08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

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.

Important special cases are<br />

|x| 0 the number <strong>of</strong> non zero entries<br />

|x| 1<br />

= |x 1 | + · · · + |x n |<br />

|x| 2<br />

= √ |x 1 | 2 + · · · + |x n | 2<br />

|x| ∞<br />

= max |x i |.<br />

Lemma 12.15 For any 1 ≤ p < q, |x| q ≤ |x| p .<br />

Pro<strong>of</strong>:<br />

|x| q q = ∑ i<br />

|x i | q .<br />

Let a i = |x i | q and ρ = p/q. Using Jensen’s inequality (see Section 12.3) that for any<br />

nonnegative reals a 1 , a 2 , . . . , a n and any ρ ∈ (0, 1), we have ( ∑ n<br />

i=1 a i) ρ ≤ ∑ n<br />

i=1 aρ i , the<br />

lemma is proved.<br />

There are two important matrix norms, the matrix p-norm<br />

||A|| p<br />

= max<br />

|x|=1 ‖Ax‖ p<br />

and the Frobenius norm<br />

√ ∑<br />

||A|| F<br />

=<br />

ij<br />

a 2 ij .<br />

Let a i be the i th column <strong>of</strong> A. Then ‖A‖ 2 F = ∑ i<br />

a i T a i = tr ( A T A ) . A similar argument<br />

on the rows yields ‖A‖ 2 F = tr ( AA T ) . Thus, ‖A‖ 2 F = tr ( A T A ) = tr ( AA T ) .<br />

If A is symmetric and rank k<br />

||A|| 2 2 ≤ ||A||2 F ≤ k ||A||2 2 .<br />

12.7.6 Important Norms and Their Properties<br />

Lemma 12.16 ||AB|| 2<br />

≤ ||A|| 2<br />

||B|| 2<br />

Pro<strong>of</strong>: ||AB|| 2<br />

= max |ABx|. Let y be the value <strong>of</strong> x that achieves the maximum and<br />

|x|=1<br />

let z = By. Then<br />

||AB|| 2<br />

= |ABy| = |Az| = ∣ ∣<br />

But<br />

∣<br />

∣A z<br />

|z|<br />

∣ ≤ max |Ax| = ||A|| 2<br />

|x|=1<br />

∣ A z<br />

|z| ∣ |z|<br />

and |z| ≤ max<br />

|x|=1 |Bx| = ||B|| 2 . Thus ||AB|| 2 ≤ ||A|| 2 ||B|| 2 .<br />

413

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

Saved successfully!

Ooh no, something went wrong!