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.

26 <strong>Convex</strong> Functions<br />

Remark. More precise measure-theoretic information on the size of the set of points<br />

at which a convex function is differentiable was given by Anderson <strong>and</strong> Klee [28]<br />

(in the guise of a result on convex bodies). Mazur [701] <strong>and</strong> Preiss <strong>and</strong> Zajíček<br />

[815, 816] employ the topological tool of Baire categories, respectively, the metric<br />

tool of porous sets to estimate the size of this differentiability set. From all three<br />

points of view, it is a large set. Compare Theorems 5.1 <strong>and</strong> 5.2 dealing with convex<br />

bodies <strong>and</strong> the discussion on first-order differentiability after these results. For a<br />

discussion of differentiability properties of convex functions on infinite dimensional<br />

spaces, see the book of Benyamini <strong>and</strong> Lindenstrauss [97].<br />

First-Order Differentiability <strong>and</strong> Affine Support<br />

Our first result is as follows:<br />

Theorem 2.7. Let f : C → R be convex <strong>and</strong> x ∈ int C. Then the following are<br />

equivalent:<br />

(i) f is differentiable at x.<br />

(ii) f has unique affine support at x, say a : E d → R, where a(y) = f (x)+u·(y−x)<br />

for y ∈ E d <strong>and</strong> u = grad f (x) = � fx1 (x), . . . , fxd (x)� .<br />

Proof. By Theorem 1.4,<br />

f − xi (x) <strong>and</strong> f + xi<br />

(x), i = 1,...,d, exist.<br />

(i)⇒(ii) By Theorem 2.4 f has an affine support a : E d → R at x, where<br />

a(y) = f (x) + u · (y − x) for y ∈ E d .Leti = 1,...,d. The restriction of f to<br />

the intersection of C with the line through x parallel to the ith coordinate axis is a<br />

convex function of one variable. This function has derivative<br />

fxi (x) = f − xi (x) = f + xi (x)<br />

at x by Theorem 2.5 <strong>and</strong> f (x)+ui (yi − xi) for yi ∈ R is an affine support at x. Thus<br />

ui = fxi (x) by Proposition 1.2. Since this holds for i = 1,...,d, the affine support<br />

a is unique <strong>and</strong> has the desired form.<br />

(ii)⇒(i) If f is not differentiable at x, then there is an index i such that fxi (x)<br />

does not exist, see Theorem 2.5. Then<br />

f − xi (x) < f + xi (x)<br />

by Theorem 1.4. By Proposition 1.2, each affine function of the form f (x) + ui(yi −<br />

xi) for yi ∈ R where<br />

f − xi (x) ≤ ui ≤ f + xi (x)<br />

is an affine support of the restriction of f to the intersection of C <strong>and</strong> the line through<br />

x parallel to the ith coordinate axis. Each of these affine supports can be extended to<br />

an affine support of f at x by the Hahn–Banach type theorem 2.3. Hence f does not<br />

have a unique affine support at x. ⊓⊔

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

Saved successfully!

Ooh no, something went wrong!