Chapter 3 Geometry of convex functions - Meboo Publishing ...
Chapter 3 Geometry of convex functions - Meboo Publishing ...
Chapter 3 Geometry of convex functions - Meboo Publishing ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
3.8. MATRIX-VALUED CONVEX FUNCTION 253<br />
Given symmetric nonnegative data [h ij ] ∈ S N ∩ R N×N<br />
+ , consider function<br />
f(vec X) =<br />
N−1<br />
∑<br />
i=1 j=i+1<br />
N∑<br />
(|x i − x j | − h ij ) 2 ∈ R (1284)<br />
Take the gradient and Hessian <strong>of</strong> f . Then explain why f is not a <strong>convex</strong><br />
function; id est, why doesn’t second-order condition (595) apply to the<br />
constant positive semidefinite Hessian matrix you found. For N = 6 and h ij<br />
data from (1357), apply line theorem 3.8.3.0.1 to plot f along some arbitrary<br />
lines through its domain.<br />
<br />
3.7.4.1 second-order ⇒ first-order condition<br />
For a twice-differentiable real function f i (X) : R p →R having open domain,<br />
a consequence <strong>of</strong> the mean value theorem from calculus allows compression<br />
<strong>of</strong> its complete Taylor series expansion about X ∈ domf i (D.1.7) to three<br />
terms: On some open interval <strong>of</strong> ‖Y ‖ 2 , so that each and every line segment<br />
[X,Y ] belongs to domf i , there exists an α∈[0, 1] such that [384,1.2.3]<br />
[41,1.1.4]<br />
f i (Y ) = f i (X) + ∇f i (X) T (Y −X) + 1 2 (Y −X)T ∇ 2 f i (αX + (1 − α)Y )(Y −X)<br />
(596)<br />
The first-order condition for <strong>convex</strong>ity (586) follows directly from this and<br />
the second-order condition (595).<br />
3.8 Matrix-valued <strong>convex</strong> function<br />
We need different tools for matrix argument: We are primarily interested in<br />
continuous matrix-valued <strong>functions</strong> g(X). We choose symmetric g(X)∈ S M<br />
because matrix-valued <strong>functions</strong> are most <strong>of</strong>ten compared (597) with respect<br />
to the positive semidefinite cone S M + in the ambient space <strong>of</strong> symmetric<br />
matrices. 3.20<br />
3.20 Function symmetry is not a necessary requirement for <strong>convex</strong>ity; indeed, for A∈R m×p<br />
and B ∈R m×k , g(X) = AX + B is a <strong>convex</strong> (affine) function in X on domain R p×k with<br />
respect to the nonnegative orthant R m×k<br />
+ . Symmetric <strong>convex</strong> <strong>functions</strong> share the same<br />
benefits as symmetric matrices. Horn & Johnson [198,7.7] liken symmetric matrices to<br />
real numbers, and (symmetric) positive definite matrices to positive real numbers.