<< Chapter < Page
  Signal theory   Page 1 / 1
Chapter >> Page >
Introduces the concepts of an epigraph and a conjugate function necessary to set up Fenchel dual problems.

Convexity and conjugate functions

We begin by reviewing two notions of convexity: for sets and for functions.

Definition 1 A subset C X is called convex if z = α x + ( 1 - α ) y C for every x , y C and α [ 0 , 1 ] ; z is called a convex combination of x and y .

Definition 2 Let C be a convex set. A functional f : C R is convex on C if f ( α x + ( 1 - α ) y ) α f ( x ) + ( 1 - α ) f ( y ) for all x , y C and α [ 0 , 1 ] . If strict inequality holds the functional is set to be strictly convex . A functional g is called concave (strictly concave) if - g is convex (strictly convex).

We will denote the region above the function f defined over a convex set C as [ f , C ] , sometimes called an epigraph , as illustrated in [link] .

Example of an epigraph.

Definition 3 Let f be a convex functional on a convex set C . The conjugate set C * is defined as

C * = { x * X : sup x C [ x , x * - f ( x ) ] < } ,

and the conjugate functional f * : C * R is defined as

f * ( x * ) = sup x C [ x , x * - f ( x ) ] .

There is a geometric intuition behind the definition of the conjugate functionals. Consider the illustration below where the horizontal axis represents the space X and the vertical axis represents the scalar field. A hyperplane in this space contains all points ( x , r ) X × R for which r = x , x * - k for some value of k R ; the vector x * determines the orientation of the hyperplane and the value k determines the shift from the origin (i.e., the intersect in the axis R ). The value of the functional f * ( x * ) corresponds to the supremum value of k for which the hyperplane intersects [ f , C ] , and is finite only for x * C * ; this is illustrated in [link] .

The conjugate function of a convex epigraph.

Note that C * is convex and f * is convex. This definition is easily extended to concave functionals.

Definition 4 Let g be a concave functional on a convex set D . The conjugate set D * is defined as

D * = { x * X : inf x D [ x , x * - g ( x ) ] > - } ,

and the conjugate functional g * : D * R is defined as

g * ( x * ) = inf x D [ x , x * - f ( x ) ] .

Note that D * is convex and g * is concave.

Fenchel duality

The following theorem will allow us to convert an optimization problem with a convex objective function into a dual problem with a concave objective function.

Theorem 1 (Fenchel) Assume that f and g are convex and concave functions, respectively, on convex sets C and D in a normed space X . Assume that C D contains points in the relative interior of C and D and that either [ f , C ] or [ g , D ] has a nonempty interior. Suppose further that

μ = inf x C D { f ( x ) - g ( x ) }

is finite. Then,

μ = inf x C D { f ( x ) - g ( x ) } = max x * C * D * { g * ( x * ) - f * ( x * ) } ,

where the maximum is achieved by some x 0 * C * D * .

In this theorem, g ( x ) is usually set to zero. From a geometrical point of view, the theorem states that there are two ways to interpret the minimum distance between the two epigraphs [ f , C ] and [ g , D ] shown below: one in terms of the original functions f , g and one in terms of the duals f * , g * : we look for the two tangent hyperplanes for f and g that are maximally separated from one another, as illustrated in [link] .

Illustration of the Fenchel dual problem on a conjugate function.

If the infimum on the left is achieved by x 0 , then

max x C [ x , x 0 * - f ( x ) ] = x 0 , x 0 * - f ( x 0 ) , max x D [ x , x 0 * - g ( x ) ] = x 0 , x 0 * - g ( x 0 ) .

Example 1 (Allocation) Assume that we have a capital x 0 available for investment with n different funds. There is a predicted gain g i ( x i ) to having stock worth x i at fund i , where the functions g i are concave. We aim to find the optimal allocation of the capital x = ( x 1 , ... , x n ) that maximizes the total gain g ( x ) = i = 1 n g i ( x i ) .

To appeal to duality, we have the concave function g ( x ) and must define a convex function, e.g., f ( x ) = 0 . The constraint set can be written as the intersection C D with

C = { x : i = 1 n x i = x 0 } = { x : 1 , x = x 0 } , D = { x : x i 0 , i = 1 , ... , n } .

Therefore, we can write our optimization problem as

min x C D { f ( x ) - g ( x ) } .

We consider the conjugate sets. First, we have

C * = { x * X : sup x C [ x , x * - f ( x ) ] < } , = { x * X : sup x C [ x , x * ] < } .

We want to define C * more explicitly. Let x * C * be written as x * = λ 1 + w , where w 1 . Then,

x * , w = λ 1 , w + w , w = w 2 ,

which can be arbitrarily large. Now let x = x 0 n 1 + λ w ; it is easy to check that x C for all λ R . If w 0 then x , x * = λ x 0 + λ w 2 , which again can be arbitrarily large. Since x * C * must hold that sup x C x , x * < , we must have that w = 0 and so x * span ( { 1 } ) . Since x * C * was arbitrary, then C * span ( { 1 } ) .

It is also easy to see that span ( { 1 } ) C * , therefore implying that C * = span ( { 1 } ) .

For D , we have a conjugate set

D * = { y : inf x D [ x , y - g ( x ) ] > - } , = { y : inf x D [ x , y - g ( x ) ] > - } ,

since g ( x ) x 0 . Now D D * since all vectors in D have nonnegative entries. Fix y D * ; if y D then there is some negative entry y i < 0 among i = 1 , ... , n . For such i let x λ = λ e i D for some λ > 0 ; then we get x λ , y = λ y i which can be arbitrarily close to (i.e., as λ we have x λ , y - . Thus y D * , a contradiction. Therefore, if y D * then y D and so D * D . We have therefore shown that D * = D .

The conjugate functionals can be written as

f * ( x * ) = sup x C x , x * = λ x 0 ,

since each x * C * can be written as x * = λ 1 . Therefore, f * can be written as a function of a single variable. Similarly,

g * ( x * ) = inf x D ( x , x * - g ( x ) ) = inf x D i = 1 n x i x i * - g i ( x i ) = i = 1 n g i * ( x i * ) ,

where we write

g i * ( x i * ) = inf x i > 0 ( x i x i * - g i ( x i ) ) .

For x * C * D * we can write x i = λ > 0 for all i = 1 , ... , n and so

g i * ( x i * ) = g i * ( λ ) = inf x i > 0 ( λ x i - g i ( x i ) ) .

Therefore, the original problem can be reformulated as the following single-variable problem:

λ * = min λ > 0 λ x 0 - i = 1 n g i * ( λ ) .

This is due to x * = λ 1 C * D * if and only if λ = 0 . Once λ * is found, we can find each x i as the minimizer in g i * ( λ * ) , cf. [link] .

Get Jobilize Job Search Mobile App in your pocket Now!

Get it on Google Play Download on the App Store Now




Source:  OpenStax, Signal theory. OpenStax CNX. Oct 18, 2013 Download for free at http://legacy.cnx.org/content/col11542/1.3
Google Play and the Google Play logo are trademarks of Google Inc.

Notification Switch

Would you like to follow the 'Signal theory' conversation and receive update notifications?

Ask