Optimal control for N-Person stochastic inclusions. II
Transcript of Optimal control for N-Person stochastic inclusions. II
This article was downloaded by: [Northeastern University]On: 15 November 2014, At: 20:18Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,37-41 Mortimer Street, London W1T 3JH, UK
Stochastic Analysis and ApplicationsPublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/lsaa20
Optimal control for N-Person stochastic inclusions. IIBui An Ton aa Department Of Mathematics , University Of British Columbia , Vancouver, B C, V6T 1Z2,Canada E-mail:Published online: 03 Apr 2007.
To cite this article: Bui An Ton (2000) Optimal control for N-Person stochastic inclusions. II , Stochastic Analysis andApplications, 18:6, 1031-1053, DOI: 10.1080/07362990008809710
To link to this article: http://dx.doi.org/10.1080/07362990008809710
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) containedin the publications on our platform. However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose ofthe Content. Any opinions and views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be reliedupon and should be independently verified with primary sources of information. Taylor and Francis shall not beliable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilitieswhatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out ofthe use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
STOCHASTIC ANALYSIS AND APPLICATIONS, 18(6), 1031-1053 (2000)
OPTIMAL CONTROL FOR N-PERSON STOCHASTIC INCLUSIONS. I1
BUI AN TON
DEPARTMENT OF ~ ~ A T H E M A T I C S , UNIVERSITY OF BRITISH COLUMBIA , VANCOUVER ,B.C. CANADA V6T 122
E - m a d address: [email protected]
ABSTRACT. The existence of a solution of a large class of nonlinear stochastic inclusions is shown . The existence of an open loop problem for those stochastic inclusions , is established. Application to an optimal strategy problem arising in the fight against drugs is given.
The purpose of this paper is to establish the existence of an open loop optimal co~ltrol problem for a large class of nonlinear stochastic inclusions.
lu the study of stochastic equations . the convergence of the approximating so- lutions is usually obtained via C'auchy criterion since there is no compact injection mapping in the natural spaces associated with the problem and thus restrictions on the type of nonlinearity a te ilnposed . Using the notion of fractional time derivative, i~~troduced in 121 by Bensoussan in the study of stochastic Navier-Stokes equations, the Prokorov theorem on the tightness of measures together with the Storokhod the- orem, we have proved in [9] the existence of an open loop control problem for a large class of stochastic inclusions .
Let H be a Hilbert space, h' be a compact convex subset of H . Let E , L be upper semi-co~~tinuons set-valued mappings of H x U into the closed convex subsets of IJ w ~ t h U being a compact convex subset of the space of controls C(0 ,T ; O;) and where U3, j = 1, ., N are Hilbert spaces .
Consider the stochastic inclusion
dy(t; w) + A(y)dt E B ( y ; u,; x,v)dt + L(y; u,; a,v)dw(t; w) on ( 0 , T ) a . s .
(1.1) ~ ( 0 ; w) = E ; E E K ; A(y) E ~ I K ( Y )
where I K ( . ) is the indicator of the set K and r,v = (v l , ., vj-1, vj+l, ., V N ) .
1991 Malhema l zc s Snbjert Classzfiralzon Primary 49 J 22, 49 N 40 , Secondary 49 N 55 , 90 D 06 , 92 H 35
103 1
Copyright O 2000 by Marcel Dekker, Inc.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1032 BUI AN TON
In 191 , we have shown that there exists a probability space (R ,A ,P ) , a f - standard Wiener process w and y such that
{(Q, A, P ) ; Y , w )
is a solution of 1.1 for a given {u, , x ,v) in U . In this paper, instead of the Prokorov and the Storokhod theorems, we shall use a
result of Von-Neumann on sections of closed graph set-valued mappings and consider the case when the probability space and the Wiener process w are given , not un- knowns of the problem. A similar approach has been used in the study of stochastic Navier-Stokes equations in Ton [ l o ]
In Section 4 , the set R S t , ( ( ; u,; x,v) of all solutions of 1.1 is shown to be non- empty .
Let { f , , gj}y=, be some given convex functionals and for y E R, to( ( ; u,; ?r,v), let
The existence of ii E U and of tj E ii,; r j i i ) such that
E { J , ( t ; y; O,; " , a ) ) 5 E{J , (E; y; v j ; r j i i ) } ; j = 1, ., N
for all v E 2.4 and all y E RSt , ( ( : v,; a,iL), is shown in Section 5 of the paper . The notations , some known results are given in Section 2 . The deterministic
case is considered in Sections 3 and 4 . The results of Sections 5 are applicable to the study of an optimal strategy for an international investment portfolio and to the study of American options as done earlier in Ton [9] . Instead of mathematical finance
we shall consider some new applications and study the problem of optimal resources allocation in the fight against drug addiction . The model studied in Section 6 seems ttrw allti ex(rnds in several directions the 2-player deter~ninistic model considered by 1l.l)awid and G.Feichtinger [GI .
2. NOTATIONS , ASSUMPTIONS, K N O L I N RESULTS
Let H and let {U,)~,, the space of controls , be real Hilbert spaces and let kr be a compact convrx subset of H with 0 E K.
Let y, be positive numbers going to 0 and let
2 = { Y : y E L 2 ( 0 , T ; H ) , y ( t ) E K a.e.; I Y 1z< m}
with
and where y = 0 outside of [O,T].
Lemma 2.1. For each positive constant c ,
{Y : 1 Y I E _ < c )
is a compact subset of L 2 ( 0 , T ; H ) Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1 1033
The lemma is s roved in the same wav as in Aubin's theorem with the fractional time-derivative replacing the usual derivative . A proof can be found in [9] .
Let U be a compact convex subset of C ( 0 , T ; $=, I J j ) and let Q, L be upper semi-continuous (u.s.c.) set-valued mappings of H X U into the closed convex subsets of H . We assume that
Assumpt ion 2.1. Let Q, L be u.s.c. set-valued mappings of H x U into the closed convex subsets of H . Suppose that there exists a constant C such that
Let (R,A, P ) be a probability space and let P be a filtration on the space . We assume that the measure P is regular in the interior , i.e. for every E , there exists a compact li, such that P ( R - K c ) 5 E and the restriction of P to I(, is continuous.
Since h' is a compact convex subset of H , it is known that IK( .) is a 1.s.c. convex function on H and hence 81K is defined .
Definition 2.1. Let (R, A, P ) be a probability space with P being regular in the interior , let F be a filtration on the space and let w be a F - standard Wiener process on the space .
Let { E ; u,, n,v) E L 2 ( R , A , P ; K ) x U , then y is said to be a solution of 1.1 if 0 y is .P measurable and
y E L 2 ( R , A, P : Lm(O,T; H ) ) ; y ( t ) E K a.e. on [O,T] a.s.
and
with
+ J,' L ( y ; u j ; ~ ~ v ) d w ( s ; w )
a.s. for all t E [O,T] and for { A ( y ) , G, L ) E 8 1 K ( y ( t ) ) x Q x L.
The equation 2.2 implies that y is in C ( 0 , T ; H) a.s. We associate with 2.2 the cost functions
where y is a solution of 2.2
Assumpt ion 2.2. Let { f j , g j ) be continuous convex functions from H x L2(0 , T ; H ) x U into R: . W e assume that
1)
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1034 BUI AN TON
2) f , , g j ( x ; . ; .rrjv) are strictly convex and that 8g,(x; . ; ~ , v ) are strictly maximal monotone.
The following theorem of Von Neumann on sections of set-valued mappings will be used throughout the paper . Theorem 2.1. Let X, Y be two separable Banach spaces and let A he a set-valued mappzng of X znto the non-empty closed subsets of Y. Suppose that the graph of A zs closed Then A has a urtzversally Radon measurable sectzoii, z e. there exzsts a slngle-valued mappzng u of X znto Y wzth u ( x ) E A ( x ) for all x E X , whzch 1s
measurable for any Radon measure on the Borel field of X .
The proof of the theorem call be found in the Appendix of [dl. For deterministic differential inclusions with singular inputs , we shall need the
following assumption.
Assumption 2.3. Let w be a cor~tznuous functzon on [O, TI and lrt N = [Tlh] , t k = I11 jor some small posztzve ~lurnber h. We assume that
(2) Thew e m t s a posztzve constant u such that
{ w ( t k ) - ~ ( t k - ~ ) } ~ <_ u 2 h ; k = 1 ,..,. v. (ii) l'here exlsts a positive constant c such that
for all real numbers a,, b k , j , k = 1, . . ,n with it 5 IV
It is easy to check that if to is a function of bounded variation on [O,T] with total variation q ( u : ) . then w satisfies Assumption 2.3. In order to apply the result; of Section 3 to the stochastic case , we need the following lenuna .
Lemma 2.2. Let (a, A, P ) be a probabzlzty space wzth Jlltratzon Ft and let ru be a Ft-standard Wzener process on the space . Suppose that E{a: + h:} < m ,then w satzsjes Assumptzon 2.3 wzth probabzlzty 1 .
Proof. 1) Since w is a Wiener process on the space , we have
E { [ w ( t k ) - ~ ( t ~ - ~ ) ] ~ } = u 2 ( t k - t k - l ) = a z h
for some positive constant u and
E { w ( t k ) w ( t 3 ) } = 0 ; j # k .
Set
and we have
E{ l 6; - u 2 h 1') = E{tGir;lk] - u4h2.
From the definition of stochastic integrals (Doob[7] , p.427-434 ) , we have the formula
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1
(2.4)
Thus,
Therefore
and hence
with probability 1 . 2) Again by the same argument as above , we get
We have
We have applied 2.4 and 2.5 in the above calculations . Therefore
with probability 1 . 0
In the proof of the above lemma , the hypothesis that o is a constant , is crucial. The Holder inequality
{ E ( f 2 . 1 2 ) } 2 I E ( f 4 ) E ( 1 ) becomes an equality when f (w; 1 , s ) = w(w; t ) - w(w; s ) and it is a consequence of the formula,used in the construction of the stochastic integral when a is a constant.
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1036 BUI AN TON
In this section we shall consider the deternlinistic case of 1.1. The main result of the section is the following theorem.
Theorem 3.1. Let G, C be as in Assun~ption 2.1 and let w be a continuous function on [O, T] satisfying Assumption 2.3 . Then for any given
{[; u,; a,v) E K x U
Similarly for v and for w.
Lemma 3.1. Suppose all the hypotheses of Theorem 3.1 are satisfied . Let 16, i) be single-valued Lipschitz coiltii~uous ntappings of H x U into H . Suppose that
l l ~ ( y ; U N H + l l i ( y ; U H H I C{1 + I I Y I I H + IIuIIc(o,T;u))
Then for each n , there exists a solution yn of 3.2 . Moreover
lI~iL(t)lI2H I '(1 + IlEIlL + l l ~ ~ l I ~ ( o , ~ . ~ / , ) + l l ~ J ~ I l ? Z ( o , ~ , u ~ - l ~ }
uizth
yh ( t ) = yn for t E [ ( n - l ) h , n h ) ; n = 1, ., N .
C zs andependent of n , u,, n,v, w . Proof. The existence of a solution yn of 3.2 is well known . We shall now establish the estimate of the lemma .
We have
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1 1037
Since ( A ( y k ) , y k ) ~ 2 0 , we obtain by taking into account our hypotheses on G , L
k=l The different constants C are all independent of n.
From the discrete analog of the Gronwall lemma , we obtain
I \Y"/~; 5 C{l + 11 ' !11$ + I I1 l~ l l i (O ,T;Ul ) + I Ia~vI I i (O ,T;U~- l ) } The lemma is now an immediate consequence of the above estimate. 0
Lemma 3.2. Suppose all the hypotheses of Theorem 3.1 are satisfied and let yh be as i n L e m m a 3.1. T h e n
The c o n s t a i ~ t C is independent of h
Proof. We have from 3.2
l l y k - y k - l 1 1 ~ 2 + h ( A ( y k ) , yk - y k - ' ) ~ = h ( ~ ( ~ ~ ; U S ; a j u k ) , yk - y k - l ) H
+ ( L ( y k ; U S ; a 3 v k ) , yk - y k - l ) H ( ~ k - W k - ' )
But
( A ( ~ ~ ) , yk - y k - l ) ~ > l K ( y k ) - I ~ ( ~ ~ - ~ ) > 0
Therefore
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
BUI AN TON
\.Ve get , by using hypothesis (i) of Assumption 2.3
5 Y * - yk-'IIL c C { l + MIL + I I u , I I & ~ , T ; ~ , ~ + I I ~ i ~ I I t ( o , r ; a ~ - l l ) k=1
The lemma is now an immediate consequence of the above estimate .
and similarly for vh.
Lemma 3.3. S ~ p p o s e all the hypotheses of Theorem 3.1 are satzsfied and let yh be as zn Lemina 3.1 . Then
{ y h ; G ( Y ~ , u h , ~ ; r 3 u h ) ; L (yh ; fLh,]; ~ 1 ~ ) ) { Y ; G ( Y ; l l j ; T ~ V ) , i ( ~ ; u ] ; K ~ U ) )
z n
( ~ ~ ( 0 , T ; H ) n (LW(O, T ; H))weak.} x ( L 2 ( 0 , T : H ) ) e p n i , as h -r O+
Moreover
I Y IzI C { 1 + IIEIILH i- I I U ~ I I & O , T , I ~ ) + / I ~ ~ ~ ~ I I ~ C ( O T L ~ ~ - ~ ~ )
luzth y ( t ) E K for almost all t E [0, TI.
Proof. From the continuity of the L2(0 , T ; H)-norm ,we have
with yh ( s ) = 0 for J outside of [O,T] Thus .
(3.3) I Y I I 125 C i l + l l E l l & + l l"~112Cco.~;~, i , , + i I n u l l Z C ( o , ~ ; ~ ~ f - i ) l
It follows from Lemma 2.1 and from the above estimate that
By hypotheses ,both c, i are Lipschitz continuous in the aprropriate spaces and it is now clear that
Lemma 3.4. Let y k be as in Lemma 3.1 . Then
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. 11
C is a constant independent of h , n.
Proof. 1) From 3.2 , we have
2 hllA(~k)~li 5 C h ( l + 1/(11?f + IIUJ(\$(O,T;~) + I l~jvl(C(o,T;U~-1))
+ (L(yk ; US; r j u k ) , A(yk))~2irk
2) Taking the summation from k = 1 to k = n and then taking the square the two sides . we obtain
Applying hypothesis (ii) of Assumption 2.3 and we have
i.
I Cho2 C lIA(Y"1l;,lli(Yk; u:'; r,vk)l12H k=l
I t follows from 3.4 , the above estimate and from the estimates of Lemma 3.1 that
( 3 . 5 ) + Ch 2 I lA(Yk)lk k=1
The different constants C are all independent of h, n. 3) Since (a - c)' 2 0 , we obtain from 3.5
Therefore
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
BUI AN TON
L e m m a 3.5. Suppose all the hypotheses of Theorem 3.1 and of Lemma 3.1 are sat- isfied . Then there exists y ( t ) 6 K a.e. on [O,T] with A ( y ) € 81r (y ) such that
y ( t ) + A ( y ( s ) ) d s = [ + 41 ~ ( y ; u,; nJu )ds + 1 &(y; u,; nJu )dw
Moreover
1 Y 12 + I I ~ Y ) I I L ~ , T : H ~ 5 C{1 + \\t\\i + ~ \ u ~ \ ~ ~ ( o . ~ ; r S ) + I I ~ ~ U I I ~ C ( ~ . T ; U N - ~ ~ } ~
P~oo j . 1 ) From 3.2 , we get
We have . by taking subsequences
Hut A(yh) E aIK(yh ) and since yh -+ y in LZ(O. T ; H ) , we obtain from a standard argument of the theory of monotone operators that x = A ( y ) .
3) Since y k E k' , we have y,,(t) E k' for almost all t and since li: is a closed convex subset of H , we deduce that y(t) is in k' for almost all t .
All the other assertions are direct consequences of Lemma 3.1 , 3.2 and 3.3.
P r o o f of T h e o r e m 3.1.1) Since Q, C are U.S.C. set-valued mappings of L2(0,T: H ) x U into the closed convex subsets of LZ(O, T; H) , we may applj ~e approximate sr- lection theorem . There exist single-valued mappings Lipschitz continuous mapping.; G,, L , of L 2 ( 0 , T ; H) x U into L 2 ( 0 , T ; H) such that
the range of G , is contained in the convex hull of the range of Q . 0 Graph G, c Graph L7 + &(ball radius 1 about the graph ). Similarly for L F . 2) From Lemma 3.4 , we have y, such that
We have, by taking subsequences
y, -+ y in L2(0 , T ; H ) n (L"(0, T; H))weak*
It is clear that y ( t ) E K for almost all t E [0, TI .
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1 1041
Since 9 is an u.s.c. set-valued mapping of L2(0, T ; H ) x U into the closed convex subsets of L2(0,T; H ) , its graph is closed a i d therefore also weakly closed . It follows from our hypotheses on G that
G(yE; ~ l j ; r j v ) + G(y; uj; r j v ) in (L2 (0 ,T ; H))weak with G E G
On the other hand , we have
GE(z ; ~ j ; X ~ V ) C G ( z ; uj; T , V ) +&(unit ball about 9) . Therefore
G,(y,; uj; r j v ) + G(y; u,; Tjv) in (L2(0 ,T; H)),,,k.
Similarly for L,(y,; u,; r j v ) . Hence
(1' G ( ~ ; u,; njv)ds; L(y; uj; r j v ) d u ~ )
in (IJZ(O, T ; H))Leak for almost all t . 1'
3) We have
(\A( .; Y ~ ) ~ ~ L ~ ( o , T : H ) 5 C Since A is maximal monotone in L2(0,T; H ) and y , + y in LZ(O, T ; H ) , we get
The estimate of the theorem is now an immediate consequence of that of Lemma 3.5 .
The theorem is proved .
4. O P E N LOOP FOR INCLUSIONS WITH SINGULAR INPUTS
In this section we shall establish the existence of an open loop equilibrium of 2.2 -2.3.
The main result of the section is the following theorem . Theorem 4.1. Suppose d l tht- hypolheses of Theorem 3.1 are satrsfied and suppose that Assumptzon 2.2 zs verzfied . Then there exzsts
{ i ; ii) E R ( [ ; i L j ; r j i i ) x U
such that
J3( t ; c; i L J ; r jC) 5 J j ( t ; y; vj; a j S ) ; j = I , ., N
for all y E R(<; v,; ~ j i i ) and all v E U. R ( [ ; uj; r j v ) denotes the non-empty set of all solutions o f 2 2 with controls {u,; x,v}
First we have the following lemma.
Lemma 4.1. Suppose all the hypotheses of Theorem 4.1 are satisfied . Then for any given (4; u,, T,U} E IC x 24 , there exists jj E R ( [ ; u,; n,v) such that
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1042 BUI AN TON
j J ( [ ; u,; K , U ) = d { J , ( E ; y; u,; K,U) : v y E R(E; u,: x J v ) j
(4.1) = J J ( [ ; y: u,; r J u )
Pvoof. Let { y n } be a minimizing sequence of the optin~ization problem 4.1 with
From the estimates of Theorem 3.1 , we have
I yrL I z I c with C independent of n . Let n -+ cc then u e obtain a subsequence, which we denote again by yn such
tha t
yn + d in L 2 ( 0 , T : H) n (Lm(O,T; H) )meak - ; i 2.
The same proof as in that of Lemma 3.4 and of Theorem 3.1 shows that j E R((: ZL,; x,u).
Let n + m in 4.2 , then with our hypotheses on f,. g, , it is clear that
.?,(E; u,; ~ , v ) = JJ(E; d ; u,; x,v)
The lemma is proved . 0
Let
with j, as in Lemma 4.1
L e m m a 4.2. Suppose all the hypotheses of Theorem 4.1 arr satzsfied . Then there exzsts u* E U such that
Proof. 1 ) Let {u'" be a minimizing sequence of the optimization problem 4.4 with
Suppose that u'" u* in C ( 0 , T; U ) , then u* E U since U is a compact convex subset of C ( 0 , T; C r ) .With the control {u;; r , v ) , we obtain by applying Lemma 4.1
j j(E; ujn; K J V ) = J j ( l ; Y;; ujn; x j ~ )
From the estimates of Theorem 3.1 , we get by taking subsequences which we denote again by y;
yjn -+ y; in LZ(O, T ; H ) n ( L m ( 0 , T ; H))weak.
A proof exactly as in that of Lemma 3.4 gives y; E R([; u;; 5 v ) . With our hypotheses on f,, g, , we deduce that
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1
j j ( ( ; u;; n j v ) 5 J j ( [ ; y;; u;; n j v )
5 liminf j j ( ( ; u?; r j v ) .
2 ) Suppose that
U J ; X ~ V ) I -6 + ~ ~ ( 4 ; x:; 6;; T ~ V )
for a given 6 > 0 and for any 47 -+ U; in U with
Then , as in the first part we get
j l ( [ ; u;; T ~ V ) 5 -6 + J ~ ( ( ; 2; u;; S ~ V )
with x E R([; uj ; x jv) . Since
R((; u;; a lv ) = n {R((; u;; a j v ) : U ; E B,(u;)} C
we obtain , by taking the illfirnuin over R([; uj ; n j v )
j i ( ( ; u;; T ~ V ) = l i m j 3 ( ( ; u?; r j v ) = J j ( [ ; Y ; ; u;; T ~ V )
Therefore N
d ( ~ ) = C j j ( ( ; u;; T ~ v ) . ]=I
The lemma is proved 0
Let
(4.5) D ( v ) = {u* : u' as in Lemma 4.2) = {u* : *(u*; v ) 5 Q(u; v ) , V u E U }
Lemma 4.3. Suppose all the hypotheses of Theorem 4.1 are satisfied . Then D ( v ) is a non-empty closed subset of U . Proof. The set D ( v ) is non-empty by Lemma 4.2 . A proof as in that lemma shows that the set is closed .We shall not reproduce it . Lemma 4.4. Suppose all the hypotheses of Theorem 4.1 are satisfied . Then there exzsts a unique 4 E U such that
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1044 BUI AN TON
(4.6) g,(E; 4,; x,v) = a, = inf{!j,([; u,; x,v) : Vu E D ( v ) } ; 1 = l , . . , N .
Plaof. Let u ' q e a minimizing sequence of 4.5 with
and un -+ u in C ( 0 , T ; [I). Since un E D ( v ) and the set D ( v ) is closed , we get 11 E D ( v ) . . With our hypotheses on g,, , we have
gj(E; u,"; ~ j v ) -+ g,(E; u j ; ~ j u )
and hence gl ( ( ; u ] ; x,v) = aj It remains to show that the solution is unique . Suppose that 4 , f i are two elements of D ( v ) with
g j ( E ; f i j ; ~ j v ) = g j ( E ; f i j ; ~ I v ) = in f{g, ([; u,; T , V ) : Vu E D ( v ) }
By hypothesk , g,(E; .; x lv ) are 1.s.c. on LZ(O,T; U,) and hence the subdifferen- tiais as,([; .; x l v ) exist . We have
T T L { g J ( ( ; w l ; T J V ) - g J ( [ ; 'J; 2 L ( ~ ( ' 1 ) > w~ - G 3 ) ~ r j d t
for all w, E LZ(O,T; U,) with ~ ( t , ) E dg,(E; .; x,v).
Since C, is a minimum of g , over D ( v ) , we have
It is clear then that
l T ( X ( d J ) 3 ' j - ' J ) H ~ ' = O
Similarly
Therefore
By hypothesis , d g j ( ( ; .; x,v) is strictly maximal monotone and hence G I = * . U ] ; 1 = 1, ..,AT.
The lemma is proved . Let 7 be the nonlinear mapping of U into U given by
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1
where Q is the unique element of D ( u ) with
In view of Lemma 4.4 , the mapping 7 is well-defined.
L e m m a 4.5. Suppose all the hypotheses of Theorem 4.1 are satisfied and let 7 be as in 4.7 ,then 7 has a fixed point .
P~oo f . Since U is a compact convex subset of C ( 0 , T ; U ) , t o apply the Schauder fixed point theorem to the mapping 7 it suffices to show that 7 is continuous .
1) Let u" E U and let 7 u " = Qn . Then it follows from Theorem 3.1 that there exists y," E R([; Q;! .rr3vn).
We obtain by takmg subsequences
i 11
L ~ ( o , T ; H ) n ( ~ " ( 0 , T ; H) )wenk . x C ( O , T ; U ) .
for j = I , . . N. A proof as in that of Lemma 3.4 gives yj E R(E; u j ; ~ , v ) . 2 ) Recalling the definition of Q , we get by taking into account our hypotheses on
f 3 I 93
WF now show that C is in D ( I I ) . First we note that D ( u ) = nn D ( v n ) . Thus if 4 is in D(un) for n 2 no , then
Thus ,
for all 4 E D ( v ) . Therefore Q E D ( u ) and 7 u = Q. The lemma is now an immediate consequence of the Schauder fixed point theo-
rem 13
Proof of Theorem 4.1. Since 7 has a fixed point ii E U , we get from the definition of the mapping 7
Q ( i ; 1) 5 Q ( v ; 6 ) for all v E U It then follows from the definition of Q that
Take u = (wk, skG) in the above inequality and we get
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
j k ( ( ; f i k ; a k 6 ) 5 . j k ( f ; w k ; a&) for all wk E Uk
BUI AN TON
Repeating the process N - 1 times and we get the theorem .
The main results of the paper are the following theorems . T h e o r e m 5.1. Let ( R , A , P ) be a probability space , w be a F-s tandard Wiener pmcess on the space and suppose that the measure P is regular in the interior . Suppose that all the hypotheses of Theorem 3.1 are satisfied . Then for any given
{ E ; U } E L , ~ ( R , A, P ; H ) x L ~ ( R , A, P ; C ( O , T ; u)) with { ( ( w ) ; u ( . ; w ) ) E K x U a.s. , there exists a solution y of 1.1 with
The set of ali solutions of 1.1 is denoted by R,to([; u,; n,v).
T h e o r e m 5.2. Suppose all the hypotheses of Theorems 4.1 and 5.1 are verzfied Then there exists 6 E L2(61; A, P ; U ) and Q E R,t,((; C,; T,G) such that
E { J , ( [ ; i ; 6,; n,C)} 5 E { J , ( ( ; y; v,; n 3 C ) } ; 1 = I . . . A'
for all y E R,t,([; v,; x,fi) and all 21, E L 2 ( R , A, P ; U j ) .
To prove T h e o r ~ m 5.1 , we shall apply Von-Neumann theorem on sections of set ~ a l u e d mappings wi th closed graphs and closed images . L e m m a 5.1. Suppose all the hypoth~ses of Theorem 3.1 are satzsjied . Then for any gzuen {(; u,, n ,v)} E h' x U , the set-valued rnappzng R ( [ ; u,; n , v ) } of H x C ( 0 . T ; U) znto L2(0, T ; H ) has non-empty closed zmages . The graph of R ( ( ; u,; x,v) zs closed.
Proof. From Theorem Theorem 3.1 , we know that R(E; uj; n j v ) has a non-empty irnage in L2(0 , T; H ) . We now show that the graph of R is closed .
Suppose that
{ln; u:; a3vn) --+ { E ; u:; a J v )
in H x C(0,T; I J ) with [" K and {u;,a3un} E U . Since k: , U are closed convex subsets of H and of C ( 0 , T; U) , we have (4 ; u,, n,v) E
K x U . Let yn E R(En; u;; a 3 v n ) . From Theorem 3.1 , we have
C is a constant independent of n . Thus we have a subsequence, denoted again by yn such that
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1
with {G, L ) E 8, L . Since 8, L are u.s.c. set-valued mappings of L2(0, T ; H ) x C ( 0 , T ; U ) into the
closed convex subsets of L2(0 ,T; H) , their graphs are closed and hence weakly closed . Therefore
With yn -r y in L2(0 ,T; H ) and A(yn) -+ cP weakly in L2(0 ,T; H ) , we get by a standard argument of the theory of monotone operators @ = A(y) .
Therefore y E R ( ( ; uj; a j v ) and the graph is closed . ,
It remains to show that the image of R((; uj; ajv) is closed in L 2 ( 0 , T ; H ) . The proof is almost identical to the above arguments and we shall not reproduce it .
Proof of Theorem 5.1. 1 ) It follows from Lemna 2.2 that w satisfies Assumption 2.3 with probability 1 . It follows from Von Neumann theorem and from Lemma 5.1 that R ( ( ; uj; a j v ) has a section which is universally Radon measurable. We shall call such section R([; u,; s j v ) .
Let A be the mapping
2) Since P is assumed to be regular in the interior , there exists a compact set Kn c R such that
1 P ( R - K n ) 5 -.
n Without loss of generality we may assume that I(,, is an increasing sequence .
Since A is a random variable , the restriction An of A to I(, is continuous. The Radon measure P induces on I(, a Radon measure Pn and thus , An(Pn) is a Radon measure on H.
With R ( ( ; 11,; a,v) being universally Radon measurable , R((; uj; s j v ) o A is 1'-111casura.ble of I<,, ill L"(O,1'; 11) .
3) Let ~ ( w ) = R ( ( ; uj; a j v ) o A(w)
and let yn(w) = y(w) if w E IC, ; yn(w) = 0 if w 4 Kn.
Then yn as mappings of R into L2(0 ,T; H ) are measurable and
yn(w) -4 y(w) VW E R - UnKn i.e. a.s w E R
since P(UnKn) = 1. 4) We have
y(w) = R ( ( ; uj; x j v ) o A(w)
E R((; uj; a j u ) . Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
BUI AN TON
The theorem is proved . To prove Theorem 5.2 we shall need the following lemma
L e m m a 5.2. Supposr all the hypotheses of Theorem 4.1 are satisfied . Let M be the set-valued mapping
M ( [ ) = { { c , G} : Q E R(E; G,; .rrj i)}
whe~e J,(<: i ; 6,; nJG) = inf{J,(E; u,; n , ~ ) : V u E U).
Then M zs a set-valued mappzng of K C H znto the closed subsets of L2 (0 , T ; H ) x C ( 0 , T ; U ) ) wzth closed graph . Proof. 1 ) It is clear from Lemma 4.1 that the set iZ/1 is non-empty . We now show that the graph of M is closed.
Let E n E K with tn + E in H. Suppose that { y n , u n ) E M(En) and that
{ y n , un } -+ { i , C } in L2(0, T ; H ) x C ( 0 , T ; U).
Since yn E R(En; uj"; n j u n ) , it follows from the estimates of Theorem 3.1 that
By taking subsequences ,we obtain as in the proof of Lemma 5.1
{ 4 y n ) , G ( y n ; u n ) , L ( y n ; u n ) } + ( 4 6 1 , G ( i : f i ) , L ( i ;
in (Lm(O, T ; H ) ) w e a p x ( L 2 ( 0 , T ; H))Leal, . Moreover i E R(E; GI; x,fi) With our hypotheses on j,, gl , we get
J l ( t n ; yn; u;; n,ujn) + J , ([ ; i; 12,; T , G ) .
2) We now show that
J,(E; 9; Gj; n1&) = i n f { J j ( E ; y; u j ; n , u ) : V u E U). Since {y'" u"} E M(En) , we have by definition
vn E U ; xn E R ( [ " ; v:; ?r1vn)
such that
J,(En; yn; u:; n3un) 5 Jj(En; xn; v;; n1vn).
From the estimates of Theorem 3.1 we get 1 xn Jz< C. As before , by taking subsequences we have
{ x n , v n } + { r , v} in L2(0 , T ; H) x C ( 0 , T ; U )
with x E R(E; vj; .rrjv). Moreover
J j ( t n ; xn; v;l; ?rjvfL) + J j ( t ; x ; v j ; n,v)
Conibining with the first part , we get
J,(E; i ; G,; x,G) 5 J , ( E ; xi v,; n p )
3 ) Let t 6 E B s ( [ ) where B s ( ( ) is a ball of radius 5 in H about the point 6 . Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. 11
Then from the previous part we get
J,([; &; 6 j ; K ~ G ) 5 E $ JJ(&; 56; v3; K ~ U )
for all xg E R(E6; vj; s jv) . Hence
J,([; 9; 6;; T,O) 5 inf{J,([s; 2 6 ; vj; K ~ V ) :
V x6 E R(&; vj; nju) ; Vv E U) 5 inf{Ji([; x; v,; K ~ U ) : Vx E R([; wj; R ~ V ) ,
vv E U}
It follows that
Therefore the graph of M is closed . 4) Now an almost identical proof shows that the image of M is closed in L2(0, T ; H ) x
C ( 0 T ; L ) 0
Proof of Theorem 5.2. From Lemma 5.2 and from Von Neumann theorem , we know that M has a universally Radon measurable section. Let A be the mapping of R -+ Iil given by
=
and set r = M o A .
Repeating the proof of Theorem 5.1 and we get the stated result .
6. OPTIMAL ALLOCATION O F RESOURCES IN T H E FIGHT AGAINST DRUG USE
In this section we shall apply Theorem 5.2 to the study of an optimal allocation of resources in the fight against drug use. Mathematical models of the drug market have been studied by Caulkin [4] ,by Dawid and Feichtinger [6] and by others in a special issue of Mathematical and Computer Modelling [5] . While most of the models are descriptive,the one introduced by Dawid and Feichtinger involved an optimization problem leading to a 2-person differential game and an open loop Nash equilibriurn. For specific functiorial forms and with 2-player, explicit calculations in the deterministic case are possible and a careful analysis of the problem was made by Dawid and Feichtinger .
As an application of Theorem 5.2 we shall consider the case of N-player with functional forms closer to the standard diffusion dynamics and with a white nolse effect . The model seems new .
Let f j be the population size of the country and let y ( t ) be the stock of drug users at the time t , 0 5 y ( t ) 5 f j ; t E [O,T]. The change in the stock of drug users is governed by : (i) the activities of the drug dealers , (ii) the death of drug users and (iii) the rehabilitation of the addicts .
We shall assume that Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
1050 BUI AN TON
r The dynamics of the stock variable is a diffusion one which is modelled by the logistic growth function y (y - y) .
r Let udealer denote the efforts of the drug dealers in attracting new costumers and in selling drugs with 0 5 ud,,,,, 5 Ed,,l,,. We shall assume that the marginal effect of the efforts of the drug dealers decrease with an increase in the level of efforts and shall model the growth term in the system dynamics by l/lldenie?-(~(y).
r 'l'he population ot add~c t s decreases with a death rate D and is reduced by medical rehabilitation . The government has two ways of keeping the stock of drug users down .
The first strategy is through medical treatment . The effect of the treatment depends on the allocated government budget on drug fighting. The governmmt effort through medical rehabilitation is measured by the control variable um,d .We shall model the effect of the medical treatment by the set-valued mapping
r ( y ; u ) = 6 4 ~ ) if umed 2 UL = X & Z S ( y ) if 0 5 L - d < U m d < ~,,d ; E [O, 11 = 0 if 0 5 um,d I Cm,d.
Thus when the budget allocated to the rehabilitation program is below certain level , we are assuming that the effect of the program is negligible and it has a significant impact only when the budget is above a level uLed. We also assumed that the marginal effect of the government effort decreases with the level of effort .
We shall denote by
The second strategy that the government can use in the fight against drug users is through law enforcements : destruction of crops , seizure raids , custom inspections
r 7 . 1 he government effort is measured through the control variable uenf,,, = u3 and obviously depends on the budget allocation and is modeled by the set-valued mapping
We assume that the effect is negligible when the budget and therefore the effort is below a critical level and has a maximum impact only when the budget is a t the level u : , , ~ ~ , , ~ Since those law enforcements measures have to be carried out in a random fashion in order to have any impact , we shall consider t the stochastic term Ldw where w is a Wiener process to model the impact of law enforcements measures .
We shall consider the stocl~astic inclusion
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1
y ( ; 0 ) = J ; E E N,yl. Throughout the section we shall take
H = R , U, = R ; j = 1 , 2 , 3 ; X: = [0, y]
and we shall make R into a Hilbert space with the usual inner product and norm
U is a compact convex subset of ( C ( 0 , T ; R ) ) 3
Lemma 6.1. Let a , -0, y be continuous positive convex functions with
for all y. . Let R , A , P ) be a probabilzty space and suppose that the probabzlzty measure P zs r~gular zn the znterzor . Let w be a p-s tandard Wzener process on the space . Then for each
{ J ; Udralerr U m e d , uenjorc} E L 2 ( f l , A, P ; H ) X ( L 2 ( a , A> ~ ( O I T ; R ) ) l 3
with { t i udealerr U m e d , ~ e n f o r c } E X U
a.e. on [O,T] and a.s. there exists y E L 2 ( R , A, P : Lm(O,T; W ) ) , solution of 6.3 . Moreover
Proof. It is trivial to check that -9, L given by 6.1-6.2 are u.s.c. set-valued map- pings of L2(0 , T; H ) x U into the closed convex subsets of L2(0 , T ; H ) . Now a direct application of Theorem 4.1 gives the stated result .
We shall associate with 6.3 cost functionals based on the following assumptions . 0 The income function g ~ ~ ~ l e T ( y ) of the drug dealers is a positive convex increasing
function. The cost function gi$,, to the drug dealers (besides their fixed costs) are due
mainly to the law enforcement actions and to their own competitions . Without police actions and without any competition among the drug dealers , the
cost to the dealer group will be zero , i.e.
The other obvious observation is that if the group is inactive ,then the cost is zero , 1.e.
We shall assume that the cost to the dealer will increase linearly with the effort of the law enforcement agencies .
With the above underlying assumptions we propose to use for cost functions gSf;tl,, ,the functions
where uen,,,,(y) > 0 measures the efficiency of the law enforcement actions , udeoler(y) is the effect of the group competition among the dealers and (1 - A)u3 is the effort of the various law enforcement agencies .
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
BUI AN TON
We shall assume that venjOTc(y) , vd,,le,(y) are of the forms
where v l , vz are two positive constants . Let
The government costs are of three different types . The medical costs due to long term medical illness caused by the use of drugs
and the medical rehabilitation costs. The loss of tax revenues to the lack of productivity of the drug addicts .
We shall assume that both costs increase with the drug addict population and take the form
where B is a positive continuous function on [O,T]. The cost of the law enforcements is assumed to be
gs(y; u l ; U Z ; u3) = & ( Y ) u ~
where S(y ) is a stritcly convex function of y..
Lemma 6.2. Let g,; j = 1,2 ,3 be as above . Then g, satisfies Assumptzon 2.2
Proof. The proof is trivial .
Let
Applying Theorem 5.2 , we obtain the following result.
Theorem 6.1. Suppose all the hypotheses of Lemma 6.1 are satisfied an,d let J, be as in (6.4) . Then for any given J E L 2 ( R , A , P ; H ) with ( E K a.s. , there exists
4 E U ; j j E R,t,(J; 4,; n j 4 )
with
j j E L 2 ( R , A, P ; Lm(O, T ; H ) ) ; E X: a.e. on [0, T1,a.s.
Moreover
E { J j ( E ; 4 j ; n,C)) 5 E { J 3 ( J ; Y ; uj; r j Q ) )
for all y E R,to(J; 0,; r j 4 ) , all u, E Uj ; j = 1 ,2 ,3 .
Thus there exists a control
= (4deale?, k e d , 4enfmc) E U Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014
N-PERSON STOCHASTIC INCLUSIONS. I1 1053
which will minimize the medical costs and maximize the impact of the law enforce- ments effort and reduce the income of the drug dealers to a minimum.
1. J.P.Aubin, Mathematical methods of game and economrc theoy,North Holland, 1979. 2. A.Bensoussan ,Results on Stochastic Navter-Stokes Equations, Control in partial differential
equations. Lecture notes in pure and applied mathematics. 165 , Marcel Dekker ,1994,pp.ll-21. 3. A.Bensoussan and R.Temam, Equations stochastiques du type Nauier-Stokes, J . Functional Anal-
ysis 1 3 (1973),195-222 4. J.P.Caulkins,Local drug markets response to focused police enforcement, Operations Research,
4 1 (1993) ,848-863. 5. J.P.Caulkins , Editor.Mathematica1 models of drug markets and drug policy, Math Com-
put.Modelling 17(1993),1-115. 6. H.Dawid and G. Feichtinger ,Optimal allocation of drug control efforts: a dafferenttal game
analysts,J,Optimization Theory and Applications .91(1996) ,279-297. 7. J.L.Doob,Stochastic Processes,John Wiley ,1953. 8. Bui An Ton, Open loop equilrbrium strategy for qutast-uartational inequalzties and for constratned
non-cooperatwe games, Numerical Functional Analysis and Optimization 7(1996),1053-1091. 9. Bui An Ton,Optimal control for N-p~rson dzfferenttal stochastzc incluszons,Stochastic Analysis
and Applications.17 ,6, (1999) ,911-935. 10. Bui An 'Ton ,Stochastic Navter-Stokes equations, International J.Applied Math.l ,(1999)
Dow
nloa
ded
by [
Nor
thea
ster
n U
nive
rsity
] at
20:
18 1
5 N
ovem
ber
2014