Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
73 views

Xxinear and Nonlinear Programming

The document discusses eigenvalues in the tangent subspace for constrained optimization problems. It makes three key points: 1) The eigenvalues of the Lagrangian matrix L restricted to the tangent subspace M (denoted LM) determine the convergence rates of algorithms for constrained problems, analogous to the role of the Hessian's eigenvalues for unconstrained problems. 2) LM can be represented by an (n - m) x (n - m) matrix obtained by projecting L onto an orthonormal basis for M. The eigenvalues of this matrix are the eigenvalues of LM. 3) Alternatively, the eigenvalues of LM can be found as the roots of a characteristic polynomial of order n+m, obtained by constructing a bordered

Uploaded by

bidbifb
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
73 views

Xxinear and Nonlinear Programming

The document discusses eigenvalues in the tangent subspace for constrained optimization problems. It makes three key points: 1) The eigenvalues of the Lagrangian matrix L restricted to the tangent subspace M (denoted LM) determine the convergence rates of algorithms for constrained problems, analogous to the role of the Hessian's eigenvalues for unconstrained problems. 2) LM can be represented by an (n - m) x (n - m) matrix obtained by projecting L onto an orthonormal basis for M. The eigenvalues of this matrix are the eigenvalues of LM. 3) Alternatively, the eigenvalues of LM can be found as the roots of a characteristic polynomial of order n+m, obtained by constructing a bordered

Uploaded by

bidbifb
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

11.

6 Eigenvalues in Tangent Subspace 335

In Example 1 of Section 11.4 it was found that x1 = x2 = x3 = 1


= −2 satisfy
the first-order conditions. The matrix F + T H becomes in this case
⎡ ⎤
0 1 1
L=⎣ 1 0 1 ⎦
1 1 0

which itself is neither positive nor negative definite. On the subspace M = y 


y1 + y2 + y3 = 0 , however, we note that

yT Ly = y1 y2 + y3  + y2 y1 + y3  + y3 y1 + y2 

= −y12 + y22 + y32 

and thus L is negative definite on M. Therefore, the solution we found is at least a


local maximum.

11.6 EIGENVALUES IN TANGENT SUBSPACE


In the last section it was shown that the matrix L restricted to the subspace M
that is tangent to the constraint surface plays a role in second-order conditions
entirely analogous to that of the Hessian of the objective function in the uncon-
strained case. It is perhaps not surprising, in view of this, that the structure of L
restricted to M also determines rates of convergence of algorithms designed for
constrained problems in the same way that the structure of the Hessian of the
objective function does for unconstrained algorithms. Indeed, we shall see that the
eigenvalues of L restricted to M determine the natural rates of convergence for
algorithms designed for constrained problems. It is important, therefore, to under-
stand what these restricted eigenvalues represent. We first determine geometrically
what we mean by the restriction of L to M which we denote by LM . Next we
define the eigenvalues of the operator LM . Finally we indicate how these various
quantities can be computed.
Given any vector y ∈ M, the vector Ly is in E n but not necessarily in M.
We project Ly orthogonally back onto M, as shown in Fig. 11.4, and the result
is said to be the restriction of L to M operating on y. In this way we obtain a
linear transformation from M to M. The transformation is determined somewhat
implicitly, however, since we do not have an explicit matrix representation.
A vector y ∈ M is an eigenvector of LM if there is a real number
such that
LM y =
y; the corresponding
is an eigenvalue of LM . This coincides with the
standard definition. In terms of L we see that y is an eigenvector of LM if Ly can
be written as the sum of
y and a vector orthogonal to M. See Fig. 11.5.
To obtain a matrix representation for LM it is necessary to introduce a basis
in the subspace M. For simplicity it is best to introduce an orthonormal basis, say
e1  e2      en−m . Define the matrix E to be the n × n − m matrix whose columns
336 Chapter 11 Constrained Minimization Conditions

Ly

LMy

Fig. 11.4 Definition of LM

consist of the vectors ei . Then any vector y in M can be written as y = Ez for some
z ∈ E n−m and, of course, LEz represents the action of L on such a vector. To project
this result back into M and express the result in terms of the basis e1  e2      en−m ,
we merely multiply by ET . Thus ET LEz is the vector whose components give the
representation in terms of the basis; and, correspondingly, the n − m × n − m
matrix ET LE is the matrix representation of L restricted to M.
The eigenvalues of L restricted to M can be found by determining the eigen-
values of ET LE. These eigenvalues are independent of the particular orthonormal
basis E.
Example 1. In the last section we considered

⎡ ⎤
0 1 1
L=⎣ 1 0 1 ⎦
1 1 0

Ly

λy y

Fig. 11.5 Eigenvector of LM


11.6 Eigenvalues in Tangent Subspace 337

restricted to M = y  y1 + y2 + y3 = 0 . To obtain an explicit matrix representation


on M let us introduce the orthonormal basis:
1
e1 = √ 1 0 −1
2
1
e2 = √ 1 −2 1
6

This gives, upon expansion,


 
−1 0
ET LE = 
0 −1

and hence L restricted to M acts like the negative of the identity.

Example 2. Let us consider the problem

extremize x1 + x22 + x2 x3 + 2x32


1 2
subject to x + x22 + x32  = 1
2 1
The first-order necessary conditions are

1+
x1 = 0
2x2 + x3 +
x2 = 0
x2 + 4x3 +
x3 = 0

One solution to this set is easily seen to be x1 = 1, x2 = 0, x3 = 0,


= −1. Let us
examine the second-order conditions at this solution point. The Lagrangian matrix
there is
⎡ ⎤
−1 0 0
L = ⎣ 0 1 1 ⎦
0 1 3

and the corresponding subspace M is

M = y  y1 = 0 

In this case M is the subspace spanned by the second two basis vectors in E 3 and
hence the restriction of L to M can be found by taking the corresponding submatrix
of L. Thus, in this case,
 
1 1
E LE =
T

1 3
338 Chapter 11 Constrained Minimization Conditions

The characteristic polynomial of this matrix is


 
1−
1
det = 1 −
3 −
 − 1 =
2 − 4
+ 2
1 3−

√
The eigenvalues of LM are thus
= 2 ± 2, and LM is positive definite.
Since the LM matrix is positive definite, we conclude that the point found is a
relative minimum point. This example illustrates that, in general, the restriction of
L to M can be thought of as a submatrix of L, although it can be read directly from
the original matrix only if the subspace M is spanned by a subset of the original
basis vectors.

Bordered Hessians
The above approach for determining the eigenvalues of L projected onto M is quite
direct and relatively simple. There is another approach, however, that is useful
in some theoretical arguments and convenient for simple applications. It is based
on constructing matrices and determinants of order n + m rather than n − m, so
dimension is increased.
Let us first characterize all vectors orthogonal to M. M itself is the set of all x
satisfying hx = 0. A vector z is orthogonal to M if zT x = 0 for all x ∈ M. It is not
hard to show that z is orthogonal to M if and only if z = hT w for some w ∈ E m .
The proof that this is sufficient follows from the calculation zT x = wT hx = 0.
The proof of necessity follows from the Duality Theorem of Linear Programming
(see Exercise 6).
Now we may explicitly characterize an eigenvector of LM . The vector x is
such an eigenvector if it satisfies these two conditions: (1) x belongs to M, and (2)
Lx =
x + z, where z is orthogonal to M. These conditions are equivalent, in view
of the characterization of z, to

hx = 0
Lx =
x + hT w

This can be regarded as a homogeneous system of n + m linear equations in the


unknowns w x. It possesses a nonzero solution if and only if the determinant of
the coefficient matrix is zero. Denoting this determinant p
, we have
 
0 h
det ≡ p
 = 0 (26)
−hT L −
I

as the condition. The function p


 is a polynomial in
of degree n − m. It is, as
we have derived, the characteristic polynomial of LM .

You might also like