JOl:K%AI.

Ok ,WPROXlhfATIOS ‘I’IIEORY 64, 343-350 ( 199 1 )

Projections onto Convex Cones

in Hiibert Space
JOHN M. INGRAM AND M. M. MARSH

Depurtmenr of Mathemuticr urd Sratistics,

Caiijorttiu Stare 1;niwristi.v. Sacramenro, C’di,fixnia 9.58 19-2694

C’ommunicarcd 1)~. Frank Deursdr

Rcccived January 17. 1990: rcvisctl July 13. 1990

Given a c!osed convex cone C’ in a Hilhcrt space H. we in%estiga!e the function

which assigns to each point x in H the ncarcst point of C to .x. We call this function
the propcrion of II onto c‘ and we give an algebraic characterization of thih
function which gcncralizes the well-known characterization of a projectmn onto
a closed subspace as an idempotcnt. symmetric linear operator. ( IV91 Ac;:dcmlc
R-C%. Inc

Let H be a real Hilbcrt space. The subset C of H is a Chrhyshet: set if

for each point x of H, there is a unique point of C which is nearest to .x;
i.e.. there is a point z in C such that I x - zI/ < !is - )!I/ for every y E C’,,{z j.
The point z is sometimes called the point of best approximation in C to x.
The set C is said to be concex if ts t (I -. t) i; E C whenever .x: 1’ E C and
0 6 t d 1. It is well-known that every closed, convex set in a Hilbert space
is a Chcbyshev set. It remains unknown if the converse (see Question 1
below) is true.

QLXXIOS 1. u C is u Chebysheti set in a Hilbert spucc. is C closed and

comex?
It is immediate that C is closed; so convexity is the real issue. A number
of people have given positive answers to Question 1 in a finite dimensional.
setting. See, for example, [4, 9, 10, 12, 15, 19 1. V. L. Kite [13] has the
most general positive result in an infinite dimensional setting. He shows
that weakly closed and Chebyshev implies convexity. G. G. Johnson [ I1 ]
has an example of a non-convex Chebyshev set in a non-complete inner
product space. It is not known if his example can bc extended to the com-
pletion of the space. Some other results related to this question can be
343
002 I-90459 1 s3.0
Copyright .I- 1991 by Acadcmlc PI%\. inc
,411 -Ights of reproduction m a!~y bm x.serwd
344 INGRAMANDMARSH

found in [ 1, 3, 16, 171. Thorough discussions of this question can be found

in [6; 19, Sect. 2; 203.
If C is a Chebyshev set, there is a natural surjective function p: H -+ C
such that, for each XE H, p(x) is the unique nearest point of C to x. We
will call p the protection of H onto C. Other authors have called p a
proximity or near-point function. The focus of this paper is to study the
algebraic characteristics of projections onto certain Chebyshev sets in
Hilbcrt spaces. In particular, we will give an algebraic characterization of
projections onto closed convex cones. This result generalizes the well-
known characterization of projections onto closed subspaces as idem-
potent, symmetric, linear operators (see [14, p. 3941 or [lS, p. 2991).
While this characterization is of interest in its own right, a similar charac-
terization, if one exists, of projections onto closed convex sets might
provide a new approach to answering Question 1.
Some other results related to projection mappings onto closed convex
cones can be found in [2, 7, 22, 23, 241.

11. CONVEX CONES AND PROJECTIONS

A Hilbert space H is a complete inner product space. A non-empty sub-

set of H is a cont’e,y cone if it is closed under addition and closed under
multiplication by positive scalars. WC will also assume that 0 is an element
of all cones under consideration in this paper. Linear subspaces are convex
cones and convex cones are clearly convex. We begin by stating some
elementary results about convex cones and projections onto closed convex
cones. These results will be of use to us later in the paper.

LEMMA 1. Let C he a concex cone.

(i) If’ XE H, y E C, and there exist r > 0 and /?< 0 such that
rx+/ly~C, then XEC.
(ii) Let x, y E H. There exist CI,b > 0 such that rx + fly E C if’ and only
[fthereexi.stO<t<l such that tx+(l-ft)y~C.
Proof (i) Now, l/r>0 and -/3!r>O, so x=(l/cx)(xx+~~)+
(-/l/lcr)YEC.
(ii) Suppose there exist x>O and /I>0 such that ax+by~ C.
Let t=a/(or+/?). Then O<t<l and tx+(l -t)y=(cr/(~+B))x+
(/I/(% + p)) y = (l/(a + p))(ax + /jy) E C. The other direction is trivial.
For CzH, let C*= (x~H[(x, y)<Oforevery YEC). (C* iscalled the
dual cone of C.) The results in the next lemma are well-known, so we omit
the proofs.
 CONVEX CONES IN HILBEKT‘ SPA(:E 345

LEMMA 2. Let Cc H.
(i) C’* is a closed, conwx cone.
(ii) C is a closed, convex cone if’ and oniy if C= c‘**.
(iii ) Let C he a closed, cuncex cone und x E H. Thrn there is CI unique
xl E C and u unique x2 E C* such that x = x, -t x2 and (x: , x2 ) = 0. (.u, ic
the neurest pninr of‘ C to x and x2 is the nearest point of C* to s.)

LEMMA 3. Let C he u closed concex cone, p the projection of’ H on!o C,

und x, J’ E H. Then
(i) p’= p (p is idempotent).
(ii) p(w) = rp(x) jbr x 2 0 (p is non-neguticelJ homogeneous).
(iii) p(x+y)=p(x)+p(y) ij‘und 0nl.y if (p(x). .r)= (p(x). p(.v))
= (.c A,,)>.
(iv) (x-p(x), p(y)> do.
(v) (x - p(x), p(x)) = 0.
(vi) 1p(x) - P(J);’ < I,.r- y 1 (p is non-e.Ypansice).
(vii) I- p is the projection qf H onto C*.
(viii) C* = p ‘(0).
Proc$ With the possible exception of part (iii), all of these are well-
known facts. Therefore, WC will prove only part (iii). Proofs for parts (i)
and (ii) are easy. Parts (iv) and (v) actually characterize the element p(.r)
in C which is the projection of x onto C (see [S, Prop. 1.12.41). That p is
non-expansive can be found in [5, Th. 31. Proofs for parts (vii) and (viii)
are straightforward.
Proof‘ of (iii). WC first note that by part (v), we have that for x. >‘E fi.

<p(y),x-p(x))+(p(.\-),Y-P(Y))
= (P(-~), x - p(x) > + (P(J), x - p(x) >
+ <pt.\-I? Y-P(J)) + <P(Y)> JC- P(J))
= (p(.r) + P(Y), x + r’- (p(x) + p(y))). (*I

Now suppose X, J E H and p(+r + ,:) = p(x) + p(y). Then by (*) and
part (v), we have

Since both (p(y). x-p(x)) and (p(x), -v- p(y)) are non-positive by
part (iv), this implies that (p(v), x - p(x)) = 0 = (p(x), y - p(y)).
346 INGRAM AND MARSH

Now suppose (p(y), x - p(x)) = 0 = (p(x), y - p(y)). Write

x + Y = P(X) + P(Y) + (x + Y - (P(X) + P(Y))),

where p(x)+p(y)~C, and x+y-(p(x)+p(y))~C* by part (vii) and
Lemma2, part(i). Also, (p(x)+p(y), x+y-(p(x)+p(~)))=O by (*)
and our assumption. Observe that

is also a representation of x + J in the form given by Lemma 2, part (iii).

By the uniqueness of this type of representation, we must have that
Ax + Y) = P(X) + P(Y).

III. IIXWOTENT, POSITIVELY HOMOGENEOUS,

FACE-LINEAR TRANSFORMATIONS

We will focus on properties (i), (ii), and (iii) of Lemma 3. They arc
analogous to the idcmpotcnt, symmetric, and linear properties which
characterize projections onto closed subspaces. In particular, property (iii)
says that a projection onto a closed, convex cone is additive with respect
to two points if and only if it is symmetric with respect to these points. We
will say that a function which satisfies this property is face-linear. For the
remainder of this paper, WC let n: H + H denote an arbitrary function
which satisfies the following three properties. For x, y E H,
(1) n2(x)=nbh
(2) n(ax) = m(x) for r > 0: and
(3) n(x+y)=n(x)+n(y) if and only if (n(x),y)=(n(x),n(y))=
(4 n(y)).
For convenience, we let K = n(H). For A4 E H, let I@ and F(M) dcnotc
respectively the topological closure and boundary of M. We now consider
the following question.

QIXSTION 2. Is K a closed, comex cone and, if‘so, is n the projection of

H onto K?
We begin by looking at some of the properties that n and K must have.

THEOREM I .l. Let x, y E H. Then

(i) (x-n(x),n(x))=O.
(ii) IIn( < IIxil. Furthermore, Iln(x)ll = llxll ifand only iJn(x) = x if
and anly I~XE K; and Iln(x)ll < llx!l ifand only if‘x$ K.
 CONVEX CONES IN HILRERT SPACE 347

(iii) n(O) = 0.
(iv) K is a cont:ex cone.
(v) Jf rz(x+y)=n(x)+n(y), then? for cccr)‘ r.j330, n(rx+/Yy)=
n(w(x) + ,by) = n(rx + [h(y)) = n(m(x) + pn(J)) = m(x) $ /h(y).
(vi) If‘ n(x) = n(y), then, jbr z, /j 3 0, iz((xx + /jJ*) = (Y i /I) n(x). In
pariicular, n(rx + /?n(x)) = (51+ /3) n(x).
(vii) If’ xEn ‘(O), y E K: and (A-, J.) = 0, then. jbr r, /I 3 0.
n( xx + /jy) = py.
(viii) Jj’x~n-l(O), ~EK, (x, y)=O, and rx+fiyL’K.for some r>O
and /I > 0, then x = 0.
(ix) &c YX + on(x) E K.for some CL> 0 und /I E R, then x E K.
(x) [fx$ K, then n(x)EI;(K).
(xi) K* cn-‘(0).

Proof: (i) By property (2), n(x + x) = n(2x) = 2n(x) = n(x) + n(x). By

property (3), this implies that (/z(x): x) = (n(x). n(x)).
(ii) By part (i), WC get that jlxll’= 11x--rz(x)I!‘+ Iln(xl12. This
implies that :in(xll < /1x1/. If I,n(x)l! = II-YII, then we get that /Ix -n(x)!1 = 0;
and thus n(x) =x. if x E K, then x = n(y) for some ?; E H. By property (I),
this implies that n(x) =n(n(y)) =n(y)= x. The other implications of (ii)
arc trivial.
(iii) By part (ii), :liz(O)il < IOIl = 0. Thus, n(0) = 0.
(iv) Suppose x, J E K and x > 0. By part (ii), n(x) =x and n(y) = y.
Hence, each of (n(x), y), (n(x), n(y)). and (x, n(J)) is equal to (s, .i; ).
We have that n(rx) =m(x) = XX by property (2), and n{x+ y)=
n(x)+ n(y) by property (3). From part (ii) again we have that
rx, x + ?: E K. Thus, K is a convex cone.
(v) BY property (3), (n(x), Y> = <n(x), n(y)) = C-K,H(Y)). BY
property (2), we get that

(n(rx), n(KvL’)> = d<n(-K). n(y) >.

(xx, n(b.v)) =d(.x, HO,)).

Thus, <4Xx), PY) > = (n(rx), n(PyJv)> = < xx, n(j3y)). By properties (3) and
(2), this implies that n(rx+ by) = n(rx) + n(/311)= m(x) + /3n(y). Using
properties (1) and (2), the other equalities in this part follow in a similar
manner.
(vi) By part (i), each of (n(x), 1~). (n(.u), n(y)). and (x, n(y)) is
348 INGRAM AS11 MARSH

equal to I n(x);12. By property (3), this implies that n(x+ y) =

n(x) + n(y) = 2n(x). This part now follows from part (v).
(vii) This statement follows from property (3) and from (v) above.
(viii) By (vii), n(crx+b~) =py. Since c(x + /I~EK, it follows that
n(xx + PJ) = cxx+ fly. Thus, xx = 0; and therefore, x = 0.
(ix) If ,420, it follows from (vi) that n(crx+pn(x))= (a +/?)n(x).
Since tlx + fin(x) E K, we get that n(rx + /In(x)) = rx + pn(x). Hence,
x = n(x), and thus x E K. If /I < 0, let t = l/( 1 - /I). Clearly 0 < I < 1. Since
K is convex, t(ax+~n(x))+(l-t)n(x)~K. Now, t(rx+/?n(x))+
(1 - t) n(x) = (tr)x. Hence, since K is a cone, x E K.
(x) Since XI$ K, it follows from (ix) that for each r >O,
xx + (1 - X) n(x) 4 K. Hence, n(x) E F(K).
(xi) Suppose XE K*. Then by part (i), I n(x) I2 = (x, n(x)) GO (since
n(x) E K). Thus, n(x) = 0, and hence x En ‘(0).
Since the image of n, namely K, must be a convex cone, we let p be the
projection of H onto X and we note two relationships between n and p.

THEOREM 1.2. Let x E H. Then

(i) !fp(x) En l(O), then, f orr,B~O,rp(x)+p(x-p(x))En '(0).
(ii) Zfp(x) EK, then n(x) = p(x).
Proof. (i) By Lemma 3, part (vii), x-p(x) E K* = K*. By Theorem 1.1,
part (xi), n(x- p(x))=O. Thus, by Theorem 1.1, part (vi), n(rp(x) +
P(x - Ax))) = (r + B) Mx)) =O.
(ii) As in (i) above, n(x- p(x)) =O. Also, by Lemma 3, part (v),
(x-p(x), p(x)) =O. Thus, by Theorem 1.1, part (vii), n(x) =n((x- p(x))
+ p(x)) = P(X).
Note that if p is the projection of H onto a closed, convex cone, then p
also has all of the properties listed in Theorem 1.1. It is not immediately
apparent whether properties (1 ), (2), and (3) imply that K is closed or that
n is the projection of H onto K. The next result shows that if K is closed,
then we do have that n is the projection of H onto K.

THEOREM 2. Suppose that n: H + H is a function, C = n(H), and C is

closed. Then C is a closed, concex cone and n is the projbction of H onto C
if and only if n satisfies properties ( I), (2), and (3).
Proof: If n satisfies properties ( 1), (2), and (3), then by Theorem 1.1,
part (iv), C is a closed, convex cone. By Theorem 1.2, part (ii), p(x) = n(x)
for all x E H, where p is the projection mapping of H onto C= C. Hence,
n is the projection mapping of H onto C.
 COI\;VEX CONES I?i HILBERT SPACF. 349

The opposite implication is immediate from Lemma 3.

COROLLARY 1. Suppose that n: H -+ H is II function und C = n(H). Then

C is a closed, conc’ex cone and n is the prqjection mapping of H onto C ij
nnd only ifn is a continuous,finction satisfj:ing properties (1 ), (2), and (3).
Proof: Suppose n is continuous and satisfies(I), (2) and (3). It foliows
from continuity and idempotentness that C is closed. Thus, by Theorem 2,
C is a closed convex cone and n is the projection of H onto C.
The opposite implication again follows from Lemma 3.
WC point out that if n is the projection mapping onto a closed convex
cone, then it follows from Lemma 3(vi) that 17is continuous. It would be
of interest to know if one must assume continuity of n to obtain the
opposite implication in the characterization given in Coro!lary 1. Thus: we
ask the following question.

QUESTION 3.1. In Corollary l? can the assumptionthut n is continuous hc

omitted.?
It is equivalent to ask the following.

QUESTION 3.2. In Theorem 2, can the ussumption thut C is ciosed be

omitted?
If properties (2) and (3) of n arc replaced with “n is symmetric and
linear,” it is well-known that these conditions characterize n as the projec-
tion onto a closed linear subspaceof H. In this setting, the assumption that
n is continuous (or that n(H) is closed) is not necessary.
The second author can show that the answer to Question 3.1 (or
Question 3.2) is YES if dim(H) < 3.

ACKNOWLEDGYENTS

The authors thank Frank Deutsch for helpful suggestions pertaining to the presentation 3f
the main result.

REFERENCES

1. E. ASPLUSI). ChebySev sets in Hilbcrt space, Twrs. ~mer. Mar/~. SW. 144 (1969).
235-140.
2. M. BARWTI AND P. L. PAPIM, Best approximation and cones in Banach spaces, ;PT
“Non-
linear Functional Analysis and Its Applications,” pp. 169-176, Reidel. Dordrecht, 1986.
3. L. N. H. BC‘NT, “Bijdrage tot de Theorie der convcxe Puntvcrzamelingen.“ Thesis, Univer-
sity of (ironingen, Amsterdam, 1934.
350 INGRAM AND MARSH

4. H. BUSIXANN, Note on a theorem on convex sets, Mar. Tidsskr. (1947): 32-34.

5. W. CHEWY AND A. A. GOLDSTEIN: Proximity maps for convex sets. Proc. ,4mer. Math.
SOC. 10 (1959), 448 450.
6. M. M. DAY, “Normed Linear Spaces,” Ergeb. Math. Gre7. 21, Springer-Verlag, Berlin,
1959.
7. P. DEUTXH, J. II. M~CABIT, AND Cr. M. PIIILLIPS, Some algorithms for computing best
approximations from convex cones, SIA!M J. Xum. Anal. 12, No. 3 (1975). 390403.
8. R. E. EDWARDS, “Functional Analysis,” Holt. Rinehart & Winston, New York, 1965.
9. N. V. EFIMOV ASD S. B. STECKIS. Chebyshev sets in Banach spaces, Dokl. Akad. Nauk.
SSSR 121 (1958), 582-585.
IO. N. V. EFIMOV ANI) S. B. STECKIX. Some properties of Chebyshev sets, Dokl. Akad. Xuuk.
SSSR 118 (1958), 17 19.
11. G. G. JOHNSON, A nonconvex set which has the unique nearest point property. J. Approx.
Theory 51, No. 4 (1987), 289-332.
12. V. L. KLF& JR.. A characterization of convex sets, Amer. Muth. Monthly 56 (1949),
247-249.
13. V. L. KI.EE, JR., Convexity of Chebyshcv sets. Math. Ann. 142 (1961), 292-304.
14. R. LAKSCN, “Functional Analysis (an Introduction),” Decker, New York, 1973.
15. T. S. MOTTKIK;, Sur quelques proprittes caracteristiques des ensembles born&s non
convexes, Rend. R. Accad. Lincei Cl. Sci. Fix Mol. Xaf. 21 (1935), 773 779.
16. R. R. ALPS, Convex sets and nearest points, Proc. Amer. Muth. Sot. 8 (1957), 790 797.
17. R. R. PHIXPS, Convex sets and nearest points, II, Proc. Amer. Marh. Sm. 9 (1958).
867-873.
lg. W. R~DIS, “Functional Analysis,” McGraw-Hill, New York, 1973.
19. I. SINGER, “Best Approximation in Normed Linear Spaces by Elements of Linear Sub-
spaces,” pp. 364 371. Springer-Verlag. Berlin, 1970.
20. F. A. VAIENTINE, “Convex Sets,” pp. 94-98 and 179.. 182, McGraw- Hill, New York, 1970.
21. L. R. WEILL, Do unique closest elements imply convexity? Unsolved Problems, Amer.
Mafh. Monthly 93, No. 8 (1986), 628.
22. E. H. ZARANTONELLO, L’algibrc des projccteurs coniques, in “Analyse convcxe ct ses
applications,” pp. 232- 242, Springer-Verlag, Berlin, 1974.
23. E. H. ZAKASTONELI.O, Projections on convex sets and spectral theory, in “Contributions
to Nonlinear Func;ional Analysis,” pp. 237 424, Academic Press, New York, 1971.
24. E. H. ZAKANTONELLO, Projectors on convex sets in reflexive Banach spaces, Rel;. L’n. 44~.
Argenrinu 29 (1984), 252 269.