Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Quadratic Programming

Insoon Yang

Department of Electrical and Computer Engineering


Seoul National University

CORE
Control + Optimization Research Lab

1 / 15
Quadratic programming (QP)

QP: minimizing a convex quadratic function over a polyhedron


1 ⊤
min z Hz + q ⊤ z + r
2 (1)
s.t. Gz ≤ w,

where H = H ⊤ ⪰ 0.

2 / 15
Geometric interpretation of QP

Figure: Geometric illustration of QP. The feasible set P , which is a polyhedron, is


shown shaded. The contour lines of the objective function, which is convex
quadratic, are shown as dashed curves. The point x⋆ is optimal.

3 / 15
Solution properties

Two possible cases (when the feasible set is nonempty):


1 The minimizer lies strictly inside the feasible polyhedron.
Q) How to find the minimizer?
2 The minimizer lies on the boundary of the feasible polyhedron.

4 / 15
Dual of QP
Consider a QP:
1 ⊤
min z Hz + q ⊤ z
2
s.t. Gz ≤ w.

What’s its dual?


When does strong duality hold?

5 / 15
KKT conditions for QP
1 ⊤
min z Hz + q ⊤ z
2
s.t. Gz ≤ w.

KKT conditions:
Primal feasibility

Dual feasibility

Complementary slackness

Stationarity

In the case of QP, KKT conditions provide necessary and sufficient


conditions for optimality.
6 / 15
Active constraints and degeneracies

Recall...
set of active constraints at z: A(z) = {i | Gi z = wi }
set of inactive constraints at z: N A(z) = {i | Gi z < wi }

Definition (Linear independence constraint qualification (LICQ))


We say that LICQ holds at z ⋆ if the matrix GA(z ⋆ ) has full row rank.

Definition
The QP is said to be
primal degenerate if there exists an optimal point z ⋆ such that the
LICQ does not hold at z ⋆
dual degenerate if its dual problem is primal degenerate.

7 / 15
Suppose the QP is not primal degenerate.

Then, λ⋆N A = 0 and GA H −1 G⊤


A is invertible.

Thus, the dual QP has a unique solution

LICQ always holds for dual QPs and thus dual degeneracy can never
occur for QPs with H ≻ 0.

8 / 15
Example: Constrained least squares problem

Constrained least spares problem or constrained linear regression problem

min ∥Az − b∥22


s.t. li ≤ zi ≤ bi , i = 1, . . . , n

is a QP.

9 / 15
Example: LP with Random Cost
min c⊤ z
s.t. Gz ≤ h
Az = b,
where the cost vector c ∈ Rn is random, with mean value c̄ and covariance
E[(c − c̄)(c − c̄)⊤ ] = Σ. Note that
E[c⊤ z] = c̄⊤ z, var[c⊤ z] = E[(c⊤ z − Ec⊤ z)2 ] = z ⊤ Σz.
A way to take variance (or risk) into account is to minimize
E[c⊤ z] + γvar[c⊤ z], γ ≥ 0.
The resulting risk-sensitive optimization problem is a QP:
min c̄⊤ z + z ⊤ Σz
s.t. Gz ≤ h
Az = b
10 / 15
Optimization Solvers

11 / 15
Idea of interior point methods

min c⊤ x
s.t. a⊤
i x ≤ bi , i = 1, . . . , 6

Figure: Central path for an LP with n = 2 and m = 6. The dashed curves show
three contour lines of the logarithmic barrier function ϕ. The central path
converges to the optimal point x⋆ as t → ∞.

12 / 15
Solvers for LP and QP

CPLEX
https://community.ibm.com/community/user/datascience/
blogs/xavier-nodet1/2020/07/09/cplex-free-for-students

Gurobi
https://www.gurobi.com/academia/
academic-program-and-licenses/

MOSEK
https://www.mosek.com/products/academic-licenses/

13 / 15
Solvers for convex optimization

CVX
http://cvxr.com/cvx/

14 / 15
Solvers for more general nonlinear optimization

Artelys Knitro
https://www.artelys.com/solvers/knitro/

SNOPT, IPOPT, etc...

15 / 15

You might also like