Convex optimization

Upcoming Lectures:!
!
Lecture 5: Convex Optimization (today / PG)"
!
Lecture 6: Duality at KKT Conditions (March 25 / PG)!
!
Lecture 7: Structured Singular Value (April 1 / RS)"
!
Lecture 8: H∞ design (April 8 / PG)"
!
!
!

Robust Control & Convex Optimization 5: !1

 References

References

These lecture notes are based to a large extent on those for the Stanford EE 364a class developed by
Stephen Boyd. His original slides can be downloaded from http://www.stanford.edu/class/ee364a/.

LMI formulations of many control problems are given in:"

S. Boyd, L. El Ghaoui, E. Feron & V. Balakrishnan, “Linear Matrix Inequalities in System and Control
Theory,” SIAM, 1994.

The most extensive text on convex optimization (and one of the best written) is:"
S. Boyd & L. Vandenberghe, “Convex Optimization,” Cambridge Univ. Press, 2004.

Robust Control & Convex Optimization 5: !2

 Convex optimization

Convex optimization

Problems to be solved:

1. µ (M ) = µ (DM D 1
) inf max (DM D
1
)
D D

2. inf Fl (G(s), K(s)) L

K(s) stabilizing

3. inf Fl (G(s), K(s)) L2

K(s) stabilizing

4. Iterative robust performance design (D, K) iteration

5. Regional pole placement, MPC, . . .

Tools to be used:

1. Linear matrix inequalities (LMIs)

2. S-procedure, linearizing transforms, Schur complements, . . .

3. cvx, yalmip, Robust Control Toolbox

Robust Control & Convex Optimization 5: !3

 Convex optimization

Convex sets:

A set is convex if the line segment between every two points is also in the set.

For all x1 , x2 C, { 1 x1 + 2 x2 | 1 + 2 = 1, 1 0, 2 0} C.

Equivalently, y = x1 + (1 )x2 , [0, 1] is in C.

x1 + (1 )x2 , [0, 1]

x2

x1

Robust Control & Convex Optimization 5: !4

 Convex optimization

Hyperplanes and halfspaces

A hyperplane is a set of the form: x aT x = b (a = 0).

a Hyperplanes are affine and convex.

a is the normal vector
x0
aT (x x0 ) = 0
(x x0 )

A halfspace is a set of the form: x aT x b (a = 0).

x aT x b
x0

x aT x b
Halfspaces are convex.

Robust Control & Convex Optimization 5: !5

 Convex optimization

Polyhedra

The solution set of finitely many linear inequalities and equalities.

{ x | Ax  b, Cx = d }

(A 2 Rm⇥n , C 2 Rp⇥n ,  is a componentwise inequality)

AT1•
AT2•

AT5•

AT3•

AT4•

A polyhedron can be viewed as the intersection of halfspaces and hyperplanes.

Robust Control & Convex Optimization 5: !6

 Convex optimization

Convex cones

Cone: A set K is a cone if, for all x 2 K, ✓x 2 K (✓ 0).

A conic combination, x, of two points, x1 and x2 has the form x1

x = 1 x1 + 2 x2 , 1 0, 2 0.

x2
0

A convex cone is a set that contains every conic combination of points in the set.

Notation: Use the notation x 2 K , x ⌫K 0.

Robust Control & Convex Optimization 5: !7

 Convex optimization

Norm balls and norm cones

A norm ball with center, xc , and radius, r is given by: {x| x xc r }.

x1
x =1
Note that the “shape” depends on the norm.
8

x =1
2

x 1=1
x2

1.0

0. 8

Norm cone: { (x, t) | x t} 0. 6

t
0. 4

0. 2

0
Norm balls and norm cones are convex. -1.0
-0.5 1.0
0 0. 5
0
x1 0.5 -0.5
x2
1.0 -1.0
The Euclidean norm cone is called a second-order cone.
Robust Control & Convex Optimization 5: !8
 Convex optimization

Positive semidefinite cone

Sn : Symmetric n n matrices.

Sn+ = { X Sn | X 0} (positive semidefinite n n matrices)

X Sn+ zT X z 0, for all z.

Sn+ is a convex cone

Sn++ = { X Sn | X 0} (positive definite n n matrices)

Sn+ is also convex with respect to its z 1

appropriately constrained components 0.8

0.6

0.4

Example: 2 2 matrices:
0.2

x y 0
0. 1
y z 0.5 1
0.8

y
0 0.6 x
−0.5 0.4
0.2
−1 0

Robust Control & Convex Optimization 5: !9

 Convex optimization

Affine functions

Consider an affine function, f : Rn Rm , ( f (x) = Ax + b, with A Rm n

,b Rm )

The image of a convex set under f is convex.

X Rn (convex) = f (X) = { f (x) | x X } is convex.

Similarly, the inverse image of a convex set is also convex.

Y Rm (convex) = f 1
(Y ) = { x Rn | f (x) Y } is convex.

Examples:

Scaling, rotation, addition, translation, ...

Linear matrix inequalities: { x | x1 A1 + x2 A2 + · · · + xn An B }, Ai , B Sp .

Robust Control & Convex Optimization 5: !10

 Convex optimization

Linear fractional transformations

f: Rn Rm ,

Ax + b
f (x) = T , dom f = x cT x + d > 0 .
c x + d

Images (and inverse images) of convex sets are convex under the linear
fractional mapping.

Example:
2z 1
Bilinear (Tustin) transform: s = f (z) = , maps the unit disk to the left half-plane.
T z+1

Imaginary Imaginary

z-plane s-plane

-1 1
Real Real

Robust Control & Convex Optimization 5: !11

 Convex optimization

Convex functions

A function, f (x) is convex if:

dom f is a convex set; and

f ( x+(1 )y) f (x)+(1 )f (y), for all x, y dom f and all [0, 1].

f (x) + (1 )f (y) f (y)

f ( x + (1 )y)
f (x)

A function, f (x), is concave if f (x) is convex.

Strict convexity (and strict concavity) are obtained by replacing with <

Robust Control & Convex Optimization 5: !12

 Convex optimization

Simple examples:

Convex functions:

Affine functions, f (x) = ax + b for any a, b R.

Exponential functions, f (x) = eax for any a R.

Powers: f (x) = x on R++ for any 1 or 0.

Powers of absolute value: f (x) = |x|p on R for p 1.

Negative entropy: f (x) = x log x on R++ .

1 n
Log-determinant: f (X) = log det(X ), for X 2 S++

Concave functions:

Affine functions, f (x) = ax + b for any a, b R.

Powers: f (x) = x on R++ for any [0, 1].

Logarithm: f (x) = log x on R++ .

n
Log-determinant: f (X) = log det(X), for X 2 S++
Robust Control & Convex Optimization 5: !13
 Convex optimization

Examples on Rn :

Convex functions:

Affine functions, f (x) = aT x + b.

n 1/p

Norms: f (x) = x p = |xi |p for p 1; x = max |xi |.

i
i=1

Examples on Rm n
:

Convex functions:

m n
Affine functions, f (X) = tr(AT X) + b = Aij Xij + b.
i=1 j=1

Maximum singular value (spectral norm):

f (X) = X 2 = max (X) = ( max (X

T
X))1/2 .

Robust Control & Convex Optimization 5: !14

 Convex optimization

Quasiconvex functions

A function, f : Rn R is quasiconvex of dom f is convex and that sublevel sets,

S = {x dom f | f (x) } are convex for all .

f (x)

f (x) is quasiconcave if f is quasiconvex.

f (x) is quasilinear if it is both quasiconcave and quasiconvex.

Note that quasiconvex functions have no non-global local minima.

Robust Control & Convex Optimization 5: !15
 Convex optimization

Optimization problems: a standard form

minimize f0 (x)
subject to fi (x) 0, i = 1, . . . , m
hi (x) = 0, i = 1, . . . , p

x Rn is the optimization variable (vector valued)

f0 : Rn R is the objective or cost function

fi : Rn R, i = 1, . . . , n are the inequality constraint functions

hi : R n R, i = 1, . . . , p are the equality constraint functions

Robust Control & Convex Optimization 5: !16

 Convex optimization

Convex optimization problems

minimize f0 (x)
subject to fi (x) 0, i = 1, . . . , m
aTi x = bi , i = 1, . . . , p

The functions, f0 , f1 , . . . , fm , are convex.

The equality constraints are affine.

A problem is quasiconvex if f0 is quasiconvex and f1 , . . . , fm , are convex.

minimize f0 (x)
subject to fi (x) 0, i = 1, . . . , m
Ax = b

The feasible set of a convex (or quasiconvex) optimization problem is convex.

Robust Control & Convex Optimization 5: !17

 Convex optimization

Semidefinite program (SDP)

minimize cT x
subject to x1 F1 + x2 F2 + . . . + xn Fn + G 0
Ax = b

where Fi , G Sk

The matrix constraint is called a linear matrix inequality (LMI)

Multiple constraints are trivially combined into a single (larger) constraint,

x1 F1 +x2 F2 +. . .+xn Fn +G 0 and x1 H1 +x2 H2 +. . .+xn Hn +M 0

if and only if

F1 0 F2 0 Fn 0 G 0
x1 + x2 + . . . + xn + 0
0 H1 0 H2 0 Hn 0 M

Robust Control & Convex Optimization 5: !18

 cvx

“Disciplined convex programming” cvx

Download cvx from www.stanford.edu/ boyd/cvx/

dx(t)
Example: proving the stability of a system: = A x(t)
dt

stable there exists P = P T 0, AT P + P A 0

there exists P = P T I, AT P + P A I

We can consider P as a matrix variable

cvx_begin sdp
variable P(n,n) symmetric
A’*P + P*A <= -eye(n)
P >= eye(n)
cvx_end

cvx status is a string returning the status of the optimization

Robust Control & Convex Optimization 5: !19

 cvx

Another example:

We want to know if the stability of two systems,

dx(t) dx(t)
= A1 x(t) and = A2 x(t)
dt dt

can be proven with a single Lyapunov function, V (s) = x(t)T P x(t)

dx(t)
= A(t) x(t) stable for A(t) = 1 (t) A1 + 2 (t) A2 , i (t) 0
dt

We want to find P = P T 0, such that AT1 P + P A1 0, and AT2 P + P A2 0

Or equivalently P = P T I, such that AT1 P + P A1 I, and AT2 P + P A2 I

cvx_begin sdp
variable P(n,n) symmetric
A1’*P + P*A1 <= -eye(n)
A2’*P + P*A2 <= -eye(n)
P >= eye(n)
cvx_end

Robust Control & Convex Optimization 5: !20

 Schur complements

Schur complements

Q(x) = Q(x)T , S(x) and R(x) = R(x)T are affine functions of x

Q(x) S(x)
0 R(x) 0 and Q(x) S(x) R(x) 1
S(x)T 0
T
S(x) R(x)

Example 1:

P (x) 0 and trace S(x)T P (x) 1

S(x) < 1

X S(x)T
trace(X) < 1, 0, (X = X T is a slack variable)
S(x) P (x)

Example 2:

tI AT
A 2 t A A T 2
t I, t 0, 0
A tI

Robust Control & Convex Optimization 5: !21

 Convex optimization

Example: matrix norm minimization (Maximum singular value)

1/2
minimize A(x) 2 = (A(x) A(x))
T

where A(x) is an LMI: A(x) = A0 + x1 A1 + x2 A2 + . . . + xn An

The equivalent SDP is:

minimize t

tI A(x)
subject to T
0
A(x) tI

The decision variables are now t and x.

The constraint equivalence follows from a Schur complement argument

A 2 t AT A t2 I, t 0,

tI A
0
AT tI

Robust Control & Convex Optimization 5: !22

 S-procedure

S-procedure

Consider a collection of m + 1 quadratic functions,

Fi (x) = xT Ai x + 2bTi x + ci , Ai Sn , i = 0, 1, . . . , m.

We want to test the condition,

F0 (x) < 0, for all x that satisfy Fi (x) 0, i = 1, . . . , m.

The S-procedure gives a sufficient convex condition,

m
If there exists i 0 such that for all x, F0 (x) i Fi (x) 0,
i=1

then F (x) < 0 for all x satisfying Fi (x) 0, i = 1, . . . , m.

Furthermore, if m = 1, and there exists an x such that F1 (x) < 0,

then the S-procedure condition is necessary and sufficient.

Robust Control & Convex Optimization 5: !23

 S-procedure

S-procedure

A b
We can show that x Ax + 2b x + c
T T
0 for all x 0
bT c
 m
X 
A0 b0 Ai bi
This gives an LMI test, ⌧i 0
bT0 c0 bTi ci
i=1

Example: find the smallest ellipse containing the union of a set of ellipses.

Suppose we have p ellipses, Ei , given by (x xi )T Ai (x xi ) 1, Ai 0, i = 1, . . . , p.

and we would like to characterize whether or not each of them is contained within an ellipse, E0 ,

x E0 (x x0 )A0 (x x0 ) 1, A0 0.

Note that determining whether or not x is in an ellipse is a quadratic constraint,

x Ei Fi (x) = xT Ai x + 2xTi Ai x + (xTi Ai xi 1) 0.

Robust Control & Convex Optimization 5: !24

 S-procedure

S-procedure

Example: find the smallest ellipse containing the union of a set of ellipses.

For each ellipse, Ei E0 for all x such that Fi (x) 0, F0 (x) 0.

By the S-procedure this is equivalent to ,

Ei E0 there exists i 0, such that for all x, F0 (x) i Fi (x) 0

A0 b0 Ai bi
there exists i 0, such that i 0.
bT0 c0 bTi ci

Finding the smallest bounding ellipse:

minimize log det A0 1

subject to A0 0
i 0, i = 1, . . . , p
A0 b0 Ai bi
i 0, i = 1, . . . , p
bT0 c0 bTi ci

The volume of an ellipse is proportional to det A0 1

Robust Control & Convex Optimization 5: !25