Conic Programming
Standard form of conic programming is as follows
min
x
c
′
x
s
.
t
.
{
A
x
=
b
x
∈
K
(1)
\min_x c'x\\ s.t. \begin{cases} Ax=b\\ x\in K \end{cases}\tag{1}
xminc′xs.t.{Ax=bx∈K(1)
where
K
K
K is a closed pointed convex cone with a non-empty interior. The conic programming can also be expressed as
min
x
c
′
x
s
.
t
.
{
A
x
=
b
A
i
x
−
b
i
∈
K
i
i
=
1
,
2
,
…
,
m
(2)
\min_x c'x\\ s.t. \begin{cases} Ax=b\\ A_ix-b_i\in K_i\quad i=1,2,\dots, m \end{cases}\tag{2}
xminc′xs.t.{Ax=bAix−bi∈Kii=1,2,…,m(2)
SOCP
Second-order cones (Lorentz cones) of dimension
k
k
k are defined as
C
k
=
{
[
u
t
]
∣
u
∈
R
k
−
1
,
t
∈
R
,
∥
u
∥
2
≤
t
}
C_k=\bigg\{ \bigg[ \begin{matrix} u\\ t \end{matrix} \bigg] \bigg| u\in\mathbb{R}^{k-1}, t\in\mathbb{R}, \lVert u\rVert_2\leq t \bigg\}
Ck={[ut]∣∣∣∣u∈Rk−1,t∈R,∥u∥2≤t}
where
∥
u
∥
2
=
u
′
u
\lVert u\rVert_2=\sqrt{u'u}
∥u∥2=u′u.
And the conic programming can be written as
min
x
c
′
x
s
.
t
.
{
A
x
=
b
∥
A
i
x
+
b
i
∥
≤
c
i
′
x
+
d
i
,
i
=
1
,
2
,
…
,
m
(3)
\min_x c'x\\ s.t. \begin{cases} Ax=b\\ \lVert A_ix+b_i\rVert\leq c_i'x+d_i, i=1,2,\dots, m \end{cases}\tag{3}
xminc′xs.t.{Ax=b∥Aix+bi∥≤ci′x+di,i=1,2,…,m(3)
where
A
i
∈
R
(
m
i
−
1
)
×
n
,
b
i
∈
R
(
m
i
−
1
)
,
c
i
∈
R
n
,
d
i
∈
R
A_i\in\mathbb{R}^{(m_i-1)\times n}, b_i\in\mathbb{R}^{(m_i-1)}, c_i\in\mathbb{R}^n, d_i\in\mathbb{R}
Ai∈R(mi−1)×n,bi∈R(mi−1),ci∈Rn,di∈R. The inequality constraint is equivalent to
[
A
i
c
i
′
]
x
+
[
b
i
d
i
]
∈
C
m
i
\left[ \begin{matrix} A_i\\ c_i' \end{matrix} \right]x+ \left[ \begin{matrix} b_i\\ d_i \end{matrix} \right]\in C_{mi}
[Aici′]x+[bidi]∈Cmi
The loss risk constraint is defined to limit the probability of portfolio return
r
p
r_p
rp blow a threshold
α
\alpha
α is denoted as
P
r
(
r
p
≤
α
)
≤
β
Pr(r_p\leq \alpha)\leq \beta
Pr(rp≤α)≤β
where
β
\beta
β sets the maximum probability. By standardizing the original value
P
r
(
r
p
≤
α
)
=
Φ
(
α
−
μ
′
w
w
′
Σ
w
)
≤
β
α
−
μ
′
w
w
′
Σ
w
≤
Φ
−
1
(
β
)
Pr(r_p\leq \alpha)=\Phi(\frac{\alpha-\mu'w}{\sqrt{w'\Sigma w}})\leq \beta\\ \frac{\alpha-\mu'w}{\sqrt{w'\Sigma w}}\leq \Phi^{-1}(\beta)
Pr(rp≤α)=Φ(w′Σwα−μ′w)≤βw′Σwα−μ′w≤Φ−1(β)
and can be reformulated into a second-order cone constraint as
Φ
−
1
(
β
)
∥
Σ
1
/
2
w
∥
2
≥
μ
′
w
+
α
\Phi^{-1}(\beta)\lVert \Sigma^{1/2}w\rVert_2\geq \mu'w+\alpha
Φ−1(β)∥Σ1/2w∥2≥μ′w+α
Set
β
≤
1
/
2
\beta\leq 1/2
β≤1/2, we have the following portfolio problem
max
x
μ
′
w
s
.
t
.
{
Φ
−
1
(
β
)
∥
Σ
1
/
2
w
∥
2
≥
−
μ
′
w
+
α
w
′
1
=
1
x
≥
0
(4)
\max_x \mu'w\\ s.t. \begin{cases} \Phi^{-1}(\beta)\lVert\Sigma^{1/2}w\rVert_2\geq -\mu'w+\alpha\\ w'\mathbf{1}=1\\ x\geq 0 \end{cases}\tag{4}
xmaxμ′ws.t.⎩⎪⎨⎪⎧Φ−1(β)∥Σ1/2w∥2≥−μ′w+αw′1=1x≥0(4)
Semidefinite Programming
SDP is the most inclusive formulation among the three types of conic optimization programs and is written in the following form:
min
x
c
′
x
s
.
t
.
{
A
x
=
b
F
0
+
x
1
F
1
+
⋯
+
x
n
F
n
⪰
0
(5)
\min_x c'x\\ s.t. \begin{cases} Ax=b\\ F_0+x_1F_1+\dots+x_nF_n\succeq 0 \end{cases}\tag{5}
xminc′xs.t.{Ax=bF0+x1F1+⋯+xnFn⪰0(5)
where symmertic matrices
F
0
,
…
,
F
n
∈
S
m
F_0,\dots, F_n\in S^m
F0,…,Fn∈Sm. An LMI defined as
F
(
x
)
=
F
0
+
x
1
F
1
+
⋯
+
x
n
F
n
F(x)=F_0+x_1F_1+\dots+x_nF_n
F(x)=F0+x1F1+⋯+xnFn
SOCP → \to → SDP
Consider a matrix
X
∈
S
n
X\in \mathcal{S}^n
X∈Sn that is partitioned into submatrices
A
,
B
A, B
A,B and
C
C
C.
X
=
[
A
B
B
′
C
]
X= \left[ \begin{matrix} A & B\\ B' & C \end{matrix} \right]
X=[AB′BC]
On basis of Schur complements, it holds for
A
≻
0
A\succ 0
A≻0 that
X
⪰
0
X\succeq 0
X⪰0 iff
C
−
B
′
A
−
1
B
≥
0
C-B'A^{-1}B\geq 0
C−B′A−1B≥0
SOCP
(
3
)
(3)
(3) can be converted to the following SDP form:
min
x
c
′
x
s
.
t
.
{
A
x
=
b
[
(
c
i
T
x
+
d
i
)
I
A
i
x
+
b
i
(
A
i
x
+
b
i
)
′
c
i
T
x
+
d
i
]
⪰
0
\min_x c'x\\ s.t. \begin{cases} Ax=b\\ \left[ \begin{matrix} (c_i^Tx+d_i)I & A_ix+b_i\\ (A_ix+b_i)' & c_i^Tx+d_i \end{matrix} \right]\succeq 0 \end{cases}
xminc′xs.t.⎩⎨⎧Ax=b[(ciTx+di)I(Aix+bi)′Aix+biciTx+di]⪰0
Reference
ROBUST EQUITY PORTFOLIO MANAGEMENT+WEBSITE