Applied Mathematics and Computation 164 (2005) 277–294

www.elsevier.com/locate/amc

Set differential equations versus

fuzzy differential equations
V. Laksmikantham
Department of Mathematical Sciences, Florida Institute of Technology, Melbourne, FL 32901, USA

Abstract

The study of fuzzy differential equations (FDEs) forms a suitable setting for a math-
ematical modeling of real world problems in which uncertainties or vagueness pervades.
In recent years, the theory of FDEs has been investigated extensively in the original for-
mulation as well as in an alternative framework, which leads to ordinary multivalued
differential inclusions. It has recently been realized that initiating the study of set differ-
ential equations in a metric space has several advantages, in addition to providing a nat-
ural setting for considering FDEs. In this paper, we present some interesting results in
this direction with the necessary background material.

2004 Elsevier Inc. All rights reserved.

Keywords: Set differential equations; Fuzzy differential equations; Existence and uniqueness; Flow
invariance; Connecting link between SDE and FDE

1. Introduction

In mathematical modeling of real world problems, we encounter two incon-

veniences, namely, the complexity and the uncertainty or vagueness. In order to
take vagueness into consideration, fuzzy set theory is considered as a suitable

E-mail address: lakshmik@fit.edu

0096-3003/$ - see front matter
2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2004.06.068
278 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

setting. In recent years, the framework for the study of fuzzy differential equa-
tions (FDEs) has been developed and several basic properties of solutions of
such equations are now available [5,7,9,11]. The formulation of FDEs has a cer-
tain disadvantage, since the diameter of solutions increases as the time increases,
because of fuzzification of the differential operator. Recently, an alternative for-
mulation is suggested to replace the study of FDEs by a sequence of ordinary
multivalued differential inclusions which can be obtained by suitable level sets
[6]. Another alternative is to go a step further to generate set differential equa-
tions (SDEs) in a metric space from the multivalued mapping and investigate
the connection between FDEs and SDEs. This setting of SDEs appears more
natural and has several advantages, if studied as an independent subject area.
Very recently [1,10], such an investigation is in progress and in this paper, we
present some ideas and results.
We give in Section 2, the needed preliminaries for the independent formu-
lation of SDEs, pointing out its advantages. Section 3, deals with the
necessary background for FDEs and the suggested alternative model. In
Section 4, we consider the existence of generalized solutions of SDEs, first
considering without any continuity assumptions and then deriving from such
results, the classical solutions when the continuity is imposed. These two cases
discussed being the extreme situations, the question that is open is to prove
when the mapping involved is upper semicontinuous. It appears that no direct
proof, such as in the case of multivalued differential inclusions, seems to be
possible. In fact, even in the special case of ordinary differential equations,
the corresponding existence result is not available. Section 4 also contains
flow invariance result for SDEs in the framework nonsmooth analysis. All
the results of this section are extensions to SDEs of the corresponding results
for ordinary differential equations via nonsmooth analysis [2]. Finally, Section
5 is dedicated to provide the relationship between FDEs and SDEs. For this
purpose, we first recall needed known results and then embark to find the
connecting links.

2. Preliminaries

Let K c ðRn Þ denote the collection of all nonempty, compact and convex sub-
sets of Rn . Define the Hausdorff metric


D½A; B
¼ max supx2B dðx; AÞ; supy2A dðy; BÞ ; ð2:1Þ

where d[x, A] = inf [d(x, y) : y 2 A], A, B are bounded sets in Rn . We note that
K c ðRn Þ with this metric is a complete metric space.
It is known that if the space K c ðRn Þ is equipped with the natural algebraic
operations of addition and nonnegative scalar multiplication, then K c ðRn Þ
 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 279

becomes a semilinear metric space which can be embedded as a complete cone

into a corresponding Banach space.
The Hausdorff metric (2.1) satisfies the following properties:
D½A þ C; B þ C
¼ D½A; B
and D½A; B
¼ D½B; A
; ð2:2Þ

D½kA; kB
¼ kD½A; B
; ð2:3Þ

D½A; B
6 D½A; C
þ D½C; B
ð2:4Þ

n
for all A; B; C 2 K c ðR Þ and k 2 Rþ .
Let A; B 2 K c ðRn Þ. The set C 2 K c ðRn Þ satisfying A = B + C is known as the
geometric difference of the sets A and B and is denoted by the symbol A  B.
We say that the mapping F : I ! K c ðRn Þ has a Hukuhara derivative DHF(t0) at
a point t0 2 I, if
F ðt0 þ hÞ  F ðt0 Þ F ðt0 Þ  F ðt0  hÞ
lim ; and lim
h!0þ h h!0þ h
n
exist in the topology of K c ðR Þ and are equal to DHF(t0). Here I is any interval
in R.
By embedding K c ðRn Þ as a complete cone in a corresponding Banach space
and taking into account the result on the differentiation of Bochner integral, we
find that if
Z t
F ðtÞ ¼ X 0 þ UðsÞds; X 0 2 K c ðRn Þ; ð2:5Þ
0

where U : I ! K c ðRn Þ is integrable in the sense of Bochner, then DHF(t) exists

and the equality
DH F ðtÞ ¼ UðtÞ; a:e: on I; ð2:6Þ
holds. Also, the Hukuhara integral of F is given by
Z Z 
F ðsÞds ¼ f ðsÞds: f is a continuous selector of F
I I

for any compact set I  Rþ .

With these preliminaries, one can consider the set differential equation
DH U ¼ F ðt; U Þ; U ðt0 Þ ¼ U 0 2 K c ðRn Þ; t0 P 0; ð2:7Þ
n n
where F 2 C½Rþ  K c ðR Þ; K c ðR Þ
.
The mapping U 2 C 1 ½J ; K c ðRn Þ
, J = [t0, t0 + a] is said to be a solution of
(2.7) on J if it satisfies (2.7) on J. Since U(t) is continuously differentiable,
we have
Z t
U ðtÞ ¼ U 0 þ DH U ðsÞds; t 2 J : ð2:8Þ
t0
280 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

Thus, we associate with the initial value problem (IVP) (2.7) the following
Z t
U ðtÞ ¼ U 0 þ F ðs; U ðsÞÞds; t 2 J ; ð2:9Þ
t0

where the integral is the Hukuhara integral.

Observe also that U(t) is a solution of (2.7) if and only if it satisfies (2.9)
on J.
We need the following properties of the Hukuhara integral.
If F : ½t0 ; T
! K c ðRn Þ is integrable, we have
Z t2 Z t1 Z t2
F ðtÞdt ¼ F ðtÞdt þ F ðtÞdt; t0 6 t1 6 t2 6 T ð2:10Þ
t0 t0 t1

and
Z T Z T
kF ðtÞdt ¼ k F ðtÞdt; k 2 R: ð2:11Þ
t0 t0

Also, if F ; G : ½t0 ; T
! K c ðRn Þ are integrable, then D½F ðÞ; GðÞ
: ½t0 ; T
! R is
integrable and
Z t Z t  Z t
D F ðsÞds; GðsÞds 6 D½F ðsÞ; GðsÞ
ds: ð2:12Þ
t0 t0 t0

We observe that
D½A; h
¼ kAk ¼ sup kak; ð2:13Þ
a2A

for A 2 K c ðRn Þ, where h is the zero element of Rn which is regarded as a one

point set.
Moreover, we define for A; B 2 K c ðRn Þ,
hA; Bi ¼ supfða:bÞ : a 2 A; b 2 Bg ð2:14Þ
so that we have
2 2 2
kA þ Bk 6 kAk þ kBk þ 2hA; Bi: ð2:15Þ
The investigation of set differential equation (2.7) as an independent subject
has some advantages. For example, when U(t) is a single valued mapping, it is
easy to see that the Hukuhara derivative and the integral reduce to the ordinary
vector derivative and the integral and therefore, the results obtained in this new
framework of (2.7) become the corresponding results of ordinary differential
systems. Also, we have only a semilinear complete metric space to work with,
in the present setup, compared to the complete normed linear space that one
employs in the usual study of ordinary differential systems. Furthermore, set
differential equations, that are generated by multivalued differential inclusions,
when the multivalued functions involved do not possess convex values, can be
 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 281

used as a tool for studying multivalued differential inclusions. See [12]. More-
over, one can utilize set differential equations of the type (2.7) to investigate
profitably fuzzy differential equations as in [3,4,11], since the original formula-
tion of fuzzy differential equations [7,8,11] suffers from disadvantage and does
not reflect the rich behavior of corresponding differential equations without
fuzziness. This is due to the fact that the diameter of any solution of fuzzy dif-
ferential equation increases as time increases because of the necessity of the
fuzzification of the derivative involved.

3. Fuzzy differential equations

Let us denote by En = [u : Rn ! [0, 1] such that u satisfies (i) to (iv) given

below].

(i) u is normal, that is, there exists an x0 2 Rn such that u(x0) = 1;

(ii) u is fuzzy convex, that is, for x, y 2 Rn and 0 6 k 6 1,
uðkx þ ð1  kÞyÞ P minðuðxÞ; uðyÞÞ;
(iii) u is upper semicontinuous (usc);
(iv) [u]0 = cl[u 2 E n : u(x) > 0] is compact.

For 0 < a 6 1, we denote [u]a = [x 2 Rn : u(x) P a]. Then from (i) to (iv), it
follows that the a-level set [u]a 2 Kc(Rn) for 0 6 a 6 1. We set

D0 ½u; v
¼ sup D½½u
a ; ½v
a
ð3:1Þ

06a61

which defines a metric in En and (En, D0) is also a semilinear complete metric
space [5]. Also, D0[u, v] satisfies similar properties as D[A, B] listed in (2.2),
(2.3) and (2.4). One can also define Hukuhara derivative DHF(t) 2 En for a
fuzzy map F : J ! En, where J = [t0, t0 + a], a > 0, similarly as before, taking
limits in the metric space (En, DR0). Moreover, if F : J ! En is continuous, it is
t
integrable, the integral GðtÞR¼ t0 F ðsÞds is differentiable and DHG(t) = F(t).
t
Furthermore F ðtÞ ¼ F ðt0 Þ þ t0 DH F ðsÞds, t 2 J.
In this setup, the IVP for fuzzy differential equations originally proposed is
of the type
DH u ¼ f ðt; uÞ; uðt0 Þ ¼ u0 2 En ; ð3:2Þ

where f 2 C [R+ · En, En], for which basic results are discussed. See [7,8,11].
This approach is based on the fuzzification of the differential operator, whose
values are in En and therefore suffers from the disadvantage, since the solution
u(t) of (3.2) has the property that diam[u(t)]a is nondecreasing as t increases.
282 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

Consequently, this formulation can not fully reflect the rich behavior of solu-
tions of corresponding ordinary differential equations.
Recently, Hüllermeier [6] has suggested an alternative formulation of fuzzy
IVPs by replacing the RHS of a system of ordinary differential equation by a
fuzzy function
f : Rþ  R n ! E n ; ð3:3Þ
n
and with the initial fuzzy set x0 2 E , so that one can consider the fuzzy multi-
valued differential inclusion

x0 2 f ðt; xÞ; xðt0 Þ ¼ x0 2 En ; ð3:4Þ

on J, where now f is defined from R+ · Rn ! En rather than R+ · En ! En as in

(3.2). However, instead of (3.4), a sequence of multivalued differential
inclusions
b
x0b 2 F ðt; xb ; bÞ; xb ðt0 Þ 2 ½x0
; 06b61 ð3:5Þ

is investigated on J, where F(t, x; b) ” [f(t, x)]b and F ðt; x; 0Þ ¼ suppðf ðt; xÞÞ.
The idea is that the set of all solutions Sb(x0, T), t0 6 t 6 T, would be b-level
of a fuzzy set S(x0, T), in the sense that all attainable sets Ab(x0, t), t0 < t 6 T,
are levels of a fuzzy set on Rn. Considering S(x0, T) to be the solution of (3.2)
thus captures both uncertainty and the rich behavior of differential inclusion in
one and the same technique.

4. Existence and flow invariance

We consider the initial value problem (IVP) for set differential equation

DH U ¼ F ðt; U Þ U ðt0 Þ ¼ U 0 2 K c ðRn Þ; ð4:1Þ

where F is any function from ½t0 ; T
 K c ðRn Þ ! K c ðRn Þ. Let

p ¼ ½t0 ; t1 ; . . . ; tN ¼ T
ð4:2Þ

be a partition of [t0, T]. Consider the interval [t0, t1]. Observe that the right hand
side of the set differential equation on [t0, t1] is constant, namely,

DH U ðtÞ ¼ F ðt0 ; U 0 Þ; U ðt0 Þ ¼ U 0 ;

which has clearly a unique solution U(t) = U(t, t0, U0) on [t0, t1]. Define the node
U1 = U(t1)and iterate next by considering on [t1, t2] the IVP

DH U ¼ F ðt1 ; U 1 Þ U ðt1 Þ ¼ U 1 2 K c ðRn Þ:

 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 283

The next node is U2 = U(t2) and we proceed this way till an arc Up = Up(t)
has been defined on all [t0, T]. We employ the notation Up to emphasize the role
played by the particular partition p in determining Up which is the Euler poly-
gonal arc corresponding to the partition p. The diameter lp of the partition p is
given by
lp ¼ max½ti  ti1 : 1 6 i 6 N
: ð4:3Þ

Definition 4.1. By an Euler solution of (4.1) we mean any arc U = U(t) which is
the uniform limit of Euler polygonal arcs U pJ , corresponding to some sequence
pJ such that pJ ! 0, where this means the convergence of the diameters
lpJ ! 0 as J ! 1.
Clearly the corresponding number NJ of the partition points in pJ must then
go to infinity. Obviously, the Euler arc satisfies the initial condition U(t0) = U0.
We can now prove the following result on existence of an Euler solution for
(4.1).

Theorem 4.1. Assume that

(i) D½F ðt; AÞ; h
6 gðt; D½A; h
Þ; ðt; AÞ 2 ½t0 ; T
 K c ðRn Þ where g 2 C½½t0 ; T

Rþ ; Rþ
gðt; uÞ is nondecreasing in (t, u).
(ii) the maximal solution r(t) = r(t, t0, u0) of the scalar differential equation
u0 ¼ gðt; uÞ; uðt0 Þ ¼ u0 P 0 ð4:4Þ
exists on [t0, T].

Then,

(a) there exists at least one Euler solution U(t) = U(t, t0, U0) to the IVP (4.1),
which satisfies a Lipschitz condition;
(b) Any Euler solution U(t) of (4.1) satisfies the relation
D½U ðtÞ; U 0
6 rðt; t0 ; u0 Þ  u0 ; t 2 ½t0 ; T
; ð4:5Þ
where u0 = D[U0, h].

Proof. Let p be the partition of [t0, T] defined by (4.2) and let Up = Up(t)
denote the corresponding arc with nodes of Up represented by U0, U1, . . . ,UN.
Let us set Up(t) = Ui(t) on ti 6 t 6 ti+1, i = 0, 1 . . . ,N1, and observe that
Ui(ti) = Ui, i = 0, 1, . . ., N1.
On the interval (ti, ti+1) we have
D½DH U p ðtÞ; h
¼ D½F ðti ; U i Þ; h
6 gðti ; D½U i ; h
Þ: ð4:6Þ
284 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

On the interval [t0, t1], we have

 Z t 
D½U 1 ðtÞ; U 0
¼ D U 0 þ F ðt0 ; U 0 Þds; U 0
t0
Z t  Z t
¼D F ðt0 ; U 0 Þds; h 6 D½F ðt0 ; U 0 Þ; h
ds
t0 t0
Z t Z t
6 gðt0 ; D½U 0 ; h
Þds 6 gðs; rðsÞÞds
t0 t0

¼ rðt; t0 ; D½U 0 ; h
Þ  D½U 0 ; h

6 rðT ; t0 ; D½U 0 ; h
Þ  D½U 0 ; h
¼ M:
Here we have employed the properties of the metric D and the integral,
monotone character of g(t, u) in (t, u) and the fact that r(t, t0, U0) P 0 is nonde-
creasing in t. Similarly on [t1, t2], we get
 Z t 
D½U 2 ðtÞ; U 0
¼ D U 1 þ F ðt1 ; U 1 Þds; U 0
t1
 Z t1 Z t 
¼ D U0 þ F ðt0 ; U 0 Þds þ F ðt1 ; U 1 Þds; U 0 Þ
t0 t1
Z t1 Z t 
¼D F ðt0 ; U 0 Þds þ F ðt1 ; U 1 Þds; h
t0 t1
Z t1 Z t
6 D½F ðt0 ; U 0 Þ; h
ds þ D½F ðt1 ; U 1 Þ; h
ds
t0 t1
Z t1 Z t
6 gðs; rðsÞÞds þ gðs; rðsÞds
t0 t1
Z t
¼ gðs; rðsÞÞds 6 rðT ; t0 ; D½U 0 ; h
Þ  D½U 0 ; h
¼ M:
t0

Proceeding in this way, we arrive at

D½U i ðtÞ; U 0
6 rðT ; t0 ; D½U 0 ; h
Þ  D½U 0 ; h
¼ M; on ½ti ; tiþ1
:

Hence it follows that

D½U p ðtÞ; U 0
6 M; on ½t0 ; T
:

Also, from (4.6) we obtain,

D½DH U p ðtÞ; h
6 gðt; rðtÞÞ 6 gðT ; rðT ÞÞ ¼ r0 ðT ; t0 ; D½U 0 ; h
Þ ¼ k; say:

 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 285

Consequently, using similar arguments, we can find for t0 6 s 6 t 6 T,

Z t Z s
D½U p ðtÞ; U p ðsÞ
6 D½F ðn; U p ðnÞ; h
dn þ D½F ðn; U p ðnÞÞ; h
dn
Zt0t Z s t0

6 gðs; rðsÞÞds þ gðs; rðsÞds

Zt0 t t0

¼ gðs; rðsÞÞds ¼ rðtÞ  rðsÞ ¼ r0 ðrÞ j t  s j 6 k j t  s j

s

for some r such that s 6 r 6 t, proving Up(t) is Lipschitz of rank k on [t0, T].
Now, let pJ be a sequence of partitions of [t0, T] such that pJ ! 0, that is
such that xpJ ! 0 and therefore NJ ! 1. Then the corresponding polygonal
arcs U pJ on [t0, T] all satisfy
U pJ ðt0 Þ ¼ U 0 ; D½U pJ ðtÞ; U 0
6 M and D½DH U pJ ðtÞ; h
6 k on ½t0 ; T
:

Hence the family fU pJ g is equicontinuous and uniformly bounded and as a

consequence Ascoli–Arzela theorem guarantees the existence of a subsequence
which converges uniformly to a continuous function U(t) on [t0, T] and thus
absolutely continuous on [t0, T]. Thus, by definition U(t) is an Euler solution
of the IVP (4.1) on [t0, T] and the claim of the theorem follows.
The inequality (4.5) in part (b) is inherited by U(t) from the sequence of
polygonal arcs generating it when we identify T with t. Hence the proof is
complete. h

If F(t, U) in (4.1) is assumed to be continuous, then one can show that U(t)
actually satisfies the IVP (4.1).

Theorem 4.2. Under the assumptions of Theorem 4.1, if we suppose in addition

that F 2 C½½t0 ; T
 K c ðRn Þ; K c ðRn Þ
, then U(t) is a solution of (4.1).

Proof. Let U pJ ðtÞðtÞ denote a sequence of polygonal arcs for IVP (4.1) converg-
ing uniformly to an Euler solution U(t) on [t0, T]. Clearly, the arcs U pJ ðtÞ all lie
in BðU 0 ; MÞ and satisfy a Lipschitz condition of the same rank k.
Since a continuous function is uniformly continuous on compact sets, for
any given
> 0, one can find a d > 0 such that
t; t 2 ½t0 ; T
; U ; U  2 B½U 0 ; M
; j t  t j< d; D½U ; U 
< d

implies D½F ðt; U Þ; F ðt ; U  Þ
<
:

Let J be sufficiently large so that the partition diameter lpJ satisfies lpJ < d and
klpJ < d. For any t, which is not one of the finitely many points at which U pJ ðtÞ is
a node, we have DH U pJ ðtÞ ¼ F ð~t; U pJ ð~tÞÞ for some ~t within lpJ < d of t. Since
D½U pJ ðtÞ; U pJ ð~tÞ
6 klpJ < d;
286 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

we get
D½DH U pJ ðtÞ; F ðt; U pJ ðtÞÞ
¼ D½F ð~t; U pJ ð~tÞÞ; F ðt; U pJ ðtÞÞ
<
:
It follows that for any t 2 [t0, T], we obtain,
 Z t 
D U pJ ðtÞ; U pJ ðt0 Þ þ F ðs; U pJ ðsÞÞds
t0
 Z t Z t 
¼ D U pJ ðt0 Þ þ DH U pJ ðsÞds; U pJ ðt0 Þ þ F ðs; U pJ ðsÞÞds
t0 t0
Z t Z t 
¼D DH U pJ ðsÞds; F ðs; U pJ ðsÞÞds
t0 t0
Z t
6 D½DH U pJ ðsÞ; F ðs; U pJ ðsÞÞ
ds 6
ðt  t0 Þ 6
ðT  t0 Þ:
t0

Letting J ! 1, we have from this,

 Z t 
D U ðtÞ; U 0 þ F ðs; U ðsÞÞds <
ðT  t0 Þ:
t0

Since
is arbitrary, it follows that

Z t
U ðtÞ ¼ U 0 þ F ðs; U ðsÞÞds; t 2 ½t0 ; T
;
t0

which implies that U(t) is C1 and therefore

DH U ðtÞ ¼ F ðt; U ðtÞÞ; U ðt0 Þ ¼ U 0 ; t 2 ½t0 ; T
:
The proof is therefore complete. h

Remark 4.1. We can extend the notion of Euler solution of (4.1) from the
interval [t0, T] to [t0, 1), if we define F and g on [t0, 1) instead of [t0, T] and
assume that the maximal solution r(t) exists on [t0, 1) and show that Euler
solution exists on every [t0, T] where T 2 (t0, 1).
Having considered the existence of an Euler solution of (4.1) without any
continuity assumptions, let us investigate the weak flow invariance using the
framework of nonsmooth analysis [2], namely, proximal normal aiming
condition. For this purpose, we need to develop the concept of proximal
normal suitably and that is what we shall describe next.
Let X  K c ðRn Þ be a nonempty, closed set. Assume that for any U 2 K c ðRn Þ
such that U and X are disjoint, and for any S  X, there exists a Z 2 K c ðRn Þ
such that U = S + Z. Then U  S is called the geometric difference. Suppose
now that, for any U 2 K c ðRn Þ there is an element S 2 X whose distance to U is
minimal, that is,
 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 287

D0 ½U ; X
¼ kU  Sk ¼ inf kU  S 0 k: ð4:7Þ
S 0 2X

Then S is called a projection of U onto X. The set of all such elements is

denoted by projX(U). The element U  S will be called the proximal normal
direction to X at S. Any nonnegative multiple n = t(U  S), t P 0, is called
proximal normal to X at S. The set of all n obtained in this way is said to
be proximal normal cone to X at S and is denoted by N PX ðSÞ.

Definition 4.2. The system (X, F) is said to be weakly invariant provided that
for all U0 2 X, there exists an Euler solution U(t) of (4.1) on [t0, 1) such that
U(t0) = U0 and U(t) 2 X, t P t0.

We can now prove the following result which offers sufficient conditions in
terms of proximal normal for the weak invariance of the system (X, F).

Theorem 4.3. Let F and g satisfy the conditions of Theorem 4.1 on [t0, 1),
t0 P 0. Suppose that U(t) = U(t, t0, U0) is an Euler solution of (4.1) on [t0, 1),
which lies in an open set Q  K c ðRn Þ. Suppose also that for every
(t, Z) 2 [t0, 1) · Q, the proximal aiming condition is satisfied: namely, there
exists an S 2 projX(Z) such that

2hF ðt; ZÞ; Z  Si 6 qðt; D20 ½Z; X
Þ; ð4:8Þ

where q 2 C½½t0 ; 1Þ  Rþ ; R
. Assume that r(t) = r(t, t0, u0) is the maximal solu-
tion of the scalar differential equation
u0 ¼ qðt; uÞ; uðt0 Þ ¼ u0 P 0;
existing on [t0, 1). Then we have
D20 ½U ðtÞ; X
6 rðt; t0 ; D20 ½U 0 ; X
Þ; t0 6 t0; 1: ð4:9Þ
If, in addition r(t, t0,0) ” 0 then U0 2 X implies U(t) 2 X, t P t0, that is, the
system (X, F) is weakly invariant.

Proof. Let Up be one polygonal arc in the sequence converging uniformly to U

as per the definition of Euler solution of (4.1). We denote as before, its nodes at
ti by Ui, i = 0, 1, . . . ,N and hence U0 = U(t0). Let Up(t) be in Q for all
t0 6 t 6 T, where T 2 (t0, 1). Accordingly, there exists for each i, an element
Si 2 projX(Ui) such that
2hF ðti ; U i Þ; U i  S i i 6 qðti ; D20 ½U i ; X
Þ:
As in Theorem 4.1, letting D[DHUp(t), h] 6 k, we find
2
D20 ½U 1 ; X
6 kU 1  S 0 k ; since S 0 2 X:
288 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

We note that U1 = U0 + Z1, where Z1 = F(t0, U0)(t1  t0) and U0 = S0 + Z0

and therefore, we get successively
2 2 2
D20 ½U 1 ; X
6 kZ 1 þ Z 0 k 6 kZ 1 k þ kZ 0 k þ 2hZ 1 ; Z 0 i
Z t1
6 k 2 ðt1  t0 Þ2 þ D20 ½U 0 ; X
þ 2 hDH U p ðt0 Þ; Z 0 idt
t0
Z t1
6 k 2 ðt1  t0 Þ2 þ D20 ½U 0 ; X
þ 2 hF ðt0 ; U 0 Þ; U 0  S 0 idt
t0

6 k 2 ðt1  t0 Þ2 þ D20 ½U 0 ; X
þ qðt0 ; D20 ½U 0 ; X
Þðt1  t0 Þ:

Since the similar estimates hold at any node, we obtain,
2
D20 ½U i ; X
6 k 2 ðti  ti1 Þ þ D20 ½U i1 ; X
þ qðti1 ; D20 ½U i1 ; X
Þðti  ti1 Þ:
and therefore it follows that
X
i
2
D20 ½U i ; X
6 D20 ½U 0 ; X
þ k 2 ðtJ  tJ 1 Þ
J ¼1
X
i
þ qðtJ 1 ; D20 ½U J 1 ; X
ÞðtJ  tJ 1 Þ
J ¼1
X
i X
i
6 D20 ½U 0 ; X
þ k 2 lp ðtJ  tJ 1 Þ þ qðtJ 1 ; D20 ½U J 1 ; X
ÞðtJ  tJ 1 Þ
J ¼1 J ¼1
X
i
6 D20 ½U 0 ; X
þ k 2 lp ðT  t0 Þ þ qðtJ 1 ; D20 ½U J 1 ; X
ÞðtJ  tJ 1 Þ:
J ¼1

Consider now, the sequence U pJ ðtÞ of polynomial arcs converging to U(t).

Since the last estimate is true at every node, lpJ ! 0, as J ! 1, and the same
k applies to each U pJ ðtÞ, we deduce in the limit the integral inequality
Z t
D20 ½U ðtÞ; X
6 D20 ½U 0 ; X
þ qðs; D20 ½U ðsÞ; X
Þds; t0 6 t 6 T ð4:10Þ
t0

for every T 2 (t0, 1). If we know that q(t, u) is nondecreasing in u, then we can
apply the theory of integral inequalities (see Theorem 1.6.1 in [9]), to arrive at
D20 ½U ðtÞ; X
6 rðt; t0 ; D20 ½U 0 ; X
Þ; t P t0 : ð4:11Þ
If, on the other hand, q(t, u) is not nondecreasing in u, we can obtain instead
of (4.10), the following integral inequality for any t0 6 t 6 t + h 6 T, h > 0,
employing a similar reasoning,
Z tþh
D20 ½U ðt þ hÞ; X
6 D20 ½U ðtÞ; X
þ qðs; D20 ½U ðsÞ; X
Þds; ð4:12Þ
t

from which we obtain, setting mðtÞ ¼ D20 ½U ðtÞ; X
, the differential inequality
 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 289

Dþ mðtÞ 6 qðt; mðtÞÞ; mðt0 Þ ¼ D20 ½U 0 ; X
; ð4:13Þ

+
where D m(t) is a Dini derivative.
Applying now the theory of Differential inequalities (see Theorem 1.4.1 [9]),
we arrive at the same estimate (4.11).
If r(t, t0,0) ” 0, then, supposing that U0 2 X implies that U(t) 2 X for t P t0
and therefore the system (X, F) is weakly invariant as claimed. The proof is
hence complete. h

The results of this section are extensions of similar results in [2] for ordinary
differential equations and so include the corresponding results as very special
cases.

5. Connection of set and fuzzy differential equations

We propose, in this section, to connect the study of fuzzy differential equa-

tions with the study of set differential equations which are generated by the
level sets of the right-hand side of (4.2), namely, F(t, x; b) = [f(t, x)]b, b 2
[0, 1] = I. To do this, we generate a mapping H : Rþ  K c ðRn Þ ! K c ðRn Þ by
defining
H ðt; A; bÞ ¼ COF ðt; A; bÞ; ð5:1Þ
n
for each A 2 K c ðR Þ and then consider the sequence of set differential equations
given by
DH U b ¼ H ðt; U b ; bÞ; U b ðt0 Þ ¼ U 0 2 K c ðRn Þ; ð5:2Þ
on [t0, T], t0 P 0, and T 2 (0,1), where DHUb is the Hukuhara derivative for
each b 2 I.
Before we proceed to the interconnection of fuzzy differential equations with
set differential equations (5.2), we need to recall some known results [5,10–12]
for the IVP (2.7) where F(t, A) is assumed to be continuous. We start with the
comparison result.

Theorem 5.1. Assume that F 2 C[R+ · Kc(Rn), Kc(Rn)] and for t 2 R+, A, B 2
Kc(Rn)
D½F ðt; AÞ; F ðt; BÞ
6 gðt; D½A; B
Þ;
where g 2 C½R2þ ; Rþ
. Suppose further that the maximal solution r(t, t0, w0) of the
scalar differential equation
w0 ¼ gðt; wÞ; wðt0 Þ ¼ w0 P 0;
exists for t P t0. Then, if U(t) = U(t, t0, U0), V(t) = V(t, t0, V0) are any two solu-
tion of Eq. (2.7) such that U(t0) = U0, V(t0) = V0, U0, V0 2 Kc(Rn) existing for
t P t0, we have:
290 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

D½U ðtÞ; V ðtÞ
6 rðt; t0 ; w0 Þ; t P t0 ;

provided D[U0, V0] 6 w0.

The next result is an existence and uniqueness theorem more general than
Lipschitz type condition the proof of which exhibits the idea of comparison
principle.

Theorem 5.2. Assume that

(a) F 2 C[R0, Kc(Rn)], where R0 = J · B(U0, b), J = [t0, t0 + a], a > 0,

B(U0, b) = [U 2 Kc(Rn) : D[U, U0] 6 b] and D[F(t, U), h] 6 M0 on R0,
where h is the zero element of Rn regarded as a one point set;
(b) g 2 C[J · [0, 2b], R+], g(t, w) 6 M1 on J · [0, 2b], g(t, 0) ” o, g(t, w) is
nondecreasing in w for each t 2 J and w(t) ” 0 is the only solution of
w0 ¼ gðt; wÞ; wðt0 Þ ¼ 0;
(c) D[F(t, U), F(t, V)] 6 g(t, D[U, V]) on R0.

Then the successive approximations defined by

Z t
U nþ1 ðtÞ ¼ U 0 þ F ðs; U n ðsÞÞds; n ¼ 0; 1; 2; . . . ;
t0
 
exist on J0 = [t0, t + a], a ¼ min a; Mb , M = max(M0, M1), as continuous
functions and converge uniformly to the unique solution U(t) = U(t, t0, U0)
of IVP (2.7) on J0.

The following global existence result is a special case of Theorem 5.2 in [10]
which serves our purpose.

Theorem 5.3. Assume that

(1) F 2 C[R+ · Kc(Rn), Kc(Rn)] and for (t, A) 2 R+ · Kc(Rn),

D½F ðt; AÞ; h
6 qðt; D½A; h
Þ;
where q 2 C½R2þ ; Rþ
, q(t, w) is nondecreasing in w for each t 2 R+ and the
maximal solution r(t, t0, w0) of
w0 ¼ qðt; wÞ; wðt0 Þ ¼ w0 ;
exists for t P t0 and for every w0 > 0;
(2) there exists a local solution U(t) = U(t, t0, U0) of (2.7) for every
(t0, U0) 2 R+ · Kc(Rn).
 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 291

Then for every U0 2 Kc(Rn) such that D[U0, h] 6 w0, the IVP (2.7) pos-
sesses a solution U(t) = U(t, t0, U0) defined for t P t0, satisfying
D½U ðtÞ; h
6 rðt; t0 ; w0 Þ; t P t0 :

A result that relates the solution of set differential equation to the attainable
set of multivalued differential inclusion is the following result. See [12].

Theorem 5.4. Assume that F 2 C[R+ · Rn, Kc(Rn)];

D½F ðt; xÞ; F ðt; yÞ
6 gðt; kx  ykÞ; ðt; x; yÞ 2 Rþ  Rn  Rn ;

and D[F(t, x), h] 6 q(t, ixi), (t, x) 2 R+ · Rn, where g and q satisfy the
assumptions listed in Theorem 5.2 and Theorem 5.3 respectively, except that con-
ditions relative to g hold for R+ · R+. Then there exist a unique solution
U(t) = U(t, t0, U0) on [t0, 1) of IVP (2.7) and the attainable set A(U0, t) of differ-
ential inclusion

x0 2 F ðt; xÞ; xðt0 Þ 2 U 0 ;

satisfying A(U0, t)  U(t), t0 6 t < 1.

Finally, we need the following representation result [5,11].

Theorem 5.5. Let Yb  Rn, 0 6 b 6 1 be family of compact subsets satisfying

(a) Yb 2 K(Rn) for all 0 6 b 6 1;

(b) Yb ˝ Ya whenever a 6 b, a, b 2 [0, 1];
Q
(c) Y b ¼ 1 i¼1 Y bi , for any nondecreasing sequence bi ! b in [0, 1]. Then there
is a fuzzy set u 2 Dn, such that [u]b = Yb. If Yb are also convex, then u 2 En.
(Here Dn denotes the set of usc normal fuzzy set with compact support and
thus En  Dn.) Conversely, the level sets [u]b, of any u 2 En, are convex and
satisfy these conditions.

Let us next list the following conditions relative to F(t, x; b) and H(t, A; b).

(1) F(t, x; b) is quasi-concave, that is,

(a) for (t, x) 2 R+ · Rn, a, b 2 I, F(t, x; a) ˚ F(t, x, b) whenever a 6 b
(b) if bn is nondecreasing
T1 sequence in I, converging to b, then for
(t, x) 2 R+ · Rn, n¼1 F ðt; x; bn Þ ¼ F ðt; x; bÞ.
(2) D[H(t, A; b), H(t, B; b)] 6 g(t, D[A, B]) for t 2 R+, A, B 2 Kc(Rn), b 2 I;
292 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

(3) g 2 C½R2þ ; Rþ
, g(t, 0) ” 0, g(t, w) is nondecreasing in w for each t 2 R+ and

w(t) ” 0 is the only solution of
w0 ¼ gðt; wÞ; wðt0 Þ ¼ 0;

for t P t0;
(4) D[H(t, A; a), H(t, A; b)] 6 Lja  bj,a, b 2 I,(t, A) 2 R+ · Kc(Rn), L > 0.

We are now in a position to prove the following result.

Theorem 5.6. Suppose that the assumptions (1) to (4) are satisfied. Then there
exists a unique solution Ub(t) = Ub(t, t0, U0b) 2 Kc(Rn), b 2 I of (5.2) and Ub(t) is
quasi-concave in b for t P t0. Moreover, there exists a fuzzy set u(t) 2 En such
that
b
½uðtÞ
¼ U b ðtÞ; t P t0 :

Proof. Since f is continuous on R+ · Rn, F(t, x; b) is also continuous for

(t, x, b) 2 R+ · Rn · I. This implies that H(t, A; b) is continuous map for
(t, A, b) 2 R+ · Kc(Rn) · I. Consequently, by Theorems 5.2 and 5.3 it follows
that there exists a unique solution Ub(t) = U(t, t0, U0b) 2 Kc(Rn) for t P t0 of
(5.2). We first show that Ub(t) ˝ Ua(t) if a 6 b for t P t0. From the definition
of quasi-concavity of F(t, x; b), it follows that H(t, A; b) is also quasi-concave in
b. Let Ua(t),Ub(t) be the solution of

DH U a ¼ H ðt; U a ; aÞ; U a ðt0 Þ ¼ U 0 2 K c ðRn Þ;

DH U b ¼ H ðt; U b ; bÞ; U b ðt0 Þ ¼ U 0 ; a 6 b;

then we find, using quasi-concavity

DH U a ¼ H ðt; U a ; aÞ  H ðt; U a ; bÞ; a 6 b:

But H(t, Ua; b) = H(t, Ub; b) because of Ua(t0) = U0 = Ub(t0) and therefore
Ua(t) ” Ub(t) by uniqueness of solutions of (5.2). Thus it is clear that
Ub(t) ˝ Ua(t), a 6 b, t P t0.
We shall next prove that if bn is a nondecreasing sequence, bn 2 I,
converging to b, then U bn ðtÞ ! U b , uniformly on compact subsets of [t0, 1).
For this purpose, set mðtÞ ¼ D½U bn ðtÞ; U b ðtÞ
and note that D½U 0bn ; U 0b
¼
mðt0 Þ. We shall assume that U 0bn ! U 0b as n ! 1. Then employing the
properties of the metric D, the definition of Hukuhara derivative and the
conditions (2) and (4), we arrive at the scalar differential inequality

Dþ mðtÞ 6 gðt; mðtÞÞ þ Ljbn  bj; mðt0 Þ ¼ D½U 0bn ; U 0b
; t P t0 :

 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294 293

Hence by Lemma 1.3.1 in [9], we obtain

mðtÞ 6 rn ðt; t0 ; gn Þ
an any compact set J  [t0, 1), where gn ¼ D½U 0bn ; U 0b
and rn(t, t0, gn) is the
maximal solution of
w0 ¼ gðt; wÞ þ Ljbn  bj; wðt0 Þ ¼ gn ; on J :
By assumption (3) rn(t, t0, gn) ! r(t, t0,0) ” 0, uniformly on J as n ! 1. Since
bn ! b as n ! 1, it follows that m(t) ” 0 on J, which in turn implies that
D½U bn ðtÞ; U b ðtÞ
! 0 as n ! 1. Thus Ub(t) is quasi-concave in b 2 I for
t P t0. Consequently, by Theorem 5.5, there exists a fuzzy set u(t) 2 En such
that [u(t)]b = Ub(t), t P t0, and this completes the proof. h

To find the connection between the solution Ub(t), of (5.2) and the attaina-
bility set Ab(U0, t) of (3.5), we have the following result.

Theorem 5.7. Let F 2 C[R+ · Rn · I, Kc(Rn)] satisfy the assumptions of Theo-

rem 5.4 for each b 2 I = [0, 1] and assume that it is also quasi-concave in b as
well. Then there exist a unique solution Ub(t) = Ub(t, t0, U0) of (5.2) for t P t0 and
the attainable set Ab(U0, t) of the inclusion (3.5) such that
Ab ðU 0 ; tÞ  U b ðt; t0 ; U 0 Þ; t P t0 : ð5:3Þ

Proof. It is easy to verify that when F(t, x; b) satisfies the assumptions required
in Theorem 5.4. the desired conditions in Theorems 5.1, 5.2 and 5.3 are also
satisfied, in view of the monotone nondecreasing nature of the functions
g(t, w), q(t, w), the definition of D and the fact H(t, A; b) is generated by
F(t, x; b). We have assumed conditions in terms of H(t, A; b) since the set differ-
ential equations are treated as an independent subject. Thus for each b 2 I,
Theorem 5.4 yields the relation (5.3). Also, both Ub(t) and Ab(U0, t) satisfy
the assumptions of Theorem 5.5, because one can prove similarly the quasi-
concavity of Ab(U0, t). Therefore, there exist fuzzy sets u(t), v(t) 2 En such that
½vðtÞ
b ¼ Ab ðU 0 ; tÞ and ½uðtÞ
b ¼ U b ðtÞ; t P t0 :
The proof is complete. h

We note that, in general, since Ab(U0, t) is only compact and not convex,
only (5.3) holds. Equality in (5.3) is valid only in some special cases.

Remark. We have only considered the connection between fuzzy differential

equations and set differential equations in the continuous case. As f in (4.2) is
studied in the continuous case only, investigating (5.2) under the condition
seems natural. However, it would be interesting to consider the usc situation in
both cases.
294 V. Laksmikantham / Appl. Math. Comput. 164 (2005) 277–294

References

[1] A.J. Brandao Lopes Pinto, F.S. De Blasi, F. Iervolino, Uniqueness and existence theorems for
differential equations with compact convex-valued solutions, Boll. Unione. Mat. Italy 3 (1970)
47–54.
[2] F.H. Clarke, Yu.S. Ledyaev, R.J. Stern, P.R. Wolenski, Nonsmooth Analysis and Control
Theory, Springer-Verlag, New York, 1998.
[3] P. Diamond, Stability and periodicity in fuzzy differential equations, IEEE Trans. Fuzzy Syst.
8 (2000) 583–590.
[4] P. Diamond, P. Watson, Regularity of solution sets for differential inclusions quasiconcave in
parameter, Appl. Math. Lett. 13 (2000) 31–35.
[5] P. Diamond, P. Kloeden, Metric Spaces of Fuzzy Sets, World Scientific, Singapore, 1992.
[6] E. Hüllermeier, An approach to modeling and simulation of uncertain dynamical systems, Int.
J. Uncertainly, Fuzzyness Knowledge Based Syst. 5 (1997) 117–137.
[7] O. Kaleva, Fuzzy differential equations, Fuzzy Sets Syst. 24 (1987) 301–317.
[8] O. Kaleva, The Cauchy problem for fuzzy differential equations, Fuzzy Sets Syst. 35 (1990)
389–396.
[9] V. Laksmikantham, S. Leela, Differential and Integral Inequalities, vol. I, Academic Press,
New York, 1969.
[10] V. Laksmikantham, S. Leela, A.S. Vatsala, Set valued hybrid differential equations and
stability in terms of two measures, J. Hybrid Syst. 2 (2002) 169–187.
[11] V. Laksmikantham, R. Mohapatra, Theory of Fuzzy Differential Equations and Inclusions,
Taylor & Francis, London, 2003.
[12] A. Tolstonogov, Differential Inclusions in a Banach Space, Kluwer Academic Publishers,
Dordrecht, 2000.