14.02.2013 Views

Gruber P. Convex and Discrete Geometry

Gruber P. Convex and Discrete Geometry

Gruber P. Convex and Discrete Geometry

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Fifth, we show that<br />

(13) G is differentiable almost everywhere on C.<br />

2 <strong>Convex</strong> Functions of Several Variables 31<br />

By (6), K = (I + G) −1 : D → E d is Lipschitz <strong>and</strong> thus can be extended to<br />

a Lipschitz function of E d into E d by (1). Denote this extension also by K .Let<br />

M be the set of points in E d where this K is not differentiable <strong>and</strong> by N the set<br />

where it is differentiable but with singular derivative. Propositions (2)–(4) then show<br />

that K (M ∪ N) has measure 0. The set L, from (7) <strong>and</strong> (8), is also of measure 0.<br />

For the proof of (13), it is thus sufficient to show that G is differentiable for any<br />

x ∈ C\ � K (M ∪ N) ∪ L � . For each such x Proposition (8) says that G is continuous<br />

at x <strong>and</strong> the above definitions of M <strong>and</strong> N imply that K is differentiable at x + u<br />

with non-singular derivative (note that x + u = (I + G)(x) = K −1 (x) �∈ M ∪ N).<br />

Hence G is differentiable at x by (9). The proof of (13) is complete.<br />

In the sixth, <strong>and</strong> final, step it will be shown that<br />

(14) f (y) = f (x) + u · (y − x) + 1 2 (y − x)T H(y − x) + o(�y − x� 2 ) for almost<br />

all x ∈ C <strong>and</strong> each y ∈ C. Hereu = G(x) <strong>and</strong> H is the derivative of G<br />

at x.<br />

The idea of the proof of (14) is to restrict f to a line segment on which it is differentiable<br />

almost everywhere <strong>and</strong> then represent it as the integral of its derivative.<br />

f <strong>and</strong> G are differentiable at almost every point x ∈ C, see (7) <strong>and</strong> (13). Let x<br />

be such a point. Then, in particular,<br />

(15) G(y) = u + H(y − x) + o(�y − x�) for y → x, y ∈ C,<br />

where u = G(x) <strong>and</strong> H is the derivative of G at x.<br />

In addition, (7) shows that, for almost every unit vector h <strong>and</strong> almost every t ≥ 0for<br />

which x + th ∈ C, the function f is differentiable at x + th.Ifh <strong>and</strong> t are such a<br />

unit vector <strong>and</strong> such a number, respectively, then<br />

f (y) = f (x + th) + v · � y − (x + th) � + o � �y − (x + th)� � as y → x + th, y ∈ C,<br />

where v = G(y + th). Thus, in particular, for y = x + sh,<br />

f (x + sh) = f (x + th) + v · h (s − t) + o(|s − t|) as s → t, x + sh ∈ C.<br />

Hence, for t ≥ 0 such that x +th ∈ C, the convex function f (x +th) of one variable<br />

is differentiable at almost every t. Its derivative is v ·h, where v = G(x +th). Since a<br />

convex function of one variable is absolutely continuous in the interior of its interval<br />

of definition by Theorem 1.1, a theorem of Lebesgue shows that integration of its<br />

derivative yields the original function, see the proof of Theorem 1.7. Hence<br />

f (x + th) − f (x) =<br />

=<br />

�t<br />

0<br />

�t<br />

0<br />

G(x + sh) · hds<br />

� u · h + h T Hhs + o(s) � ds = u · ht+ 1<br />

2 hT Hht 2 + o(t 2 ) as t → 0

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!