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.

32 <strong>Convex</strong> Functions<br />

by (15), uniformly for all h. Hence<br />

(16) f (y) = f (x) + u · (y − x) + 1<br />

2 (y − x)T H(y − x) + o(�y − x� 2 ) as<br />

y → x, y ∈ C,<br />

uniformly for y of the form x + th <strong>and</strong> thus for almost all y ∈ C. Since f (y) <strong>and</strong><br />

f (x) + u · (y − x) + 1 2 (y − x)T H(y − x) depend continuously on y, (16) holds for<br />

all y ∈ C. This concludes the proof of (14) <strong>and</strong> thus of Alex<strong>and</strong>rov’s theorem. ⊓⊔<br />

In contrast to the fact established in Theorem 2.8 that the existence of all first partial<br />

derivatives of a convex function implies that the function is of class C 1 , the existence<br />

of all second partial derivatives or the second order differentiability of a convex function<br />

does not guarantee that it is of class C 2 .<br />

2.3 A <strong>Convex</strong>ity Criterion<br />

As for functions of one variable, it is of interest to ascertain when a function f :<br />

C → R of several variables is convex.<br />

Below we give a simple, yet useful convexity criterion due to Brunn [174] <strong>and</strong><br />

Hadamard [460].<br />

Hessians<br />

Let f : C → R have partial derivatives of second order, fxi ,xk (x), i, k = 1,...,d,<br />

at x ∈ C. As in Section 2.2, the d × d matrix<br />

⎛<br />

⎞<br />

H = H(x) =<br />

fx1,x1<br />

(x) ... fx1,xd (x)<br />

⎜ fx2,x1 ⎜ (x) ... fx2,xd (x)<br />

⎝ ....................<br />

fxd ,x1 (x) ... fxd ,xd (x)<br />

⎟<br />

⎠<br />

is called the Hessian matrix of f at x. Corresponding to it is the Hessian (quadratic)<br />

form of f at x, defined by<br />

y → 1<br />

2 yT Hy = 1<br />

2<br />

d�<br />

i,k=1<br />

The <strong>Convex</strong>ity Criterion of Brunn <strong>and</strong> Hadamard<br />

fxi ,xk (x) yi yk for y ∈ E d .<br />

The following result for convex functions of d variables will be proved by retracing<br />

it back to the one-dimensional case <strong>and</strong> using the fact that a convex function in one<br />

variable of class C 2 has non-negative second derivative.<br />

Theorem 2.10. Let C be open <strong>and</strong> f : C → R of class C 2 . Then the following<br />

statements are equivalent:<br />

(i) f is convex.<br />

(ii) For any x ∈ C the Hessian form H(x) of f at x is positive semi-definite.

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

Saved successfully!

Ooh no, something went wrong!