 1976
[9] U. Brauer and W. Brauer, A new approach to the jeep problem, Bull. Euro. Assoc. for
Theoret. Comp. Sci. 38(1989), 145–154
[10] J.W. Craggs, Calculus of Variations, Problem Solvers # 9, George Allen & Unwin,
London, 1973
[11] A.K. Dewdney, Computer recreations, Scientific American, June 1987, 128–131
[12] S.E. Dreyfus and A.M. Law, The Art and Theory of Dynamic Programming, Academic
Press, 1977
[13] L. Elsgolts, Differential Equations and the Calculus of Variations (English Translation),
MIR, Moscow, 1970
[14] G.M. Ewing, Calculus of Variations with Applications, Norton, 1969 (reprinted by
Dover, 1985)
[15] N.J. Fine, The jeep problem, Amer. Math. Monthly 54(1947), 24–31
[16] C. Fox, An Introduction to the Calculus of Variations, Oxford University Press, 1950
(reprinted by Dover, 1987(
[17] J.N. Franklin, The range of a fleet of aircraft, J. Soc. Indust. Appl. Math. 8(1960),
541–548
[18] D. Gale, The jeep once more or jeeper by dozen, Amer. Math. Monthly 77(1970), 493–
501
[19] J. Gilbert and F. Mosteller, Recognizing the maximum of a sequence, J. Amer. Statist.
Assoc. 61(1966), 35-73
[20] H. Goldstein, Classical Mechanics, Addison-Wesley, Reading, 1950
[21] A. Hausrath, B. Jackson, J. Mitchem, E. Schmeichel, Gale’s round-trip jeep problem,
Amer. Math. Monthly 102(1995), 299–309
59
 View publication stats

60 BIBLIOGRAPHY

[22] J. Kelly, A new interpretation of information rate, Bell System Tech. J. 35(1956), 917–
926
[23] C. Lanczos, The Variational Principles of Mechanics (4th Edition), University of
Toronto Press, 1970. Reprinted by Dover, 1987.
[24] L.D. Landau and E.M. Lifshitz, Course of Theoretical Physics, Vol 1: Mechanics (Eng-
lish Translation), Pergamon Press, Oxford 1960
[25] C.G. Phipps, The jeep problem, Amer. Math. Monthly 54(1947), 458–462
[26] D.A. Pierre, Optimization Theory with Applications, Wiley, New York, 1969 (reprinted
by Dover, 1986)
[27] W. Ritz, Über eine neue Methode zur Lösung gewisser Variationsprobleme der mathe-
matischen Physik, J. Reine Angew. Math. 135(1908)
[28] S.M. Ross, Introduction to Stochastic Dynamic Programming, Academic Press, New
York, 1983
[29] M. Sniedovich, Dynamic Programming, M. Dekker, 1992
[30] H. Taylor, Evaluating a call option and optimal timing strategy in the stock market,
Management Sci. 12(1967), 111-120
[31] D.M. Topkins, Minimizing a submodular function on a lattice, Operations Res. 26(1978),
305-321
[32] R. Weinstock, Calculus of Variations with Applications to Physics & Engineering, Mc-
Graw Hill, New York, 1952 (reprinted by Dover, 1974)
[33] H.S. Wilf, Mathematics for the Physical Sciences, Wiley, New York, 1962 (reprinted by
Dover, 1978)