Control and Cybernetics

vol. 30 (2001) No. 3

Existence in optimal control problems of certain

Fredholm integral equations
by

Tomas Roubicek 1 •2 and Werner H. Schmidt 3

1
Mathematical Institute, Charles University, Sokolovskci 83
CZ-186 75 Praha 8, Czech Republic

2
Institute of Information Theory and Automation , Academy of Sciences
Pod vodarenskou vezf 4, CZ-182 08 Praha 8, Czech Republic

3 Institut
fiir Mathematik und Informatik
Ernst-Moritz-Arndt-U niversi tat G reifswald
Friedrich-Ludwig-Jahn-Strai3e 15a, D-17 487 Greifswald, Germany

Abstract: Existence of an optimal control for certain systems

governed by nonlinear Fredholm integral equations that arc of a
Hammerstein type with respect to the control is proved under con-
vexity of t he orientor field by using a relaxed problem. Then, this
convexity assumption is put away by finer analysis of the maximum
principle. Illustrative examples are presented.
Keywords: optimal control, integral equation , existence, relax-
ation , orientor field , Young measures.

1. Problem formulation, first existence results

Existence theory in optimal control, which started from classical works by Filip-
pov (1959) and Roxin (1962), has been successfully applied to various problems
governed by ordinary differential eq uations, see Cesari (1983) or Mordukhovich
(1977) for a survey. This theory is essentially based on convexity of the sets
of admissible velocities, so-called orientor fields . Such convexity is often neces-
sary, as demonstrated by Brunovsky (1968) and, as Bittner (1994) remarked , it
seemed to be the reason why no direct gc nerali ~ation to integral processes was
available. Actually, Schmeling ( 1979, 1981) gave a sophisticated example for
This research has been partly covered by Deutsche Forschungsgemeinschaft through SFD
255 and from the C;;cch Republic through the grants 201/00/07G8 (GA CR), A 107 5707,
A 107 5005 (GA AV CR), and MSM 11320007 (MSMT Crt) . The authors are thankfu l to
304 T. ROUBiCEK, W.H. SCHMIDT

nonexistence of optimal solutions for a control pro ce~Js with Volterra equation
J;
of the type y(t) = Yo(t) + j(t,r,y(r),u(r))dr , u(t) E S, with the orientor
field f(t, r, r, S) always convex and compact, see (5.2) below, though for some
rather special problems of this kind the existence can be proved, see Schmeling
(1981) or Roubfcek and Schmidt (1997).
Nevertheless, the Filippov- Roxin t heory was extended for the nonlinear in-
tegral equations of the Fredholm (sometimes also called Urysohn) type that are
of Hammerstein type with respect to control by Balder (1993), Bennati (1979),
Cowles (1973) and, for less general equations, also by Zolezzi (1972). For the
Volterra-type equations, even some more references exist; see Remark 2 below.
Besides, the Fredholm case was also addressed by the authors, Roubfcek and
Schmidt (1997), but in completely different situations, via the Bauer's extremal
principle.
In this paper, we want to present the Filippov- Roxin theory for this special
class of Fredholm equations that are of Hammerstein typ e with respect to con-
trol, similarly as already done in Balder (1993), Bcnnati (1979) , but by using
an auxiliary relaxation by Ro ubfcek (1998) similarly as proposed in Mniio~ and
Pedregal (2001) for problems governed by ordinary differential equations and
in Roubfcek (1999) for general situations. This provides a deeper insight and
enables us to refine the existence results to cover also problems with nonconvex
orientor fi eld like it was done in Gabasov and Mordukhovich (1974) , Ioffe and
Tikhomirov (1974), Mordukhovich (1988, 1999), Munoz and Pedregal (2001) for
problems governed by ordinary differential equations.
To pursue this goal, we will treat the following isop crimetrically constrained
optimal control problem:

Minimize k <p(x, y(x), u(x)) dx (a cost fun ctional)

subject to y(x) = Yo(x) + k K(x, Cy(0)f(~, y(0, u(0)d~,

(state equation) (P)
k19( x, y(x), u(x)) d.r :S 0, (state/control constrai nts)
u(x) E S(x) for a.e. X En, (control constraints)
y E Lq(O; JR"), u E LP(D; JRm),

where n is a subset in a Euclidean space with a finite Lebesgue measure and the
functions K: 0, X 0, XJR" -+ JR" xl, j: 0, XJR" XJRm -+ JR1, tp: 0, X JR" X]Rm-+ JR,
19 : n x JR" x JRm -+ JRk with JRk ordered by a closed convex cone (to give a sense
to "{) :S 0" ), Yo : D -+ JR" and the multivalued mapping S : 0, ::4 lRm will be
subjected to the following basic data qualification:

<p, K, j, 19 are Caratheodory mappings, i.e.

measurable in x, ~ E n and continuous in the other variables, (1.1a)
Optimal control of Fredholm integral eq uations 305

l<p(x, 1·1, s) - <p(x,r2,s) l

~ (aqf(q-1)(x) + bhlq- 1 + bhlq- 1 + clslp(q- 1)/q)lr1- r2l, (l.lc)
IJ(x, r, s) l ~ a,(x) + clrlqfr + cls iPh, (l.ld)
IJ(x, r1, s)- f(x, r2, s) l
~ (an/(q- 1 )(x) +blrdq-,)h +blr2l(q-rJh + clslp (q - ,)/(nl )lr1 -r2 l, (l.le)

IK(x,~,r-) 1 ~ a~/( 1 - 1)(0 + cx lr-lq(r- 1Jh, (l.lf)

J
I[K J]( x, ~ ' r, s) l ~ L aq ,j(x)a1 ,j(0 + aq(x)( lr lq-£ +l s i ~'-£ ), (l.lg)
j == 1
I[Kj](x,~ ,r- 1, s)- [Kf](x,~,rz,s) l ~ R(x,O ir-1- r-21 , (l.lh)
119(.7:, T, s) l ~ a1(x) + c( lrlq + ls iP- £), (l.li)
l19(x, r1, s) -19(x, r-2, s) l
~ (aqf(q-l)(x) + blrtl q- l + bhl'~ - 1 + cls lp (q - 1 )/q) lr-t- r2l, (1.1j)
s: n :::::1 ~m is measurable and S(x) is closed for a.a ..TEn, (1.1k)
Yo E Lq(n ; ~n) , (1.1 1)

where n , m.,l , k , J EN and p E [1,+oo), q, / E (l , + oo), q 2 /, b,c E ~'

£ E U(f2; Lqf(q-t)(f2)), c > 0, and a<> E L<> (f2), i. e. the lower indices in a's
indicate integrability. The assumptions (1. 1a,b,g,i) ensure existence of all three
integrals appearing in (P) and together with (1.1 c-f, h) make possible to do a
correct relaxation by using Roubicek (1998) and also to rewrite th e relaxed
problems in terms of Young measures.
We will confine ourselves to the case when, for every admissible control 7L, the
existence and uniqueness of the solution y = y( u) to t he involved integral equa-
tion is ensured by the Banach contraction principle, which can be guaranteed
by the assumption

(1.2)
with£ referring to (1.1h), though other principles which are standard in integral-
equation theory can be used, too. Of course, we need also a feasibility and
coercivity of the whole problem (P). For simplicity, we can assume

:J(u, y) E LP(n; ~"') x L 9 (f2; ~

11
) : i 19(x, y( x), u(x)) d1; ~ 0, u(x) E S(x) ,

y(x) = Yo( x) + i K(x,~, y(O)J(Cy(~),n(O)d~ a.e. , (1.3a)

<p(x, r-, s) 2 c ls iP for some c > 0. (1.3b)

It turns out that the relevant orientor field has the form
. {( qo,qt,qz ) E
Q( x,T ) .= ml+l+k.
IN> ,
306 T. ROUBICEK, W.H. SCHMIDT

for x En andrE .!Rn , see also Balder (1993), Bennati (1979), Zolezzi (1972) .
From the assumptions (1.1a) and (1.3b) it follows that Q(x, r) is closed for a. a. x
and for all r. A solution u to (P) will be called stable if any sequence of controls
converging to some optimal control is minimizing asymptotically feasible, i.e.

lim lluk- uiiLP(f!;IR") = 0 => lim

k-+oo k-+oo
I } r!1 '!9(x,yk(x),uk(x))d.1: ! = 0

& lim sup { <p(x, Yk(x) , ·u k(x)) dx ~ inf(P) , (1.5)

k-+oo Jo
where Yk is a response to the control Uk, i.e. the solution to yk(x) = yo(x) +
fo K(x, ~' Yk(~))f(~, yk(0, V-k(0) d~.
Now we can formulate the generalization of the Filippov and Roxin theorem,
which follows directly from Propositions 1- 3 given further on:

THEOREM 1 Let (1.1)-(L'3) hold and let

Q(x,r) be convex (1.6)

for a. a. X E n and all r E R" . Then (P) has an optimal solntion. Moreover,
any solntion to (P) is stable in the sense (1.5).

REMARK 1 The above statement is similar to the existence theorem. by Balder

(1984, 1993) and Bennati (1979) wheTe, howeveT, no isoperimetTic ineqnalities
constmints aTe consideTed (in fact, Balder·, 1984 and Bennati, 1979 consideT th e
state constmint of the type y(x) E A not involving e.r,plicitly the contml) and,
becanse of weakeT assnmptions used in BaldeT (1984, 1993) and Bennati (1979) ,
no stability is obtained. Yet, having not guamnteed that a perturbed control is
f.-optima l in a snitable sense seems to have a little pmctical nsage, although the
existence of a (possibly 1mstable) optimal contml is tJworetically interesting.
Let us also mention that Cesari's Q-type property, i.e. Q(x , r)
n o>O cl CO Ullr-,'11~ 6 Q(x, r), is valid because of ( 1. 6) and continuity of the data
with respect tor-variable as assumed, in padicular, in (1.1a).

REMARK 2 This convexity imposed on Q(x, r) is trivially satisfied if <p(x, r, ·)

is convex, f(x,r,·) is linear·, '!9(x,r,·) is convex (with respect to the or·dering
of .IRk), and S(x) is conve.r, for· a. a. x E n and all r E .IRn . Th en, however,
the standard diTect method based on weak compactness can be applied. In s·uch
linear/convex case, other classes of integml equations can also be handled, see
Pelczewsl.:i (1989) , Yusifov and Kamgezov (1990) , etc.

REMARK 3 (Volterra integral equations) The special case (of th e Volt er'Ta type)
of our integml equations has been already investigated by Angel (1976), Balder
(1993), Gadson (1990} and Yeh (1978}; they used n := [0, T], l := 2, K(t, T, y)
Optimal control of Fredholm integra l equations 307

constmints, and ensured existence by assuming convexity of an 01-ientor· field

like (1.4) . For some special Volterm case see also Schmeling {1981}.
Moreover, for the even more special case of n = l and K(t, r , y) = identity
matrix for t ;::: T otherwise J( = 0, we can co ver basically the original results by
Filippov ( 1959) and Roxin {1962) b1tt here with isopcrimctrical state constmints;
in this case the existence of a unique response of the controlled system is cnsurwl
without (1.2).
REMARK 4 (Semilinear elliptic equations) For the special case of n being a
sm ooth domain, J( = K( x, 0 the Green fun ction of th e Laplace operator -ll
on D, and Yo a harmonic function , i.e . llyo = 0 on D, our integral equation
represents the Dirichlet b01mdary value problem fo r the scmilincar equation
-fly(x) = j(x,y( x),u(x )) fo r a.a. X E fl, (1.7a)
y(x) = Yo(x) for a.a. :r E Dn (th e boundary of D), (1. 7b )
as pointed out already by Pachpatte (1981), sec also Stua1't {1914} fo r the Neu-
mann boundary conditions. Our assumption ( 1. 2) then bo·unds Lipschitz con-
tinuity of J (x, ·, s) in c01-rclation with the integrability of Green's function J(
depending on the dim ension of n, and Th eorem 1 gives then existence of an
optimal control of {1 . 7) under· integml state/control constraints. For such sort
of results we refer also to Papageorgi01t {1991a, 1991b} and a one-dimensional
illustrative Example 5.3 fu rther on.

2. Auxiliary relaxed problem

In this section we will construct and analyze a continuously extended (so-called
relaxed) problem. Essentially, we can usc directly the results from Roub!Cek
(1998) modifi ed to our special case or, if K(:r,~, ·) is constant like in Re-
mark 3, we can even directly usc Roubfcek (1997), Section 4.6. We first con-
struct a suitable convex locally (sequentially) compact envelope of the Lebesgue
space LP(D; !Rm) of controls. To do this, we ta ke a sui table linear space of
Caratheodory integrands containing all possible nonlinearitics, e.g. one can con-
sider
H := span {rp o y + a · (J o y) + {) i o y + hs;
a E p/h-ll(D ; IR 1) , y E L'~(D; !Rn), j = 1, ... , k} , (2.1)
where J o y : nxiR"' __... IR 1 is defined by [foy]( x,s) := f(x , y(x),s) a nd si milarly
also rp 0 y : n X !Rm __... lR and 19 0 y : n X !Rm __... IR"' a nd moreover
hs(x, s) := max(1 , min
uES(x)
10'- si) . (2.2)

Note that hs is a Caratheodory function because S is measurable, see (1.1k).

In view of (1.1b,d,i) , it is nat ural to equip H with
308 T. IlOUBICEK, W.H. SCHMIDT

which is a norm, see Roubfcek (1997), Example 3.4.13. Since 1 > 1 and q < +oo,
both p/(1'-l)(fl.) and U(fl.) are separable, and t hus also H is separable if
1
equipped with the norm (2.3) . For a ranging C(fl.; ffi. ), this separability was
shown in Roubfcek (1999), Lemma 1. Here however, instead of the obvious
inequality \\a· h\\H ~ 1\a\\L"' (fl) \\h\\H , we must use \\a· h\\H ~ \\a\!l()i~!1l(fl) +
1\\h\'\\H, which allows us to estimate
1h- l )
\\a·(! o y)- a· (f o y) \\H -< b1\ llf o vi' IIH + c6, \ a- all'£1/(>-l)(fl;u-..) TlJ)I

+ 62\ l all;!-,~~~~;l (n;ffi.l > + Co2 III (J o y) - (f o y)l' 1\H

(2.4)

with b1, b2 > 0 arbitrarily small and C 0 ,, C6 2 depending on b's . Then, for a given
c > 0, a E L'/(1-ll(fl.; ffi. 1) andy E U(fl.; ffi.n) , one can take b1 small enough so
a
that It ~ c I 4, then take from a (chosen fixed) dense countable subset close
enough to a so that h ~ c I 4, further take b2 sma ll enough so that h ~ c I 4,
and finally y from a (chosen fixed) dense countable s ubset close enough toy so
that 14 ~ c I 4 by using (1.1e). This proves the separa bility of the set {a· (f o y );
a E p/(-r-l)(fl.;ffi.1), y E U(fl.;ffi.")}. From this, the separability of the whole
space H follows easily by using also (1.1c,j).
Furthermore, we define a (norm,weak*)-continuous (possibly not injective)
embedding i: LP(fl.; ffi.m) --t H* by

(i(u) , h) := fnh(x,u(x))dx. (2.5)

Of course, if h is valued in ffi." (or in ffi."' ), then (i(u), h) is defined by (2.5)

component-wise. Let us a bbreviate also Kx(~,r) := K(x,~,r). Then if Kxf:
fl. X ffi.n X ffi.m --t ffi.n and (Kx f) o y defined by [(K xf) o y];(~, s) ·-
L~=l K;j(X, ~ ' y(0)1J(~, y(O , s) belongs to Hn due to (2.1) with (l.lf). We
can rewrite the original problem (P) in the equivalent form

Minimize (i(u),cpoy) }
subject to y(x) = y0 (x) + (i(u), (Kx f) o y) for a .a. x E 0., (P')
(i(u), {) o y) ~ 0, (i(u), hs) = 0,
y E Lq(fl.; ffi.")) , u E LP(fl.; ffi.m).

Note that, as S(x) is closed (see (1.1 k)), we have hs(x, s) > 0 for s E ffi.m \ S(x)
while hs(x, s) = 0 for s E S(x), and therefore (i(u), hs) = 0 is indeed equivalent
to u(x) E S(x) for a .a. x E 0. .
Furthermore, we define the set of the so-called generalized Young functionals
by Y);(0.; ffi.m) := w* -cl-i(LP(fl.; ffi.m)). It is known from Roubfcek (1997) that , as
" rrm<:PnllPnrP nf (? 1) wit.h (1 .::\l . Y~ (fl.: ~rn) is a convex locallv (seauentiallv)
Optimal co nt rol of Fredholm integral equations 309

us remind that cp o y E H , (K"' f) o y E H" (for a. a. x E rl), 19 o y E Hk, and

hs E H because of the choice (2.1) and (l.lf). Then, in view of ( P'), we can
define t he relaxed problem as follows:

Minimize (17, cp o y) }
subject to y.(x) = y 0 (x) + (17, (.K "' f) o y) for a.a. x E rl,
(RP)
(17, 19 o y) ::; 0, (ry, hs) = 0,
y E Lq(rl; JRn)), 17 E Y};(H; lRm) .

PROPOSITION 1 Let {1.1} - {1.3} hold. Then the relaxed problem (RP ) always
has a solution 17 E Y};(rl ;lRm). Moreove1·, min( RP )::; inf(P).
Proof. It just follows from Roubfcek (1998), Proposition 4.1 , together with Re-
mark 3.2 . Note t hat it can be explicit ly seen from Roubfcek (1998), Remark 3.3
that Ro ubfcek (1998), Condition (10) is ind eed fulfilled. •
Let L';'(rl ; rca(JR"')) denote the space of weakly measurable essentially
bounded fun ctions on rl with values in t he space of Borel measures on JRm ,
and rcai(JRm) the set of all probability measures on JR'" . Further, let us define
the set of t he so-called LP- Young measures as

YP(fl; JR"') := { (x f---+ v,.) E L';'(rl; rca(JRm));

Vx E rcai(JRm) for a.a. X Erland /' /' js jP IJx(ds)d:r < + oo } . (2.6)

.JI! .JIR"'
The natural imbedding LP(rl; JRm) ---+ Y~'(rl; JR"') is t hen defined by u f---+ v with
V.c = 01,(.-c) • where Os denotes the Dirac measure supported at s E JR"'.
We will also need a relaxed problem written in terms of D ' -Yonng measures,
which looks as follows:

Minimize { { cp(x,y(x) ,s) vc(ds)dx

./nJ.R"'
s ubj ect to y(:r) = Yo(x)

+ l K(x,~, y(0) (.~"' f(~, v(O , s) v€(ds)) d~, ( RP' )

r /'
.JI! ./'Rm
19(x, y(x), s) l!x(cls) dx ::; 0,
supp(vx) C S(x) for a.e. x E rl ,
y E Lq(rl; JR"), v E Y~'(rl; JR"') .

We call 17 E Y};(rl ; JRm) p-nonconcentrating if it is attainable by a sequence

{1lk}kEN C LP(rl; JR"') (i.e. i(1Lk:}---+ 1] weakly* in H *) s uch t.hat {iukj~' ; kEN}
is relatively weakly compact in U(rl). Since His separable, any such 11 possesses
at least one D'-Young measure representation u E Y 1'(rl; JRm) in the sense that
for all hE H:
rr
310 T. ROUBICEK, W.H. SCHMIDT

and, conversely, any v E YP(r2; IR"') determines by t he formula (2.7) some p-

nonconcentrating 17 E Y};(n; !Rm) , see Roubicek (1997) , Proposition 3.4 .15.

PROPOSITION 2 Under the assumptions of Proposition 1, any solution 7) E

Y};(r2; !Rm) of (RP) is p-nonconcentmting and any LP- Young-meas1tre represen-
tation 11 E YP(r2 ;!Rm) of such 7) solves (RP') . Moreo ver, min(RP') = min(RP).

Proof. The p-nonconcentration of any 7) solving (RP ) follows from Roubicek

(1998), Propositions 4.2 with 4.1. Thus we do not change the set of minimizers
by restricting (RP) only to p-nonconcentrating functionals in Y};(r2; !Rm).
By using LP- Young measure representation of t hese functionals , we can
equivalently write this modified relaxed problem in the form:

Minimize r r rp(.r, y(x), s) Vx(ds) dx:

.Jn .JJR"'
subject to y(x) = Yo(x)
+ r r [Kx f](~ , y(0 , s)vdds)d~ for a .a.
.Jl! .fiR"'
X En,
(RP")
{ { iJ(x, y(x), s) Vv(ds) dx:::; 0,
./n }Rm
{ { hs(x , s)r;x (ds)dx=O ,
.Jl! .fiR"'
y E Lq(n; IRn), v E YP(r2; IR"').

In particular, min(RP") = min(RP).

As K(x , ~,v(O) is simply a constant in terms of s for a .a. ~En , we get

.In L"' [Kx f](~, y(~), s) v~(ds) d~

= f f K"'(~, y(~) ) f(~, y(O, s) vE(ds) d~
.Jl! .f.R"'
= l K(x,~,y(0)(.~., f(~,y(0,s)vdds)) d~.
Also, j11 frr~"' hs(x, s) vx(d s) dx = 0 is equivalent to the condition supp(v,,) C
S(x) (for a.e. x: E r2) beca use it holds hs(:r, s) > 0 for s E !Rm \ S(:r) while
hs(x , s) = 0 for s E S(x).
Altogether, we can see that (RP" ) is equivalent to (RP'). •

3. Proof of Theorem 1
Now we are ready to give a quite simple proof of Theorem 1, following Ro ubfcek
(1999), Lemma 2. Let us remark that the optimal Young-measure solution
and the measurable-selection technique has already been used in Munoz and
Optimal contro l of Fredholm integral equations 311

but in a bit different arrangement relying on the maximum principle. Inspired by

Munoz and Pedregal (2001), it will be more suitable for the proof and especially
for further modification in Section 4 to reformulate the convexity condition (1.6)
into the form

co[cp x f x 19](x, r, S(x)) C Q(x, r). (3.1)

Indeed, (3.1) implies (1.6) because, takingq 1 ,q 2 E Q(x,r), one has s 1 ,s 2 E S(x)
such that qt 2: cp(x,r,si), q~ = f(1:,r,si), and q1 2: cp(x,T,si) for 'i = 1, 2,
and then (3.1) guarantees existence of s 3 E S(x) such that :Z:i=l,Z ~(cp(x, r, s t
f(x, r, si), cp(x, r, si)) E Q, which eventually results in :Z:i=l, 2 ~qi E Q(x, r). Con-
versely, (1.6) implies (3.1) because always co[cp x f x 19](x, r, S(x)) C coQ(x, r).

PROPOSITION 3 Assume {1.1) - {1.3) hold. Let v be the solution to (RP') and
let also {1.6) hold for a. a. x E 0 and aliT E lRn. Then there is u E LP(O; lRm)
such that

u(x) E U(x) := { s E S(x); cp(:r, y(.1:), s) :S: .L"' cp(x, y(x), a) v,c(da),
f(x, y(x), s) = { f(x, y(x), a) Vc(da),
}Rm
19(.7:, y(x), s) :S: { 19(x, y(x), a) llx(da)}, X E f2, (3.2)
J'R"'
and any such u is an optimal control for (P).

Proof. For a .a . x En, llx can be approximated weakly* by a sequence {v{}JEN of

convex combinations of Dirac measures supported on supp(vx)· Then JR"' [cp x
f x 19](x, y(x), s) v{(ds) E co[cp x f x 19](x, y(x) , supp(l/x)), so that by passing to
the limit and using (1.6) in the form (3.1), one gets

r [cp
km X f X 19](x,y(:r),s)l/x(ds) =lim /
J-oo km [cp X f X 79](x , y(x),s)l/~(ds)
E clco[cp X f x 19](x,y(x),supp(vx))
C clco[cp x f x 19](.7:, y(.1:), S(x)) C Q(x, y(x)); (3.3)

this limit passage is indeed correct at each 1: E n for which JR"' !s iP 1/x(ds) is
finite, sec Roubicck (1999), Lemma. 2 for details.
This enables us to show that the set U(:r) defined by (3.2) is nonempty.
Indeed, because of definition (1.4), for any (qo , q1 , qz) E Q(x, y(x)) there is
s E S(x) such that qo 2: cp(x, y(:r), .s), q1 = f(x, y(x), s) and q2 2: 19(:r, y(x), s).
Hence, for the particular choice

r -
312 T. ROUBICEK, W.H. SCHMIDT

the inclusion (3.3) implies that qo 2: 'P(J;, y(x), s), q1 = f(x, y(x), s) and q2 2:
?'J(x, y(x), s) for somes E S(x).
Moreover, by Aubin and Frankowska (1990) the multivalued mapping U :
n =l !Rm defined by (3.2) is measurable because s is measurable, 1/ is weakly
measurable, and 'P, f and ?') are Caratheodory mappings, see , again, Roubicek
(1999), Lemma 2 for details.
0 bviously, U (X) is closed for a.a. X E f2 because S (X) is closed and 'P( X, r, ·),
f(x, r, ·)and ?'J(x, r, ·)are continuous, see (1.1k) and (1.1a), respectively. Then,
by Aubin and Frankowska (1990), Theorem 8.1.4, the rnultivalued mapping U
possesses a measurable selection u( x) E U (x).
In particular, u(x) E S(x). Moreover, in view of (3.2) with (3.4),

fo ?'J(x, y(x), u.(x)) dx

: ; Jnrq2(x) dx = Jnr }Rmr ?'J(x, y(x), s) 1/x(ds) dx::; 0 (3.5)

and also

f(x, y(x), u(x)) = q1 (x) = 1, f(x, y(1:), s) Vx( ds) (3.6)

for a.a. X E n, so that

fo K(x,~,y(0)(1m f(~,y(0,s)vt:(ds)) d~
y(x) = Yo(x) +

= kK(x, ~, y(0)ql(~) d~ = fo K (x, ~' y(O)f(~, y(0, u(~)) d~. (3.7)

By using also Propositions 1 and 2, we get

fo 'P(x,y(x),u(x))dx :S fo qo(x)dx
= r r ¢(x, y(x), s) Vx(ds) dx = min(RP') = min(RP)::; inf(P).
Jn }Rm
(3.8)

In particular, (3.8) and the coercivity (1.3b) together with (1.3a) and (1.1b)
imply that

E fo iu(x)IPdx :S fo 'P(x,y(x),u(x))dx :S inf(P) < +oo (3.9)

so that u E £P(f2; !Rm).

AJtnp·pt.hPr . (~..')) (~ . 7 ) . :wrl (~ . 9) show t hil.t, tlu~ nair (u.1!) is admissible
Optimal control of Fredholm integral equat ions 313

REMARK 5 (Zero relaxation gap) Under the assumptions (1.1) - (1..'3) and (1.6),
Proposition 3 implies that the first term in (.'1.8) equals to inf(P) , so that (3.8)
gives min(RP) = inf(P) = min(P) , i.e. th ere is no relaxation gap. Let us em-
phasize that this is a nontrivial fact for problems involving state constraints.

Proof of Theorem 1. By Proposition 1, we get some 17 solving (RP). Then

Proposition 2 yields its Young-measure representation I/ solving (RP'). Finally,
usi ng t his v , Proposition 3 gives a solution u to (P). The stability (1.5) of
any solution to (P) is a consequence of zero relaxation gap (see Remark 5) and
the correctness (in mathematical sense) of the relaxed problem, sec Roubicek
(1 997), Sect. 4.1. •

4. Refined existence results

The method used above allows easily to refine the existence results on condition
that one has so me a priori information about the support of the Young measure
representing some optimal relaxed control. Such information can be obtained
sometimes by analyzing the maximum principle. In the context of optimal
control of ordinary differential equations, we refer to Gabasov and Mordukhovich
(1974), Ioffe a nd Tikhomirov (1974, Sect. 9.2.2, Proposition 1 with Theorem 3],
Mordukhovich (1988), Sect. 20 or Mordukhovich (1999), and also to Munoz and
Pedregal ( 2001).
To use the standard maximum principle, we must still assume certain smooth-
ness of the data with respect to the state variable to ensure the relaxed problem
to be smoot h. Namely, we assume that, for q 2: 2, the data K, f , <p, a nd '13
satisfy, beside ( 1.1 )- ( 1. 2) , also

<p~, K;., f:., '13~ are Caratheodory mappings, (4.1a)

l <p~ ( x,r 1 ,s )-<p;. (x,r2,s )l
:S: (aq/(q -2) (x ) + blr1lq- 2 + bhlq- 2 + clsiP(q- 2)/q) lrl- r2l, (4.1b)
I[Kf] ~(x,Crl,s)- [KJ];.(x, ~ , r2 ,s )l :S: f1(x , Oir1- r2 l, (4.1c)
l'13;.(x, r1, s)- '13~(x, r2 , s) l
:S: (aq/(q-2)(x) + blr1lq- 2 + bhl'~- 2 + clslp(q- 2)/q)lrl- r2 l (4.ld)

with £1 E L'~(D.; Lqf (q- 2 l(D. )) ; here <p;. denotes the differential of J( x, ·, s), K~ is
the differential of K(x, ~' ·) , etc., and the remaining notation is as in (1.1) .
Moreover, (l.lf) is to be strengthened a bit:

IK(x, ~ , r)l :S: a(x , O + cirl q(l-l)h (4. 1e)

for some a* E Lb - llh(D.;L'~(D.)) with a * ( x, ~) = a(~,x); equivalently, this

requires fnUn Ia(~ , x ) I '~ d0"Y/(q('Y-l)) dx < + oo.

PROPOSITION 4 Let the assumvtions (1 .1 )-(1.3 ) and (.4..1 J hold. o.nd {p_f. f.h.P.
314 T. ROUBICEK, W.H. SCHMIDT

and v E YP(D; IRm) be a Young-measure representation of 17 in the sense {2. 7).

Then, for some multipliers >. 0 2:: 0 and >. 1 2:: 0 not vanishing simultaneously,
the following maximum principle holds

r 1i>. ,>. ,>.,y(X, s) Vx(ds) =

}R"'
0 1 max 1i>. 0 ,>. 1 ,>.,y(X, s) for a. a. X E f2,
sES(x)
(4.2)

where the Hamiltonian 1i>. 0 ,>. 1 ,>.,y is defined by

1i>. 0 ,>. 1 ,>.,y(x, s) :=in >.(~)K(~,x, y(x)) d~ f(x, y(x), s)

- Ao<p(x, y(x), s)- >.J'19(x, y(x), s) (4.3)

and ). E Lqf(q-l)(D; IRn) solves the adjoint l-inear integral equation

>.(x) = f f [Kf];.(~,x,y(x),s)T >.(Ovx(ds)d~

Jn }Rm
- Ao r
}IRm
r '19~(x, y( x ), s) Vx(ds)
<p~(x, y(x), .s) Vx(ds)- >.J }IRm (4.4)

for a.a. X E n with (-)T denoting the transposition . Moreover, the following
transversality condition holds:

>.T r r 19(x, y(x), s) Vx(ds) dx

Jn }R"'
= 0. (4.5)

Proof. Let us enlarge H from (2.1) so that also

(<p;. oy)yEH, ('19~oy)yEHk, ( [Kf]~oy)yE H 11

for ally, yE Lq(D; IRn). Moreover, (4.1e) ensures that x ~---+ fn >.(OK(~, x, y(x)) d~
belongs to p/('y-l)(D;IR1) for any). E Lqf(q-l)(D;IRn) so that also 1i>. 0 ,>. 1 ,>. ,y
E H due to (2.1). In such a way, we get a refined relaxation scheme yielding a
smooth relaxed problem (RP). Moreover, we can assume H separable; it just
suffices to modify the arguments of continuity (2.4) and to use (4.1a- d) and the
separability of Lq(D; lR11 ) . Then, our assertion just fo llows as a special case from
Roubicek (1998), Proposition 5.1 with Remark 3.2. •
Now we can state a rather strong existence principle which does not require
the orientor field Q to be convex. The usage of such result for nonconvex Q is,
however, not straightforward because it requires a certain information about at
least one solution of the relaxed problem and the corresponding adjoint state.
Yet, in particular cases a relevant information can sometimes be isolated, see
Section 5 below.

'THF.nRRM?. T.Pf th P n.RRnm.nfinnR (1.1 )- (1 . .'1) and (Ll) hold. and let the cone
Optimal control of Fredholm integral eq uations 315

with c01-responding >. 's from Proposition 4, and let the following condition hold
for a.a. X E [2;
co [<p x f x 19](.r-,y(x),M(x)) C Q(x,y(:r;)), (4.6)
where M(x) C S(x) is an estimate of the set of maximizers of the Hamiltonian
H ,\ 0 ,.x 1 ,.X,y, i.e.
M(x) :J {s E S(x); 7i.x 0 ,.x 1 ,.x,y(x,s) = ~~(:) 7i.x 0 ,.x 1 ,.x,y(x,s)}.
8 (4.7)

Then ( P) has at least one solution and, moreover, every solution to ( P) is stable
in the sense (1.5).
Proof. Let I/ denote some Young-measure representation of the optimal relaxed
control 11 in question. By Proposition 4, for a.a. X E n, 1/x must be supported
on M(x) . T hen, Proposition 3 gives a solution ·u to (P) if used in a refined
way, namely (3 .3) can simply exploit M(x) and (4.6) instead of S(x) a nd (3.1),
respectively. The stability (1.5) is again the consequence of zero relaxation gap
(by arguments as in Remark 5) and of the correctness of the relaxed problem. •
REMARK 6 There are two extreme situations. First, the maximum principle
does not give any specific information, see Section 5.1 below, or the particular
problem often is so complicated that one is unable to extract such informa-
tion; then we can say that M(x) = S(x) only, and (4.6) coincides with (3.1),
i. e. with the Pilippov- Roxin-type condition (1.6}. Second, sometimes it may
happen that one can a pr-iori guarantee, by a specific analysis, that the Hamil-
tonian 7i.x 0 ,.x 1 ,,\,y (x, ·) is maximized only at a single point for a.a. x E n, see
Remark 7 below, i.e. the set on the right-hand side of (4. 7) is a singleton for
a.a. x En, and by choosing M(.r,) equal to this set we make the condition (4JJ)
trivially satisfied.

5. Examples
In this last section we present three concrete, rather simple, illustrative ex-
amples .

5.1. Schmeling's example of nonexistence in Fredholm equations re-

visited

l
To present the example by Schmeling (1979, 1981) in our context, we must
modify it slightly. For example, we consider the following problem:

Minimize 1T (2y(t)- t2? dt

0
subject to y(t) = 1 t

[(T- T)U(T) + (t- T)u(T) 2 dT,

J (PI)
u(t) E [-1.11 for a.a. t E [0 , TL
316 T. ROUBICEK, W.H. SCHMIDT

As usual for Volterra equations, we put n := (0, T) and write t and T instead
of x and ~, respectively.
For reader's convenience, let us remind the slightly modified arguments from
Schmeling (1981) to show t he nonexistence of solutions of (PI): Taking a fast os-
cillating sequence {uk: (0, T) ___, {1, -1} hEN converging to 0 weakly in LP(O, T),
we get a sequence of corresponding states {YdkEN converging to y(t ) = ~t 2
in C(O, T), which shows inf (Pr) = 0. Therefore, if there exists a solution (11, y)
to (Pl), then inevitably y(t ) = ~t 2 would hold. By differentiating the Volterra
equation involved in (Pl), one would get

t = (T- t)u(t)- tu( t)

2 d rt
+ dt Jo tu(T) 2 dT

= (T- t)u(t) + 1t u(T) 2 dT (5.1)

for a.a. t E [O,T]. Note that, since u is bounded, t f-+ J;[(T- T)u(T ) + (t-
T)u( T) 2 ] dT is Lipschitz continuous and thus a.e. differentiable. If possibly (5.1)
does not hold just fort:= T, we can pass to the limit with t ___,Tin (5.1), which
J
gives T = 0T u(7) 2 dT. T his would be possible only if lui = 1 a.e. on [0, T].
Coming back to (5.1), we would get t = (T- t)u(t) + t, which would give u = 0,
a contradiction showing that a solution (u,y) to (PI) cannot exist.
Of course, ( P 1) must somehow violate the assumptions of the above presented
theory. We consider n = m = 1, p, q arbitrary, cp(t, 7', s) := (2r- t 2 ) 2 (we
can add the term like c-max(ls!P, 1) t o satisfy formally (1.3b)), {) := 0, and
S(t) := [-1, 1]. Then, we can think of choosing l = 1 and thus have to p ut

f( t, T, r, s ) ·-
.- (T - T)s + (t - T) s,2 , K ( t, T, 1·) -_ { 1 'f t
if 2 T, (5.2)
0 1 t < T.
This makes the orientor field f(t, T, r, S) always convex compact, but evidently
such f does not have the form required in Theorems 1 and 2 because it depends
also on t. Alternatively, we can choose l = 2 and then put

·-( 2) ( -{(T-t,t-T) ift2T,

f (t,r,s ) .- s,s , K t,T,r)- (O,O) ift < 7 _ (5.3)

It has already the above considered form, but the orientor field

Q(t, r) = {(qo, s, s 2 ) E IR 3 : qo 2 0, s E [-1, 1]} (5.4)

is evidently nonconvex and ( 1.6) is violated so that Theorem 1 cannot be used .

Analyzing the maximum principle (4.2)-- (4.5), we get simply >.o = 1, >. 1 = 0,
and>.= 4(2y- t 2 ). For the (unique) relaxed optimal control v 1 = !8 1 + !L 1 ,
one gets y = !t2
, and hence>.= 0 and the Hamiltonian H(t, ·)constant for a .a.
f t= rn 7'1 <::n t.h::l.t. nPrP<::<::::l.rilv M(f) = ,c;(f) ::l.nrl (4 ()) tim<:: (~nineirlP<:: with (1 ())
Optimal con trol of Fred holm integral equatio ns 317

5.2. Mordukhovich's example with a nonconvex orientor field mod-

ified
To present the examples from Mordukhovich (1988), Sect . 20 or Mordukhovich
(1999) in context of integral equ ations , we consider the following problem:

Minimize k y(x) 2 dx
s ubj ect to y( x) = Yo(x)

+ .£ (1(1(x , O'u1(0 + K 2(x, Ouz(~)) d~,

n(:~: ) := ('u1(x),nz(x)) E {1 , -1} X [-1 , 1] for a.a. X E !1,
y E L 2 (!1), u E L 00 (!1; JR 2 ),

This is a very special case of (P), with m = l = 2, n = 1, k = 0, 'P (x,r, s ) = r 2 ,

f( x , 1', s) = 8, K(x, ~ ' r) = (K1(:r , ~), Kz( x, 0) E JRnxl := rre, fJ = 0, a nd S(x) =
{1, -1} x [-1 , 1] having two disconnected convex components. Our ass umptions
a re ra ther trivia lly fulfilled with q = 2, p E [1 , + oo) arbitra ry a nd C = £1 = 0,
provided]{ satisfies (l.lf,g) , which means here K 1,2 (x , ·) E p /h- 1>(!1) for a.a.
:1: E nand IK1 ,2(x, OI ::; 'Lf= 1 aq ,j(x)a1,j(O, the growth with respect to s as
well as the coercivity (1.3b) being irrelevant since S(x) is a priori bounded.
As t here a re no state constraints, we have simply Ao = 1 a nd A1 = 0, and
t he adjoint equa tion (4.4) then gives A = -2y. Then , the Hamiltonian (4 .3)
results in 7-ly( J.:, s)=7-ly( x, s 1, sz) = -2 L;=
1,2 Si .J~ y(OK;(~ , x) d~. On certain
conditions , it m ay happ en that the set of maximizers of 7-ly( x, ·) is contained
only in one component of S (x) . E .g. , ass uming

K 1 (:r , O { < 0 ~f yo(x) > 0, (5 .5)

> 0 1fyo(x) < 0
for a.a. ~ E !1, we have t he following chai n of implicat ions

lvo(:r) l > k[IK1 (:r,~) l + IKz(x, OIJd~:::} ly (x) l > 0

:::} 't:/8z : 81 f--+ 7-lu( x, 81, sz) increasing
or 't:/s2: s 1 f--+ 7-ly( x,s 1,sz) decreasing,

for x fixed (but a rbitra ry). This ensures that, for a.a . x En, t he set My(x) :=
{8 E S(.1:) ; 7-ly(x,s) = max 7-ly (x,S( :r ))} satisfi es

My(x) C {1} x [-1 , 1] or My(x) C {-1} x [-1 , 1].

R eali11ing t hat 7-ly (:c, ·) is affine, we can see t hat My (:r ) is convex fo r a.a. :r E n ,
a nd t hus ( 4.6) satisfied . Then , Theorem 2 with l'vf = My yields existence of the
solution to (P2) provided (5 .5) holds and lvo l > ./~( IK1 (-, ~) I+ IKz(- , 01) d~ is
318 T. ROUBICEK, W .H. SCHMIDT

Let us emphasize that Q(x ,r ) = (r 2 +lR+) x S(x) is nonconvex in our case

so that the classical Filippov-Roxin-type condition (1.6) cannot be used. Also,
the cost functional is strictly convex with respect to state and thus the relaxed
problem to (P2) has no concave structure, so that also the results from Roubfcek
and Schmidt (1997) relying on Bauer 's principle cannot be used. Moreover,
contrary to the original example Mordukhovich (1988, 1999) , My(x) need not be
a singleton so that we cannot apriori say that the Young-measure representation
of the optimal relaxed control is a.e. a Dirac measure, which would yield the
existence of an optimal control of (P2) straightforwardly. Besides, nontriviality
of this example relies also on the fact that the solution to (P 2) indeed does not
exist if, e.g. , Yo = 0 and K 1 2: 0 does not vanish, which in fact covers also the
well-known Bolza example.
To obtain the original Mordukhovich's (1988, 1999) example, one can put
simply n := (O,T), K2(t,T) = 0, K1 (t,T) = 1 fort 2: T, otherwise K1(t,T) = 0,
and yo constant.

5.3. Example with a boundary-value problem and an isoperimetric

inequality
This example illustrates a simple, but rather nontrivial application of Theorems
1 and 2 to optimal control of elliptic equa tions (as announced in Remark 4).
Let us consider the following problem :
·1
Minimize u 2 (x) dx
.o
/
subject to y( x) = a( 1 - :r) + bx
+ 1x (1- x)~(sin u(O- cy(0 ) d~
+1 1
x(1- ~)(sin u(O- cy(~)) d~,
1 1
y( x) 3 dx:::; 0,
y E L3 (0, 1), u E L 2 (0, 1).

This falls into the previous considerations if /,; = l = m = 1, p = 2 (determined

by the growth of <p(x, r, ·)), q = 3 (in agreement with the growth of 'l9(x, ·, s)),
Yo( x) = a(1- x) + bx, f( x,r,s) = -cr + sins , <p(x , r,s) = s 2 , 'l9(x , r,s) = 1· 3 ,
S = JR, and the kernel J( is symmetric:

T/( ~ ) { (1 -Ox for .1: :::; ~' (5.6)

1\ :z:, , r = c( 1 _ x) c c
" tor :r: > <.·
The constant c should satisfy

(r; 7\
Optimal control of Fredholm integral equations 319

in order to fulfill (1.2). However, using the monotonicity argument instead of

the Banach-fixed-point one, we can admit arbitrary c positive, too.
The integral equation is then equivalent to the two-point boundary-value
problem for the 2nd-order linear elliptic ordinary equation:
d2y
- d.1: 2 + cy =sin u on (0, 1), y(O) =a, y(1) = b, (5.8)

which can be checked just by direct calculation, if the short-hand notation

w =sin u- cy is used:

d·)
0
r
d.~;
d (
= d::2 a(1- x) + bx + ./o (1- x)~w(O d~ + 1
x
1
x(1- Ow(O d~
)

d ( b-a-)
=d.?.: r ~w(Od~+ 11"'(1-~)w(Ocl~ )
0
= -:z:w(x)- (1- x)w(x) = -w(x).
Let us mention an illustrative interpretation of (5.8) as a deflection of stretched,
homogeneous, elastically supported (c determines the linear response of the
support) string with fixed end points, loaded in a perpendicular direction by the
force sin1t. This also shows that the solution y E L 2 (0, 1) is, in fact, smooth,
namely y E W 2 •00 (0, 1) if one uses the standard notation for Sobolev spaces. It
is an interesting observation that, although f(x, r, ·) is nonlinear, the condition
(l.G) is valid, i.e. Q(x, r), here independent of (x, T), is always convex. Thus
Theorem 1 yields existence on an (unspecified) optimal control.
Let us now modify (P3) by restricting the admissible control values to the set
S(x) = [-37!', -21r] u [-1r, OJ u [21r, 37r]. (5.9)
Then Q(x, T), again independent of (x, r ), is no longer convex so that Theorem 1
does not apply. Let us analyze the optimality conditions. The adjoint equation
(4.4) now looks as

.\(x) = -c 1 1
K(.1:, 0.\(~) d~- 3.\ 1 y(0 2 (5.10)

for K(x, 0 = K(x, ~, T) defined by (5.6) and for some .\o, .\ 1 non-negative, .\o+.\1
> 0. We claim that
·1
A(x) := K(~, x).\(0 d~::; 0 for a.a ..1: E [0, 1]. (5.11)
./ 0

For c::; 0, we can see directly from (5.10) that even A ::; 0 as a consequence of
non-negativity of]{ and /\ 1. Using again K;::: 0, we get (5.11). For c > 0, we
must usc finer argument: A defined by (5.10)- (5.11) solves, in fact, the adjoint
two-point boundary-value problem
320 T. ROUBICEK, W.H. SCHMIDT

and then (5.11) follows from the maximum principle (in the sense used in theory
of 2nd-order differential equation) by use of a cont radiction argument, saying
that, if maxxE[D,l] A(x) > 0, then at some x E [0, 1] there must be A(x ) > 0 and
~A(.T):::; 0 simultaneously, which, however, contradicts (5.12).
Now, the Hamiltonian (4.3) takes the form 'H>.o)q ,>.,y(x, s) = A(x) sin s -,\ 0 s 2
(up to a function constant ins-variable) and, having (5.11) at our disposal, we
can see that its maximum cannot be attained for s positive. Thus, from the
maximum principle (4.2), we can see that any solu tion v to (RP3) satisfies

supp(vx) C [-1r, OJ U [-31!', -211'] =: M(x) for a.a. x E [0, 1]. (5.13)

However, taking this M, the condition (4.6) is satisfied! Therefore, Theorem 2

can be applied even without any specific knowledge about a solution to the
relaxed problem (RP 3 ). Also, let us emphasize that we do not have any specific
knowledge about the adjoint state A (except for (5.11)) nor about the Lagrange
multipliers -'o and -'1·

REMARK 7 (Existence in ordinary-differential-equation problems) One can find

in the literature furth er examples for situations covered by Theorem 2 in conte.rt
of systems governed by ordinary differential equations, which can be viewed as a
very special Volterra integral equation. We refer to Gabasov and Monlv).;hovich
(1974), Joffe and Tikhom irov (1974), Sect. 9.2.2, Proposition 1with Theorem 3,
Mor·dukhovich (1988, 1999) or Munoz and Pedregal (2001}. A non-academic
example with M(x) a singleton for a. a. x E Sl and thus (4.6) satisfied trivially
was presented by Bittner ( 1998) for a flight optimal contml problem.

References
ANGELL, T.S. (1976) On the optimal control of systems governed by nonlinear
Volterra equations. J. Optim. Theory Appl., 19, 29- 45.
AUBIN, J.P. and FRANKOWSKA, H. (1990) Set-val-ued Analysis. Birkhanser.
BALDER, E.J. (1984) On existence problems for the optimal control of certain
nonlinear integral equations of Urysohn type . .J. Optimization Th . Appl.,
42, 447- 465 .
BALDER, E. (1993) Exist ence of optimal solutions for control and variational
problems with recursive objectives. J. Math. Anal. Appl., 178, 418- 437.
BENNATI, M.L. (1979) Un theorem di esistenza per controlli ottimi di sistemi
definiti da equazioni integrali di Urysohn. Ann. Mat . Pura Appl. (IV) ,
121, 187- 197.
BITTNER, L. (1994) Existence of optimal control for integral processes . In: XI.
Herbstschule "Variationsrech., opt. Prozesse und Anwendungen", Schmidt ,
W .H .. r.d .. Prenrint- Reihe Mathematik No. 1. Universitiit Greifswald.
Optimal control of Fredholm integral equations 321

BITTNER, L. (1998) On the convexification of the optimal control problems of

flight dynamics. In: Variational Calculus, Optimal Control and Applica-
tions, Schmidt, W.H., Heier, K., Bittner, 1., Bulirsch, R., eds., ISNM 124,
Birkhauser, Basel, 87- 96.
BRUNOVSKY, P. (1968) On the necessity of a certain convexity condition for
lower closure of control problems. SIAM .J. Control, 6, 174- 185.
CARLSON, D.A. (1990) Infinite-horizon optimal controls for problems governed
by a Volterra integral equation with state-and-control dependent discount
factor. .J. Optim. Th. A ppl., 66, 311- 336.
CESAR!, L. (1983) Optimization: Theory and Applications. Springer, New York.
COWLES, D.E. (1973) An existence theorem for optimization problems involving
integral equations. SIAM .J. Control, 11, 595-606.
GABASOV, R. and MoRDUKHOVICH, B.SH. (1974) Individual existence theo-
rems for optimal equations. Doklady AN SSSR, 215 (in Russian). Engl.
trans!. Soviet Math. Dokl., 15, 576-581.
FILIPPOV, A. F. ( 1959) On certain questions in the theory of optimal control
(in Russian). Vestnik Moskov. Univ., Ser. Math. Mech. Astr. Fiz. Khim.,
2. Engl. Trans!. SIAM .J. Control, 1 (1962), 76-84.
IOFFE, A.D. and TIKHOMIROV, V.M. (1974) Theory of Extremal Problems (in
Russian). Nauka, Moscow. Engl. trans!.: North-Holland, Amsterdam, 1979.
MORDUKHOVICH, B.SH. (1977) Existence of optimum controls . .J. Soviet Math.,
7, 850- 886.
MORDUKHOVICH, B.SH. (1988) Methods of approximation in problems of opti-
mization and control (in Russian). Nauka, Moscow.
MORDUKHOVICH, B.SH. (1999) Existence theorems in nonconvex optimal con-
trol. In: Calculus of Variations and Optimal Control, Ioffc, A., Reich, S.,
Shafr·ir, I., eels., Chapman & Hall, CRC Res. Notes in Math., 411, CRC
Press, Boca Raton, FL.
MUNOZ, .J. amd PEDREGAL, P. (2001) A refinement on existence results for
optimal control problems. Nonlinea1 · Anal., Th., Meth. Appl., 46, 381--
398.
PACHPATTE, B.G. (1987) Existence theorems for Hammerstein type integral
equations. Analele stiintifice ale Univ. Al.I. Cuza din Ia§i, 33, 197·-201.
PAPAGEORGIOU, N.S. (1991a) Optimal control and relaxation of nonlinear el-
liptic systems. Japan.!. Indust. Appl. Math., 8, 525- 550.
PAPAGEORGIOU, N.S. (1991b) An existence theorem for a class of nonlinear
elliptic optimal control problems. Commentat. Math. Univ. Carolinae , 32,
273- 279.
PELCZEWSKI, .J. (1989) Optimization of the system described by Fredholm's
integral equation. Demonstmtio Mathematica, 22, 773- 787.
RousfCEK, T. (1997) Relaxation in Optimization Theory and Variational Cal-
culus. W. de Gruyter, Berlin.
RousfcEK, T. (1998) Optimal control of nonlinear Fredholm integral equations.
322 T. ROUBICEK, W.H. SCHI"v!IDT

RousiCEK, T. (1999) Convex locally compact extensions of Lebesgue spaces

and their applications. In: Calculus of Variations and Optimal ContTOl,
Ioffe, A., Reich, S., Shafrir, I., eds., Chapman & Hall, CRC Res. Notes in
Math. 411, CRC Press, Boca Raton, FL, 237-250.
RousiCEK, T. and SCHMIDT, W.H. (1997) Existence of solutions to certain
nonconvex optimal control problems governed by nonlinear integral equa-
tions. Optimization, 42, 91-109.
RoxrN, E. (1962) The existence of optimal controls. Michigan Math. J., 9,
109- 119.
SCHMELING, H. (1979) Zur Existenz der optimalen St euerungen bei Aufgaben
mit nichtlinearen Volterraschen Integralgleichungen. Preprint No.4, Ernst-
Moritz-Arndt Universitat Greifswald.
ScHMELING, H. (1981) Zur Existenz der optimalen Steuerungen bei Volter-
raschen Integralgleichungen. In: Nonlinear Analysis (Berlin 1979), Abh.
Akad. Wiss. DDR, Abt . Math. Naturwiss. Tech., Akademie Verlag, Berlin,
395- 398.
STUART, C.A. (1974) Existence theorems for a class of nonlinear integral equa-
tions. Math. Z., 137, 49- 66.
YEH, L.S. (1978) Existence theorems for optimal control problems characterized
by Volterra integral equations. Rev. Ro-umaine Math. Pnres Appl., 23 ,
1249- 1271.
YUSIFOV, B.M. and KARAGEZOV, Z.A. (1990) On the problem of the existence
of optimal controls. Diff. Uravneniya, 26, 1437- 1441 (in Russian). Engl.
transl.: Diff. Eq., 26, 1065- 1068.
ZOLEZZI, T. (1972) Esistenza e condizioni necessaire per controlli ottimi di
sistemi definiti da cquazioni intcgrali. Ann. Mat. Pnra Appl. (IV) , 91,
97--118.