Professional Documents
Culture Documents
Convex Cardinality Optimization
Convex Cardinality Optimization
Junyu Zhang
1 / 26
Outline
Case studies
`1 -regularized methods
2 / 26
Convex Cardinality Optimization
supp(x) := {i : xi 6= 0}
3 / 26
More about cardinality
4 / 26
Convex-cardinality problems
minimize card(x)
s.t. x ∈C
minimize f (x)
s.t. x ∈C
card(x) ≤ k
5 / 26
Sparse portfolio selection
The problem set up:
n assets, each with expected return ri , 1 ≤ i ≤ n.
The covariance matrix of the the returns Σ.
Each assets invest wi .
minimize w T Σw
s.t. r T w = µ, 1T w = 1.
minimize w T Σw
s.t. r T w = µ, 1T w = 1,
card(w ) ≤ k.
6 / 26
Sparse portfolio selection
minimize w T Σw + γ · card(w )
s.t. r T w = µ, 1T w = 1.
In practice, solve
minimize w T Σw + λkw k1
s.t. r T w = µ, 1T w = 1.
7 / 26
High dimensional statistics
lim β̂n = β ∗
n→∞
8 / 26
High dimensional statistics
In practice, solve
n
X
β̂n = arg min (xiT β − yi )2 + λn kβk1 .
β i=1
9 / 26
Sparse coding / Dictionary learning
Unsupervised methods for learning sets of over-complete bases to
represent data efficiently.
Belief: natural images are generated by linear combination of bases.
11 / 26
Exact solution to convex-cardinality problems
minimize
n
f (x) + λ · card(x)
x∈R
s.t. x ∈C
12 / 26
Binary linear programming
Binary LP problems:
minimize cT x
x
s.t. Ax ≤ b, x ∈ {0, 1}n .
reduction:
minimize cT x
x
s.t. Ax ≤ b,
card(x) + card(1 − x) = n.
13 / 26
`1 -norm heuristic
Tune λ to get desired sparsity. The larger λ is, the sparser the
solution is.
P
Or sophisticated versions: i wi |xi |.
14 / 26
The bias in the solution
15 / 26
Solving the `1 regularized problem
Suppose f is convex and smooth, consider
...
16 / 26
Iterative shrinkage-thresholding algorithm (ISTA)
minimize f (x) + γ · 1T z
x,z
s.t. |xi | ≤ Rzi , for i = 1, ..., n
x ∈ C, z ∈ {0, 1}n
18 / 26
Interpretation as convex relaxation
minimize f (x) + γ · 1T z
x
s.t. |xi | ≤ Rzi , for i = 1, ..., n
x ∈ C, z ∈ [0, 1]n
19 / 26
Sparse design
Find design vector x with smallest card(x) such that some set of
specifications are satisfied.
Problem formulation
minimize card(x)
s.t. x ∈C
Examples
antenna array beamforming (zero coefficients correspond to unneeded
antenna elements)
truss design (zero coefficients correspond to bars that are not needed)
...
20 / 26
Truss design
Figure: Such truss contains a huge number of steal bars, potentially many of
them are not necessary.
21 / 26
Sparse modeling / regressor selection
minimize
n
kAx − bk2
x∈R
s.t. card(x) ≤ k
Reformulations:
minimize
n
kAx − bk2 + λ · card(x)
x∈R
minimize
n
kAx − bk22 + λ · card(x)
x∈R
minimize
n
card(x) s.t. kAx − bk2 ≤
x∈R
minimize
n
kAx − y k22
x∈R
s.t. card(x) ≤ k
23 / 26
Estimation with outliers
minimize ky − X β − w k22
β∈Rd ,w ∈Rn
s.t. card(w ) ≤ k
24 / 26
Robust principle component analysis∗
minimize kM − Lk
L∈Rn×n
s.t. rank(L) ≤ k
minimize kM − L − W k
L,W ∈Rn×n
s.t. rank(L) ≤ k, card(W ) ≤ k 0 .
25 / 26
Robust principle component analysis∗
minimize kM − L − W k
L,W ∈Rn×n
s.t. rank(L) ≤ k, card(W ) ≤ k 0 .
26 / 26