~A"Ig2 54 CALIFORNIA L)NIV SANTA BARBARA INST FOR THE INTEROIS--ETC F/0 121
SEP 82 M *OLO(ER6. E 6 STRAVS AFOSR-79-0127
1(0
UCASSIFIE 8F018Tf62-OS.3 N
AFOSRTR. 8-2-0843
ON GENERALIZATIONS OF THE PERRON-FROBENIUS THEOREM
Moshe Goldberg* E.G. Straus -*
Department of Mathematics Department of Mathematics
Technion - Israel Institute University of California
of Technology Los Angeles, California 90024
Haifa 32000, Israel
and
Institute for the Interdisciplinary
Applications of Algebra
and Combinatorics
0University of California
Santa Barbara, California 93106
SELEC
DTIC
OCT 1 4 I1Mll
C"
L* Research sponsored in part by the Air Force Office of Scientific
Research, Air Force System Command, Grant AFOSR-79-0127.
La_. •* Research supported in part by NSF Grant HCS-79-03162.
Approved for Publlo releaso ;
distribution unlItited.
fool
0~J
ABSTRACT
The Perron-Frobenius Theorem states that a matrix with
nonnegative entries has at least one nonnegative eigenvalue
of maximal absolute value and a corresponding eigenvector
with nonnegative components., An- this paper-we discuss
generalizations of this celebrated theorem that locate an
eigenvalue of maximal absolute value and the components of
a corresponding eigenvector within a certain angle of the
complex plane depending on the angle which contains the
entries of the matrix. A complete description of the 2 x 2
case as well as partial results for the general case are given.
Accession For
NTIS GRA&I
DTIC TAB
Unannounced 0
Justificatio
BY-
Distribution/
Availability Codes A T M,
O7FCOW SCrrOTIC ZAW,4 A'
Avail and/or NDTICE OF TMR ITTAL TO DTIC
Diet ] Special This technical report has been review-d -id is
*1 approved for public releisse IAW AP% 13')-12.
Distribution isunlimited.
XA&IH J. J ?W
Chief, Tehnioal Information Division
i"I
:1.1
1. Introduction and Statement of Results.
The Perron-Frobenius Theorem states that a matrix with nonnegative
elements has at least one nonnegative eigenvalue of maximal absolute value.
It seems natural to ask for generalizations that locate an eigenvalue of
maximal absolute valud within a certain angle of the complex plane depending
on the angle which contains the elements of the matrix.
More precisely, let us consider the family An(a) of all n X n
complex matrices A = Cajk) whose entries are contained in a sector
(a) {z: jarg zj a; 0 6 a 9 n fixed).
For each A E An(a) let P= P(A) denote the minimal (nonnegative) angle
for which the sector S(P) contains an eigenvalue of A with maximal
absolute value. Thus, defining
p(a) = sup p(A),
A E A (a)
we pose the following
PROBLEM 1. Give Pn(a) as a function of a and n.
The Perron-Frobenius Theorem states that
Pn(0) = 0, n = 1,2,3.
Another simple observation is that for any m < n and A E A (c),
we have A 9 O. E An(a), where 0 is the k Xk zero matrix. Thus,
(1.1) n(a) is a nondecreasing function of n.
1.2
Obviously,
(1.2) (or) a;
so (1.1) and (1.2) yield
Pff(a) P 100v = or, n=12,
Since
An(a) Q Aa(W') for a' > a,
we also find that
Pn(a) is a nondec'reasing function of a.
In Section 2 we evaluate p2 (CO and obtain some partial results for
arbitrary n, as stated in the following theorem:
THEOREM 1.
A* (i) If n 2 then
af fora
fora>
2-2
(ii) If n k3 is odd then
P~()=nfora>
(iii) If k 4 is even then
iI
1.3
Pn a +n for 2n < a 9=22n- 2'
and
pn(a) = n for a> 2--2
Note the discontinuity of 02 (a) at a =/4.
We also note that part (i' of the theorem implies that in general, A
in A (a) does not have a square root in A (a/2).
Recall tha. the Perron-Frobenius Theorem not only proves that for
matrices with nonnegative entries an eigenvalue of largest absolute value
is positive, but also that a corresponding eigenvector can be chosen with
nonnegative components.
With this in mind, for each A E A (a) we denote y = y(A) to
be the minimal (nonnegative) angle such that the sector S(Y) contains
the components of an eigenvector corresponding to an eigenvalue of maximal
absolute value which lies in S(P(A)). Hence, defining
yn(a) = sup y(A),
A E (a)
we pose:
PROBLEM 2. Determine Yn(a) as R function of a and n.
The Perron-Frobenius Theorem tells us that
n(O)= 0, n = 1,2,3,...,
and we obviously have
Y1 (a) = 0.
|
I I.III I -. -. .. . -- - "
•I •i
i
I • "I• i i I
1.4
Also, as for n (a), it is not hard to see that
yn(a) is a nondecreasing function of n,
and
yn(a) is a nondecreasing function of a.
Finally, since any eigenvector (x1 ,...,xn)' E n (prime denoting
the transpose) can be rotated so that its components are embedded in the
sector S(n-n/n) it follows that
In analogy with the Perron-Frobenius Theorem it is perhaps natural
to ask whether
Yn(a) = on(a), n 2,3,4,...,
or whether there is any nontrivial bound for yn(a) when a > 0. This
turns out to be incorrect, at least for all n a 4, as stated in the
following theorem which is proven in Section 2:
THEOREM 2.
(i) If n = 2 then
a+ for O<aS
Y2(0)=
for a > .
(ii) If n 3 then
1.5
Y3 (a) k 2a + for 12
and
-
Y3 (a ) = 2 for a >
(iii) If n 2&4 then
(1.3) = n- n-1 for a> 0.
For n _2 note the discontinuity of yn(a) at a = 0.
We also note that in fact, for n 2, a k n/4; n = 3, a > 7t/12;
and n k 4, a > 0; yn(a) attains its maximal possible value.
Theorems 1 and 2 and the inequality in (1.3) can be summarized in
the following two tables:
n 2 2 2 3 3 3
a 0 < a S<a a of>g 0 <a 1S < a S>
2 2> O ~~12 12 6
n(a) a a++ ir aa X
22a
yn(a) a k 2a +
Table 1. Cases n = 2,3
I .l l l , ,'l , 1l ' 4 IB-ll I ' II J l
I
1.6
n n 4 even n k 4 even n k 4 even n k 5 odd n k 5 odd
0~ ~ < < n a >
2n 2n 2n-2 2n-2 2n 2n
_n_ a _
n- n-1 n-1 i n-1
"n- - n n n n n
Table 2. The case n k 4
We have not settled the following questions:
QUESTION 1. If 0 5 a S n/2n and n k 3, is n(a)=a?
QUESTION 2. What is pn(a) if a > 4 is even and n/2n < a S n/(2n-2)?
QUESTION 3. If 0 < a S n/12, what is
i.
a-p
2.1
2. Proofs.
Prco of Theorem 1. Since for all n and a we have
a On (a) 5,
and since Pn(a) is a nondecreasing function of a, it remains to
obtain the upper bounds
(2.1) 2 (a) a, .n
(2.2) P2() ;5 a + <a
and the lower bounds
(2.3) P (a) + n-a
+ 7 n2> 2 even, -<a 2n-
(2.4) = n, n - 3 odd, a > 2n'
pn(a) k n
(2.5) n a)it, n >4 even, a > 7-
n•- 2n-2
We start with the upper bounds.
Since multiplication of a matrix by a constant multiplies the
eigenvalues by the same constant, we may assume that the spectral radius
of A, i.e., the maximal absolute value of the eigenvalues, is 1.
Moreover, since the eigenvalues of the conjugate matrix A - (aij) are
the complex conjugates of the eigenvalues of A, we can assume that
! eip(A)
= is an eigenvalue of A.
- -- .,|= ,. - . ,
II
2.2
With these assumptions we first prove that
(2.6) P(A) < a + ic/2 for A E A2 (oe), 0 <a -<n/2.
Let A = (ajk) E 42 (a) have eigenvalues X = eip(A) and '. Since
X +X'= tr A = a1 a22 E k(CO
and since
K'EU
where U is the unit disc centered at the origin, then
X. E u n (-x + sa)).
Thus,
(2.7) + X' E (X + U) fl () - T
so that T is the intersection of the unit disc centered at A and-
the sector S(a). Therefore, if
p(A) k a + R/2,
then T = (0), and by (2.7), X' = -X. Hence, I'!with
= 1
larg X'I < arg X = P(A) in contradiction to the definition of P(A).
L Thus we have (2.6), and (2.2) follows.
Now, fix a with
0 S a 5 t/4,
and let us prove (2.1).
2.3
Suppose that for some A = (ajk) E A2 (a) with eigenvalues X = eip(A)
and ', Ix'I < 1, we have
(2.8) a< P(A) < a +n/2.
Since P(A) < a + n/2, then T t {0). Further, since P(A) > a,
then if max. arg a.. < a, there exists a positive number r so that
max arg(a.. = a.
j=1,2
We can thus replace A by the matrix
A I- (A + rAI) E A2 (a)
whose eigenvalues are X and X" = (X' + rk)/(l + r). Hence, IA"I < 1,
__ I
unless ' = A = 2 tr A E S(A) which contradicts the assumption that
P(A) > a. So, we may use A0 instead of A, drop the subscript, and
assume without loss of generality that
(2.9) arg a 1 a[.
By (2.7) we have
(2.10) a +a 22 = tr A E T;
so
a2 2 E -all + T.
Also, by (2.9), (2.10), and since a22 E S(a), we have all E T; so if
we denote by ail the point symmetric to all in T, then (see Figure 1)
- 7 -- -
I I III
2.4
= 2e cos(P - a) - all E -a11 + T.
We may not have
a2 2 a11.
For in that case
A' = tr A - 1 1 + all
a
hence Mx'I = 1 and by (2.8), larg 'I <Iarg X( which contradicts the
definition of P. Therefore,
a2 2 11
so
arg(A - a22 ) < arg(A - a
and
(2.11) arg(X - a.1)(X - a22 ) arg(X - a11 ) + arg(A - a22 )
< arg(A - a) + arg(A - a'l) (0 + a) + (n - 0 + a) = n + 2a.
Since A E A (a) we also have
(2.12) arg(X - a11 )( - a22 ) = arg(X - a11 ) + arg(X - a22 ) Z 2p > 2a,
and
(2.13) -2a arg a12 a21 2a.
But now, since 0 S a 6 x/4, equations (2.11)-(2.13) contradict
the characteristic equation of A,
(A- -2.5
al 8a22) =a,2a2 1 *
Thus, assumption (2.8) fails; so in view of (2.6) we must have
and (2.1) follows.
Figure 1.
2.6
Figure1
2.7
The main tools in obtaining the lower bounds in (2.3)-(2.5) are the
n x n circulants
a0 a1 . . an- 2 an-1
(2.14) A an-I a0 an-3 an-2
a1 a2 an- 1 a0
whose eigenvalues are
2 + + A 1
a + aw + aw
j ~ 0 2 + a U- j j n,
j
where
2
w. =e nij/n I ...,nU
are the n-th roots of unity.
Assume now that
oa> x/2n,
and set the a. in (2.14) to zero except for the entries
J
+" ' ~ia iann
a= e , aM+ 1 = re
r > O, 0 < e < nl2n, 0 S am n-2,
where c is chosen sufficiently small so that A E Aa(a) and m will be
determined later. The eigenvalues are
,
(a+
a= w a )w j = I..,,
and we see that their maximal absolute value is obtained for the j that
minimize$
_ .
2.8
+
Iarg a* " arg(w."a+ )I = larg a - arg - arg a 1 - 2n
Since 0 < e < n/2n, this occurs at j = 1; thus,
IA1 1 > IXl , j -
and consequently
P(A) = Iarg A1i .
Moreover, since the argument of am + wla,+ goes from argam=a to
arg wIam+1 = a + n/n - 8 as r goes from zero to infinity, then
arg = arg(a, + wla.+l)wl = arg(a* + wla,~l) + 2nm/n
runs through the interval
.. I
(2.15) i=( + ,a +i - .
Now, if n is odd and
n/2n < a < n/n,
we pick
*l = n-i
2 ;
so the interval I in (2.15) contains x. Hence, there exists a circulant
A E A,(a) such that
O(A) =arg l=
I
' '
i i ... ..- ,
2.9
Therefore,
Pn(a) n, n 3 odd, !- < a<y -
and by the monotonicity of Pn(a) in (1.4) we get (2.4).
If n is even and
X/2n < a S n/(2n-2),
we choose
n-2
2
so that the interval I
-U contains the point a + n - n/n - 2. Thus,
we can find a circulant A E A (a) with
P(A) = jarg A 11 = a + n - f/n - 28;
and since 9 was arbitrarily small, we have
Pn(a) k a + n k2 even, <
so (2.3) is established.
Finally, if n is even and
a > n/(2n-2),
then
a > R/21
where A n-i is odd. Thus, by (2.4),
2.10
~I
and since Pn(a) is a nondecreasing function of n, we obtain (2.5). 03
Proof of Theorem 2. Since for all n and a we have
yn(a ) - n
n_1
and since yn(a) is a nondecreasing function of a, it suffices to obtain
the upper bound
(2.16) 0 < a ,
and the lower bounds
l (2.17) Y2( ; + 4 0 < a W
[
(2.18) ¥30 a+ ,0 < a< -
(2.19) y(a) n n. n, n a 4, a>0.
Starting with the upper bound in (2.16), we let
a ;5 n/4,
and take any A = (ajk) E A2 (a). As in the previous proof, we may
assume without loss of generality that the eigenvalues of A are A
and A' with
J-I INNNl'
2.11
I .1 1.
e
S
Now let (x1 ,x2 )' be an eigenvector corresponding to X. If one of
the components vanishes, the other may be taken to be 1; so in this case
y(A) = 0. If the eigi'nvector does not have a zero coordinate we may
assume that it is of the form (1,x)' where x 0,
0 thus
(2.20) y(A) = llarg xl.
We have
a11 + a1 2x = ,
a21 +a = AX,
and therefore
A - al a2
A-a 21
(2.21) x11
a12 -a 22
Since a 5 n/4, then by part (i) of Theorem 1 we have P(A) 6 a
so the interval connecting the origin and the point A lies in s(a),
and the line r through A perpendicular to the interval [0,K)
(see Figure 2) must intersect the positive real axis. Now, if al1
and a22 are located on the right of r then
1al1+ a221 > 2 1A = 2,
in contradiction to the fact that
I11 + a22 x IS + IA'! S 2.
Thus, say, a11 is on the (closed) left side of r, and
2.12
Iarg(X-a 1 1 )I 9 larg XI + =
So finally, by (2.20) and (2.21), and since 5 a,
t we have
y(A) = glarg xl =farg
2 a 12 :
! Iarg(X-a 1 )I + !Iarg a1 2 1 (p + + C) _a +
and (2.16) follows.
Fiur
I@
2.13
Figure2
2.14
For the lower bounds, our tools are n x n lower triangular
matrices of the form
x
a At 0
A a A' ,I l < IAl
a At
where evidently, the eigenvector belonging to A is given, up to an
arbitrary factor, by
(2.22) (1,x, x2, ... , xn'l) ' with x = a
With these matrices, we now make suitable choices for the quantities
A, A' and a for different values of n.
First, for n = 2 and
0 < a S n/4,
set
= le, ' - i (a+n / 2 )
a=
where E > 0 is small and then 6 > 0 is chosen so small that A'j < I
and A E An(a). The vector in (2.22) is now
(I,x)', -i(2Gafl/2- ).
thus
0_ 1-
22.15
y(A) = jargxlI a+i-
and since £ is arbitrarily small we obtain (2.17).
For n =3 and
0 < a n/l2,
we make the same choice of X, V' and a as for n = 2, so
the eigenvector in (2.22) is
,x 2) ' , x
X(1 e-i(2a+n/2-6)
Therefore,
y(A) = Iarg x2 l = 2a + n/2 - C,
and again, since c is arbitrarily small, we obtain (2.18).
For n = 4 and a > 0 we set
A = e , A' = cos 9, a = sin E, 0 < C < a.
The corresponding eigenvector in (2.22) is (1, i, -1, -i)'; so
y(A) n
and (2.19) holds for n = 4.
a Finally, if n > 4 and a > 0, our choice is
, - > -2ni/
= 1, issufc 1 , a l,thO, w1 _E f/ n,
whr gin f£is sufficiently small thenAEA()an tKI(1
where
if again, A E~-n~a n [<
!A
2.16
The eigenvector in (2.22) is now
f
2n-I)
(I, W, Wit ... P W1
so
y(A)=
n
and the proof of (2.20) is complete. o
Note that the lower bounds in (2.17)-(2.19) were obtained for matrices
in A (a) whose eigenvalues are all contained in S(a). In fact, for
n k 4 we found matrices A such that y(A) obtained the maximal possible
value, while all the eigenvalues were arbitrarily close to 1.
Il2
UNMTARRTFTM)f
SECURITY CLASSIFICATION OF THIS PAGE ("on.. Dims EnforedO
1. RPORTNUMER V ACESSIN .3-RPIERMINGAORG.RP NUMER
7. AuTNOR(s) 9. CONTRACT OR GRANT NLJMBER(a)
Moshe Goldberg and E.G. Straus AFOSR-79-0127
9. PERFORMING ORGANIZATION NAME AND ADDRESS SO. PROGRAM ELEMENT. PROJECT, TASK
Institute for the Interdisciplinary Applications AREA & WORK UNIT NUMBERS
of Algebra & Cojubinatorics, University of PE611O2F; 2304/A3
California, Santa Barbara CA 93106
It. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE
Directorate of Mathematical & Information Sciences September 1982
Air Force Office of Scientific Research 13. NUMBER OF PAGES
Balling AFE DC 20332 24
14. MONITORING AGENCY NAME A ADORESS(Ii different bo Controlling Office) IS. SECURITY CLASS. (of tis report)
-UiCLASSIFIED
1SA. DECL ASSI F9CATI ON/ DOWN GRAOIN G
SCHEDULE
1S. DISTRIBUTION STATEMENT (of tis,Repor)
Approved for public release; distribution unlimited.
17. DISTRIBUTION ST. 4ENT (of ~ abstrt entered in Block 2.it differentI from Report)
1S. SUPPLEMENTARY TEs
IS. KEY WORDS (Cotnaa on reverse side It necessary and Identify by Weoek numnber)
Eigenvalues; eigenvectors.
* 20. ABSTRACT (Contlw nueverso~e aid. It ncsary and idetify'by Moc~k fnsw)
The Perron-Frobenius Theorem states that a matrix with nonnegative entries has
at least one nonnegative eigenvalue of maximal absolute value and a correspond-
ing eigenvector with nonnegative components. In this paper the authors discuss
generalizations of this celebrated theorem that locate an eigenvalue of maximal
absolute value and the components of a corresponding elgenvector within a cer-
tain angle of the complex plane depending on the angle which contains the
entries of the matrix. A complete description of the 2x2 case as well as par-
tial results for the general case are iven.
DD I, 01N, 147 UNCLASSIFIED
S ECURITY CLASSIFICATION OF THIS PAGE (119amData En'tered)
82 10 12 142
a-
