05.01.2013 Views

COMPUTATIONAL PROBLEMS IN ABSTRACT ALGEBRA.

COMPUTATIONAL PROBLEMS IN ABSTRACT ALGEBRA.

COMPUTATIONAL PROBLEMS IN ABSTRACT ALGEBRA.

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

396 R. E. Kalman Invariant factors and control theory 397<br />

We say that Y has a finite-dimensional realization if and only if there exist<br />

matrices F, G, H over K such that<br />

Yk=HFkG, k=O,l,.... (11)<br />

The matrix F is required to be nX n (then H is pXn and G is nXm). We<br />

say that the realization is minimal if and only if n is the smallest integer for<br />

which (11) can be satisfied. The following theorem is fundamental:<br />

Zf Fy and & belong to minimal realizations or the same Y, then they are<br />

similar.<br />

A proof may be found in [5] or [6].<br />

It is now clear that the theory of realizations generalizes the theory of<br />

elementary divisors: if A is some given matrix, then the set of all minimal<br />

realizations of the sequence {Yk} = {Ak} is identical with the similarity<br />

class of A, since all triples of the form (F, T, T-l), with T-IFT = A, are<br />

minimal realizations. So the following is a well-defined problem: Given<br />

{Yk} possessing a finite-dimensional realization, determine the similarity<br />

invariants of F belonging to some minimal realization. A rather detailed<br />

examination of this problem built around the classical machinery of invariant<br />

factors and elementary divisors is given in [7], to which the reader is<br />

referred also for additional motivation and background material.<br />

In complete generality, that is, in terms of exhibiting efficient “functors”<br />

to linear algebra, the solution of the problem is definitely not known at<br />

present.? Let us review briefly what is known.<br />

The simplest invariant of F (minimal) is given by the following result,<br />

which is new, turns out to be quite simple, and seems to be fundamental:<br />

Let S;denote the blockwise NXNgeneralized Hankel matrix<br />

[Y#J Yl . . . YN-1 -<br />

Yl Yz . . . YN<br />

sy=. .<br />

[&-1 Y, . . . y,-11<br />

induced by the matrix sequence Y. Then dim F = rank ST for N suflciently<br />

large. In other words, a finite-dimensional realization exists if and only if<br />

the rank of S; is eventually constant, and then<br />

n(Y) = rank SG = fl deg yi (vi+1 1 yip i = 1, . . ., q-l),<br />

where yi are the invariant factors of the square matrix F belonging to any<br />

t The computational experience of numerically determining minimal realizations has<br />

been summarized in [8], pp. 373-405. Four methods have been compared there: (i) classical<br />

elementary divisor theory (see [71); (ii) partial fraction expansions and rank computations<br />

([4], Section 8); (iii) direct application of elementary linear algebra ([41 Sections<br />

7 and 8); (iv) Ho’s algorithm via Hankel matrices 161.<br />

minimal realization of Y. In particular, rank S; = n(Y) for all r =- n(Y) or<br />

even r * deg ~1.<br />

Referring to the module language mentioned at the end of 0 2, we can<br />

rephrase the preceding statements also as follows: The module induced by<br />

any F belonging to a minimal realization of Y is of dimension n(Y); this<br />

module is the direct sum of precisely q cyclic pieces, each with annihilating<br />

polynomial pi. In short, the (numerical) sequence Y may be used to determine<br />

module invariants (n, q, the YJ with exactly the same ease (or difficulty)<br />

as it can be used to determine similarity invariants for F.<br />

It is clear that q, deg ~1, and even the vi could be determined by a combinatorial<br />

procedure examining the linear dependences of subsets of the<br />

matrix 5;. The detailed prescriptions are easily inferred from [4-81. It is,<br />

however, not yet clear if a “functorial” procedure can be obtained for<br />

this purpose. The clarification of this problem, in view of the situation<br />

sketched above, is clearly one of the outstanding present research problems<br />

in linear algebra.<br />

This problem is closely related also to other unsolved elementary problems<br />

of a linear-algebraic type. Let us mention the following interesting<br />

CONJECTURE. (“Parametrization of minimal realizations.“) Let (F, G, ZZ)<br />

be a minimal realization of its own sequence {Yk = HFkG, k = 0, 1, . . .}.<br />

Let X be the corresponding K[z]-module, with q cyclic pieces. Let ~1, . . . , ye<br />

be the invariant factors of X, with vi+1 ] yi and deg vi = n!. (Thus y1 = minimal<br />

polynomial of X and nl + . . . + n, = dim X.) Finally, let m = p = q<br />

and rank G = q.<br />

Then: For fixed q and fixed (nl, . . . , n4) the set of all such triples plus<br />

a set of measure zero (corresponding to triples which are not minimal)<br />

is a linear space over K whose dimension is precisely equal to<br />

5 min {ni, ?Zi,1 -ni}+ 2 (2j-1)nj (no = -).<br />

i=l j=l<br />

The first term in the above expression represents the minimal number<br />

of parameters necessary to specify all the invariant factors (given their<br />

degrees).<br />

The second term is the dimension of the linear space of transformations<br />

leaving the rational canonical form of F invariant. (This number was<br />

first determined by Frobenius.) It can be shown that the dimension of<br />

the linear transformations leaving F invariant is precisely the same as the<br />

number of parameters in G which can be fixed over the whole class: For<br />

instance, if q = 1 or if nl = . . . = nq all elements of G can be fixed.<br />

This is closely related to canonical forms of completely controllable pairs<br />

(F, G). See [7].

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!