Mathematical Programming 24 (1982) 70-91

North-Holland Publishing Company

NECESSARY AND SUFFICIENT OPTIMALITY CONDITIONS

FOR A CLASS OF N O N S M O O T H MINIMIZATION PROBLEMS

A. B E N - T A L

Faculty of Industrial Engineering and Management, Technion-Israel Institute of Technology,

Hal[a, Israel

J. Z O W E

Department of Mathematics, University of Bayreuth, 8580 Bayreuth, Federal Republic of

Germany
Received 16 December 1980
Revised manuscript received 21 August 1981

The purpose of this paper is to derive, in a unified way, second order necessary and
sufficient optimality criteria, for four types of nonsmooth minimization problems: the discrete
minimax problem, the discrete lrapproximation, the minimization of the exact penalty
[unction and the minimization of the classical exterior penalty .function. Our results correct
and supplement conditions obtained by various authors in recent papers.

Key words: Necessary and Sufficient Second Order Conditions, Nonsmooth Optimization,
Minimax Problem, h-Approximation, Penalty Functions, Directional Derivatives.

1. Preliminaries

1.1. Introduction

This p a p e r p r e s e n t s s e c o n d o r d e r o p t i m a l i t y c r i t e r i a f o r a c l a s s of n o n s m o o t h
p r o b l e m s w h i c h i n c l u d e s t h e f o l l o w i n g f o u r c l a s s i c a l e x a m p l e s of n o n s m o o t h
optimization

(P1) T h e discrete minimax problem

m i n i m i z e f ( x ) :-- m a x { g l ( x ) . . . . . gs(x)}. (1.1)

(P2) T h e discrete lrapproximation problem

m i n i m i z e f ( x ) : = ~ Igi(x) I. (1.2)
i=1

(P3) T h e minimization of the exact penalty function

s'

m i n i m i z e f ( x ) := ~go(x) + 2 ~ max{0, g,(x)} + ~ Ig,(x)l. (1.3)

i=l i=s'+l

7O
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 71

(P4) The minimization of the exterior penalty function

minimize f(x) := go(x) + 1 P ~_~ [max{0, gi(x)}] p, p I> 1. (1.4)

i=l

Here s' and s are natural numbers, ix and p are positive reals and the gi's are
smooth functions defined on a real normed vector space X.
Problem (P1) is widely studied in the literature, see e.g. the books by
D e m ' y a n o v and Malozemov [7] and b y Danskin [6]. In the former a second order
sufficient condition is given [7, Chapter III, Theorem 4.2].
(P2) was treated by Chralambous in [4], where necessary first order con-
ditions, and sufficient second order conditions are derived, however with a
wrong proof for the latter.
The exact penalty function f in (P3) is associated with the non-linear pro-
gramming problem
rain go(x),
subject to gi(x)<~O for l~<i~<s' (1.5)
gi(x)=0 for s ' + l < - i ~ s .

Obviously, a local minimizer of (P3) which satisfies the constraints in (1.5), is an

optimal solution for (1.5). Pietrzykowsky [14] showed that vice versa, for all
sufficiently small positive/x, an optimal solution for (1.5) is also a local minimizer
of (P3) (under the 'linear independence' constraint qualification). N e c e s s a r y
conditions for (P3) were given by Coleman and Conn [5]. H o w e v e r , their
conditions do only hold under additional regularity assumptions.
The problem (P4) is associated to the nonlinear programming problem (1.5)
with inequality constraints only. Auslender [1] derived necessary second order
conditions for p = 2, 3. Here we supplement these results for general p ~> 1 and
establish sufficient conditions as well.
The necessary and sufficient conditions for all four problems are derived in a
unified way. To this we treat (P1)-(P4) as special cases of the more general
nonsmooth problem (P):

minimize f ( x ) : qi(max{gil(x) ..... giki(x)}). (1.6)

i=l

It is clear that f is not differentiable at every x even if all the data qi and gii are
smooth. To see that (P) contains (P1)-(P4) as special cases put for example
q i ( t ) : = t, k i : = 2 , gil: =g~ and g~2:= - g i for i = 1. . . . . m. Then problem (P)
reduces to the discrete ll-approximation problem.
In Section 2 we summarize the primal and dual necessary and sufficient
optimality criteria for (P). These conditions are free of any additional regularity
assumptions. Moreover, the gap between the sufficient conditions and their
necessary counterparts is minimal, in the sense that they differ only in a strict
72 A. Ben-Tal and J. Zowe/Optimality of nonsmooth problems

inequality (>) versus a weak one (I>) in the second order inequality. The primal
conditions will be phrased in terms of first and second order directional deriva-
tives of f. Section 3 is devoted to the computation of these directional deriva-
tives, the main tool being a chain rule. In Section 5 we will give a proof for the
optimality conditions stated in Section 2. A theorem of the alternative, proved in
Section 4, enables us to transform the primal necessary conditions into dual
conditions. The optimality criteria for the special cases (PI)-(P4) are easy
corollaries of our general theorems. These corollaries are collected in Section 6.
In Section 7 we add some remarks about the nonsmooth problem with con-
straints.

1.2. Statement of the general problem

Throughout this paper X is a real normed vector space with topological dual
X*. The canonical bilinear form on X* × X will be denoted by ( . , . ) . We write
g'(x) and g"(x) for the first and the second Fr6chet-derivative of a function
g : X ~ R. To simplify the notation we often use g'(x)d instead of (g'(x), d). For
fixed x, g"(x) is interpreted as a bilinear functional on X × X. In case X is the
n-dimension Euclidean space R", then g'(x)= Vg(x) t and g"(x)= V2g(x) the
gradient and Hessian of g at x, respectively.
Let m and ki, 1 ~<i ~< m, be fixed natural numbers and let q i : 0 ~ k ~ R and
g~ : x ~ R for 1 ~< i ~< m and 1 ~<j ~< k~, be C2-functions. We always denote by
the optimal point and assume henceforth that

q'i({/) ~> 0 where {/= max gii(~) for i = 1. . . . . m. (1.7)

l~j~k i

Further we put

Ki:={1,2 . . . . . ki} f o r i = l . . . . . m.
With these data we define our general minimization problem

(P)
minimize f(x) := ~ qi(max gij(x)).
i=l j~K~
In addition to the sets K i we will use the following sets of indices for given x
and d:
Ki(x) = {j ~- Ki [ gii(x) = max gil(x)}
lEK~

Ki(x, d) := {j E Ki(x) [ g'ij(x)d = max g~l(x)d}. (1.8)

l~Ki(x)
Further we introduce an abbreviation which helps to simplify the notation. Let i
be a natural number, K a finite set of indices and aii, j E K, real numbers. Then
we put
ai.K := max {aij}. (1.9)
j~K
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 73

Using this notation the max-terms in (1.8) can be written in a much shorter form

max gil(x) = gi. K,(x),

I~:K i

max g~l(x)d = g'i,n,(x)(x)d.

IEKi(x)

It should be noted that g'~,K~(x)(x)d does not stand for the directional derivative of
gi, K~(x)(') at x applied to d.
In the sequel we will use the above notation without further mentioning.

2. Optimality conditions for the general problem

Recall the standard definition of the first directional derivative of a function ~,
from the real normed space X into another such space Y, at x in the direction d
(if this limit exists)

~'(x; d) = lim t-~[q~(x + t d ) - q~(x)]. (2.1)

t$0

Suppose q~'(x; d) exists. Then we introduce as second directional derivative of

at x in the direction d and z (whenever this limit exists; compare Ben-Tal and
Zowe [2])
q~"(x; d, z) = lim t-2[q~(x + td + t2z) - ~(x) - tq~'(x; d)]. (2.2)
t$0

For z = 0 definition (2.2) coincides with the proposal for a second directional
derivative due to D e m ' y a n o v and P e v n y i [8].
(2.2) reduces to (2.1) for d = 0:

~o"(x; 0, z) = ~'(x; z). (2.3)

We have the following expansion of q~ in terms of the directional derivatives

(provided q~'(x; d) and q~"(x; d, z) exist):

q~(x + td + t2z) = q~(x) + tq~'(x ; d ) + t2qg"(X; d, z) + o(t ~) for t t>0.

(2.4)
If ¢ is Fr6chet-differentiable at x, then, obviously, the limit under (2.1) exists and

~'(x, d) = ~'(x)d for all d E X. (2.5)

If q~ is even twice continuously ditterentiable in a neighborhood of x, then

q~(x + td + t2z) - q~(x)
= q~'(x)(td + t2z) + ½q~"(x)(td + t2z, td + t2z) + o(t 2)
= t#(x)d + t2[q~'(x)z + ½q~"(x)(d, d)] + o(t2).
Hence, the limit (2.2) exists and

~"(x; d, z) = q~'(x)z + ½q~"(x)(d, d) for all d, z ~ X. (2.6)

74 A. Ben-Tal and J. Zowe[Optimality of nonsmooth problems

In the sequel we will use (2.4)-(2.6) without further mentioning.

We can now formulate optimality criteria for our general minimization prob-
lem (see Section 1.2)
(P)
min f ( x ) := ~ qi(max &(x)).
i=1 jEK i

Theorem 2.1 [Primal necessary and sufficient conditions for (P)].

(a) If ~ is a local minimizer of f, then
q"(~z; d, z) >i 0 .
f'(:~ ; d) <~0 ~ ~for every z E ix. (2.7)

(b) Suppose dim X < oo. If

d) Oo} . d,z)>O (2.8)

and d # ~ o r every z ~ X',

then ~ is an isolated local minimizer of f.

Here

f'(~; d) = ~ q'i(max gi~(,2)), max g'ii(~)d, (2.9)

i=1 jEK i jEKi(.~)

f"(~; d, z) = ~ q'i(max &(:~)) • max {glj(:~)z +½g~;(~)(d,d)}

i= I j~K i jEKI(,,~, d)

1 ~ q'/(max &(:~))[max g'ii(~)d] 2. (2.10)

"}- 2 i= 1 jEK i JeKi(e)

Theorem 2.2 [Dual necessary and sufficient conditions for (P)].

(a) A necessary condition for ~ to be a local minimizer of f is that for every
d E X satisfying
(o)
q'i(max gij(x)) " max glj(2)d <<-O,
i=1 j~K~ JeKi(~)
there correspond real multipliers Yij = yij(d) I> 0 for j E Ki(~, d) and i = 1, 2 ..... m
such that
(i)
~, Yij = q~(max &(~)) for i = 1, 2 . . . . . m,
j~Ki(.~, d) jEK~

(ii) ~ E yijg;j(~)=0,
i=1 j~Ki(~, d)

(iii)
~ yi~g'[j(~7)(d,d) + ql'(max g~j(~))[ max glj(~)d] 2/> 0.
i=1 i EKi(~,d) = jEK i i.~Ki(~)

(b) Suppose dim X < ~. Then a su1~icient condition for Y, to be an isolated

 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 75

local minimizer of f is that for every d ¢ 0 satisfying (o) there are yij >I 0 for
j E K~(Y,, d) and i = 1..... m such that (i)-(iii) hold with strict inequality in (iii).

A proof of the a b o v e theorems will be given in Section 5. In Section 6 we will

show how the a b o v e conditions specialize for (P1)-(P4). We add some remarks.

R e m a r k 1. Fix d = 0. Then (2.7) implies f ' ( ~ ; z) = f"(~; 0, z)/> 0 for all z E X, i.e.
our primal n e c e s s a r y second order condition contains the n e c e s s a r y first order
condition. Further note that (2.8) is automatically satisfied w h e n e v e r the
sufficient first order condition holds, i.e. w h e n e v e r f ' ( ~ ; d ) > 0 for e v e r y d ~ 0.
If we choose d = 0 in T h e o r e m 2.2(a), then our dual n e c e s s a r y second order
condition reduces to the necessary first order condition: there are multipliers
y~j i> 0 satisfying (i) and (ii).

R e m a r k 2. It is an easy exercise to show the equivalence of (2.7) and (2.8),

respectively, to the standard second order conditions for the differentiable
problem:

Necessary condition:
f'(~) = 0 and f"(~)(d, d) ~ 0 for all d ~ X.

Sufficient condition (dim X < ~):

Vf(g)=0 anddtV2f(~)d>0 for a l l d U X , d~0.

R e m a r k 3. The additional hypothesis dim X < oo in T h e o r e m s 2.1(b) and 2.2(b) is

essential (compare Maurer and Zowe [13]). Consider the real sequence space 12
a n d p u t m := 1, k I := 1, ql(t) := t a n d g u ( x ) := ~ = 1 (s~x~- x4)where s is a n y e l e m e n t
in 12 with s~ > 0 for all i. L e t ~ be the zero-element of 12. Then

f'(~;d)=g~l(~;d)=0 for a l l d E l 2 ,

f"(~; d, z) = g;,(~; d, z) = ~, s~d~ > O for d # 0 and arbitrary z.

i=1

H e n c e (2.8) holds. Nevertheless ~ is not a local minimizer. Consider in 12 the

sequence {(xi)"}n=l,2,... with (xi) n = 2s, for i = n and (xi)" = 0 for i ~ n. Obviously
x" -~ ~ but

f(x")=4s~-16s4.<O=f(~) forn=l,2 .....

L e t us point out an important feature of T h e o r e m 2.2: In the n e c e s s a r y

condition we do not assert--conversely, in the sufficient condition we do not
require---that (i), (ii), (iii) hold for all d with a fixed vector of multipliers Yir If we
want independence of the multipliers f r o m d, then we have to impose a
76 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

regularity assumption on (P). One such condition is

A~j = 0 for i = 1, ..., m | }kij : 0

(CQP) for i = 1. . . . . m

~ ~.ijg'ij(X) : 0 and j E K i ( ~ ).
i= l j~.Ki(~ )

Corollary 2.3. Consider the general problem (P) and assume that (CQP) holds at
g. Then a necessary condition for ~ to be a local minimizer o f f is the existence of
a unique set of multipliers Yij ~> 0, i = 1, 2 .... , m and i E K~(2), such that

(1) ~ Yii = q'i(max glj(x)) for i = 1, ..., m,

jEKi(~) JEK i

(li)
~ Yljg'ij(.~) = 0,
i=l j~Ki(~ )

and for all d satisfying (o) in Theorem 2.2(a),

(i~i) ~,
yog'[j(~)(d, d) + ~ q'[(max gii(g))[ max g'ii(x)d] 2 >! O.
i=1 J~Ki(£) i= 1 jEK i JEKi(J~)

Proof. L e t ~ be optimal and suppose d z and d 2 satisfy (o). By T h e o r e m 2.2, we

can c h o o s e Y~ijfor i = 1 . . . . . m, j E Ki(g, d ~) and l = 1, 2 such that conditions (i)
and (ii) of T h e o r e m 2.2(a) are satisfied. P u t ytij := 0 for j E K ~ ( ~ ) \ K i ( g , d~). T h e n
use (CPQ) to show that the e x t e n d e d vectors (Y,~)i=l...... ;Jeri(e) and
(Y,~)i=l...... ;j~KI(~) coincide.

3. Computation of the directional derivatives of f

The following chain rule constitutes the b a c k b o n e of o u r caculus.

Lemma 3.1 [Chain rule]. Let q~(. ) = h(k(. )) with functions k : X ~ Y and h : Y - >
Z where X, Y, Z are real normed vector spaces. Let x, d, z E X be given. Put
y := k(x) and suppose v := k'(x; d), w := k"(x; d, z) and h'(y; v), h"(y; v, w)
exist. Further, suppose that h is locally Lipschitz continuous at y = k(x), i.e.
there exists a neighborhood U of y and some real L such that

]lh(yl) - h(y91l ~< L Ily,- y211 for all Yl, Y2 E U. (3.1)

Then q~'(x; d) and q~"(x; d, z) exist and
(i) q~'(x; d) = h'(y; v),
(ii) q~"(x; d, z) = h"(y; v, w).

Proof. We only p r o v e (ii). T h e (easier) p r o o f of (i) can be copied f r o m this. B y (i)

 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 77

and by an e x p a n s i o n of k (see (2.4)),

q~"(x; d, z) = lim t-2[h(k(x + td + t2z)) - h ( k ( x ) ) - th'(k(x); k'(x; d))]

t$o

= lim t 2[h(k(x) + tk'(x; d) + t2k"(x; d, z) + o(t2))

t~o
- h ( k ( x ) ) - th'(k(x); k'(x; d))]

= lim t-2[h(y + tv + t2w + o(t2))- h(y) - t h ' ( y ; v)]

t$0

= lim{t-2[h(y + tv + t2w) - h ( y ) - th'(y; v)]

t$o
+ t-2[h(y + tv + t2w + o(t2)) - h(y + tv + t2w)]}.

B e c a u s e of the L i p s c h i t z continuity of h at y the s e c o n d t e r m in b r a c k e t s tends

to 0 for t ~ 0. This s h o w s

q~"(x; d, z) = lim t-2[h(y + tv + t 2 w ) - h ( y ) - th'(y ; v)]

t~0
= h"(y; v, w).

W e need a f u r t h e r auxiliary result.

L e m m a 3.2. Consider the f u n c t i o n l : R" ~ R defined by l(yl . . . . . Yn) := m a x Yi- The

l<.i<.n
directional derivatives I'(y ; v) a n d / " ( y ; v, w) exist f o r all y, v and w E R n and

(i) /'(y; v) = max{vl I i such that Yi =/(Y)},

(ii) /"(y ; v, w) = max{wi I i such that Yi = l(y) and vi = l'(y, v)}.

Proof. (i) is easily verified. T o p r o v e (ii) fix y, v, w, and c h o o s e j such that

yj = l(y) and vi = l'(y; v). Since f o r all sufficiently small positive t,

max{yi + tvi + t2wi} = yj + tvj + tEwj,

l~i~n

one obtains

lira t-E[l(y + tv + t2w) - l(y) - tl'(y ; v)] = w r

t$0

This p r o v e s (ii).

N o w c o n s i d e r the f u n c t i o n f of our general m i n i m i z a t i o n p r o b l e m (P) (see

Section 1.2):

f ( x ) = ~ qi(max git(x))
i= 1 jEKi
78 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

with C2-functions q~ and g~j. W e use the n o t a t i o n i n t r o d u c e d in (1.8) and (1.9).

Proposition 3.3. The directional derivatives of f exist and

(i) f'(x ; d) = ~, q'i(gi, K,(x))gl, K,~x)(X)d,

i=1

(ii) f"(x; d,z) = ~, q,(g,,r,(' x)) • m a x {g~(x)z + lgTS(x)(d,d)}

i= 1 jEKi(x, d)

1 ~, q ~'(gi,Ki(X))[g I,K,~x)(x)d] 2.
+2i=1
Proof. Fix i and put k ( . ) : = gi, K~('), i.e.

k(. ) = l(g(. ))
with l as defined in L e m m a 3.2 and g ( ' ) = ( g H ( ' ) ..... g~k~('))t. N o t e that l is
locally L i p s c h i t z continuous. H e n c e L e m m a s 3.1 and 3.2 s h o w

k'(x; d) = l'(k(x); k'(x; d))

= m a x g'ij(x)d = gl, r~(x)(x)d. (3.2)
JEKi(x)

Similarly we get

k"(x; d, z) = l"(k(x); k'(x; d), k"(x; d, z))

= max g';i(x; d, z)
JEKi(x, d)

= m a x {g'ii(x)z + ½g'~j(x)(d, d)}.

J•Ki(x, d)

The assertion follows if we apply the chain rule o n c e m o r e to q~(k(x)) and sum
up o v e r i.

The a b o v e result p r o v e s equalities (2.9) and (2.10) in T h e o r e m 2.1.

4. A theorem of the alternative

In this section we establish a theorem of the alternative w h i c h will be used to

p r o v e T h e o r e m 2.1(b) and T h e o r e m 2.2(a). T h r o u g h o u t let m be a natural
n u m b e r ,B1 . . . . . Bm finite index sets and n : = ]Bll + -.. + [Bml. F o r i = l, 2 . . . . . m
and j E Bi let aij be given elements of X * and let /3, al and /3~j be given real
numbers.

Theorem 4.1. Suppose cti >t 0 for i = 1..... m. Then the following two statements
hold.
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 79

(a) One and only one of the s y s t e m s (I) and (II) has a solution.
(I) Find z ~_ X such that

ai m a x {(air, z) +/30} +/3 < O.

i=1 JEB i

(II) Find Y~r >~0 .for i = 1. . . . . m and j E B~ such that

(i)
~ , Yo = ai .for i = l, 2 . . . . . m,
fEB i

(ii) ~, ~ yiiaii = O,
i=l jCB i

(iii) ~ ~ yJ3,r + / 3 / > O.

i= I i•B i

(b) S u p p o s e dim X < oo. Then one and only one of the s y s t e m s (i) and (II) has a
solution.
(I) Find z E X such that

a i max{(air, z) +/3it} + 13 ~< O.

i=l r~-B i

(iI) Find Yir >~0 .for i = 1. . . . . m and j E Bi such that (i) and (ii) hold and

(i~i) ~, ~, Yij/3ir+/3 > O.

i=1 jEB i

Proof. (a) O b v i o u s l y , (I) is insolvable if and only if

inf ~ ai m a x { a 0, z) +/3it}/> - / 3 ,
z~-X i=l jEB i

or, in other t e r m s (note that b y a s s u m p t i o n all ~ I> 0),

inf aiui I (z, u) ~ X x R m and

(air, z) - ui <~ -/31r for i = 1. . . . . m and j E Bi} I> - / 3 .

In t e r m s of the linear functions c : X × R m ~ R, A : X x R m -~ R n and the v e c t o r

b ~ R n defined b y

c(z, u):= ~ ~,u~,

i=l

(A(z, u)) o := - (aij, z) + ui,~

fori=l ..... m andj~B~,
bij ---/3ij, J
80 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

the last inequality reads

(tOp :=) inf{c(z, u) l A(z, u) >i b, (z, u) E X × Rm}~
> -/3. (4.1)
Obviously a Slater type condition is satisfied for the above linear programming
problem: there exists (~, t~) E X x R m such that
(A(~, ti))ij > bij for i = 1..... m and j E Bi. (4.2)
Then a well-known duality theorem for linear programming in infinite dimen-
sional spaces (see Duffin [9] or Krabs [12, Ch. II, Theorem 4.14]) tells us that
whenever tOp is finite (which is the case by (4.1) and (4.2)) the dual problem
(tOp =) sup{ytb [ y E R", y i> 0 and ytA(z, u) = c(z, u)
for all (z, u) E X x 0~r~} (4.3)

has an optimal solution ? and toD = top/> --/3. Feasibility means that ? i> 0 and

~ ?ii(-(a,i, z) + ui) = ~ a,u, for all (z, u) ~ X × R ' .

i=l i EB i i=l

It is easily seen that this implies (i) and (ii). In addition

tOp = = t> - / 3 ,
i=l jEB i

i.e. y is a solution of (II).

Conversely, if (II) is satisfied, then (4.3) is feasible. Hence by the weak duality
theorem
-/3 <~ tOp ~< tOe. (4.4)
But as we know already, -/3 ~<~o~ means that (I) has no solution. This proves (a).
(b) The proof is just a copy of the above proof, the only difference being the
fact that we have strict inequality in (4.4) (-/3 < COD---<top) and in (4.1), due to the
fact that for finite-dimensional X the infimum in (4.1) is attained and is strictly
greater than -/3.

5. Proof of Theorems 2.1 and 2.2

We start with the easy part (a) of the first theorem.

Proof of Theorem 2.1(a). Suppose there are d, z with

f'(x; d) <~0, f"(~ ; d, z) < 0.

Then for all sufficiently small positive t (we use (2.4)),
.f(~ + td + t 2 z ) - f(~) = tf'(~; d)+ t2f"(~; d, z)+ o(t 2) < 0,

contradicting the minimality of 2.

 A. Ben-Tal and J. Zowe/Optimality of nonsmooth problems 81

The proof of Theorem 2.1(b) will be given after the proof of Theorem 2.2.

Proof of Theorem 2.2. For a fixed vector d ~ X put for i = 1 . . . . . m and j E

Ki(~, d):

Bi :-= Ki(~, d),

ai := q'i(gr,(~)),

a~j := g'~i(~), (5.1)

& : = ½g'[j(YO(d, el),

/ 3 : = ~ 1/~__m
= l qi(grl
" ( ~ ))[gi,' nit,)( ~ d) ] 2.

Note that, as required in Theorem 4.1, the above ai's are nonnegative by
assumption (1.7). With the above notation we have (see Proposition 3.3(ii))

f"(*; d, z) = k cq max{(ali, z} +/3i,} +/3. (5.2)

i=1 JEB i

(a) Let ~ be optimal. Fix some d satisfying (o), i.e. f'(~; d)~<0 by Proposition
3.3(i). Hence, by Theorem 2.1(a) and by (5.2),

k cti max{(aii, z) +/3ij} +/3 ~ 0 for all z E X.

i=1 j~-B i

In other words: system (I) of the theorem of the alternative has no solution and
thus (II) has a solution. This proves (a).
Co) Without loss of generality we may take m = 1. We write K for K1, q for ql
etc. Our proof is by negation. Suppose ~ is not an isolated local minimizer, i.e.
there exists a sequence {x,},=1,2..... x , ~ ~, with
x, n--~
~ butf(x.)<~f(~) for alln.

We put (here I" II is any norm on the finite-dimensional X )

a. := x, - t. := IId, ll.
A compactness argument yields the existence of some d ¢ 0 in X such that (for a
suitable subsequence)

d = lim t~d,.
n~

L e t us see that d satisfies (o). By assumption, f(x,) <~f(~) and thus for all n,

0 >1 t;'[f(x,) - f(~)]

= t;'[f(~ + t,d) - f(~)] + t;'[f(x,) - f ( ~ , + t,d)].
The first term of the right-hand side converges to f'(~; d) as n ~oo. For the
82 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

second bracket we have with some Lipschitz constant L (f is Lipschitz con-

tinuous since q and gr = max(gj} are Lipschitzian),

]t~[f(xn) - f ( ~ + tnd)]] ~< LlltX1d, - dll.

Obviously, the right-hand side converges to 0 for n ~ oo. H e n c e (o) is satisfied by
the above d. We choose Yii = Yo(d) as in the theorem. The proof will be finished
if we can show that (iii) does not hold with strict inequality for the above d and
the associated yii(d)'s. A Taylor expansion of the C - f u n c t i o n q at gK(g) shows

0 i> f ( x ~ ) - f ( ~ )
= q(gK(x~))- q(gr(.~))
= q ' ( g ~ ( g ) ) [ g r ( x . ) - gr(~)]
+ lq"(gK(~) + O.[gK(x.) - gK(~)])[gr(x.) - g~¢(~)]2 (5.3)

with 0 ~< 0, <~ 1. For convenience let us write An and B,, respectively, for the first
and the second part of the right-hand side of the last equation. From (i) and a
Taylor expansion of the gj's we obtain with suitable 0 ~< trni ~< 1,

An = q'(gr(,x)[gK(x~) - gK(x)]

= ~, Yj[gr(x~) - gK(-~)]
jEK(~, a)

/> ~ Yj[gj(xn)-gj(x)]
j~K(~, d)
1
yig';(~ + trjndn)(d,, an).
jE K(.~, d) j(~K(~,d)

Exploiting (ii) and dividing by tEn we have for n = l, 2 . . . . .

1 , yjg'~(~+(~fln)(t~dn, t~dn). (5.4)

For the term Bn we get

t ~2B n = ½q"(gr ( x ) + On [gr (xn) - gr (x)])

[gK(~ + tnd) - gr(~) q gK(x.) - gr(~ + tnd)] 2.
×
[ (5.5)
tn tn J
Since gr, is Lipschitz continuous at ~,

gr(xn)-gr(x + tnd) 90 as n-->~.

tn
Further (see (3.2) in the proof of Proposition 3.3),

gr(~ -1-t ~ d ) -
gK(x)->g'rt~)(~)d as n --)~.
tn
 A. Ben-Tal and J. Zowe/Optimality of nonsmooth.problems 83

This together with (5.3), (5.4) and (5.5) shows

0/> lim sup(t~2An + t~2B.)

n~

1> 1 ~ yjg'~($)(d, d) + ~q"(gr(~))[g'K~)($)d]2.

Z i~K(~, d)

We end up with a contradiction to the assumption that (iii) holds with strict
inequality.

Proof of Theorem 2.1(b). We show that (2.8) is equivalent to the (sufficient)

conditions of Theorem 2.2(b). By Proposition 3.3, f'(27; d)~<0 if and only if d
satisfies condition (o) in Theorem 2.2. We use once more Proposition 3.3, the
notation (5.1) and Theorem 4.1(b) to see that f"()7; d, z) > 0 for all z E X if and
only if (i), (ii) and (iii) with strict inequality hold for some suitable y I> 0.

6. Optimality conditions for problem (P1)-(P4)

The general problem (P) reduces to the minimax problem (P1) by setting
re=l, qt(t)=t, K 1 = { 1 , 2 ..... s}, glj(x)=gj(x)forjEKl. (6.1)
Note that (1.7) is satisfied. Using (6.1), Theorem 2.2 reduces to the following
statement.

Theorem 6.1 [Necessary and sufficient conditions for problem (P1)].

(a) A necessary condition for ~ to be a local minimizer of (P1) is that for every
d E X satisfying

max g;(g)d <~0 (6.2)

jEK(~)

there correspond real multipliers yj = yi(d) for j E K(~, d) satisfying

yj>10 for all j and ~ yj=l, (6.3)

j~K(~, a)

yig;(~) = 0, (6.4)
i~K(~, d)

yig~(£)(d, d) I> 0. (6.5)

jEK(~, d)

(b) Suppose dim X < ~. Then a sufficient condition for ~ to be an isolated

local minimizer of (P1) is that for every d ~ 0 satisfying (6.2) there correspond
yfs satisfying (6.3), (6.4) and (6.5) with strict inequality.
84 A. Ben-Tal and J. Zowe/ Optimality o[ nonsmooth problems

Here

K(~) := {1 ~< j ~< s [ gj(~) = max gi(~)},

l~i<.s
(6.6)
K(~, d ) : = {j E K(~) ] g;(,~)d = max g~(X)d}.
i~_K(~)

Example. W e d e m o n s t r a t e here the use of the n e c e s s a r y condition to easily

establish the n o n o p t i m a l i t y of ~ = (0, 0) t for the m i n i m a x p r o b l e m

min{gl = - X l - x~, g2 = 3xl + x ~ + x~}.

x 1, x 2

This s a m e c o n c l u s i o n had to be derived b y a specialized g e o m e t r i c a l a r g u m e n t in

[7, pp. 69-71].
At g = (0, 0) t, K ( g ) = {1, 2} and (6.2) b e c o m e s

m a x { - d l , 3dl} ~< 0.

T h u s the set of d ~ •2 satisfying (6.2) is

and, f o r e v e r y such d, K ( £ , d) = {1, 2}. T h e r e f o r e (6.3)-(6.5) r e d u c e s to

Yt ~>0, Y2 ~>0, Yl+y2=I,

° 0 0

i.e.

Yl >~ 0, Y2 I> 0, Yt + Y2 = 1, - Y l + 3Y2 = 0, - Y t + Y2 ~> 0. (6.7)

Clearly no such Yl, Y2 exist, so the n e c e s s a r y condition does not hold at

= (0, 0) t c a n n o t be a local minimizer. N o t e incidently that this c o n c l u s i o n could
not h a v e b e e n o b t a i n e d b y the first o r d e r n e c e s s a r y condition (i.e. (6.7) without
the last inequality) alone.

W e n o w turn to the I t - a p p r o x i m a t i o n p r o b l e m (P2), which is a special case of

the general p r o b l e m (P) with

m = s, qi(t) = t, K i = {1, 2}, gil(x) = gi(x),

giz(x) = - gi(x), f o r i = 1. . . . . s.
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 85

It is convenient to define the following index sets:

A := {i ] g~(X) = 0}, A* := {i ] g~(X) # 0},

(6.8)
A + := {i [ gi(~) > 0} A- := {i [ g~(~) < 0};
V(d) := {i E A [ gi(Y~)d = 0}, V*(d) := {i ~ A [ gi(g)d~ 0},
(6.9)
V+(d) := {i E A [ g',(g)d > 0}, V - ( d ) := {i E A [ g',(x)d < 0}.

Moreover, we set
gi := sign gi(g), O~= Oi(d) := sign g'i(X)d. (6.10)

Theorem 6.2 [Necessary and sufficient conditions for problem (P2)].

(a) A necessary condition for ~ to be a local minimizer of (P2) is that for every
d ~ X satisfying

Ig'~(X)dI + ~ o',g'~(~)d~0 (6.11)

ifiSA lEA*

there correspond real multipliers u~ = ui(d) for i E V ( d ) satisfying

lu~I ~< 1, i ~ V(d), (6.12)

~, u,g'~('2)+ ~, O,g'~(g)+ ~_, o',g'~(g)=0, (6.13)

iEV(d) iEV*(d) iEA*

~, u,g'[C~)(d,d)+ ~_, O,g'[(g)(d,d)+ ~, o-g[(~)(d,d)>-O. (6.14)

iEV(d) iEV*(d) lEA*

(b) Suppose dim X <oo. Then a sufficient condition for Y, to be an isolated

local minimizer of (P2) is that for every d ¢ 0 satisfying (6.11) there correspond
ui's satisfying (6.12), (6.13) and (6.14) with strict inequality.
Here the index sets A, A*, V(d), V*(d) are given in (6.8), (6.9) while tr~ and O~
are defined in (6.10).

Proof. First note that for the ll-case the index sets Ki(g) and K~(£, d) in
Theorem 2.2 reduce to

{1} if i E A +, [ {1} if i E A + u V+(d),

Ki(~) = {2} if i E A-, Ki(~, d) = l {2} if i E A- u V-(d),
{1,2} if l E A , {1,2} ifiEV(d).

Hence condition (o) in Theorem 2.2 becomes

g~(Y~)d + ~, (-g'i(:g)d) + ~_, max{g~(g)d, -g[(g)d} ~<0

i + iEA- iEA
86 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

which is equivalent to (6.11). Condition (i) becomes

Yll 1, i E A + U V+(d);
= Yi2: 1, i ~ A- U V-(d);
Yll + Yi2 = 1, i ~ V(d). (6.15

(ii) says

(-g'i)($)) + ~ (Y,l- Y,2)g;(x) = 0

iEA+~UV+(d) gPi('~) + iEA EUV-(d) i~ V(d)

which further reduces to

E N Oig'i(x) + ~, (Yil Yigg'i(-~) = O.

-
(6.16)
iEA* iEV*(d) iEV(d)

Finally, (iii) becomes

N ~r,g';(~)(d, d) + Oig'~(£)(d, d) + (Yil- y~2)g:(~)(d, d) i> 0.

lEA* i~EV*(d) iEV(d)
(6.17)

Using Yil, Yi2~>0 for all i and Yll + Yi2= 1 for i E V(d) and setting ui := 2y~1- 1, it
is seen that (6.15)-(6.17) convert to (6.12)-(6.14).

Using the expression for K~(g) in the above proof, one finds that the regularity
condition (CQP) simplifies for the /t-problem (P2) to the classical constraint
qualification
(CQP2) the set {g'~(g) ] i ~ A} is linearly independent.
Parallel to Corollary 2.3 one obtains

Corollary 6.3. Consider the l~-problem (P2) and assume that at ~, (CQP2) holds.
Then a necessary condition for 2 to be a local minimizer is that there exists
(unique) multipliers
[uil ~< 1, i~A,
such that

~, ulg'i(g) + ~ trig'i(~) = 0 (6.18)

i~A lEA*

and for every d E N, where

=0 ifiEA and lull< 1
N:={dlg}(~)d >lOifi~Aandui=l ,

<~O if i E A and u~ = - I
the following inequality holds

~, u,g"(g)(d, d) + ~, cr,g"(2)(d, d) ~>O. (6.19)

i•A lEA*
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 87

Proof. The only new thing which needs a proof, is the appearance of the set N
instead of the set

M := {d I ~'ieA]g'i(~)d[+ leA*
~" o'ig'~(YOd~<0}.
To see that M = N , let first d E M. Use (6.18) to obtain

~, (Ig'i(g)d - uig~(f,)d) <~O.

iEA

As [ui[ ~< 1, this can only hold if each term in the summation is zero; d ~ N
follows easily. Let now d E N. Then

([g~(~)d I - u,g~(~)d) = 0
lEA

and again by (6.18),

[g',(~)dl + ~ o-~g'~(~)d = O,
i~A i~A*

i.e. dEM.

Remark. With strict inequality in (6.19) and dim X < oQthe optimality conditions
in Corollary 6.3 become sufficient. This is a special case of Theorem 6.2(b). Such
a result was stated in [4], however with a wrong proof (see the counterexample
in [3]).
We close this part of the section with a simple example for an application of
Theorem 6.2 (in fact the corollary).

Example. Consider

min Igl(x)l + Ig2(x) I = I - a x E+ 4x21 + Ix 2 + (x2 - 1)21.

X~R 2

We ask: for what values of the real parameter ct is ~ = (0, 0) t optimal?

Here A = {1}, and (CQP2) trivially holds. We use Corollary 6.3. Condition
(6.18) is satisfied with the unique multiplier ul = I:

)=(Oo)" (6.20)

The set N is simply {(ao9 [ dl E ~} and (6.19) becomes

(-a+2)d~>0 for a l l d l C R .

Hence ~---(0, 0) t will be an isolated local minimizer for a < 2 but not a local
minimizer for a > 2. Note that the necessary first order condition (6.20) holds
regardless of the value of a, and hence does not provide any information.
88 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems

The exact penalty f u n c t i o n in problem (P3) is obtained for the general problem
(P) by setting m = s + 1 and

for i = 1. . . . . s', q~(t) = 2t, Ki = {1, 2}, gil(X) = O,

gi2(X) = gi(x),
fori=s'+l . . . . . s, q~(t)=t, K~={1,2}, g , ( x ) = g~(x),
gi2(X) = -- gi (X),
for i = s + 1 = m, % ( 0 = t, K m = {1}, g,.l(x) = I~go(X).
A straightforward substitution in Theorem 2.2 yields:

Theorem 6.4 [Necessary and sufficient conditions for problem (P3)].

(a) A necessary condition f o r ~ to be a local minimizer o f (P3) is that f o r every
d E X satisfying
s'
Ixg;(~)d + ~ , g',()i)d + ~ ~,,g',(e)d + ~, g',(~)d <<-0
lEA lEA* i=1

there correspond real multipliers wi = wi(d) f o r i ~ V ( d ) satisfying

Iw, L~ 1, i ~ V(d),
s'
~,g;(~) + Y. w,g',(~) + Y. o~g',(,~)+ 5". ,~,g',(~) + ~, g:(x) = o,
i E V(d) i E V*(d) iE A* i= 1

Ixg;(~)(d, d) + ~ w,g'[(i)(d, d) + ~, O,g'[(Yc)(d,d)

iEV(d) i~V*(d)
s'
+ ~, ~,,g'[(x)(d, d) + ~, g'[(x)(d, d) >10. (6.21)
lEA i=1

(b) S u p p o s e dim X < ~. Then a sufficient condition f o r ~ to be an isolated

local minimizer of (P3) is as above, but with strict inequality in (6.21) f o r d # O.
Here A, A * , V ( d ) , V * ( d ) are defined as in (6.8), (6.9) and ~ri, Oi = Oi(d) as in
(6.10).

Remark. For problem (P3) the regularity assumption guaranteeing the in-
dependence of the multipliers from d is also (CQP2). In the simplified version
which results under such a constraint qualification one can also replace the set M
with the set N (see Corollary 6.3). This simplified necessary condition was
obtained in [5, Corollary 2]. The sufficient counterpart of our simplified condition
is the correct version of the conditions given in [5, Corollary 3].

For the classical penalty f u n c t i o n we choose in the general problem m =

s + 1, qm(t) = t, K m = {1}, g,,l(x) = go(x) and

qi(t) = ~pt p, Ki = {1, 2}, g , ( x ) = O, gi2(x) = gi(x) for i -- 1.... , s.

Note that assumption (1.7) holds f o r all p t> 1.
 A. Ben-Tat and J. Zowe/ Optimality of nonsmooth problems 89

We illustrate the optimality conditions for the case p = 2, so we consider the

problem

rain f(x) := go(x) + ~Pl ~ [max{o, gi(x)}] 2. (6.22)

i=l

Theorem 6.5 [Necessary and sufficient conditions for problem (P4); the case
p = 2].
(a) A necessary condition for ~ to be a local minimizer of problem (6.22) is
that

g~(~) + p ,e~a+g,(~)g'i(Y,) = 0 (6.23)

and for every d E X,

gg(,Y)(d, d) + p ~ gi($)g'[(Y,)(d, d) + P ~ [max{O, g](2)d}] 2

lEA + i~:A

+ p ~_, [g'~(~)d]z>t O. (6.24)

lEA +

(b) If dim X < ~, a sufficient condition for ~ to be an isolated local minimizer for
problem (6.22) is as above but with strict inequality in (6.24) for d ~ O.
Here A, A + are defined as in (6.8).

ProoL The multipliers yij in Theorem 2.2 are independent from d for problem
(6.22). They are uniquely determined by condition (i) (note that at least d = 0
satisfies (o)):

Ys,+l,l = 1, Yi2= pgi(x) for i E A + and Y0 = 0 otherwise.

(i) and (ii) reduce to (6.23) and (6.24) and (o) becomes

g~(.~)d + p ~, g,(2)g',(.~)d <~O.

lEA +

Because of (6.23), this holds for every d in X.

7. Nonsmooth constraints

A general theory of second order necessary conditions for non-linear prob-

lems with nonsmooth data, whether in the objective function or in the con-
straints, was developed in [2]. These conditions are stated in terms of
(i) the cone of directions of quasi-decrease (quasifeasibility) of f : X ~ R, at
~, denoted by Ds(g),
(ii) the set of second order directions of decrease (feasibility) of f at ~ with
respect to the direction d, denoted by QI(~, d),
90 A. Ben-Tal and J. Zowe[ Optimality of nonsmooth problems

(iii) the support function 8"(. I Qi) of Qt(~, d),

8*(" I Qs) := sup (-, z).

z~qI

For the nonlinear problem

(NLP) min{fo(X) [ f~(x) <~O, i ~ I}

the necessary optimality conditions at a feasible solution ~ are (see [2] Theorem
9.1.):
For every d E ni:s~(~)=0Ds~(£)nDf0(£) there correspond continuous
linear functionals li E dom 8*(- I QI), not all zero, satisfying
(B-Z) lo + ~, l~ = 0 (the Euler-Lagrange equation),
i :/~(~)=o

8"(101 Q10)+ ~-- 8"(!, ] Qii) ~<0 (the Legendre inequality).

i : [i(£)=0
f~.~)d = 0

The results of Sections 3 and 4 of this paper can be interpreted as follows: For
the functional

f(x) - qi(max gij(x)) (7.1)

i=1 j~K i

we have
(i) Df(~) = {d I f'(x; d) ~< 0; f'(g; d) given by Proposition 3.3},
(ii) for d ~ Dt(~),
Qs(£, d) = {z I f"(£; d, z) < 0 ; f"(g; d, z) given by Proposition 3.3},
(iii) for Qr given in (ii) and d E Dt(~);

- ~ li~=m~ ~, YOg'~(x)(d,d)-~1/__~1 q,(g,,K,(

" g ))[g,,r,(e)(
' £ d) ] 2
'= JEKi(£, d ) "=

ify=~ ~ y0g'ii(£) for some Yij ~>0

~*(Y I Qs)= i=1 jEKi(~,d)
satisfying ~ Yo = q~(gi, n~(x)), i = 1 . . . . . m ,
jEKi(£, d)
oo otherwise.

Thus all the necessary elements to deal with objective functions and constraints
involving functionals of the general form (7.1), are given in (i)-(iii), and can be
directly incorporated into the (B-Z) optimality conditions.
 A. Ben-Tal and J. Zowe/ Optimality of nonsmooth problems 91

References

[1] A. Auslender, "Penalty methods for computing points that satisfy second order necessary
conditions", Mathematical Programming 17 (1979) 229-238.
[2] A. Ben-Tal and J. Zowe, "A unified theory of first and second order conditions for extremum
problems in topological vector spaces", Mathematical Programming Study 19 (1982) 39-76.
[3] A. Ben-Tal and J. Zowe, "Discrete /i-approximation and related nonlinear nondifferentiable
problems", Preprint, Institute of Mathematics, University of Bayreuth, Bayreuth (1980).
[4] C. Charalambous, "On the condition for optimality of the non-linear/l-problem", Mathematical
Programming 19 (1980) 178-185.
[5] T.F. Coleman and A.R. Conn, "Second order conditions for an exact penalty function",
Mathematical Programming 19 (1980) 178-185.
[6] J.M. Danskin, The theory of max-rain (Springer, Berlin, 1967).
[7] V.F. Dem'yanov and V.N. Malozemov, Introduction to minimax (Wiley, New York, 1974).
[8] V.F. Dem'yanov and A.B. Pevnyi, "Expansion with respect to a parameter of the extremal
values of game problems", USSR Computational Mathematics and Mathematical Physics 14
(1974) 33-45.
[9] R.J. Duffin, "Infinite programs", in: H.W. Kuhn and A.W. Tucker, eds., Linear inequalities and
related systems (Princeton University Press, Princeton, NH, 1956) pp. 157-171.
[10] R. Fletcher and G.A. Watson, "First and second order conditions for a class of nondifferenti-
able optimization problems", Mathematical Programming 18 (1980) 291-307.
[11] S.P. Han and O.L Mangasarian, "Exact penalty functions in nonlinear programming",
Mathematical Programming 17 (1979) 251-269.
[12] W. Krabs, Optimization and approximation (Wiley, New York, 1979).
[13] H. Maurer and J. Zowe, "First and second order necessary and sufficient optimality conditions
for infinite-dimensional programming problems", Mathematical Programming 16 (1979) 98-110.
[14] T. Pietrzykowski, "An exact penalty method for constrained maxima", SIAM Journal on
Numerical Analysis 6 (1969) 299-304.