Symmetric Matrix In: Manchester Ml3 Opl, Engeanc
Symmetric Matrix In: Manchester Ml3 Opl, Engeanc
Nicholas J. Higham
Dqx@nent SfMathemutks
Unioersitg 0fMafwhmtfs
Manchester Ml3 OPL, EngEanc
Sdm%sd by G. W. Stewart
ABSTRACT
1. INTRODUCTION
where the spectral radius p(B)= max{lhl:det(B- XI)= 0), and the
Frobenius norm
8 W)
m-llAll
iNA)
llAll
&3tB=UH~i2polm~ti(UTU==I, H=HT>,@).2%enX,=
positiwtzppmdmti
(B + FE),/2 is the unique of A in the Fmbeniusmmz,
and
Let X -bepsd. Fmm the fact that IIS+ Klli- I[!# + llKllt if
IkqC
S=STand K= -KT,wehave
= Z Y$+ c(xi-Yid)s
i#j t
A. x,20,
di=
(
ot’
9 Xj<O,
@=l)
that is,
x,= z&g(a,)zT. (2-a)
TEIEDREM
3.1. For A E
LEMMA
3.2. &JG(r)) is a stddy munotonejhction of r (r 2 p(C)).
108 NIGEIOLASJ. HIGHAM
We have
G(r+Sr)-G(r)= [(r+&)21+C2]"2-(r21+C2)1/8
=D-‘[(r+6r)2
- r2]I = (2rSr + Gr2)D-1,
usingthe fact the two matrices
that whose sum is D commute. Thus
G(r + 6r) -G(r) is positivedefinite, which implies that h,,,JG(r + Er)) -
X&(G(r)) > 0, as requirei< q
LEMMA
3.4. l[A - G(r)li2 = r (r 3 p(c))=
Proof:
= r. n
LEMMA
3.5. &(A) 6 IlA- XFl12< 28,(A).
SYMMETRIC- !smmDElimlTEMATlux 109
=p(c)*=(o, -&&Q),
~~tinspecialcases(suchas~nAisnonnal)~~isno~tw~y
to compute i&(A) from (3.I), “becausethe eigemalues of G(r) = B + ( rel +
CB)lfl 8z5 nonlinear functio~~sof r. Therefore we turn to iterative methods.
~e~~~tPlis~ontbat88(A)istobeestlmated~modestaccuracy.
Suppose G(p(C)) is not positive SemidRfioife[if it is, then &(A) = P(C)&
Corollary 3.3 suggests the follow@ bisection app&. Find an interval
[a,@] containing P =&(A). If G(y) is positive definite9 where y =
(a+ jQ/2, accept the interval [a,yjcontaining r*; othawise accept [y,Bl*
Repeat the promo u& the desired accmacy is obtained.
A suitable convergence test is (since 0 < u < j3)
wherefclisa
whichtalceseffect choice for to1 is
pllAllF, where cc should be
110 NICHOLAS J. HIGHAM
then
c( 2) = B + z( rBI + nyz=
-Z[Z=Bz+(r~~+h)“s]Z=,
and it suffices to form and test for positive definitena the matrix fi + (rs1+
A)‘? where s = ZTBZ, and whe!rethe square root is trivial to compute.
An initial kterval contaiuing P is given by the bounds
where
M=max(O,-X*(B)).
The upper bound is from the proof of Lemma 3.5, and the second lower
bound follows from the fact that a psd matrix has nonnegative diagonal
elements. These bounds differ by no moxethan a factor 2, and they are exact
when A is symmtic (since C= 0). We mention that computation of
X&(B) could be avoided by using the altemative, but pote&ially much
weaker, upper bound jfA]l,.
3. B := Z=BZ.
4. Form an interval [a, j?] bracketing &s(A) using (4.2).
5. IfB+(as~+A)‘/BiS~SetB:=aand~~step7. ,_
6. While (#I- a)/2 > max{ fqtol) .i_
P:= (a + /3)/2,
G = B +($l + A>l/z,
ifGispsdthen~=dsea=r.
7. x := Z[B +{/!?sI -t A)qE?:
Thetestin~p5~n~~tIrecraseswh~thelowerboundiserractand
hence sectioning is unnv.
Algorithm nsr works entirely in reai arithmetic. In step 2 one can either
form C2 explicitly and apply a symmetric eigensolver, or ccmpute a real
Schur decomposition C=2DZT (preferably using a routine that takes ad-
vantage of skewmetry) and take A=Ds. The former approach is
numeri4y stable, since ]]C]]% = ]lC2]iP
A remainhg question is how to test the definiteness of G in the iteration
ofstep6.Thisisconsideredinthenextsecticn.
Theoem 5.1 immediably yields 3ne half of the quired stability result.
ForifCholesldfactorizationofG~down,thenGcaanotbeposi~~
definite with ~JG)c,tl < 1 as this would contradict part (a). Thus either G is
not positive definite or it is positive definite with K&)c,,u > 1. In the kter
case ‘G is within !&normdistance c,ullGl12of a singular, symmetric matrix,
and the answer “not positive definite” is therefore the correct one for a
matrix close to G.
Considering now the “po&ive definite” answer, we note that Theorem
5.1 apparently is not applicable to genesal symmetric matrices G. However,
~~naf~proafsin[l5,~l]~thatthe~e~result
(5.1) bkls without any assumption on the definiteness of G-it requiresonly
that the algorithm M to completion. It follows from (5.1) that if Choleski
factorization of a symmea+ matrix G succsds, then G is near to a positive
definite mats& n~~~1yIAT.
We con&de that the Choleski positive definiteness test is numerically
stable.
6. ACXNRATEDIRERhUNATION OF $(A)
If~A)is~ape~~to~y~~tfigures,thenitisdesIrable
to use a method that converges rnc~re
rapidly t&n Algorithm EST.To obtain
such B .-method we assume i&(A)> p(C) and we uss Theorem 3.1,
bmma 3.2, and Gxnlby a3 to reformulate the problem of computing
8~A)asthatoffindingthetmique~ofthefunction
= rs(r)“(rSH
+ c”) - %(‘), (62)
114 IWHOLAS J. HIGHAM
~RJSof Qgnificant figures will occur in the subtraction when re * p(C), and
SYMMETRICPOSl’IlVE SEMIDEFINITE
MATELX 115
may oeeur in the addition: for example when B is diagonal, so that [if G(r*)
is not positive definite] &&G(P)) = g,,(r*) = 0 (some i). The aceumey of
tile computti function and derivative values may be adversely deeted (see
Example 5 in the next se&ion). Unfortunately there does not seem to be any
way to avoid these losses of sign&ant figures~but at feast they are easily
detected.
7. NUMJBICiLEXAMPLES
TABLE1
Example ta Algorithm steps (see text)
sefpeme
10 (3,% 3,3,3,% 3,% 293)
5 (NNNNIq
9 (5,&S, 5,5,3,3,4,5)
10 (BBBBNNNNNN)
9 (4,4,3,4,4,4,39 493)
7 (NNNNNNN)
4 10 EST 10 (10, 1, 1, 10,1, 10,1,103il@,10)
RTSrn 5 (NN
5 5 EST 3 (4,494)
RTSAFE 22 (NNBBNB..
. NBBNB)
NIGHOLAS J. HIGHAM
0 0
I
01
10 0.
A= 0 1 01
Hahn= [12] states that i&(A) = i(l+&)‘fi = 0.8994. The computed value
of &(A) was correct to 15 digits. P = G(&dA)) has one zero e@annhe and
fwooioTder1,anditsupper~isgiventofaursigniflcantfignresby
‘!M55!a
P= I
0.4370
0.6871
0.5559
0.1689
0.7682
.1
For mmparison, &(A) = 1.2247and
1
0.1768 OSSMI 0.1768
Exums2. Aisaperturbationd&e5x5Hihrtmatrix((i+j-
1)~‘),ob~~by~~the(4,5)element~zero.~eslower~~~
with the initial SeqWU!eofbiSf&h~~~,~~by~2~sne9sof
$(A) = O.W32trJp(C) = O.Q625[see note (4) oh Section 0].
-1 -1 -1
To five figures the computed qA) is 1.27+. G&(A)) has one zero
eigeinYh.q t!he others being of order I. The closeness of &&A) to p(C) =
1.2071 does not unduly affect the convexgencq although we did observe that
many mare steps were _ when using a weaker initial bra&t t&an that
given by (4.2).
6. CONCLUSIONS
IntheF&eniusnorm,theuniquepositiveappr&untofA,andthe
distancei!~A),am~evhtedusingasinglespectmldezompcsition.
IIhos’s formula (3.1) for &A) poses some interesting computational
~~~Esr~~eflident~y~~~~~AXanda~~~
app~&~t,toIowaccu@y.Forhighaocuracycomnplltatiansthezero
fWingMmiqueoHection6wodcswelhmostcases,andwehavenot
encounW0dany~in~~~~pasitive~proximant
G&(A)) has mukple zero e@wahes. The usehhss of the formula (3.1)
~~accunrcy islimitedbyitspotentialhrlossofsig&-
cant~whichcanahttheatWnableaccuncyandslowtheconver=
geu~~ofaniterati~~.Foat?mately,suchlossofsignifi~seems
REmmucEs
118 NICHOLASJ. HIGHAM