J. Math. Anal. Appl.

288 (2003) 47–68

www.elsevier.com/locate/jmaa

Optimality conditions for age-structured control

systems ✩
Gustav Feichtinger,a Gernot Tragler,a and Vladimir M. Veliov a,b,∗
a Institute for Econometrics, Operations Research and Systems Theory, Vienna University of Technology,
Argentinierstrasse 8/119, A-1040 Vienna, Austria
b Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 1113 Sofia, Bulgaria

Received 28 August 2001

Submitted by A. Cellina

Abstract
We consider a fairly general model (extension of the Gurtin–MacCamy model of population
dynamics) of an age structured control system with nonlocal dynamics and nonlocal boundary condi-
tions. A necessary optimality condition is obtained in the form of Pontryagin’s maximum principle,
which is applicable to a number of practically meaningful models where the previously known results
fail. We discuss such models (an epidemic control, and a capital accumulation model) as illustrations.
 2003 Elsevier Inc. All rights reserved.

Keywords: Age-structured systems; Population dynamics; McKendrick equation; Optimal control of distributed
systems; Pontryagin’s maximum principle

1. Introduction

Optimal control problems for age-structured systems are of interest for many areas of
application, as harvesting [2,16,17,21,24], birth control [3,7,8], epidemic disease control
and optimal vaccination [14,20,22,23], investment economic models [4,5,11,13,18,26], and
for a variety of models in the social area [1,19]. Many of the above papers present opti-
mality conditions for particular models, usually in the form of a maximum principle of
Pontryagin’s type. A general maximum principle for nonlinear McKendrick-type systems

✩
This research was partly supported by the Austrian Science Foundation under contract No. P15618.
* Corresponding author.
E-mail addresses: or@e119ws1.tuwien.ac.at (G. Feichtinger), tragler@e119ws1.tuwien.ac.at (G. Tragler),
vveliov@eos.tuwien.ac.at (V.M. Veliov).

0022-247X/$ – see front matter  2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2003.07.001
48 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

is obtained in [6]. However, a number of extensions of the McKendrick and Gurtin–

MacCamy [15] models arose recently, where the existing optimality conditions are not
applicable. Below we sketch two such extensions, which are versions (aimed just to out-
line the critical points) of models investigated in [1,13,22].
The first one arises in epidemic control problems, as well as in the social area, where
contagious factors are responsible for the evolution (viruses, fashion, drug consumption,
crime, etc.). The model is described by the McKendrick-type system


∂ ∂    
+ y(t, a) = S(t, a)Φ p(t, a) − u(t, a)Θ y(t, a) , (1)
∂t ∂a
ω
p(t, a) = m(a, a  )y(t, a  ) da , (2)
0
with the following meaning: y(t, a) is the number of the individuals of age a who belong
to a specific group (the group of infected individuals, in the epidemic context; the group
of drug users, or offenders, in a social context) at time t. The quantity p(t, a) measures
the total impact towards becoming infected, or becoming a drug user, etc., to which a
nonmember individual of age a is exposed at time t. The function m measures, therefore,
the impact of a member of the group of age a  on a nonmember of age a. Individuals of
age more than ω are disregarded. The “reputation” function Φ transforms the total impact
p(t, a) to a rate at which the nonmember individuals of age a (S(t, a) is their number at
time t) become infected (users). Eq. (1) is a limit form of the fact that the members of age
a + h at time t + h are just the members who were of age a at time t, plus those of age
a who have joined the group in the interval [t, t + h] (the first term), minus those of age
a who have left (the second term). The function Θ(y) represents a state-feedback factor,
u(t, a) is an age-specific control (treatment) that is to be chosen to minimize the objective
function
T ω
  
d(a)y(t, a) + c u(t, a) da dt, (3)
0 0
possibly subject to control constraints. Here d(a) is the damage caused by a member of
age a, while the second term represents the cost of the control.
The known optimality conditions are not applicable to this problem at least because of
the age-dependency of the integral (nonlocal) quantity p(t, a). On the other hand, kernels
m(a, a ) depending in a nonseparable way on a and a  are typical in the area.
Another specific case of an age-dependent kernel m(a, a  ) (where m is an indicator
function) arises when the fertility/mortality rate at an age a depends on the number of
individuals younger (and/or older) than a [9]. One of our goals in the paper is to cope with
such situations.
The second example is a simplified version of a capital accumulation model presented
in [13]. Below, K(t, a) is the stock of capital goods of age a at time t, u(t, a) is the intensity
of buying machines of age a at time t, v(t) is the purchase rate of new machines, δ(a) is
the depreciation rate, Q(t) is the output at time t, I (t) is the total purchase rate of old
machines at t. The basic equations are
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 49


∂ ∂
+ K(t, a) = −δ(a)K(t, a) + u(t, a), (4)
∂t ∂a
K(0, a) = K0 (a), K(t, 0) = v(t),
ω ω
Q(t) = m(t − a)d(a)K(t, a) da, I (t) = u(t, a) da. (5)
0 0

Here d(a) is the productivity of the capital goods of age a, corrected by the factor m(t − a)
reflecting the technological progress (notice that t − a is the vintage of the machine), i.e.,
m
0. The maximal age of the capital goods is ω, r is the discount rate. The objective
function to be maximized is
T 
 
e−rt R Q(t) − α0 v(t) − βv(t)2 − γi I (t)2
0
ω
 
− α(a)u(t, a) + γe (a)u(t, a) da dt,
2

where R(Q) is the revenue (which is usually assumed to be either a linear or a quadratic
function of Q, depending on whether a competitive or monopolistic economy is consid-
ered), α0 v and αu are the acquisition costs of machines, βv 2 is the adjustment cost for new
machines, γi I 2 and γe u2 are adjustment costs and implementation costs for old machines,
respectively.
Two are the specific features for which the general result in [6] is not applicable. The
first is the presence of a boundary control, v(t). The second is more essential: the distrib-
uted control u(t, a) appears both in the differential equation and in the equation for I ,
where the integration is with respect to a. There is a deep obstacle to treat such sys-
tems with the abstract approach from [6]. The measure theoretical constructions needed
to implement this approach require that the controls in the differential equation, and in the
equations where integration in a is involved, are independent; in particular, they cannot
coincide or be subjected to joint constraints.
Notice also, that it makes sense to consider K0 (a) as another control, if for example,
a new enterprise is to be created. Therefore we include initial value control in the general
model below.
In the present paper we consider the following general optimal control problem:
ω
 
minimize l a, y(T , a) da
0
T ω
 
+ L t, a, y(t, a), p(t, a), q(t), u(t, a), v(t), w(a) da dt, (6)
0 0

subject to the equations

50 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68


∂ ∂  
+ y(t, a) = f t, a, y(t, a), p(t, a), q(t), u(t, a) , (7)
∂t ∂a
ω
 
p(t, a) = g t, a, a  , y(t, a ), u(t, a  ) da  , (8)
0
ω
 
q(t) = h t, a, y(t, a), p(t, a), q(t), u(t, a) da, (9)
0
the initial condition
 
y(0, a) = y 0 a, w(a) , (10)
the boundary condition
 
y(t, 0) = ϕ t, q(t), v(t) , (11)
and the control constraints
u(t, a) ∈ U, v(t) ∈ V , w(a) ∈ W. (12)
Here t is the time, running in a given interval [0, T ], a ∈ [0, ω] is a scalar variable inter-
preted as age, y(t, a) = (y1 (t, a), . . . , ym (t, a)) ∈ Rm , p(t, a) = (p1 (t, a), . . . , pn (t, a)) ∈
Rn , and q(t) = (q1 (t), . . . , qr (t)) ∈ Rr are the states of the system, u(t, a) ∈ U , v(t) ∈ V
and w(a) ∈ W are distributed, boundary and initial controls, respectively, U , V and W are
subsets of finite-dimensional linear normed spaces, l, L, f, g, h, y 0 , ϕ are given functions,
ω and T are given positive numbers. The strict formulation and the suppositions are given
in the next section. We refer to [10,25] for more detailed discussion of the age-structured
systems.
The main goal of the paper is to obtain a necessary optimality condition (maximum
principle) for the problem (6)–(12), in a form readily applicable to the numerous models
of this type that arise. Since the system is nonlinear and nonconvex, and since, on the other
hand, our aim is to obtain a global maximum principle, the general approach from [12] is
appropriate, in principle. We adapt this approach (in fact, the classical Weierstrass method
in the calculus of variations) using the specific structure of the system, choosing appropri-
ate variations, and utilizing the precise sensitivity estimates obtained in Section 3.
The paper is organized as follows. In Section 2 we give a precise formulation of the
problem and the suppositions. In Section 3 we estimate the dependence of the solution
of system (7)–(11) on perturbations, which is essential for the subsequent proofs. The
maximum principle for the problem (6)–(12) is formulated and discussed in Section 4 and
proven in Section 5. In the last section we briefly analyze the two examples presented
above, utilizing the maximum principle.

2. Problem statement and suppositions

def
We denote by D = [0, T ] × [0, ω] the domain in which we consider the problem (6)–
(12). Thus
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 51

y : D
→ Rm , p : D
→ Rn , q : [0, T ]
→ Rr , u : D
→ U,
v : [0, T ]
→ V , w : [0, ω]
→ W, l : [0, ω] × R
→ R,
m

L : D × R × R × R × U × V × W
→ R,
m n r

f : D × Rm × Rn × Rr × U
→ Rm , g : D × [0, ω] × Rm × U
→ Rn ,
h : D × Rm × Rn × Rr × U
→ Rr , y 0 : [0, ω] × W
→ Rm ,
ϕ : [0, T ] × Rr × V
→ Rm .
Admissible control is any triple (u, v, w) of measurable functions u : D
→ U , v :
[0, T ]
→ V , w : [0, ω]
→ W .

Standing suppositions. The sets U , V and W are compact, V and W are convex.
The functions l, L, f, g, h, y 0 , ϕ are Carathéodory (that is, measurable in t, a, a  and
continuous in the rest of the variables), locally essentially bounded, differentiable in
(y, p, q, v, w), with locally Lipschitz partial derivatives, uniformly with respect to u ∈ U
and (t, a) ∈ D, a  ∈ [0, ω]. The ith component of h is independent of qj for each
i = 1, . . . , r and j
i. Moreover, we suppose1 that there is a compact set Z ⊂ Rm+n+r
such that for every admissible control the system (7)–(11) has a unique solution (y, p, q)
on D (in the sense described below) and the solution takes values in Z.
For a fixed admissible control (u, v, w) we shall define the notion of solution to the
system (7)–(9). Our definition is equivalent to that in [3] (see Lemma 1 below) and, es-
sentially, also to those in [6,25]. Since in the next sections we shall need this notion for
systems with different side conditions, we consider now the following somewhat more
general problem for Eqs. (7)–(9). Let Γ ⊂ D be an arbitrary continuous curve joining the
points (0, ω) and (T , 0), and such that every characteristic line t − a = const intersects Γ
at a single point. Instead of the initial condition (10) and the boundary condition (11) we
consider the condition
y(γ ) = ȳ(γ ), γ ∈ Γ, (13)
where ȳ is a given measurable bounded function on Γ . Let us denote by e the vector (1, 1).
For γ ∈ Γ we denote by S(γ ) the interval of all values s such that γ + se ∈ D (see Fig. 1).

Definition 1. Solution of (7)–(9), (13) is any triple of measurable and bounded functions
y, p, q on D ([0, T ], respectively) such that the equality
s
 
y(γ + se) = ȳ(γ ) + f γ + τ e, y(γ + τ e), p(γ + τ e), q(t0 + τ ), u(γ + τ e) dτ
0
(14)

1 The last supposition is redundant. It is included (recommended by one of the referees) for the sake of sim-
plification of the proofs. In fact, uniqueness and local existence can be proven (see [6,25]), while global existence
is required for the optimal solution only.
52 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

Fig. 1. The “side” conditions are given on Γ .

holds for a.e. γ = (t0 , a0 ) ∈ Γ and a.e. s ∈ S(γ ), (8) holds for a.e. (t, a) ∈ D, and (9) holds
for a.e. t ∈ [0, T ].

Obviously every function that differs on a set of measure zero from a solution is
also a solution, therefore the solution can be considered as an element of L∞ (D; Rm ) ×
L∞ (D; Rn ) × L∞ ([0, T ]; Rr ). Similarly, the controls u, v and w can be considered as
elements of L∞ (D; U ), L∞ ([0, T ]; V ) and L∞ ([0, ω]; W ), respectively.
Eq. (14) implies that for a.e. γ ∈ Γ the solution y is (equivalent to) an absolutely con-
tinuous function on the characteristic line through γ .

Remark 1. As a consequence of the last fact one easily obtains that for any solution
(y, p, q) ∈ L∞ (D; Rm ) × L∞ (D; Rn ) × L∞ ([0, T ]; Rr ) and for any continuous curve Υ
that intersects each characteristic line at most finitely many times, the restriction y|Υ is
an well-defined (trough the values of the absolutely continuous representatives along the
characteristic lines) element of L∞ (Υ ; Rm ) and y|Υ L∞ (Υ )  yL∞ (D) .

In particular, y(t, ·) is a well defined L∞ -function for every t ∈ [0, T ], therefore the
terminal term in (6) makes sense.

Lemma 1. If (y, p, q) ∈ L∞ (D; Rm ) × L∞ (D; Rn ) × L∞ ([0, T ]; Rr ) is a solution of (7)–

(9), (13), then the mapping
 
[0, T ]  t → y(t, ·) ∈ L1 [0, ω]
is Lipschitz continuous.

Proof. We have
   
y(t + ε, a) − y(t, a) = y(t + ε, a + ε) − y(t, a) + y(t + ε, a) − y(t + ε, a + ε) .
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 53

The first term in the right-hand side is proportional to ε thanks to the absolute continuity of
y along the characteristic lines, and the boundedness of all functions involved. The second
term, integrated on [0, ω − ε] is also proportional to ε since y is bounded. This implies in
an obvious way the claim. ✷

3. Sensitivity of the age-structural system

The following proposition plays a key role in the proof of the optimality condition in
Section 5. Since here we consider the problem with the initial and boundary conditions
(10)–(11), the set Γ consists of the left and the lower bounds of D. We shall use the
def
notation D  = D × [0, ω].

Proposition 1. There exists a constant C such that for every two admissible controls
(û, v̂, ŵ) and (u, v, w) the corresponding trajectories (ŷ, p̂, q̂) and (y, p, q) satisfy the
inequalities
 
p − p̂L∞ (D) + q − q̂L∞ ([0,T ])  C ∆∞ (u) + ∆1 (v, w) , (15)
 ∞

y − ŷL∞ (D)  C ∆∞ (u) + ∆ (v, w) , (16)

y(t, ·) − ŷ(t, ·)
+ p − p̂L1 (D) + q − q̂L1 ([0,T ])
L1 ([0,ω])
 1 
 C ∆ (u) + ∆1 (v, w) ∀t ∈ [0, T ], (17)
where the following notations are used:
  
∆f (t, a) = f t, a, ŷ(t, a), p̂(t, a), q̂(t), u(t, a)
 
− f t, a, ŷ(t, a), p̂(t, a), q̂(t), û(t, a) ,
    
∆g (t, a, a  ) = g t, a, a , ŷ(t, a  ), u(t, a  ) − g t, a, a , ŷ(t, a  ), û(t, a  ) ,
  
∆h (t, a) = h t, a, ŷ(t, a), p̂(t, a), q̂(t), u(t, a)
 
− h t, a, ŷ(t, a), p̂(t, a), q̂(t), û(t, a) ,
    
∆y 0 (a) = y 0 a, w(a) − y 0 a, ŵ(a) ,
    
∆ϕ (t) = ϕ t, q̂(t), v(t) − ϕ t, q̂(t), v̂(t) 
and also
∆1 (u) = ∆f L1 (D) + ∆g L1 (D  ) + ∆h L1 (D) ,
∆1 (v, w) = ∆y 0 L1 ([0,ω]) + ∆ϕ L1 ([0,T ]) ,
  ω
∆∞ (u) = max esup ∆f (γ + se) ds, esup ∆h (t, a) da,
γ ∈Γ t ∈[0,T ]
S(γ ) 0
ω
esup ∆g (t, a, a  ) da  ,
(t,a)∈D
0
∆∞ (v, w) = ∆y 0 L∞ ([0,ω]) + ∆ϕ L∞ ([0,T ]) ,
and esup means “essential supremum.”
54 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

Proof. Let M be a Lipschitz constant of f , g, h and ϕ with respect to (y, p, q) ∈ Z (the

set Z is introduced in the end of the standing suppositions), uniformly in (t, a) ∈ D, u ∈ U ,
v ∈ V (respectively (t, a, a ) ∈ D  for the function g). Below d1 , d2 , . . . , C1 , C2 , . . . denote
constants depending only on M, T and ω. The following calculations are routine, but
somewhat lengthy, therefore we skip some details.
Denoting
def   def  
∆y (t, x) = y(t, x) − ŷ(t, x), ∆p (t, x) = p(t, x) − p̂(t, x),
def  
∆q (t) = q(t) − q̂(t)
we obtain successively
ω ω
 
∆p (t, a)  M ∆y (t, a ) da + ∆g (t, a, a  ) da , (18)
0 0
ω ω ω
∆q (t)  d1 ∆y (t, a  ) da  + d2 ∆g (t, a  , a  ) da  da 
0 0 0
ω
+ d3 ∆h (t, a  ) da , (19)
0

where the second inequality is obtained iteratively using the first one and the structural
supposition about h. Using the definition of a solution to (7) first with γ = (0, a0 ) and then
with γ = (t0 , 0) we obtain
s

∆y (s, a0 + s)  ∆y 0 (a0 ) + M∆y (τ, a0 + τ ) + M∆p (τ, a0 + τ ) + M∆q (τ )
0

+ ∆f (τ, a0 + τ ) dτ (20)

and
s

∆y (t0 + s, s)  ∆ϕ (t0 ) + M∆q (t0 ) + M∆y (t0 + τ, τ ) + M∆p (t0 + τ, τ )
0

+ M∆q (t0 + τ ) + ∆f (t0 + τ, τ ) dτ. (21)
Then utilizing (18) and (19), changing appropriately the variables and reordering the terms
we come up with the following inequality, valid for a.e. (t, a) ∈ [0, T ] × [0, ω]:
t  ω
∆y (t, a)  ∆0 (t, a) + M∆y (s, a − t + s) + d2 ∆y (s, a  ) da 
θ(t −a) 0
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 55

ω ω
 
+ ∆f (s, a − t + s) + M ∆g (s, a − t + s, a ) da + M ∆h (s, a  ) da 
0 0
ω ω
+ M2 ∆g (s, a  , a  ) da  da  ds, (22)
0 0

where
def
θ (α) = max{0, α},

 ∆y 0 (a − t)

 ω
if a > t,
− a, a  ) + M∆ − a, a  )
def
∆0 (t, a) = ∆ϕ (t − a) + d3 ∆y (t h (t


0
ω 
+ 0 M 2 ∆g (t − a, a , a  ) da  da  if a < t.
Integrating (22) with respect to a in [0, ω] and changing the order of integration where
appropriate, we obtain the inequality
ω
def
δ(t) = ∆y (t, a) da
0
t


 ∆y 0 L1 ([0,ω]) + ∆ϕ L1 ([0,T ]) + d4 δ(s) +

∆f (s, ·)
L ds
1 ([0,ω])
0





+
∆g (s, ·, ·)
L ([0,ω]×[0,ω])
+
∆h (s, ·)
L ds.
1 1 ([0,ω])

According to Lemma 1, δ is a continuous function and one can apply the Grunwall in-
equality to obtain a constant C1 such that
ω
 
∆y (t, a) da  C1 ∆1 (u) + ∆1 (v, w) .
0

For each t ∈ [0, T ] we obtain the estimation for the first summand in (17). Then inte-
grating in t the above inequality and (19) and integrating in (t, a) the inequality (18), we
obtain (17).
Obviously
 
∆1 (u) + ∆1 (u, v)  C2 ∆∞ (u) + ∆∞ (v, w) .
Then from (18) and (19) we obtain (15). The substitution of this estimate in (20) gives
s
  
∆y (s, a0 + s)  ∆y 0 (a0 ) + M∆y (τ, a0 + τ ) + MC3 ∆∞ (u) + ∆1 (v, w)
0

+ ∆f (τ, a0 + τ ) dτ.
56 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

Since the function s → ∆y (s, a0 + s) is absolutely continuous we can apply the Grunwall
inequality to obtain
 
∆y (t, a0 + t)  C4 ∆y 0 (a0 ) + ∆∞ (u) + ∆1 (v, w)
for a.e. a0 ∈ [0, ω] and for a.e. t ∈ [0, ω − a0 ]. Similarly we obtain from (21) that
  
∆y (t0 + a, a)  C5 ∆ϕ (t0 ) + ∆∞ (u) + ∆1 (v, w)
for a.e. t0 ∈ [0, T ] and for a.e. a ∈ [0, min{ω, T − t0 }]. Combining these two inequalities
we obtain
 
∆y (t, a)  C6 ∆∞ (u) + ∆∞ (v, w) ,
which implies (16). ✷

4. The necessary optimality condition

The issue of existence of an optimal solution of the problem (6)–(12) (cf. [3]) is outside
the scope of this paper, therefore we suppose it.

Condition E. There exists an optimal solution (ŷ, p̂, q̂, û, v̂, ŵ) ∈ L∞ (D; Rm ) ×
L∞ (D; Rn ) × L∞ ([0, T ]; Rr ) × L∞ (D; U ) × L∞ ([0, T ]; V ) × L∞ ([0, ω]; W ).

Let us introduce the following

Notational convention: We abridge the notations by skipping those arguments of func-

def
tions that are fixed at a value with a “hat.” For example, f (t, a) = f (t, a, ŷ(t, a), p̂(t, a),
def
q̂(t), û(t, a)), while f (t, a, u) = f (t, a, ŷ(t, a), p̂(t, a), q̂(t), u), etc. Especially, by defi-
def
nition g(t, a, a  , u) = g(t, a, a  , ŷ(t, a  ), u), and similarly for g(t, a, a  ).
Below ∇z denotes differentiation with respect to the variable z. We introduce the follow-
ing adjoint system for the adjoint functions (ξ, η, ζ ), considered as row-vector functions
(while y, p and q are column-vectors) from L∞ (D; Rm ) × L∞ (D; Rn ) × L∞ ([0, T ]; Rr ):


∂ ∂
− + ξ(t, a) = ∇y L(t, a) + ξ(t, a)∇y f (t, a) + ζ(t)∇y h(t, a)
∂t ∂a
ω
+ η(t, a  )∇y g(t, a  , a) da ,
0
 
ξ(T , a) = ∇y l a, ŷ(T , a) , ξ(t, ω) = 0, (23)
η(t, a) = ∇p L(t, a) + ξ(t, a)∇p f (t, a) + ζ(t)∇p h(t, a), (24)
ω
 
ζ (t) = ξ(t, 0)∇q ϕ(t) + ∇q L(t, a) + ξ(t, a)∇q f (t, a) + ζ(t)∇q h(t, a) da. (25)
0
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 57

Remark 2. The notion of solution to the adjoint system in D is a particular case of Defin-
ition 1, where Γ consists of the right and upper bounds of Γ .

For the solution (ŷ, p̂, q̂), and the corresponding solution (ξ, η, ζ ) of the adjoint system
(we shall prove that it exists) we define the initial, boundary, and distributed Hamiltonians:
T
def
H0 (a, w) = ξ(0, a)y (a, w) +
0
L(s, a, w) ds,
0
ω
def
Hb (t, v) = ξ(t, 0)ϕ(t, v) + L(t, b, v) db,
0
ω
η(t, a  )g(t, a  , a, u) da 
def
H (t, a, u) = L(t, a, u) + ξ(t, a)f (t, a, u) +
0

+ ζ(t)h(t, a, u).

Theorem 1 (Pontryagin’s maximum principle). Under supposition E the adjoint system

(23)–(25) has a unique solution ξ, η, ζ and for a.e. t0 ∈ [0, T ], a0 ∈ [0, ω] and (t, a) ∈ D
∂H0   
a0 , ŵ(a0 ) w − ŵ(a0 )
0 ∀w ∈ W,
∂w
∂Hb   
t0 , v̂(t0 ) v − v̂(t0 )
0 ∀v ∈ V ,
∂v  
H (t, a, u) − H t, a, û(t, a)
0 ∀u ∈ U. (26)

Notice that the maximum principle is local with respect to the side controls and global
with respect to the distributed control. The reason is the discontinuity of the operator “side
control → solution” considered in the spaces L1 → L∞ (cf. (16)).

5. Proof of the maximum principle

Proposition 2. The adjoint system (23)–(25) has a unique solution in D.

def
Proof. For λ = (ξ, η, ζ ) we denote by F2 (λ) and F3 (λ) the right-hand sides of (24) and
(25), respectively. For γ = (γ1 , γ2 ) ∈ Γ and s ∈ S(γ ) = [−s(γ ), 0] we define
0 
def
F1 (λ)(γ + se) = ξ̄ (γ ) + ∇y L(γ + τ e) + ξ(γ + τ e)∇y f (γ + τ e)
s

+ ζ(γ1 + τ )∇y h(γ + τ e)

58 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

ω
  
+ η(γ1 + τ, a )∇y g(γ1 + τ, a , γ2 + τ ) da dτ,
0

where ξ̄ (γ ) is the side condition in (23) (here Γ consists of the upper and the right sides
of D). We shall prove existence of a fixed point of the operator λ → (F1 (λ), F2 (λ), F3 (λ))
by the Banach contraction mapping theorem. For this purpose we introduce the follow-
ing norms in the spaces L∞ (D; Rm ), L∞ (D; Rn ), L∞ ([0, T ]; Rr ), and in their product,
respectively:
   
ξ  = esup e−Mt ξ(t, x), η = esup dp e−Mt η(t, x),
def def
(t,x)∈D (t,x)∈D
−Mt 
 
ζ (t),
def def
ζ  = esup dq e λ = ξ  + η + ζ ,
t ∈[0,T ]

where M, dp and dq are constants that will be appropriately fixed later.

Since F is an affine operator, it is enough to prove that F (·) − F (0) < 1. We have

0 


 

(F1 (λ) − F1 (0)
 esup e−M(γ1 +s) ∇y f ∞ ξ(γ + τ e)
γ ∈Γ, s∈S(γ )
s
ω
    
+ ∇y h∞ ζ(γ1 + τ ) +   
∇y g∞ η(γ1 + τ, a ) da dτ
0

 esup Ce−M(γ1 +s)

γ ∈Γ, s∈S(γ )

0
 
× eM(γ1 +τ ) ξ  + dq−1 ζ  + ωdp−1 η dτ
s
 
1 1 ω
C ξ  + ζ  + η ,
M dq M dp M
where C majorizes all the coefficients ∇y f ∞ , . . . , ∇q h∞ in (23)–(25).
Similarly we obtain
 

F2 (λ) − F2 (0)
 C dp ξ  + dp ζ  .
dq
Finally, using Remark 1 and the structural condition for the function h we obtain

F3 (λ) − F3 (0)
 C1 dq ξ ,

where C1 = (1 + ω)C(1 + ωC + · · · + (ωC)q−1 ). It is easy to verify that one can chose the
constants dq , dp and M (in this order) in such a way that all the coefficients multiplying
the norms of ξ , η and ζ in the above estimations for F1 , F2 and F3 are less than 1/3. This
means F (·) − F (0) < 1 and implies the claim of the proposition. ✷
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 59

Lemma 2. Let y and ξ be solutions in D to the equations



∂ ∂ ∂ ∂
+ y(t, a) = F (t, a) and − + ξ(t, a) = Ξ (t, a),
∂t ∂a ∂t ∂a
respectively, where F, Ξ ∈ L∞ (D). Then the function
ω
def
ν(t) = ξ(t, a)y(t, a) da
0

is Lipschitz continuous and for a.e. t ∈ [0, T ]

ω
 
ν̇(t) = ξ(t, 0)y(t, 0) − ξ(t, ω)y(t, ω) + −Ξ (t, a)y(t, a) + ξ(t, a)F (t, a) da.
0

Proof. The Lipschitz continuity follows from Lemma 1 and the boundedness of y and ξ .
Changing the variable a = t − x we have
t
def
ν(t) = ξ(t, t − x)y(t, t − x) dx.
t −ω

Here one can differentiate in t, since for a.e. t and for a.e. x ∈ [t − ω, t] both y and ξ are
differentiable in the direction e. Differentiating and changing back the variable x we obtain
the claim. ✷

Now we proceed with the proof of Theorem 1. Below we shall use the notations intro-
duced in Proposition 1, with the difference that ∆y , ∆p , . . . , ∆ϕ are defined without taking
the norms.
The standing supposition ensures that the partial derivatives of l, L, f, g, h, ϕ are Lip-
schitz continuous in Z (this set is introduced in the last of the standing suppositions), uni-
formly with respect to (t, a) ∈ D (respectively (t, a, a ) ∈ D  and (u, v, w) ∈ U × V × W ).
From (8) and (9) we obtain
ω
 
∆p (t, a) = ∇y g(t, a, a  )∆y (t, a  ) + ∆g (t, a, a ) + ep (t, a, a ) da  , (27)
0
ω

∆q (t) = ∇y h(t, a)∆y (t, a) + ∇p h(t, a)∆p (t, a) + ∇q h(t, a)∆q (t)
0


+ ∆h (t, a) + eq (t, a) da, (28)

where the functions ep , eq can be estimated (almost everywhere) using Proposition 1 as

follows:
60 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

     
ep (t, a, a )  C ∆∞ (u) + ∆∞ (v, w) ∆y (t, a  ) + Θg (u; t, a, a ) ,
   
eq (t, a)  C ∆∞ (u) + ∆∞ (v, w)
      
× ∆y (t, a) + ∆p (t, a) + ∆q (t) + Θh (u; t, a) ,
where
   
Θg (u; t, a, a ) = ∇y g t, a, a  , u(t, a  ) − ∇y g(t, a, a  ),
   
Θh (u; t, a) = max ∇y h t, a, u(t, a) − ∇y h(t, a),
   
∇p h t, a, u(t, a) − ∇p h(t, a),
   
∇q h t, a, u(t, a) − ∇q h(t, a) .

Here and below C stays for a constant, independent of u, v, w. Similarly,

def  
δ(t, a) = L t, a, y(t, a), p(t, a), q(t), u(t, a), v(t), w(a) − L(t, a)
= ∆L (t, a) + ∇y L(t, a)∆y (t, a) + ∇p L(t, a)∆p (t, a) + ∇q L(t, a)∆q (t)
+ eL (t, a), (29)
def
where ∆L (t, a) = L(t, a, u(t, a), v(t), w(a)) − L(t, a). Also
def      
∆l (a) = l a, y(T , a) − l a, ŷ(T , a) = ∇y l a, ŷ(T , a) ∆y (T , a) + el (a),
where
   
eL (t, a)  C ∆∞ (u) + ∆∞ (v, w) + w − ŵL ([0,ω]) + v − v̂L ([0,T ])
∞ ∞
      
× ∆y (t, a) + ∆p (t, a) + ∆q (t) + ΘL (u; t, a) (30)
    
el (a)  C ∆∞ (u) + ∆∞ (v, w) ∆y (T , a), (31)
where ΘL (u; t, a) is defined similarly as Θh (u; t, a) but with the function L instead of h.
(For the final part of the proof it is crucial to keep track of all the remainders e and the way
they are integrated below.)
Utilizing Lemma 2 (and the notations there) we obtain
ω
 
ξ(T , a)y(T , a) − ξ(0, a)y(0, a) da
0
T
= ν(T ) − ν(0) = ν̇(t) dt
0
T 
= ξ(t, 0)y(t, 0) − ξ(t, ω)y(t, ω)
0
ω
 
+ −Ξ (t, a)y(t, a) + ξ(t, a)F (t, a) da dt.
0
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 61

Subtracting the same equality applied to ŷ and using that ξ(t, ω) = 0 we obtain
ω
def  
I = ξ(T , a)∆y (T , a) − ξ(0, a)∆y (0, a) da (32)
0
T  ω
 
= ξ(t, 0)∆y (t, 0) + −Ξ (t, a)∆y (t, a) + ξ(t, a)∆F (t, a) da dt, (33)
0 0
where
def  
∆F (t, a) = f t, a, y(t, a), p(t, a), q(t), u(t, a) − f (t, a)
= ∇y f (t, a)∆y (t, a) + ∇p f (t, a)∆p (t, a) + ∇q f (t, a)∆q (t)
+ ∆f (t, a) + ef (t, a)
and, similarly as above,
   
ef (t, a)  C ∆∞ (u) + ∆∞ (v, w)
      
× ∆y (t, a) + ∆p (t, a) + ∆q (t) + Θf (u; t, a) ,
where Θf (u; t, a) is defined similarly as Θh (u; t, a) but with the function f instead of h.
We represent also
   
∆y (t, 0) = ϕ t, q(t), v(t) − ϕ t, q̂(t), v̂(t) = ∇q ϕ(t)∆q (t) + ∆ϕ (t) + ey0 (t)
with
 0     
e (t)  C ∆∞ (u) + ∆∞ (v, w) + v(t) − v̂(t) ∆q (t).
y
From the adjoint equation for ζ the first term in the right-hand side of (33) can be expressed
as
ξ(t, 0)∆y (t, 0) = ξ(t, 0)∇q ϕ(t)∆q (t) + ξ(t, 0)∆ϕ (t) + ξ(t, 0)ey0 (t)
ω

= ζ (t)∆q (t) − ∇q L(t, a)∆q (t) + ξ(t, a)∇q f (t, a)∆q (t)
0

+ ζ (t)∇q h(t, a)∆q (t) da + ξ(t, 0)∆ϕ (t) + ξ(t, 0)ey0 (t).
Also we add to the right-hand side of (33) the quantity η(t, a)∆p (t, a) − (. . .)∆p (t, a) = 0,
where (. . .) is the expression for η(t, a) in the adjoint equation. Thus the right-hand side
of (33) takes the form
T  ω
 
I= ζ (t)∆q (t) − ∇q L(t, a)∆q (t) + ζ(t)∇q h(t, a)∆q (t) da
0 0

+ ξ(t, 0)ey0 (t) + ξ(t, 0)∆ϕ (t)

62 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

ω
+ −∇y L∆y (t, a) − ζ(t)∇y h∆y (t, a)
0
ω
− η(t, a  )∇y g(t, a  , a)∆y (t, a) da  + ξ ∆f (t, a) + ξ ef (t, a)
0
+ η∆p (t, a) − ∇p L∆p (t, a) − ζ(t)∇p h∆p (t, a) da dt. (34)

Now we multiply (27) by η(t, a) and (28) by ζ(t), and integrate on D and on [0, T ],
respectively. After changing the order of integration in the triple integral we add the so two
obtained equalities to (34) and obtain
T 
I= ξ(t, 0)∆ϕ (t) + ξ(t, 0)ey0 (t)
0
ω
+ −∇y L∆y (t, a) − ∇p L∆p (t, a) − ∇q L(t, a)∆q (t)
0
ω
+ ξ ∆f (t, a) + η(t, a  )∆g (t, a  , a) da  + ζ(t)∆h (t, a)

ω
0

  
+ ξ ef (t, a) + η(t, a )ep (t, a , a) da + ζ(t)eq (t, a) da dt. (35)
0

From the optimality of (û, v̂, ŵ, ŷ, p̂, q̂) and the definition of δ(t, a) in (29) we have
ω T ω
0 ∆l (a) da + δ(t, a) da dt.
0 0 0
Adding to this inequality the equality (35), subtracting (32), and taking into account the
end-time condition for ξ we obtain
ω T
0 ξ(0, a)∆y (0, a) da + ξ(t, 0)∆ϕ (t) dt
0 0
T ω ω
+ ∆L (t, a) + ξ(t, a)∆f (t, a) + η(t, a  )∆g (t, a  , a) da
0 0
0

+ ζ (t)∆h (t, a) da dt + e, (36)

where
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 63

ω T
def
e= el (a) da + ξ(t, 0)ey0 (t) dt
0 0
T ω ω
+ eL (t, a) + ξ(t, a)ef (t, a) + η(t, a  )ep (t, a  , a) da 
0 0
0

+ ζ (t)eq (t, a) da dt.

According to the estimations for each remainder and Proposition 1 we have

def  
|e|  ε = C ∆∞ (u) + ∆∞ (v, w) + w − ŵL∞ ([0,ω]) + v − v̂L∞ ([0,T ])
  
× ∆1 (u) + ∆1 (v, w) + meas (t, a) ∈ D: u(t, a) = û(t, a) . (37)
Expressing the summands in (36) in terms of the Hamiltonians we obtain
ω T
         
0 H0 a, w(a) − H0 a, ŵ(a) da + Hb t, v(t) − Hb t, v̂(t) dt
0 0
T ω
    
+ H t, a, u(t, a) − H t, a, û(t, a) da dt
0 0
T ω
    
+ L t, a, u(t, a), v(t), w(a) − L t, a, w(a)
0 0
   
− L t, a, u(t, a), v(t) − L(t, a) da dt

T ω
    
+ L t, a, u(t, a), v(t) − L t, a, v(t)
0 0
   
− L t, a, u(t, a) − L(t, a) da dt + ε. (38)

6. Construction of needle variations

Below we fix arbitrary representatives of all L∞ -functions involved. Let us prove the
first variational inequality in Theorem 1. Let a0 ∈ (0, ω) be a Lebesgue point of the func-
tions ∂H ∂H0
∂w (a, ŵ(a)) and ∂w (a, ŵ(a))ŵ(a). We fix an arbitrary w ∈ W , and define for
0

h ∈ (0, 1) (small enough, so that the interval [a0 − h, a0 + h] is contained in [0, ω])
def  
w(a) = ŵ(a) + h w − ŵ(a) χ[a0 −h,a0 +h] (a),
64 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

where χ[a1 ,a2 ] denotes the characteristic function of [a1 , a2 ]. Then w(·) is an admissible
control, since W is convex. We take also u = û and v = v̂. Obviously for some constant d
the nonzero terms in (37) can be estimated as
∆∞ (v, w)  dh, w − ŵL∞ ([0,ω])  dh, ∆1 (v, w)  dh2 .
Then we obtain
ω
    
0 H0 a, w(a) − H0 a, ŵ(a) da + Cd 2 h3 .
0
Using the particular form of w(·) and the Lipschitz differentiability of H0 in w we obtain
a0 +h
1 ∂H0   
0 a, ŵ(a) w − ŵ(a) da + O(h),
h ∂w
a0 −h

which gives the desired inequality, since a0 is a Lebesgue point of the integrand.
The proof of the second inequality in Theorem 1 is completely analogous.
The last inequality will be proven by contradiction. Assume that
   
H t, a, û(t, a) > inf H (t, a, u): u ∈ U
on a set of positive measure in D. Since H is continuous in u and U is compact, by a
standard measure theoretic argument, there exist ū ∈ U , δ > 0 and a set Ω ⊂ D of positive
measure, such that
 
H (t, a, ū) − H t, a, û(t, a)  −δ (39)
for every (t, a) ∈ Ω. Almost every point in the interior of D is a Lebesgue point of both
H (t, a, ū) and H (t, a, û(t, a)), therefore there exists such belonging to Ω. Denote it by r =
(t¯, ā), and let R(r, h) be the coordinate box of size h × h centered at r. For all sufficiently
small h > 0 we have R(r, h) ⊂ D. Define the controls
u(t, a) = χR(r,h) ū + (1 − χR(r,h) )û(t, a), v(t) = v̂(t), w(a) = ŵ(a).
Obviously for an appropriate constants C1 and C2 we have for the nonzero terms in (37)
∆∞ (u)  C1 h, ∆1 (u)  C2 h2 ,
 
meas (t, a) ∈ D: u(t, a) = û(t, a)  h2 .
Then (38) is fulfilled for all sufficiently small h and we obtain

1   
0 2 H (t, a, ū) − H t, a, û(t, a) da dt + CC1 C2 h,
h
R(r,h)

which contradicts (39) since r is a Lebesgue point of both H (t, a, ū) and H (t, a, û(t, a)).
The proof of Theorem 1 is complete.

 ω of a nondistributed control u(t, a) = u(t) or in pres-

Remark 3. We mention that in case
ence of “budgetary” constraints 0 u(t, a) da  B for u one has to use different classes
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 65

of needle variations, but (38) can still be applied. The corresponding maximum principle
in the first case is straightforward, while the case of budgetary constraints require some
additional work.

7. Some applications

In this section we apply the maximum principle to the two “stylized” models discussed
in the introduction. For the first model we obtain some qualitative features of the optimal
control, while for the second model we outline an analytic representation of the optimal
controls. More detailed analysis and numerical implementations based on the maximum
principle for models in the social area (drug control and investment models) will be pre-
sented in [1,13,18] and elsewhere.
(1) First we consider the model (1), (2), (3). The adjoint system from Theorem 1 reads
now
ω
 

−(ξt + ξa ) = d(a) − ξ(t, a)u(t, a)Θ y(t, a) + η(t, a  )m(a  , a) da ,
0
 
ξ(T , a) = 0, ξ(t, ω) = 0, η(t, a) = ξ(t, a)S(t, a)Φ  p(t, a) ,

and the minimization condition of the Hamiltonian gives c (u) = ξ(t, a)Θ(y(t, a)). Further
we simplify the consideration taking Θ(y) = 1. Moreover, the following natural conditions
will be supposed: d(a), m(a, a  ), S(t, a) are nonnegative and twice differentiable, Φ(·) is
monotone increasing, c(·) is strictly convex, d(·) and m(a, ·) are concave. Then the adjoint
equation becomes
ω
 
−(ξt + ξa ) = d(a) + ξ(t, b)S(t, b)Φ  p(t, b) m(b, a) db.
0

Since c (·) is strictly monotone increasing and differentiable, the following properties of
the optimal control can be deduced from the same properties of ξ(t, a):

(i) for every fixed t the optimal distribution u(t, ·) is a concave differentiable function; in
particular it has a single peak a(t) ∈ [0, ω];
(ii) the magnitude u(t, a(t)) of the peak decreases with t.

It is not straightforward that ξ has the above two properties. The proof is technical,
therefore we just sketch the main points. First, from the adjoint equation and the boundary
conditions one can prove that ξ satisfies the equation

ω ω
 
ξt + ξa = ξa (t, ω) − ξt a (t, ω) + ξaa (t, ω) (ω − a) − β(t, b) db da ,
a a
66 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

where
ω

 
β(t, a) = d (a) + ξ(t, b)S(t, b)Φ  p(t, b) m2 (b, a) db
0

(m2 denotes the second derivative with respect to the second argument of m). The suppo-
sitions imply β(t, a)  0. The above differential equation has the explicit solution
min{t +ω−a,T
 }
 
−ξa (s, ω) + ξt a (s, ω) + ξaa (s, ω) (ω − a + t − s)
t
ω ω
  
+ β(t, a ) da da ds. (40)
a−t +s a

From here one can calculate all derivatives of ξ that are of interest. In particular
 T
β(s, a − t + s) for a ∈ [0, ω + t − T ],
ξaa (t, a) = t  t +ω−a
ξaa (t + ω − a, ω) + t β(s, a − t + s) for a ∈ (ω + t − T , ω].
It is easy to check that the above function is continuous (as well as ξa ). Since the values of
β are nonpositive, to prove concavity of ξ it remains to show that ξaa (t + ω − a, ω)  0.
This can be done making use of the boundary condition and the inequality ξa (t, ω)  0,
which also follows from the boundary condition ξ(t, ω) = 0 together with ξ(t, a)  0.
The second conclusion (ii) follows from
d            
ξ t, a(t) = ξt t, a(t) + ξa t, a(t) ȧ(t) = ξt t, a(t) = ξt t, a(t) + ξa t, a(t)
dt
 0.
Using the representation (40) one can obtain also other qualitative properties of ξ , there-
fore also of the optimal control.
(2) Now let us apply Theorem 1 to the vintage capital model described in the introduc-
tion. Here we have the specifications y = K, q = (Q, I ), p and w are missing,
 
−rt 1
   
L(t, a, K, Q, I, u, v) = −e R(Q) − α0 v − βv − γi I − α(a)u + γe u ,
2 2 2
ω
l = 0, f (t, a, K, u) = −δ(a)K + u, h1 (t, a, K) = m(t − a)d(a)K,
h2 (u) = u, ϕ(v) = v.
The adjoint system becomes

−(ξt + ξa ) = −ξ(t, a)δ(a) + ζ1 (t)m(t − a)d(a),

 
ξ(T , a) = 0, ξ(t, ω) = 0, ζ1 (t) = −e−rt R  Q(t) , ζ2 (t) = 2e−rt γi I (t).
Maximizing the Hamiltonians corresponding to u and v and substituting ζ2 from the above
equation we obtain
 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68 67

1  rt 
u(t, a) = − e ξ(t, a) + 2γi I (t) + α(a) , (41)
2γe (a)
1  rt 
v(t) = − e ξ(t, 0) + α0 . (42)
2β
Integrating (41) on [0, ω] one can express I (t) as
ω
1  rt 
I (t) = − e ξ(t, a) + α(a) da. (43)
2(γe + ωγi )
0

Substituting λ = −ert ξ ,
the equation for ξ becomes
   
λt + λa = λ(t, a) δ(a) + r − R  Q(t) m(t − a)d(a),
λ(T , a) = 0, λ(t, ω) = 0.
A routine calculation (we suppose T
ω) gives
min{t +ω−a,T
 }
 a−t+s  
λ(t, a) = m(t − a) e− a (δ(θ)+r) dθ
R  Q(s) d(a − t + s) ds. (44)
t

Notice that the solution λ, therefore also ξ , is negative. The quantity −λ(t, a) represents
the current shadow price at time t of a unit of capital of age a.
Eqs. (44), (43), (41) and (42) express in an analytic form the optimal control in terms
of R  (Q(t)). If R is linear (as in the case of nonmonopolistic economy) then R  (Q(t)) is
a known constant and the problem has an explicit analytic solution. In the monopolistic
case with quadratic R, Eq. (44) is linear in Q(·), therefore the optimal controls are linear
functionals of Q(·). In this case Q(t) can be shown to satisfies an appropriate Fredholm
equation.

References

[1] Ch. Almeder, J.P. Caulkins, G. Feichtinger, G. Tragler, An age-structured single-state drug initiation
model—cycles of drug epidemics and optimal prevention programs, Socio-Economic Planning Sciences,
in presss.
[2] S. Anita, M. Iannelli, M.Y. Kim, E.J. Park, Optimal harvesting for periodic age-dependent population dy-
namics, SIAM J. Appl. Anal. 58 (1998) 1648–1666.
[3] V. Barbu, M. Iannelli, Optimal control of population dynamics, J. Optim. Theory Appl. 102 (1999) 1–14.
[4] E. Barucci, F. Gozzi, Investment in a vintage capital model, Res. Econ. 52 (1998) 159–188.
[5] E. Barucci, F. Gozzi, Technology adoption and accumulation in a vintage capital model, J. Econ. 74 (2001)
1–38.
[6] M. Brokate, Pontryagin’s principle for control problems in age-dependent population dynamics, J. Math.
Biol. 23 (1985) 75–101.
[7] W.L. Chan, B.Z. Guo, Optimal birth control of population dynamics, J. Math. Anal. Appl. 144 (1998) 532–
552.
[8] W.L. Chan, B.Z. Guo, Optimal birth control of population dynamics, II: Problems with final time, phase
constraints, and minimax costs, J. Math. Anal. Appl. 146 (1990) 523–539.
[9] J.M. Cushing, The dynamics of hierarchical age-structured populations, J. Math. Biol. 32 (1994) 705–729.
68 G. Feichtinger et al. / J. Math. Anal. Appl. 288 (2003) 47–68

[10] J.M. Cushing, An Introduction to Structured Population Dynamics, in: CBMS-NSF Regional Conf. Ser. in
Appl. Math., Vol. 71, SIAM, Philadelphia, 1998.
[11] N. Derzko, S.P. Sethi, G.L. Thompson, Distributed parameter systems approach to the optimal cattle ranch-
ing problem, Optimal Control Appl. Methods 1 (1980) 3–10.
[12] H.O. Fattorini, A unified theory of necessary conditions for nonlinear and nonconvex control systems, Appl.
Math. Optim. 15 (1987) 141–185.
[13] G. Feichtinger, R.F. Hartl, P.M. Kort, V.M. Veliov, Anticipation effects of technological progress on capital
accumulation: a vintage capital approach, in preparation.
[14] D. Greenhalgh, Some results on optimal control applied to epidemics, Math. Biosci. 88 (1988) 125–158.
[15] M.E. Gurtin, R.C. MacCamy, Nonlinear age-dependent population dynamics, Arch. Rational Mech. Anal. 54
(1974) 281–300.
[16] M.E. Gurtin, L.F. Murphy, On the optimal harvesting of age-structured populations: some simple models,
J. Math. Biosci. 55 (1981) 115–136.
[17] M.E. Gurtin, L.F. Murphy, On the optimal harvesting of persistent age-structured populations, J. Math.
Biol. 13 (1981) 131–148.
[18] R.F. Hartl, P.M. Kort, V.M. Veliov, G. Feichtinger, Capital accumulation under technological progress and
learning: a vintage capital approach, in preparation.
[19] A. Haurie, S. Sethi, R. Hartl, Optimal control of an age-structured population model with applications to
social services planning, Large Scale Systems 6 (1984) 133–158.
[20] H.W. Hethcote, Optimal ages of vaccination for measles, Math. Biosci. 89 (1988) 29–52.
[21] N.G. Medhin, Optimal harvesting in age-structured populations, J. Optim. Theory Appl. 74 (1992) 413–423.
[22] J. Müller, Optimal vaccination patterns in age-structured populations, SIAM J. Appl. Math. 59 (1999) 222–
241.
[23] J. Müller, Optimal vaccination patterns in age-structured populations: endemic case, Math. Comput.
Model. 31 (2000) 149–260.
[24] L.F. Murphy, S.J. Smith, Optimal harvesting of an age-structured population, J. Math. Biol. 29 (1990) 77–90.
[25] G.F. Webb, Theory of Nonlinear Age-dependent Population Dynamics, Dekker, New York, 1985.
[26] A. Xepapadeas, A. de Zeeuw, Environmental policy and competitiveness: the Porter Hypotheses and the
composition of capital, J. Environ. Econ. Manage. 37 (1999) 165–182.