Optimization

ISSN: 0233-1934 (Print) 1029-4945 (Online) Journal homepage: https://www.tandfonline.com/loi/gopt20

On the convergence of the proximal algorithm for

saddle-point problems

A. Moudafi

To cite this article: A. Moudafi (1995) On the convergence of the proximal algorithm for
saddle-point problems, Optimization, 33:3, 191-200, DOI: 10.1080/02331939508844076

To link to this article: https://doi.org/10.1080/02331939508844076

Published online: 20 Mar 2007.

Submit your article to this journal

Article views: 65

View related articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gopt20
Optimization, 1995, Vol. 33, pp. 191-200 0 1995 OPA (Overseas Publishers Association)
Reprints available directly from the publisher Amsterdam B.V. Published under license by
Photocopying permitted by license only Gordon and Breach Science Publishers SA
Printed in Malaysia

ON THE CONVERGENCE OF THE

PROXIMAL ALGORITHM FOR
SADDLE-POINT PROBLEMS

A. MOUDAFI
LACO, URA 1586, Universitk de Limoges 87060 Limoges Cedex, France
(Receiued 16 March 1994; in final form 14 September 1994)

The proximal algorithm for saddle-point problems min,,, max,,, L(x,y), where X, Yare Hilbert spaces and
L:X x Y-+R is a proper, closed convex-concave function in X x Y is considered. Under a minimal
assumption on L, 6:= inf sup L= sup inf L, it is proved that the proximmai point algorithm (PPA), with
positive parameters {I.,), converges in general provided that on:= XI:=, I.,+ + co.

KEY WORDS Convex-concave functions, saddle-points, proximal point algorithm.

Mathematics Subject Classification 1991:
Primary: 90C30, Secondary: 49M37

1. INTRODUCTION

Let X, Y be two real Hilbert spaces, we consider the convex-concave saddle-point

problem

min max L(x, y)

xex yeY

where L X x Y+ @ be a convex-concave function which is closed and proper in the

sense of Ref. 11.
Many problems from mathematical economics, equilibrium problems, location
problems and solutions of games can be formulated as problem (1). One method for
solving (1) is the proximal point algorithm (PPA)introduced by Rockafellar [lo]. This
method is based on the notion of proximal mapping J,,
1 1
J,(x,y)=(x,,y,)=argminmax{L(u,u)+-~u-xl2--lv-y12)
usx ueY 2A 22
it starts at a point (x,, y&X x Y, and generates recursively a sequence of points
{(xk,Y,)) by the rule

where {iis
,a)sequence of positive numbers.
191
192 A. MOUDAFI

U p to now the PPA was known to converge under the assumption that the generated
sequence is bounded or, equivalently, under the existence of saddle-points (th. 5,
Rockafellar [lo]). The aim of this work is twofold. First (perhaps the more important) is
to investigate the convergence of the PPA in the case where L has no saddle-points. The
second is to give a direct proof in the case when Argminmax L # 0, actually theorem
5 (in [lo]) is obtained as an application of a result in the more general framework of
maximal monotone inclusions.
We use the following notation in this paper. A function L X x Y + [W is convex-
concave if it is convex in the x variable and concave in the y variable. For E > 0 and
(x,, y,)eX x Y, TE,(x,, yo) is defined as the set of (u,, u,) satisfying

When E = 0, TE, reduces to T,, the maximal monotone operator introduced by

Rockafellar [I 11.
For E > 0, the E - Argminmax of a function L is defined as the set of (x,, yo) verifying

This set reduces to the exact Argminmax of L when E = 0.

Finally, d(x, C) will denote the distance from the point x to the set C, we set a,:=
,
C; = A, and we assume 6:= inf sup L = sup inf L.
The organization of the rest of the paper is as follows. Since for practical purposes
{(x,, y,)) should be obtained according to some approximation criteria rather than
exactly as in (2), it is proposed in section 2 an approximate version of the PPA. Its
convergence is proved without boundedness assumption and is sharpened in the case
when L admits saddle-points. In section 3 a perturbed version of the PPA is proposed
and its convergence result is established under mild conditions on the so-called
marginal functions associated with L. An application to penalization in constrained
saddle-point problem is also given.

2. THE MAIN CONVERGENCE RESULTS O F THE PPA

2.1. An Approximate Version of the PPA

From a practical point of view it is essential to replace the iteration scheme (2) by
a looser relation which takes into account an approximate calculation of the iterates.
To this end, we propose the following procedure:

T::= a,,,, L x a,,,,(- L)

where { E ~is}a sequence of positive numbers which go to 0 and dl,, (resp. a,,,) stands for
the E-subdifferential of L with respect to the first (resp. second) variable.
 PROXIMAL ALGORITHM 193

Let us first observe that if ek = 0, V ~ N,

E the iterative scheme (3) reduces to the PPA.
The next result contains the fundamental estimates from which we derive conver-
gence result of the approximate PPA.
Lemma 2.1: Let {%,) be an arbitrary sequence of positive numbers and {(x,, y,)) the
sequence generated by (3). Then for any (u,V ) E Xx Y ,

and

Proof: Relation (3)gives

which combined with definition of T? yield

Since
2
2(x,~,-x,,u-x,)=Ix,~,-x,l +lu-~,l~-lu-x~-,1~ (7)
and according to (6),we obtain
2)LkL(u,yk) 2 2AkL(xk,yk) + IU - xkI2- I U -xk- + Ixk- -xk12 - 2Akck.
Similarly, we have
+
2 i k L ( x kV, ) < 21, L(xk,yk) Iv - ykI2 - Iv - yk- 11 2 - Iyk- 1 - ykI2 + 2'kEk'
Summing the last inequalities for k = 1 to n, we obtain

and

Dividing by on and taking into account the fact that

we deduce the desired result. .

L(u,yk) 6 sup L(u,y) and L(xk,v) 2 inf L ( x , v),
Y ~ Y xsx
194 A. MOUDAFI

Now, let us recall the following result which is due to Silverman-Toeplitz and which
will be needed in the proof of the next theorem.
Lemma 2.2 (see for example [3]): Let {I,,), {a,) be two sequences of numbers with
A, >, 0 for all n and limn, ,a, exists. Then
+

The two next results give convergence properties of the approximate PPA.
Theorem 2.3: Let {(x,, y,)} be a sequence generated by (3) and assume
(i) lim E, = 0,
n-+oo
(ii) lim an = + co,
n-+m

then

Proof: Passing to the limit in (4) and (9,we can write

We also have
lim sup "" L(Xk'
< inf sup L(X,y) + lim -Iu - x0l2 + lim C;: = 1 A,&,
n-+m an XPX YEY n-+ m 2an n-+a On

From condition (i), (ii) and lemma 2.2, we obtain the announced result (8). H
Let us consider now the case when Argminmax L # 0.
Theorem 2.4: Let {(x,, y,)) be a sequence generated by (3)and suppose Argminmax L # 0,
then the following estimate holds true:
d2((x0,yo),Argminmax L)
2an
+ =,
an
i,~,
(9)

If in addition,
(i) A k 2 A > 0 ,
(ii) Akek< + CO,

then

(a) lim lx,-x,-,l=0 and lim lyk-y,-,l=O

k-+m k-+m
(b) lim L(x,, y,) = 6
k-+m
(c) (x,, y,) converges weakly to (x*, y*)~ArgminmaxL.
 PROXIMAL ALGORITHM 195

Proof: Formula (9) follows by taking u = 2 in (5) and v = j in (4), where (2, j) is the
point in Argminmax L closest to (x,, yo) and taking into account the fact that when
(2, j)~ArgminmaxL, we have

L(2, j) = sup L(2, y) = inf L(x, j).

YEY xsx

Let us first show that sequence {(x,, y,)) is bounded. Indeed, let (2, j)~ArgminmaxL,
thanks to (3) and (7) we get the relation (10):

which combined with

L(2, y,) < sup L(2, y) = L(2, j ) = 6

YEY
(11)
L(x,, j ) 2 inf L(x, j ) = L(2, j ) = 6
xEX

yield

and

Adding the two last inequalities, we obtain the relation (12):

exist and is finite (see for example [7], proposition 5.6.1.2.).

In view of (12) and (13), we get easily that the sequence {(x,, y,)) is asymptotically
regular, that is

Since {A,) is bounded away from zero, we also have

196 A. MOUDAFI

Taking u = 2, v = J in (6) and according to ( 1 I), we get

Passing to the limit in the two last inequalities and thanks to (14) and the fact that
{(x,,y,)} is bounded, we obtain

lim sup L(x,, y,) <6 and lim inf L(x,, y,) 2 6,
,-+a k-+co

thus
lim L(x,, y,) = 6.
k++m

Now let {(x*,y*)) be a weak cluster point of {(x,,y,)). There exists a subsequence
{(x,,y,)) which converges weakly to (x*,y*) and
( X V; x., Yv- I
4
-Y v
1
E TY (xv,Y,).

Since E , -0 and the graph of the maximal monotone operator T Lis weakly-strongly
closed, we get at the limit
(0,0 )TL(x*,Y*)
~
or (x*,y*) is a saddle-point of L.
It remains to prove that there is no more than one weak cluster point, our argument
follows that given in (Ref. 10, p. 885) and is presented here for completeness.
Let (2,y")be another weak cluster point of {(x,,y,)), we will show that 2 = x* and
y" = y*. This is a consequence of (13).Indeed,

-
we see that the limit of ((x,, y,) - (x*,y*), (x*,y*) - (2,j ) ) as k + cc must exists. This
limit has to be zero because (x*,y*) is a weak cluster point of {(x,,y,)). Hence, at the
limit

Reversing the role of (2,j) and (x*,y*) we also have

P(x*,Y*)= p(%y") + I(x*,Y*)-(2,Y")l2,
That implies (2,y") = (x*,y*), which completes the proof.
 PROXIMAL ALGORITHM 197

3. PERTURBATION O F THE APPROXIMATE PPA

3.1. A Perturbed Version of the Approximate PPA

Here, as suggested by Lemaire [5] in the convex case, we add another perturbation by
replacing, in the iteration scheme, L by a convex-concave function Lk which is proper
and closed. That is the PPA is not applied to the given problem itself but rather to
a fixed approximation of it (for instance by discretization or penalization). The
resulting procedure is

The next theorem gives a convergence result for the iterative scheme (15).

Theorem 3.1: Let {(x,, y,)) be a sequence generated by (15) and assume

(i) lirn ck= 0, lirn on= + co.

k'+ m n-+m
(ii) Mk:= supLk(.,y) pointwise converges to f with dom Mk 3 dom f or Mk <f .
YEY
(iii) mk:= inf Lk(x,.) pointwise converges to g with dom mk 3 dom g or mk 2 g.
xsx
(iv) 6 := inf f (x) = sup g (y),
xsx YEY

then

lim X i = 1 ) ~ k ~ k ( ~~ kk>-) 6.
n-+m cn

Proof: By the same arguments as in the proof of lemma 2.1, we have

and

from which we deduce

and
198 A. MOUDAFI

Passing to the limit in the above inequalities and taking into account lemma 2.2, (i), (ii)
and (iii), we obtain

and

lim sup Xi= I (I,Lk(xk2Y,) lim sup 1 'kMk(~) = lim Mn(u)=f(u)

n-+m 'Jn n++m 'Jn n-+ m

the desired result follows according to (iv). H

3.2. Example
We give an application to penalization in constrained saddle-point problems.
Let C (resp. D) be a nonempty closed convex sets of X (resp. of Y) and Lo:X x Y - +R be
a continuous convex-concave function. It is well known that finding a saddle-point of
Lo on C x D and finding a saddle-point of L on X x Yare equivalent problems, where
L is given by

Now, let us consider a sequence { L k }of special approximations of L defined as

where 4,:X -,[Wi and $,: Y-, 54' are convex lower semicontinuous penalty functions
such that

It is the case for exterior penalty functions (see for example [2], [ 5 ] or [9]). Now, let M,,
mk be the marginal functions associated with L,, we have

and
 PROXIMAL ALGORITHM

Clearly

and
Mk(4 Gf (XI =
i+Mo(x) i f x ~ C
cc otherwise

i
mo(y) if YED
m,(y) > g(y) = - cc otherwise.

So hypothesis (ii), (iii) and (iv) of theorem 3.1 are satisfied.

We note that in the case when X:= Rn,Y:= Rmand C, D are specified by the convex
inequalities

and

there are two useful penalty approximations, namely

(gTl2, O<r,'t+CO.
Quadratic exterior penalty
(hf)2, O < r k ? + CO.

Acknowledgments
Thanks are due to Professor Bernard Lemaire for his many insightful comments.

References
[I] Auslender A. (1985)Two general methods for computing saddle points with applications for decompos-
ing convex programming problems, Appl. Math. Optim. 13,79-95
[2] Auslender A,, Crouzeix J. P. and Fedit P. (1987) Penalty proximal methods in convex programming, J.
Optimization Theory Applications, 55(1), 1-21
[3] Dunfor N. and Shwartz J. (1988) Linear operators, part I, general theory, New York
[4] Giiler 0. (1991) On the convergence of the proximal algorithm for minimization problems. S I A M J.
Control Optimization, 29,403-419
[5] Lemaire B. (1988) Coupling optimization methods and variational convergence, Trends in Mathemat-
ical Optimization
163 Lemaire B. About the convergence of the proximal method, 6th French-German conference on
Optimization
[7] Martinet B. (1972) Algorithmes pour la resolution des problemes d'optimisation et minimax, These
d'Etat universite de Grenoble
[8] Moudafi A. (1993) Coupling proximal methods and variational convergence, ZOR-Methods and
Models of Operations Research, 269-280
[9] Mouallif K. (1989) Convergence variationnelle et methodes perturbees pour les problemes d'optimisa-
tion et de point selle, these d'Etat Universite de Liege
200 A. MOUDAFI

[lo] Rockafellar R. T. (1976) Monotone operators and the proximal point algorithm, SIAM J . Control
Optimization, 14, 877-898
[l 11 Rockafellar R. T. (1970)Monotone operators associated with saddle functions and minimax problems,
Nonlinear functional analysis, Part I, F. E. Browder ed. Symposia in Pure Math., vol. 18, Amer Math.
Soc., Providence, 397-407