Week 6: Semidefinite programming
1 Linear algebra review
Square matrix eigenvalue and eigenvectors
Square matrix
M
M
M, has eigenvector
v
v
v and eigenvalue
λ
\lambda
λ if
M
v
=
λ
v
Mv=\lambda v
Mv=λv
Compute
λ
\lambda
λ
d
e
t
∣
M
−
λ
I
∣
=
0
det|M-\lambda I|=0
det∣M−λI∣=0
Symmetric matrix
M M M is symmetric ⇒ \Rightarrow ⇒ eigen values are real, and eigen vectors are orthogonal. This is consistent with the theory: symmetric matrices have orthogonal eigenvectors, where as non-symmetric matrices do not.
Spectral decomposition of symmetric matrix
M
=
∑
λ
i
v
i
v
i
T
M=\sum\lambda_i v_iv_i^T
M=∑λiviviT
M
=
V
Λ
V
T
M=V \Lambda V^T
M=VΛVT
v
i
v_i
vi is normalized eigenvector, orthogonal to each other.
2 Positive semidefinite matrix (PSD) (must be Hermitian)
Definition of PSD:
- ⇔ \Leftrightarrow ⇔ λ i ≥ 0 \lambda_i\geq 0 λi≥0 and M = M T M=M^T M=MT.
- ⇔ \Leftrightarrow ⇔ x T M x ≥ 0 x^TMx\geq0 xTMx≥0, ∀ x ∈ R n \forall x\in \mathbb{R}^n ∀x∈Rn
- For ∀ A \forall A ∀A, A A T AA^T AAT and A T A A^TA ATA are always PSD
3 Semidefinite programming (SDP)
3.1 Definition
min x c T x s . t . F 0 + ∑ i = 1 m x i F i ≥ 0 \begin{aligned} \min_x \quad & c^Tx\\ s.t. \quad & F_0+\sum_{i=1}^m x_iF_i \geq0 \end{aligned} xmins.t.cTxF0+i=1∑mxiFi≥0 Where the objective is linear function, and F i F_i Fi are symmetric. Eigenvalues are positive.
3.2 With Linear Programming (LP)
min x c T x s . t . A x ≥ 0 \begin{aligned} \min_x \quad& c^Tx\\ s.t. \quad&Ax\geq0 \end{aligned} xmins.t.cTxAx≥0 min x c T x s . t . ∑ i A i x i ≥ 0 \begin{aligned} \min_x \quad & c^Tx\\ s.t. \quad&\sum_i \mathcal{A_i}x_i \geq0 \end{aligned} xmins.t.cTxi∑Aixi≥0 Where A i = d i a g ( A i ) \mathcal{A_i}=diag(A_i) Ai=diag(Ai), A i A_i Ai is the i t h i_{th} ith column of A A A.
3.3 Schur complement
X
=
A
=
(
A
B
B
T
C
)
X= \begin{aligned} A= \begin{pmatrix} A&B\\ B^T& C \end{pmatrix} \end{aligned}
X=A=(ABTBC) is symmetric.
Then,
- X > 0 X>0 X>0 ⇔ \Leftrightarrow ⇔ A > 0 A>0 A>0, and C − B T A − 1 B > 0 C-B^TA^{-1}B>0 C−BTA−1B>0
- X > 0 X>0 X>0 ⇔ \Leftrightarrow ⇔ C > 0 C>0 C>0, and A − B C − 1 B T > 0 A-BC^{-1}B^T>0 A−BC−1BT>0
- If A > 0 A>0 A>0, then X > 0 X>0 X>0 ⇔ \Leftrightarrow ⇔ C − B T A − 1 B > 0 C-B^TA^{-1}B>0 C−BTA−1B>0