Structured Matrix Variables

Fairly complex matrix variable structures and interdependencies can be specified with lmivar. Recall that the symmetric block-diagonal or rectangular structures are covered by Types 1 and 2 of lmivar provided that the matrix variables are independent. To describe more complex structures or correlations between variables, you must use Type 3 and specify each entry of the matrix variables directly in terms of the free scalar variables of the problem (the so-called decision variables).

With Type 3, each entry is specified as either 0 or ±x_n where x_n is the n-th decision variable. The following examples illustrate how to specify nontrivial matrix variable structures with lmivar. The following examples show variable structures with uncorrelated and interdependent matrix variables.

setlmis([]) 
[X,n] = lmivar(1,[3 1]);
n

n = 
6

Y = lmivar(3,n+[1 2 3;2 1 2;3 2 1]);

Y = lmivar(3,toeplitz(n+[1 2 3]));

lmis = getlmis; 
decinfo(lmis,X)

ans = 3×3

     1     2     4
     2     3     5
     4     5     6

decinfo(lmis,Y)

ans = 3×3

     7     8     9
     8     7     8
     9     8     7

setlmis([]); 
[X,n,sX] = lmivar(1,[1 0;1 0]);
[Y,n,sY] = lmivar(1,[1 0;1 0]);

sX

sX = 2×2

     1     0
     0     2

sY

sY = 2×2

     3     0
     0     4

[Z,n,sZ] = lmivar(3,[0 -sX(1,1);-sY(2,2) 0]);

sZ

sZ = 2×2

     0    -1
    -4     0

Complex-Valued LMIs

The LMI solvers are written for real-valued matrices and cannot directly handle LMI problems involving complex-valued matrices. However, complex-valued LMIs can be turned into real-valued LMIs by observing that a complex Hermitian matrix L(x) satisfies

L(x) < 0

if and only if

This suggests the following systematic procedure for turning complex LMIs into real ones:

For LMIs without outer factor, a streamlined version of this procedure consists of replacing any occurrence of the matrix variable X = X₁ + jX₂ by

and any fixed matrix A = A₁ + jA₂, including real ones, by

For instance, the real counterpart of the LMI system

reads (given the decompositions M = M₁ + jM₂ and X = X₁ + jX₂ with Mj, Xj real):

Note that X = X^H in turn requires that $$X_1=X_1^H$$ and $$X_2+X_2^T=0$$. Consequently, X₁ and X₂ should be declared as symmetric and skew- symmetric matrix variables, respectively.

Assuming, for instance, that M ∊ C^5×5, the LMI system (Equation 1) would be specified as follows:

M1=real(M), M2=imag(M) 
bigM=[M1 M2;-M2 M1] 
setlmis([])

% declare bigX=[X1 X2;-X2 X1] with X1=X1' and X2+X2'=0:

[X1,n1,sX1] = lmivar(1,[5 1]) 
[X2,n2,sX2] = lmivar(3,skewdec(5,n1)) 
bigX = lmivar(3,[sX1 sX2;-sX2 sX1])

% describe the real counterpart of (1.12):

lmiterm([1 1 1 0],1) 
lmiterm([-1 1 1 bigX],1,1) 
lmiterm([2 1 1 bigX],bigM',bigM) 
lmiterm([-2 1 1 bigX],1,1)

lmis = getlmis

Note the three-step declaration of the structured matrix variable bigX,

See Structured Matrix Variables for more details on such structure manipulations.

Specifying c^Tx Objectives for mincx

The LMI solver mincx minimizes linear objectives of the form c^Tx where x is the vector of decision variables. In most control problems, however, such objectives are expressed in terms of the matrix variables rather than of x. Examples include Trace(X) where X is a symmetric matrix variable, or u^TXu where u is a given vector.

The function defcx facilitates the derivation of the c vector when the objective is an affine function of the matrix variables. For the sake of illustration, consider the linear objective

where X and P are two symmetric variables and x₀ is a given vector. If lmsisys is the internal representation of the LMI system and if x₀, X, P have been declared by

x0 = [1;1] 
setlmis([]) 
X = lmivar(1,[3 0]) 
P = lmivar(1,[2 1]) 
	: 
	: 
lmisys = getlmis

the c vector such that $$c^{T}x=\text{Trace}\left(X\right)+x_0^{T}P{x}_0$$ can be computed as follows:

n = decnbr(lmisys) 
c = zeros(n,1)

for j=1:n, 
	[Xj,Pj] = defcx(lmisys,j,X,P) 
	c(j) = trace(Xj) + x0'*Pj*x0 
end

The first command returns the number of decision variables in the problem and the second command dimensions c accordingly. Then the for loop performs the following operations:

In our example the result is

c = 
	3 
	1 	
	2
	1

Other objectives are handled similarly by editing the following generic skeleton:

n = decnbr( LMI system ) 
c = zeros(n,1) 
for j=1:n, 
	[ matrix values ] = defcx( LMI system,j,
matrix identifiers) 
	c(j) = objective(matrix values) 
end

Feasibility Radius

When solving LMI problems with feasp, mincx, or gevp, it is possible to constrain the solution x to lie in the ball

x^Tx < R²

where R > 0 is called the feasibility radius. This specifies a maximum (Euclidean norm) magnitude for x and avoids getting solutions of very large norm. This may also speed up computations and improve numerical stability. Finally, the feasibility radius bound regularizes problems with redundant variable sets. In rough terms, the set of scalar variables is redundant when an equivalent problem could be formulated with a smaller number of variables.

The feasibility radius R is set by the third entry of the options vector of the LMI solvers. Its default value is R = 109. Setting R to a negative value means “no rigid bound,” in which case the feasibility radius is increased during the optimization if necessary. This “flexible bound” mode may yield solutions of large norms.

Well-Posedness Issues

The LMI solvers used in the LMI Lab are based on interior-point optimization techniques. To compute feasible solutions, such techniques require that the system of LMI constraints be strictly feasible, that is, the feasible set has a nonempty interior. As a result, these solvers may encounter difficulty when the LMI constraints are feasible but not strictly feasible, that is, when the LMI

L(x) ≤ 0

has solutions while

L(x) < 0

has no solution.

For feasibility problems, this difficulty is automatically circumvented by feasp, which reformulates the problem:

Find x such that

as:

Minimize t subject to

In this modified problem, the LMI constraint is always strictly feasible in x, t and the original LMI Equation 2 is feasible if and only if the global minimum t_min of Equation 2 satisfies

t_min ≤ 0

For feasible but not strictly feasible problems, however, the computational effort is typically higher as feasp strives to approach the global optimum t_min = 0 to a high accuracy.

For the LMI problems addressed by mincx and gevp, nonstrict feasibility generally causes the solvers to fail and to return an “infeasibility” diagnosis. Although there is no universal remedy for this difficulty, it is sometimes possible to eliminate underlying algebraic constraints to obtain a strictly feasible problem with fewer variables.

Another issue has to do with homogeneous feasibility problems such as

A^TP + P A < 0, P > 0.

While this problem is technically well-posed, the LMI optimization is likely to produce solutions close to zero (the trivial solution of the nonstrict problem). To compute a nontrivial Lyapunov matrix and easily differentiate between feasibility and infeasibility, replace the constraint P > 0-by-P > αI with α > 0. Note that this does not alter the problem due to its homogeneous nature.

Semi-Definite B(x) in gevp Problems

Consider the generalized eigenvalue minimization problem

Minimize λ subject to

Technically, the positivity of B(x) for some x ∊ Rⁿ is required for the well-posedness of the problem and the applicability of polynomial-time interior-point methods. Hence problems where

with B₁(x) > 0 strictly feasible, cannot be directly solved with gevp. A simple remedy consists of replacing the constraints

A(x) < B(x), B(x) > 0

by

where Y is an additional symmetric variable of proper dimensions. The resulting problem is equivalent to Equation 3 and can be solved directly with gevp.

Efficiency and Complexity Issues

As explained in Tools for Specifying and Solving LMIs, the term-oriented description of LMIs used in the LMI Lab typically leads to higher efficiency than the canonical representation

This is no longer true, however, when the number of variable terms is nearly equal to or greater than the number N of decision variables in the problem. If your LMI problem has few free scalar variables but many terms in each LMI, it is therefore preferable to rewrite it as Equation 4 and to specify it in this form. Each scalar variable x_j is then declared independently and the LMI terms are of the form x_jA_j.

If M denotes the total row size of the LMI system and N the total number of scalar decision variables, the flop count per iteration for the feasp and mincx solvers is proportional to

While the theory guarantees a worst-case iteration count proportional to M, the number of iterations actually performed grows slowly with M in most problems. Finally, while feasp and mincx are comparable in complexity, gevp typically demands more computational effort. Make sure that your LMI problem cannot be solved with mincx before using gevp.

Solving M + P^TXQ + Q^TX^TP < 0

In many output-feedback synthesis problems, the design can be performed in two steps:

It turns out that a particular solution Xc of Equation 5 can be computed via simple linear algebra manipulations [1]. Typically, Xc corresponds to the center of the ellipsoid of matrices defined by Equation 5.

The function basiclmi returns the “explicit” solution Xc:

Xc = basiclmi(M,P,Q)

Since this central solution sometimes has large norm, basiclmi also offers the option of computing an approximate least-norm solution of Equation 5. This is done by

X = basiclmi(M,P,Q,'Xmin')

and involves LMI optimization to minimize ∥X ∥.