31.12.2013 Views

Lecture 2 Piecewise-linear optimization

Lecture 2 Piecewise-linear optimization

Lecture 2 Piecewise-linear optimization

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.

necessity: suppose A does not satisfy the nullspace condition<br />

• for some nonzero z ∈ nullspace(A) and support set I with |I| ≤ k,<br />

‖P I z‖ 1 ≥ 1 2 ‖z‖ 1<br />

• define a k-sparse vector ˆx = −P I z and y = Aˆx<br />

• the vector x = ˆx+z satisfies Ax = y and has l 1 -norm<br />

‖x‖ 1 = ‖−P I z +z‖ 1<br />

= ‖z‖ 1 −‖P I z‖ 1<br />

≤ 2‖P I z‖ 1 −‖P I z‖ 1<br />

= ‖ˆx‖ 1<br />

therefore ˆx is not the unique l 1 -minimizer<br />

<strong>Piecewise</strong>-<strong>linear</strong> <strong>optimization</strong> 2–19

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

Saved successfully!

Ooh no, something went wrong!