See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220349867

Componentwise Distance to Singularity

Article  in  Journal of Universal Computer Science · January 1998

Source: DBLP

CITATION READS

1 54

1 author:

H.G. Rex
University of Leipzig
25 PUBLICATIONS   114 CITATIONS   

SEE PROFILE

All content following this page was uploaded by H.G. Rex on 31 December 2015.

The user has requested enhancement of the downloaded file.

 q
\J
U
§
.Y

\.;
Componentwise Distance to Singularity
d
-{ G. Rex
§ § (Institute of Mathematics, University of Leipzig
\ u rex@mathematik.uni-leipzig. de)

§ I
o
\, §o
Abstract: A pertr-rrbation matrix "4: A * 4 is considered, where A € IR',' and
\o O < A € IR'''. The matrix .Ä is singular ifl .4 contains a real singular matrix. A
problem is to decide if is singular or nonsingular, a NP-hard problem. The decision
"4
.i \' can be made by the computation of the componentwise distance to the nearest singu-
9 t\ lar matrix defined on the basis of the real spectral radius, and by the sulution of 4"
eigenvalue problems.
Theorem 6 gives a new computation basis, a natural way to the "componentwise dis-
h tance ..." definition, and a motivation to rename this in lodius of singularity denoted
\ q by sir(A, A).
This new way shows: (i) - sir results from a real nonnegative eigensolution of a non-
\ ür linear mapping, (ii) - sir has a norm representation, (iü) - sir can be computed by
§F 2'-1 nonnegative eigensolutions of the nonlinear mapping, (iv) - for the special case
I\ A : WT ,O I p,g € R-' a formula for a computation of sir is given, also a trivial
\ algorithm for the computation, and some examples as demonstration.
\
o Key \Mords: perüurbation matrix, interval matrix, componentwise distance to the
nearest singular matrix, radius of singularity, NP-hard
\
§ § Introduction
§
In this paper a perturbation matrix ..

§-^
§ A_ A+A
with A € IR''' and 0 ( A e IFf'" or equivalent to this an interval matrix
§§
A=IA_ A, A+ Al
is considered.

The following two problems are the subject of this paper:

o to decide where is the nearest singular matrix with regard to the matrix A;
o to decide if "4 is singular or regular.
A matrix of the type above is called singular iff a real singular matrix is included
in ,4, and a matrix is said to be regular iff it is nonsingular.

It is well-known that the given decision problems can be solved by the compu-
tation of d(A, ^4) denoted as "radius of regularity" I see Poljak , Rohn (93) ] or
as "componentwise distance to the nearest singular matrix" I see Demmel (92)]
and IRump (97)].
 On the other hand it has been shown by I Poljak , Rohn (93) ] that the
computation of d(A,A) is a NP-hard problem; there are 4' linear eigenvalue
problems to solve (see IDemmel ], IHigham ], I Chaitin-Chatelin , Frayss6 ]).

Theorem 6 in I Section 2 ] gives a new computation basis and "a natural way" to
/), and a motivation to rename this in radius of singularity
the definition of d(A,
denoted by sir{A,/), also to make a distinction between both computation
formulae.

In ISection 4 ] it is shown that d(A,A) = sir(A,A), but for a computation of

sir are "only" 2'-1 nonnegative eigensolutions to compute.
In the following it should be explained what is to be understood by " a natural
way": At first in ISection 1.1] it is shown by using of interval mathematics that
the representation of "4 as perturbation matrix is equivalent to the midpoint-
radius representation of an interval matrix.

Theorem 3 gives equivalent formulations for the singularity of an interval matrix.

The assumption that the midpoint matrix A of A is regular allowed the study
of a nonlinear mapping F5 : IRI -+ Ri with Fs(z) :- AIA-t§zl where
IRI := {r e IR" I 0 < rr,I < i( n} denotes the cone of theR",and S o
signature matrh Isee Section 1.1.1 ]. Note there are 2n-r signature matrices,
and this set is denoted by O".

The basis statement of this paper is (see Theorem 6):

3(.\;z) € IR1 x IRI ( .Fs(z) = \2, llzll = 1 ) .

The proof based on Brouwers fired point theorem given in ISection T.1.2].
Note ll . ll stands for any vector norm.

On the set of the nonnegative eigensolutions is

zl := j[S* max{ A(^9) = ll4lA-lszlll I Fs(z) = Ä(s) z , llzll: t}
defined, see (17).

Let now, for a S, z and A the equation Fs(z) = A z be fulfiled, then it can
be shown that this is equivalent to A(A) := At(llDA is singular if A+ 0
Isee Section 3 ]. Furthermore in ISection 3 ]"4(r) :-- A* 14 is discussed
dependent of ,21. This way is new and leads also to the well-known definition
min{t > 0l "4(t) singular} =: sir(A,4), here denoted by sir(A,A) to make a
distinction to d{A,/) and between both ways. That the ways are really different
is shown in I Section 3 ] and I Section 4 ].

The term radius of singularity for sir(A,4) is motivated by the following: for
t = sir(A, a) is "4(t) the closure of A(t) regular for t € 10, sir(A, a)) ; u singular
matrix s e 0A exists, where ä"4 is the set of boundary matrices of "4, and last
but not least sir(A,4) has with.zl .ll.ll- representation.
The special case .4 - pqT is in ISection 5 ] considered. The application of theorem
6 gives here a new representation formula for sir and Ieads to an algorithm for the
computation; examples are given. Especially it is shown that p is an eigenvector
of the mapping Fs and

sir(A,pqr) - rnac s E6n-t {cr lA-L Spl}

This formula contains also the special case where all elements A;,j = 1 consid-
ered by IRohn (96) ] , IDemmel (92) ] , IRump (97) ]. A representation of sir
in a subordinate matrix norm is also given by (32).

l-.1 Definitions, notations, and preliminaries

In this first section brief survey connections between interval and perturbated
matrices are given, since the interval analysis gives an easy access to the subject
of this paper.
The symbolIIR:= {X lX = [X,X] i X,X € IR} denotes the set of all
closed real intervals. The following equivalent notations for an X € IIR are used
.
+ rad(x)l
*',',1*,x\'mid(x) (1)

with the definitions:=n1iä\

mid{X)'= (2( +X)12 for the midpoint of X and
rad(X)'- (X - X)12 for the radius of X.It is obvious to see that rad(X)
is always nonnegativ. The representation (1) is called the midpoint-radius rep-
resentation of an interual X. In this context rad(X) can be interpreted as a
perturbation of mid(X).
For the following some rules for intervals are used.
LemmaL. Let fl € IR andY e IIF.,, then
Yr = mid(Y)a +. rad(Y)lrl (2)
0e Y<=:+lmid(Y)l < rad(Y) (3)

Proof. To (2): The application of the multiplication for intervals I see Moore
(79) l, and (1) gives
Yr : [min{ (mid(Y) rad(Y)) r, (mid(Y) + rad(Y)) r} ,
-
n,ax{(mid(Y) rad(Y)) r,(mid(Y) + rad(Y)) r})
-
= mid(Y) r *. rad(Y) lxl.
To (3):
0e Y <+ mid(Y) - rad(Y) < 0 < mid(Y) + rad(Y) a lmid(Y)l < rad(Y).
tr
 An extension of intervals to interval matrices, interpreted as a perturbation
matrix, is useful in this context. A n-by-n interval matrix (A;,i) can be generated
by a componentwise perturbation (/;,i) of a real matrix(At,i). The componen-
twise representation is given by using (1) with

At1=A;,itA;,i, 0 a A;,i
where A;,7 is the midpoint, and A;,i is the radius of A;,i or
A= A+A, A€ IR''', 0 < A€ IR''' (4)

a well-known notation for a perturbation matrix. This shows that because of (1)
the representation (4) is equivalent to the following interval matrix
A=IA- A, A + 4l (5)

For the followingthe absolute value l.l, and the relations (, €, 9, J are to
be used element- or componentwise, respectively.

As next a definition and some equivalent formulations for the singularity of an

interval matrix (5) are given.

Definition2.

"4 singular <- ls€IR'''fr€lR"(se"4, §r=0, n*0) (6)

and an in this context important

Theorem3.
A singular * 1r € ß." ( 0 € As, t * 0 ) (7)
+:+ 3r € tr," ( lA*l 3 Alxl, r { 0 ) (8)

Proof. The elementewise application of (2) to Ac with an c € IR' and "4 from
(5) gives
Ax:{arlaeA} (e)
= Ax * Alxl. (10)

To (6) e (7): With (9) it is obvious that (6) <+ (7), then
fs€IR"'"lr€lR"(se A, §r=0, *+0) s (0e Ar).
To (7) <+ (8): Consider 0 e "4r from (7), use (10), and apply (3) with Ax =
mid(A r), Alcl : rad(A r) then
0e Ar = Ar * Alal e lAcl ! Alal
proves the equivalence. tr
 For the examinations in this paper it is important to note that the equivalence
relation (8) can be specified with the midpoint matrix A = mid,(A); there are
two cases in
Theorem4.
Asinsular § lc€m."(lAxl= 0<4lcl,c$0) (11)
(A singular, A regular) e 1y € R" (lvl < AIA-'yl, y *O) (I2)
Proof. To (11): This is obvious with (8), and A = mid(A) is singular. To (12):
This is equivalent to (8) because of the regularity of A and y := Ax € x =
A-ry. tr

1.1.1 Signature matrices

Let r eIR*, then r =,Slrl with
S := diag(sign(rl), . . ., sign(o"))
where S is called the signature matrir of the uector r.
The following properties are obvious with the definition of S:

S =.97 =.9-1 ; 52 = 1; lsl = I , I denotes the n-by-n identity matrix.

The set of all these signature matrices is denoted with O" . Note the cardinality
of O" is 2".

L.t.z Brouwer's fi*"d poi*rt theorem

In this context we use the following version of Brouwers ficed point theorem
Theorem5. Let 0B be the sphere of a closed unit ball B:= {r € IR' I llrlls
1) o/IR", andletT: B -+ IR" äe s continuous mapping on B withT(AB) gB,
then

1r€B(Tr=x).
Proof. See I Riedrich (i6) ] . D
2 Basic theorem

On the basis of (12) a nonlinear mapping is defined and studied in

Theorem6. Let S €O" and Fs : IRI -+ IR,i with Fs(z) := AIA-rSzl, then
l(.\;z)€IR+xIRI( Fs(z) -\r,llzll=1) (13)

and

l = ll4lA-t S zlll (14)

Such a pair (); z) is called a nonnegatiue eigensolution of Fs.

Proof. For the nonlinear mapping ,Es it is useful to define the kernel of Fs by
ker(fu):= {z eIR[ | rs(z) = 0, z + 0].
The proof of this theorem is divided in two parts:
Part 1: Assume 0 < 4 € IR"'' , then it is obvious to see that &er(F5) = 0
t The intersection of a closed unit ball with IRir
(where 0 denotes the empty set).
is defined by
Bqi= {ze IR} lllrll S t}.
Now, a nonlinear functional fs : B+ -+ IR,i can be defined by

"fs := ll AIA-1 Szlll

and since ker(Fs) =0 f s has the property
O < /s(z) Yz€Ba.
This property allows the definition of a nonlinear mapping
Gs : Ba -+ IRf with

frkl rtrn.
1

Gs(z) :=

Properties of Gs are

(i), Gs e c0(B+) and (ii): Gs@B+) g aB+ I B+

where Co denotes the set of all continuous functions. (ii) is clearly by construc-
tion.
Put now B := Bq and 7 := Gs, and apply theorem 5 then
1ze Ba(Gs(z) = z)
( Gs has a fixed point) and furthermore for the mapping F5 is
Fs(z) = ), z

with
A := ll Ala-r Srlll, llzll = 1

satisfied.
This means the nonlinear mappiug Fs has at least a real nonnegative eigen-
solution (^; ,).
Part2 (generalcase): Let 0 < 4.Thendefinewithanarbitrary 0 < C €IR*''
A(t):- A + tC, t € IR.," .

Since 0 < A(t) fort ) 0 part 1of this proof can be applied, and

()(t); z(t))e IR1 xlR[ existswith llz(t)ll = 1

for
Fs(z(t)) =.\(t)z(t)
with
,\(t) := ll^(t) lA-L S r(r)lll .

Since ä.B1 is compact, there is an accumulation point z e )Ba for each zero
sequence { t* } -+ *0 . Because of the convergence of the sequences

{r's(r({ tr }) } -+ Fs(r) and { .\( { t* } ) } -+ r

for each {tr} -+ *0, any such äccumulation point satisfied Fs(r) = lz with
.\:= ll Al4-r ^9zlll and llzll : 1 . o

The statement of this theorem is, that for Fs exists at least a real nonnega-
tive eigensolution. This statement is very important for the following definition
of the radius of singularity.
 3 Radius of singularity - sir(A, A)
Theorem 6 gives the basis for the definition of the radius of singularity.
Consider

Fs(z) = ),2 - Ale-r S zl (1b)

and (15) can be transformed with r :-- A-L S z and.\ l0 into

lAxl: lo4l, (16)

on the other hand use (8):

A singular +:+ lr € ß," (lArl S Alrl, r + 0) .

Equation (16) shows that l A r l < Al r l is sharp for the "smallest pair" (1/.\; r).
That leads with (16), and (14) under the consideration that,\ = )(S) to the
following definition
A,: pB* max{ )(S) = ll4l/-1^Srlll I ps(4 = Ä(s) z ,llzll= r} (17)

Remark /: This definition depends on the eigensolutions of F5. Note that also z
depends on S, z: z(S).

Put .tl in (16) then is because of (8) the perturbation matrix "4(|) singular.
That is for
( regular if t e [0, ])
A(t):= ArtA< (18)
I singular if , € [i, *]
where .4(t) be closed by ,4(oo) ihis will be convenient in this context, see also
' below.

Furthermore there are two special cases to be considered:

(u),

e fr € R" ( lArl = O = j|ful, x I

,4 singular 0)

That is for "4(t): set | - 0 then "4(t) is singular Vt e [0,m].

(b): Define
ker(F):= Useo.{Äer(F5)} = {c € IR"(dl1-1cl = 0 , t +O}.
Letker(F)t'0and
FQ,d € IRa x IR"\ßer(F) (Älrl = Aln't rl, c l0) (19)

then
.\lrl= 0- AIA-rcl Vc € ker(F) (20)
is solvable only for .\ = 0, that is A = 0.

It can be seen immediately that (19) and (20) are also true if / is replaced by
tA. That means
V, € [0,oo)(Är € IR"(lol <tAlA-rrl, * * 0) s /(7) resular). (21)
The equivalence (21) follows from the negation of (12).

The statement of (b) for

"4(t)
is:
"4(t) is singular for t = cp (21= 0).

In consideration of these three cases (18), (a), and (b) it is evident to define
min{t € [0,m] lA(t)si,nsular] =i sir(A,A) (22)
here denoted as radius of singularity of a matrir A with
( O if A is singular (see (o))
sir(A,d) ={ooif A=0 (see(Ö)) (23)
[ ] otherwise (see ( 18))

For the following is sir as abbreviation of sir(A,4) to be understood.

The way to sir has shown:
* sir results from a real nonnegative eigensolution of Fs;

- sir has with (17) a normrepresentation for any vectornorm on IR";

- because of lA-1Szl = 1,4-1(-,S)zl VS € O" are "only" 2'-1 eigensolu-

tions of ,Fs to compute;

- a singular matrix s e 1A{sir) exists where 1A(sir) denotes the boundary

set of (sir) (see section 5 examples) and further
"
A(sir) = int{ A(sir) } u 0"1,(str)
with int{A(sir) } = {o € IR''' I ae A(t) t € [0,sir)],

Remarle 2: From (b) follows

Ä=0 + ker(F)t'O
or equivalent
ker(F) =fi =+ A> 0 .

An example is given in the next section.

For a better understanding of sir some examples are given in I Section 5 ].
3.L A nontrivial example for sir = oo

Let
A:= D * A (24)
with D is a regular real diagonal matrix, and
/0*,. .r,\
{oo*.
*I .
t......1
a:=lo..o**l
I o. . .o* I

\oo. o)
where * stands for arbitrary nonnegative real elements.

With (24) is Fs(z) = Al D-r.Sz I and the application of the Perron/Frobenius

theorem, see e.g. IRiedrich (76) ], "gives

Fs(z) = AlDl-'z = p(AlDl-' ), = 0 z YS e O"-r

where p(.) denotes the spectral radius. It is obviously to see that z1 = 0, and
therefore is sir(.A, Ä) - * .

RemarlcS: It is easy to see that ker(F)f 0 since e1 := (1,0,..., 0)" €. ker(F).

4 sir(Ar 4) is equivalent to d,(A, A)

J.Rohn defined in I Rohn (89) ], and IPoljak , Rohn (93) ] the radius of resularity
by
d(A,A) :* inf{i > 0 I lA - t A, A + t Al sinsular} . (25)

With (22) is shown that the inl is achieved.

The computation formula given by J. Rohn I see Rohn (89) ] for d(A, A) is
equivalent to sir(A, A).

Then an equivalent transformation of

Fs(z) = Ale-rSzl- ),2 {26)
with
u=A-LSz
glves

),lAal = Alxl ,
and with

TyAn - lArl, Tr, = lcl, 71,72 QOn

is (26) transformed into
A-tTrAT2r = .\r, (27)

a real eigenvalue problem.

The definition of the real spectral radius for a .B € IR'''

po(B),-{ I max{ l,\ | } if B has real eigenvalues

' : 0 1,
t otherwrse

applied to (27) gives

Tr ATz)} = A (2s)
,,ä*ä"{ Po(
"A-r
The left hand side of (28) was given by J. Rohn in [Rohn (89)], and is because
of (27) equivalent to zl from (17). Therefore is d(.4, A) = sir(A,Ä) * given in
(23).

For a computation of. A are 2'-1 eigensolutions of Fs to compute instead of 4n

linear eigenvalue problems for ps( A-r Tr 4Tz).

The computation of sir is also a NP-hard problem since in [Poljak and Rohn
(93)] was shown that the computation ol d(A,A) is a NP-hard problem.

In the following section applications of theorem 6 are given.

5 A dyad as a special perturbation matrix

Let
A:=pQT , p,q4_IF,",
4 is called dyad and A be the regular midpoint matrix of the perturbation
matrix A- A+. A.

Then on the basis of the proof of theorem 6 for a,S € O"

'pq:lA-Lszl -- 1 -
,yt-n;flfir§;ffl
G,(z\=
ilpilo
This means

,= 1

llollp, llzll=t (29)

is a fixed point of Gs. On the other hand is (29) also an eigenvector for

*(# p) = qr lA-' sp,

# r=
^# o

with the eigenvalue

.\ = )(S) = qr 1A-r S pl .

Finally there is the following representation for sir

sir(A,pqr)=#. (30)

In the following there are some other representations for (30) given:

Define e := (L, .. .,1\' € R", the diagonal matrices

Do :- diag(p) , Ds := diag(q) then p: Dp e , Q : D, e ,

and

H := DrA-t Do '
With these definitions and llr ll1 :- er lc I is
sir(A,pqr) = (31)

equivalent to (30).

sir can be represented alsoby a'subordinate norm Isee Golub, Van Loan (89)]
on afinite set, especially the corners of the unit cube on IR', then

sir(A,pqr) = (82)
||IIim
where

ll fl ll-,, = fl s, ll,} = ll /{se l11 .

sääXt,{ll rr#r31,
Because.S € O"-1 is the cardinality 2"-1, furthermore it was shown in IRohn
(96) ] that the computing of (30) is NP-hard.

It is worthy to stress, that the formulae (30) - (32) are suitable for a computa-
tion of a parallel computer I Rex ].

The following examples are given for a demonstration of the statements of this
paper. The basis for the computation gives the following algorithm.
5.1 ALGORITHM
given: A,p,Q
set: zl := 0
k:=l
repeat:

^9 €
O"-1 formal notation
Aw = Sp solve a linear system of equations (exactly)
\ := qT lwl
Test: If ,\ > ,4 then A:= ),

k:=(x41
until lr > 2"-L

5.1.1 Example 1

LetA=A*pqr with

o .- ( 1-1\
o'=\r /t\ /s\
,'=(n/'
r)' P
'=(r/'
then sir := sir(A,pqr) = 2ll5 computed with Algorithm 1.

The matrix

."a}i :=A* psirpqr isregularfor 0 ( p ( 1. (33)

A(i = (l-i) .,* (ä ä)

Furthermore

.4(1)=+(3 ;;] i:l,J,i) eA,- ([;,;][_1,3])

There is only one singular matrix s on ä"4

1 ('J e oA := :i ;l])
"'= -l) + (1 Z',1lrl
5.L.2 Example 2
Let A= ALpqr with
/l 0 -4\
' '= (.?-; il' o'= (
)
, '= (3!u) ,

then sir := sir(A,pqT) = 40138 computed with Algorithm 1.

And analogous to (33) is

/I 0 -4\ 40 / 0.5 0.25 2 \
A(p) = (l _; *,s8 lo.s 025 z I
S,) \o.s o.2b z )
Furthermore

, ll 9,2911-5 , 5l [-116,36]\
A(t\ =*rö I [-10,10] [ 33, 43] [ 74,154] | ),4
\[ 9,29] [-43 ,-33] [ -40, 40]/
t /1e.5, 28.51 [ -4.75 , 4.75]l-rr4 381 \ ,
A= *,o I [-9.5, 9.5] [33.25,42.75)1 76 rszl I , .

\[-9.s, 2B.s1l-42.75,-33.25] [ -88, 381 /

There is a singular matrix s on 0A

:- i-q / I -5-116\ , /{ 9,28}{-5, 5}{-116,36}\

[-'3 -i3 -'^il '0A:= * (i '3 ; l3l I i3 ; ä3i I 1t','1fi)
§

References

IChaitin-Chatelin , Frayss6 (96) ] Chaitin-Chaielin, F., Frayss6, V.: "Lectures on Fi-

nite Precision Computations"; SIAM, Philadelphia (1996)
I Demmel (92) ] Demmel, J.: "The Componentwise Distance to the Nearest Singular
Matrix"; SIAM J. Matrix Anal. Appl. ,13 (1992) 10-19.
I Golub , Van Loan (89) ] Golub,G. H., Van Loan, C. F.: "Matrix Computations"; The
Johns Hopkins University Press, Baltimore (1989)
[Higham (96) ] Higham, N.J.: "Accuracy and Stability of Numerical Algorithms";
SIAM, Philadelphia (1996)
I Moore (79) ] Moore, R. E.: "Methods and Applications of Interval Analysis"; SIAM,
Philadelphia (1979)
 Poljak , Rohn (93) ] Poljak, S., Rohn, J.: "Checking Robust Non-singularity is NP-
Hard"l Math. of Control, Signal, and System, 6 (1993), 1-9.
Rex ] Rex, G.: "The Radius of Singularity for Matrices"; in preperation
Riedrich (70) ] Riedrich, T.: "Vorlesungen ueber nichtlineare Operatorgleichungen";
Teubner VG, Leipzig (1976)
Rohn (89) ] Rohn, J.: " Systems of Linear Interval Equations" I Linear Algebra Appl.,
126 (1e8e) 3e-78.
Rohn (96) ] Rohn,' J.: 'Complexity of Some Linear Problems with Interval Data"l
Tech. Report No. 687, Institute of Computer Science, Academy of Sciences, Prague
(1ee6)
Rump (97) ] Rump, S. M.: "Bounds for the Componentwise Distance to the Nearest
Singular Matrix"; SIAM J. Matrix Anal. Appl. 18, 1 (1997) 83-103.

This article was processed using the ffif macro package with JUCS style

View publication stats