Remark 2.1.9 This result gives one more argument in favour of our relative topology: if we take a closed convex set C, open it (for the topology of aff C), and close the result, we obtain
2 Functional Operations Preserving Convexity
2.1 Operations Preserving Closedness ( a) Positive Combinations of Functions
Proposition 2.1.1 Let II, ... , 1m be in Conv IRn [resp. in Conv IRn], tl, ... , tm be positive numbers, and assume that there is a point where all the
/j s
are finite. Thenthe function
m 1:= Ltj/j
j=1 is in Conv IRn [resp. in Conv IRn ).
PROOF. The convexity of
I
is readily proved from the relation of definition (1.1.1).As for its closedness, start from
liminftjlj(Y)
=
tj liminf Ij(Y) ~ tj/j(x)y-+x y-+x
(valid for tj > 0 and
/j
closed); then note that the lim inf of a sum is not smaller thanthe sum of lim inf's. 0
As an example, let
f
E Conv Rn and C C ]Rn be closed convex, with domf
n Ci=
0.Then the function f
+
Ie of Example 1. 3 (a) is in Conv IR.n • This trick can be used to simplify the notation for constrained minimization problems:inf {f(x) : x E C} and inf{(f
+
Ie)(x) : x E ]Rn}are clearly equivalent in the sense that they have the same infimal value and the same solution-set.
(b) Supremum of Convex Functions
Proposition 2.1.2 Let {fj }jE} be an arbitrary family of convex [resp. closed con-vex] functions.
If
there exists Xo such that sup} Ij(xo) < +00, then their pointwise supremumI
:= sup{/j :
j E J}is in Conv]Rn [resp. in Conv IRn).
PROOF. The key property is that a supremum of functions corresponds to an intersec-tion of epigraphs: epi I
=
njE} epi Ij' which conserves convexity and closedness.The only needed restriction is nonemptiness of this intersection. 0 In a way, this result was already announced by Proposition 1.2.8. It has also been used again and again in the examples of § 1.3.
Example :U.3 (Conjugate Function) Let f : JRn ~ iR U {+oo} be a function not identi-cally +~, minorized by an affine function (i.e., for some (so, b) E JRn x JR, f ~ (so, .) - b on ]Rn). Then,
f* :]Rn 3 s 1-+ sup {(s,x) - f(x) : x E dom!}
is called the conjugate function of f, to be studied thoroughly in Chap. X. Observe that I*(so):::; band I*(s) > -00 for all s because domf
!
121. Thus,1*
E Conv]Rn; this is true without any further assumption on f, in particular its convexity or closedness are totallyirrelevant here. 0
Example 2.1.4 Let S be a nonempty set (not necessarily convex) and take ]Rn 3 x 1-+ IPS(x) :=
! [lIxlI2 -
d~(x)] ,where dS is the distance function to S, associated with the Euclidean norm 11·11. A surprising fact is that IPs is always convex. To see it, develop
d~(x)
=
infIIx - cll2 = IIxll2 -
sup [2(c, x)-lIcU2]
cES ceS
to obtain
IPS(x) = sup {(c, x) -
!lIcll2 :
c ES} ;
IPS thus appears as the pointwise supremum of the affine functions (c, .) - 1/211c1l2, and is closed and convex. In view of the previous example, the reader will realize that IPs is the
conjugate of the function 1/211 . 112
+
Is. 0(c) Pre-Composition with an Affine Mapping
Proposition 2.1.5 Let
f
E Conv]Rn [resp. Conv]RnJ
and let A be an affine mapping from ]Rm to ]Rn such that 1m An
domf
=1= 0. Then the functionf 0 A: ]Rm 3 X ~ (f 0 A)(x)
=
f(A(x»is in Conv]Rm [resp. Conv]Rm
J.
PROOF. Clearly (f 0 A)(x) > -00 for all x, and there exists by assumption y
=
A(x) E ]Rn such that f(y) < +00. To check convexity, it suffices to plug the relation A(ax
+
(1 - a)x') = aA(x)+
(1 - a)A(x')into the analytical definition (1.1.1) of convexity. As for closedness, it comes readily
from the continuity of A when
f
is itself closed. 0Example 2.1.6 With f (closed) convex on lRn , take Xo E dom f, d E ]Rn and define A: lR 3 t 1-+ A(t) = Xo
+
td;this A is affine, its linear part is t 1-+ Aot := td. The reSUlting f 0 A appears as (a parametriza-tion of) the restricparametriza-tion of f along the line Xo
+
Rd, which meets dom f (at xo).This operation is often used in applications: think for example of the line-search problem, considered in §II.3. Even from a theoretical point of view, the one-dimensional traces of f
are important, in that f itself inherits many of their properties; Proposition 1.2.5 gives an
instance ofthis phenomenon. 0
160 IV. Convex Functions of Several Variables
Remark 2.1.7 With relation to this operation on f E ConvlRn [resp. ConvlRn ], call V the subspace parallel to aff dom f. Then, fix Xo E dom f and define the convex function fo E Conv V [resp. Conv V] by
fo(Y) := f(xo + Y) for all y E V.
This new function is obtained from
f
by a simple translation, composed with a restriction (from ]Rn to V). As a result, dom fo is now full-dimensional (in V), the relative topology relevant for fo is the standard topology of V. This trick is often useful and explains why "flat"domains, instead of full-dimensional, create little difficulties. 0
(d) Post-Composition with an Increasing Convex Function
Proposition 2.1.8 Let f E Conv IRn [resp. Conv IRn] and let g E Conv IR [resp.
Conv lR] be increasing. Assume that there is Xo E IRn such that f (xo) E dom g, and set g(+oo) :=
+00.
Then the compositejunction g 0 f : x f-+ g(f(x)) is in ConvlRn [resp. in Conv IRnj.PROOF. It suffices to check the inequalities of definition: (1.1.1) for convexity, (1.2.3)
for closedness. 0
The exponential g(t) := exp t is convex increasing, its domain is the whole line, so exp f (x) is a [closed] convex function of x E IRn whenever f is [closed1 convex. A function
f :
IRn ~ ]0, +00] is called logarithmically convex when logf
E Conv IRn (we set again log(+oo) = +00). Because f=
explog f, a logarithmically convex function is convex.As another application, the square of an arbitrary nonnegative convex function (for ex-ample a norm) is convex: post-compose it by the function g(t) = (max{O, t})2.
2.2 Dilations and Perspectives of a Function For a convex function
t
and u > 0, the functionfu : IRn 3 x f-+ fu(x)
=
uf(x/u)is again convex. This comes from Propositions 2.1.1 and 2.1.5 but can also be seen geometrically: since fu(x)/u = f(x/u), the epigraphs and sublevel-sets are related by
epi fu
=
u epi f, episfu=
u epist,
Sr (fu)=
U Sr /u (f), which express that fu is a "dilated version" of f.More interesting, however, is to study fu as a function of both variables x and u, i.e. to consider the set of all dilations of f. We therefore define the perspective of f as the function from IR x IRn to IR U {+oo} given by
j(u, x) := {uf(x/U) if u > 0,
+00
ifnot.Proposition 2.2.1
If t
E Conv IRn, its perspective j is in Conv IRn+
I.PROOF. Here also, it is better to look at
j
with "geometric glasses":epij = {(u,x,r)ElRtxlRnxlR: f(x/u)~r/u}
= {u(l,x',r'): u > 0, (x',r') E epif}
= Uu>o{u({l} x epif)} = lRt({l} x epif) and epi
j
is therefore a convex cone.~(I
Fig.2.2.1. The perspective of a convex function
o
Figure 2.2.1 illustrates the construction of epi
j,
as given in the above proof. Embed epi f into IR x JRn x JR, where the first JR represents the extra variable u; shift it horizontally by one unit; finally, take the positive multiples ofthe result. Observe that, following the same technique, we obtaindomj = 1R~({l} x domf) . (2.2.1)
Another observation is that, by construction, epi j [resp. dom j) does not contain the origin oflR x lRn x IR [resp.lR x lRn).
Convexity of a perspective-function is an important property, which we will use later in the following way. For fixed Xo E dom f, the function d ~ f (xo
+
d) - f (xo) is obviously convex, so its perspectiver(u, d) := u(f(xo
+
d/u) - f(xo») (for u > 0) (2.2.2) is also convex with respect to the couple (u, d) E lRt x ]Rn. Up to the simple change of variable u ~ t = l/u, we recognize a difference quotient.The next natural question is the closedness of a perspective-function: admitting that
f
itself is closed, troubles can still be expected at u = 0, where we have brutally set j(O, .)=
+00 (possibly not the best idea ... ) A relatively simple calculation of clj
is in fact given by Proposition 1.2.5:162 IV. Convex Functions of Several Variables
Proposition 2.2.2 Let
I
E Conv]Rn and let x' E ri domI.
Then the closure clj
of its perspective is given as follows:{ ul(x/u) if u > 0,
(cl j)(u, x) = lillla-l-oal(x' - x
+
x/a) if u = 0,+00
if u < O.PROOF. Suppose first u < O. For any x, it is clear that (u, x) is outside cl dom
j
and, in view of (1.2.8), cl j(u, x)= +00.
Now let u ? O. Using (2.2.1), the assumption on x' and the results of §I11.2.1, we see that (1, x') E ri dom
j,
so Proposition 1.2.5 allows us to write(cl j)(u. x) = lim j (u, x)
+
a [(1 , x') - (u, x)]) a-l-o= lim [u
+
a(1 - u)] I(x+a(xl-x»)a-l-o u+a(l-u) .
If u = 1, this reads cl j(l, x) = cl I(x) = I(x) (because I is closed); ifu = 0, we
just obtain our claimed relation. 0
Remark 2.2.3 Observe thatthe behaviour of j(u, -) for u
+
0 just depends on the behaviour of f at infinity. If x = 0, we havecl ](0,0) = lirnaf(x') = 0 [f(x') < +oo!].
a./-O
For x =1= 0, suppose for example that domf is bounded; then f(x' - x
+
x/a)= +00
for a small enough and cl ](0. x) =+00.
On the other hand, when domf is unbounded, cl ](0, .) may assume finite values if, at infinity, f does not increase too fast.For another illustration, we apply here Proposition 2.2.2 to the perspective-function r of (2.2.2). Assuming Xo E ri dom f, we can take d' = 0 - which is in the relative interior of the function d H- !(xo
+
d) - !(xo) - to obtain(cl r)(O, d)
=
lim !(xo - d+
.d) - !(xo)r-++oo •
Because (. - 1)/. ~ 1 for. ~
+00,
the last limit can also be written (in lR U {+oo}) (clr)(O. d) = lim f(xo+
td) - f(xo)1-++00
We will return to all this in §3.2 below.
o
2.3 Infimal Convolution
Starting from two functions
II
and/Z,
fonn the set epi11 +
epih c IRn x IR:
C := {(XI
+
X2, rl+
'2) : rj ? !.i(Xj) for j = 1. 2}.Under a suitable minorization property, this C has a lower-bound function le as in (1.3.5):
lc(x)
=
inf {rl+
r2 : rj ~ f.j(Xj) for j = 1,2, Xl+
X2 = x}.In the above minimization problem, the variables are rl, r2, Xl, X2, but the rj's can be eliminated; in fact, le can be defined as follows.
Definition 2.3.1 Let II and f2 be two functions from IRn to IR U {+oo}. Their infimal convolution is the function from IRn to IR U {±oo} defined by
(fl
t
h)(x) ;= inf {fl (Xl)+
h(X2) ; Xl+
X2 = X}= infyERn[fI(Y)
+
hex - y)]. (2.3.1) We will also call "infimal convolution" the operation expressed by (2.3.1). It is called exact at X=
Xl+
X2 when the infimum is attained at (Xl, X2), not necessarily~~. 0
We refer to Remark 1.2.1.4 for an explanation and some comments on the termi-nology "infimal convolution". To exclude the undesired value -00 from the range of an inf-convolution, an additional assumption is obviously needed: in one dimension, the infimal convolution of the functions X and -x is identically -00. Our next result proposes a convenient such assumption.
Proposition 2.3.2 Let the functions II and
12
be in Conv IRn. Suppose that they have a common affine minorant:forsome (s, b) E IRn x RIj(x) ~ (s, x) - b lor j
=
1,2 and all X E IRn . Then their infimal convolution is also in Conv IRn.PROOF. For arbitrary X E IRn and x], X2 such that Xl +X2
=
x, we have by assumption II (Xl)+
h(X2) ~ (s, X) - 2b > -00,and this inequality extends to the infimal value (fl
t
h)(x).On the other hand, it suffices to choose particular values Xj E dom f.j, j
=
I, 2, to obtain the point XI+
X2 E dom(flt h).
Finally, the convexity of IIt 12
results from the convexity of a lower-bound function, as seen in § 1.3(g). 0Remark 2.3.3 To prove that an inf-convolution of convex functions is convex, one