Abstract. The eigenvectors and eigenvalues of symmetric block circulant matrices had been found, and that method is extended to general block circulant
matrices. That analysis is applied to Stephen J. Watsons alternating circulant
matrices, which reduce to block circulant matrices with square submatrices of
order 2.
1. Circulant Matrix
A square matrix in which each row (after the first) has the elements of the
previous row shifted cyclically one place right, is called a circulant matrix. Philip
R. Davis (1979, p.69) denotes it as
..
. bn2 bn1
b
b
b
0
1
2
..
bn1 b0
.
b1
bn3 bn2
..
.
b
b
b
b
b
def
n2
n1
0
n4
n3
B = circ(b0 , b1 , , bn1 ) =
. (1)
.
.
.
.
.
.
..
..
..
..
..
..
..
b
.
b3
b4
b0
b1
2
..
. bn1 b0
b1
b2
b3
Thomas Muir (1911, Volume 2, Chapter 14) had denoted its determinant as
C(b0 , b1 , , bn1 ).
1.1. Complex nth roots of 1. Denote
def
2
.
n
(2)
(j = 0, 1, 2, , n 1).
(3)
(4)
(5)
196
GARRY J. TEE
and
cos(n j)f = cos 2(n j)f /n
= cos(2f 2f j/n)
= cos 2f j/n = cos jf .
(6)
(7)
j = b0 + b1 j + b2 2j + b3 3j + + bn1 n1
(8)
w(j) = j
,
j
..
.
n1
j
for j = 0, 1, 2, , n 1.
1.2.1. Duplicated eigenvalues. From this, it follows that the circulant matrix B is
real and has eigenvalues j , if and only if [Davis (1979), p.80, Problem 15]1
j = nj ,
(j = 1, 2, , n 1).
(9)
This does not relate to 0 (which is real for real B), and for even n = 2h it implies
that h is real.
If B is real symmetric then its eigenvalues are real, and (9) reduces to
j = nj ,
(j = 1, 2, , n 1).
(10)
197
mentioned only briefly by Philip R. Davis (1979, p.67, Problem 1 & p.81, Problem
16), and the following result seems to be new.
Theorem 1: Every complex symmetric circulant matrix of order n has a single
eigenvalue with odd multiplicity if n is odd, but it has either two eigenvalues or
none with odd multiplicity if n is even. All other eigenvalues occur with even
multiplicity.
Proof: For j = 0, 1, , n 1, the formula (8) gives the eigenvalue
j = b0 + b1 j + b2 2j + + bn2 n2
+ bn1 n1
j
j
1
= b0 + b1 j + b2 2j + + b2 2
j + b1 j
h
i
h
i
= b0 + b1 j + j + b2 2j + 2j +
h
i 0
if n = 2h 1
h1
h1
+bh1 j + j
+
bh hj if n = 2h .
(11)
h1
X
bf cos jf +
f =1
0
bh (1)j
if n = 2h 1
if n = 2h .
(12)
Replacing j by n j, we get
nj = b0 + 2
h1
X
bf cos(n j)f +
f =1
= b0 + 2
h1
X
f =1
bf cos jf +
0
bh (1)2hj
0
bh (1)j
if n = 2h 1
if n = 2h
if n = 2h 1
if n = 2h
= j ,
(13)
in view of (6).
Thus, with h = (n + 1) 2, the sequence of eigenvalues 1 , 2 , , h2 , h1 is
duplicated as the sequence n1 , n2 , , nh+2 , nh+1 . Every distinct eigenvalue in the first sequence occurs an even number of times within that pair of
sequences.
With odd n = 2h 1, B has the eigenvalue
0 = b0 + 2(b1 + b2 + + bh1 )
(14)
which is a simple eigenvalue with odd multiplicity 1; unless 0 equals any j for
1 j h 1, in which case 0 occurs with odd multiplicity greater than or equal
to 3.
With even n = 2h, B has the eigenvalues
0 = b0 + 2(b1 + b2 + + bh1 ) + bh ,
(15)
If 0 = h then they have even multiplicity, whether or not they equal any j for
1 j h 1. Otherwise 0 either has multiplicity 1 or else it equals some j for
1 j h 1, in which case 0 occurs with odd multiplicity greater than or equal
to 3. And similarly for h .
198
GARRY J. TEE
2 v
w =
(16)
,
..
.
n1 v
where v is a non-null -vector and is any nth root (3) of 1.
2The Forder Collection in Auckland University Library has a copy of that book by Muir & Metzler,
with a letter from Metzler to Henry George Forder (1930 June 5) pasted into it. Metzler explained
that, in revising and enlarging Muirs Treatise, My object was to give the mathematicians of the
world a chance to benefit by more than thirty years of labor in this particular field without having
to pay twice as much for the book if put up by a regular publisher. I am sorry that there are so
many misprints but more than 50% occurred after the last proof went back.
That copy includes Metzlers Errata sheet listing about 150 misprints: but the book contains very
many further misprints.
199
(17)
= v,
n1
)v
= v,
n1
)v
= 2 v,
(bn1 + b0 + b1 + b2 + + bn2
(bn2 + bn1 + b0 + b1 + + bn3
(18)
and so forth.
Each of the n compound rows in (18) reduces to the first of them, and that can
be rewritten as the eigenvector equation
Hv = v,
(19)
(20)
For vectors of the form (16) the set (20) of n eigenvector equations of order , for
the n values of , is equivalent to the single eigenvector equation (17), of order n.
For each of the n values of , the equation (19) can be solved to give eigenvectors
v (or fewer if the eigenvectors of H happen to be defective), each with its eigenvalue
. Each eigenvector v of H, with its corresponding , gives an eigenvector w of
the block circulant matrix B, with that eigenvalue .
For any , if H has m linearly independent eigenvectors v(1) , v(2) , , v(m) ,
then it follows from the structure of (16) that the corresponding eigenvectors
w(1) , w(2) , , w(m) of B are linearly independent. Conversely, if B has linearly independent eigenvectors w(1) , w(2) , , w(m) of the form (16), each with the same ,
then the first elements of those compound vectors are a set of linearly independent
eigenvectors of H (for that ).
2.2. Orthogonal eigenvectors. Consider any
pound vectors
j s
js ,
x =
y =
..
.
jn1 s
The scalar product of x and y is
t
k t
2k t
..
.
(21)
n1
t
k
x y = s t + s (j k )t + s (j k )2 t + + s (j k )n1 t
= 1 + (j k ) + (j k )2 + + (j k )n1 (s t)
= 1 + u + u2 + + un1 (s t),
(22)
200
GARRY J. TEE
For each , H has at least 1 eigenvector, and hence the eigenvectors of B span
at least n dimensions.
2.3. Hermitian block circulant matrices. The block circulant matrix B in (17)
is Hermitian if and only if bj = bnj for j = 1, 2, , n 2 and b0 is Hermitian.
Note that, if n = 2h then bh = bh , so that both b0 and bh must be Hermitian.
The matrix H then (20) reduces to
H = b0 + b1 + b2 2 + + bn2 n2 + bn1 n1
= b0 + b1 + b2 2 + + b2 2 + b1 1
h
i h
i
= b0 + b1 + b1 + b2 2 + b2 2 +
h
i 0
if n = 2h 1
h1
h1
+ bh1
+ bh1
+
bh h if n = 2h .
(23)
h1
X
f =1
bf fj
bf fj
0
bh (1)j
if n = 2h 1
if n = 2h .
(24)
2.3.1. Real symmetric block circulant matrices. Equation (24) for Hermitian Hj
now reduces to
Hj = b0 +
h1
X
f =1
0
bf j f + bTf fj +
bh (1)j
if n = 2h 1
if n = 2h ,
(25)
and so H0 and (for even n = 2h) Hh are real-symmetric, with real orthogonal
eigenvectors with real eigenvalues. But the other Hj are not real-symmetric in
general, and so their eigenvectors are complex in general. But each Hj is Hermitian
with all eigenvalues real. And the eigenvalues of B are the union of the n sets of
eigenvalues for all the Hj .
201
= b0 +
h1
X
bf fj
bTf j f
f =1
= Hj
0
bh (1)j
if n = 2h 1
if n = 2h
(26)
Hnj v = v
(j = 1, 2, , (n 1) 2).
(27)
Thus, Hnj also has eigenvalue , with eigenvector v. The entire set of eigenvalues
of H1 , H2 , , H(n1)2 is duplicated by the relation (27).
But that duplication does not apply to H0 , nor (for even n = 2h) to Hh .
Let be any of the eigenvalues of H0 , or of the 2 eigenvalues of H0 and
of Hh when n = 2h. If the sum of the multiplicities of in those 2 submatrices
is odd (e.g. if is a simple eigenvalue of H0 but not an eigenvalue of Hh ), then
B has as an eigenvalue of odd multiplicity, whether or not = j for some
j = 1, 2, , (n1)2. But if the sum of the multiplicities of in those submatrices
is even, then B has as an eigenvalue of even multiplicity, whether or not = j
for some j = 1, 2, , (n 1) 2.
Thus, we have proved the following generalization of Theorem 1:
Theorem 2: Every real symmetric block circulant matrix of type BC n, has no
more than 2 eigenvalues with odd multiplicity if n is even, but it has no more
than eigenvalues with odd multiplicity if n is odd. All other eigenvalues occur
with even multiplicity.
Non-extension from real-symmetric to Hermitian case. Does the relation (27) generalize from block circulant matrices which are real-symmetric to Hermitian block
circulant matrices?
No, for a counter-example is given by Hermitian B BC 4,2 with
0
2+i
0 i
b0 =
, b1 =
, b2 = 0, b3 = b1 .
(28)
2i
0
0 0
Since 1 = i,
H1 = b0 + i b1 b2 i b3 =
0
2 + 2i
2 2i
0
0 2
H3 = b0 i b1 b2 + i b3 =
,
2 0
(29)
(30)
202
GARRY J. TEE
f
f
T f
f
(31)
bf j + bf j = bf j + j = 2bf cos f j,
which is a real symmetric submatrix. Accordingly, the expression (25) simplifies to
h1
X
0
if n = 2h 1
(32)
Hj = b0 + 2
bf cos f j +
bh (1)j if n = 2h ,
f =1
v
v
j v
j v
2v
2
,
j
j
w =
w =
(34)
,
..
..
.
.
jn1 v
n1
v
j
which are orthogonal by (22). Those may be replaced by their linear combinations
(w + w)/2 and i(w w)/2:
v
0
cos j v
sin j v
cos
2j
v
sin
2j v
<(w) =
=(w) =
(35)
,
,
..
..
.
.
cos(n 1)j v
sin(n 1)j v
which are an orthogonal pair of real eigenvectors of B for the double real eigenvalue
.
203
J1
J2
def
J3
L = dJ1 , J2 , J3 , , Jm c =
(1 m ). (38)
.
.
.
Jm
Each Jordan block is either of the form [q ], where the corresponding column of U
is an eigenvector with eigenvalue q ; or else it is of the form Jq = q I + R, where
R and the unit matrix I are of the same order as Jq , and
0 1
0 1
0 1
R =
(39)
.
.
.
.. ..
0 1
0
In this case, the first of the corresponding sequence of columns of U is an eigenvector
with eigenvalue q , and the other columns of U in that sequence are generalized
eigenvectors with eigenvalue q . L is unique, apart from permutation of the Jordan
blocks. The eigenvectors (and generalized eigenvectors) in U are not uniquely
specified any eigenvector (and its associated generalized eigenvectors) can be
scaled by any nonzero factor; and if H has multiple eigenvalue with linearly
204
GARRY J. TEE
independent eigenvectors v(1) , v(2) , , v(g) , and if z(1) , z(2) , , z(g) are a linearly
independent set of linear combinations of v(1) , v(2) , , v(g) , then those g vs could
be replaced by the g zs.
H has defective eigenvectors, if and only if L has at least one Jordan block of
order greater than 1. For any eigenvalue q , its eigenvectors span space whose
dimension equals the number of Jordan blocks with q . The eigenvalue q has
multiplicity equal to the number of times that it occurs on the diagonal of L, which
is the sum of the orders of the Jordan blocks with q .
Consider the compound vector T, whose n elements are square matrices of order
(cf. (16)):
U
T =
(40)
.
..
.
n1 U
The columns of U are linearly independent (since U is non-singular), and hence
the columns of T are linearly independent.
The first few compound rows of BT are:
(b0 + b1 + b2 2 + b3 3 + + bn1 n1 )U =
2
(bn1 + b0 + b1 + b2 + + bn2
n1
)U =
HU = UL,
HU = (U)L,
2 HU = (2 U)L,
(41)
(42)
(43)
def
[T0 T1 . . . Tn1 ]
def
dL0 , L1 , L2 , , Ln1 c =
(44)
L0
L1
L2
..
(45)
Ln1
Within each Tk the columns are linearly independent, and (22) each column of Tk
is orthogonal to every column of every Tj , for j 6= k. Therefore the columns of W
are linearly independent, and so W is non-singular.
For k = 0, 1, , n 1, the eigenvector equations (43) can be represented as
BW = W.
(46)
205
The matrix is a direct sum of Jordan blocks (of order ), and W is nonsingular. Hence (46) gives the Jordan Canonical form of B.
= W1 BW.
(47)
bj = jI R (j = 1, 2, , n 1),
(48)
where R is a Jordan block of order > 1, with a single eigenvector v and all
eigenvalues are zero:
0 1
1
0
0
1
0
0 1
R =
,
v
=
(49)
.. .
..
..
.
.
.
0
0 1
0
0
For each ,
H =
=
b0 + b1 + b2 2 + b3 3 + + bn1 n1
I( + 22 + + (n 1)n1 ) R( + 2 + + n1 ).
(50)
+ 22
2
+ 2
+ 33
23
+
3
+ + (n 1)n1
(n 2)n1
+ +
n1
(n 1)n
(n 1)n .
(51)
(52)
Thus, for j > 0, Hj is a Jordan block with the single eigenvector v as in (49), with
all eigenvalues equal to n/(j 1). Indeed, since Hj is its own Jordan Canonical
Form Lj with Hj I = ILj , the unit matrix I is an orthogonal matrix of the single
eigenvector and the 1 generalized eigenvectors of Hj .
With 0 = 1, we get H0 = I(n(n 1)/2) R(n 1), so that H0 has the single
eigenvector v as in (49), with all eigenvalues equal to n(n 1)/2.
Thus, the matrix B (as in (48)) has an orthogonal set of n eigenvectors, each
with an eigenvalue of multiplicity .
206
GARRY J. TEE
1
which is the Jordan Canonical Form of H0 . Thus H0 D = D 2 n(n 1)I + R ,
with D as an orthogonal matrix of the single eigenvector and the 1 generalized
eigenvectors of H0 .
Therefore, we have found the complete Jordan Canonical Form (45) =
dL0 , L1 , , Ln1 c of B, with
n
L0 = 21 n(n 1)I + R,
Lj =
I + R (j = 1, 2, , n 1),
(54)
j 1
and the complete orthogonal matrix W (44) of n eigenvectors, each with 1
associated generalized eigenvectors, where (cf. (40))
D
I
I ...
I
D
1 I
2 I . . . n1 I
2
D
I
22 I . . . 2n1 I
1
W =
(55)
.
..
..
..
..
..
.
.
.
.
.
D 1n1 I n1
I ...
2
n1
n1 I
This class (48) of matrices of type BC n, could be used as a test for procedures
for computing eigenvectors and eigenvalues of matrices with defective eigenvectors.
As with any test matrix, it could also be used with any symmetric permutation of
the rows and columns.
3. Alternating Circulant Matrices
Stephen J. Watson considered a matrix which he calls an alternating circulant
matrix. Specifically, consider a 2n by 2n matrix where each row is obtained from
the preceding row by the simple cyclic permutation (1 2 3 . . . 2n) followed by
multiplication by 1 [Watson]. He reported that he has discovered a way to
characterize the spectrum of such matrices, and he wondered whether that was a
known fact.
Watsons alternating circulant matrix has rows 1 and 2 of the form:
c0
c1
c2
c3
c2n2
c2n1
.
(56)
c2n1 c0 c1 c2 c2n3 c2n2
That represents a matrix B of type BC n,2 where the square submatrices are
c0
c1
c2j c2j+1
b0 =
, bj =
, (j = 1, 2, , n 1). (57)
c2n1 c0
c2j1 c2j
In this case, for each of the n values of , the matrix H is
d
e
H =
,
e d
(58)
where
d
e
= c0 + c2 + c4 2 + c6 3 + + c2n2 n1 ,
= c1 + c3 + c5 2 + c7 3 + + c2n1 n1 .
(59)
207
and 01
could be used as an orthogonal basis for the eigenspace of that null
matrix.
Otherwise, for each of the n values of , the two eigenvalues of H are
p
1 =
d2 e2 , 2 = 1 ,
(60)
with eigenvectors
v(1) =
e
d + 1
v(2) =
e
d 1
(61)
e
v =
,
(62)
d
so that the eigenspace of H has dimension 1 rather than 2.
For each and v we construct (16) an eigenvector w of B, with corresponding
eigenvalue . Thus, the eigenvectors and eigenvalues of Watsons alternating circulant matrix have been found. The eigenvectors span 2n dimensions, except that
for each value of such that d2 = e2 6= 0 the eigenspace loses 1 dimension.
3.1. Bound for eigenvalues. For any square matrix A and any norm kAk, each
eigenvalue has modulus || kAk. The spectral radius (A) is defined as the
maximum modulus of any eigenvalue, and so 0 (A) kAk.
For the alternating circulant matrix B in (56), using the rowsum norm we get
that
2n1
X
(B) kBkrow =
|ck | .
(63)
k=0
2
2
2
2
2
|| = d2 e2 |d| + || |e| = |d| + |e| .
From (59),
n1
n1
n1
X
X
X
j
|d| =
c2j j
|c2j | || =
|c2j | ,
j=0
j=0
j=0
(64)
(65)
and similarly
|e|
n1
X
|c2j+1 | .
(66)
j=0
||
|d| + |e|
n1
X
j=0
|c2j | +
n1
X
j=0
2
|c2j+1 | .
(67)
208
GARRY J. TEE
(68)
j=0
3.1.1. Sharp bound for spectral radius. The inequality (68) may be rewritten as:
2
2
n1
n1
X
X
2
(B)
|c2j | +
|c2j+1 |
j=0
j=0
n1
X
|c2j | +
j=0
2
= kBkrow 2
n1
X
|c2j+1 | 2
n1
X
j=0
|c2j |
j=0
j=0
n1
X
|c2j |
n1
X
n1
X
|c2j+1 |
j=0
|c2h1 | .
(69)
h=0
This gives an upper bound for (B) less than or equal to (63), with those bounds
being equal if and only if ck = 0 either for all even k, or for all odd k.
Indeed, we shall shew that (68) gives the smallest possible bound for (B), in
terms of the moduli |ck |.
For alternating circulant matrix B with first row [c0 , c1 , . . . , c2n1 ], consider the
complex alternating circulant matrix B0 with first row [c00 , c01 , . . . , c02n1 ], where
c0k = |ck | 0 for even k and c0k = i |ck | for odd k. Hence, each element of B0 has
the same modulus as the corresponding element of B.
P 0
P
|ck | =
|ck |
c02n2 =
c04 + +P
For B0 with = 1, we get that d = c00 + c02 +
P
|c0k | = i
|ck | for odd k. Hence,
for even k, and e = c01 + c03 + + c02n1 = i
for the real eigenvalues with = 1,
2
2
n1
n1
X
X
c02j +
c02j+1 ,
2 = d2 e2 = d2 e2 =
(70)
j=0
j=0
and so the inequality (67) for eigenvalues reduces to an equality when = 1 for
such B0 . Hence, the inequality (68) reduces to an equality for such B0 .
Therefore the inequality (68) is the sharpest bound for (B), in terms of the
moduli of the elements of B.
3.1.2. Real alternating circulant matrices.
If B is real then with = 1 it fol
lows from (59) and (58) that H = ed de
is real, and hence its eigenvalues
d2 e2 are either both real or both imaginary. Therefore, every real alternating
circulant matrix B has at least one pair of eigenvalues which are either both
real or both imaginary (or both zero), with eigenvectors (or eigenvector) of the form
(16) with = 1. If d = e 6= 0 then the double zero eigenvalue
has a single real
1
1
eigenvector (16), with (cf. (62)) v= 1
if d = e, but v= 1
if d = e.
If B
is real with even n, then with = 1 it follows from (59) and (58) that
209
circulant matrix B with even n has at least one pair of real orthogonal eigenvectors
(16) with = 1, and with real eigenvalues .
3.1.3. Sharp bound for spectral radius with even n. For any complex alternating
circulant matrix B with first row [c0 , c1 , . . . , c2n1 ] with even n, consider the real
alternating circulant matrix B0 with first row [c00 , c01 , . . . , c02n1 ]. Here
Similarly, with = 1 the inequality in (66) then reduces to equality, since for
each j, c2j1 (1)j 0; and hence
e =
n1
X
c2j+1 .
(73)
j=0
Thus, if the coefficients of B0 satisfy (71), then the inequalities in (67) and (68)
reduce to equalities. In that case, the pair of real eigenvalues for = 1 each
attain the bound (67):
2 = d2 e2 = d2 + e2
= (c00 c02 + c04 c07 + + c02n4 c02n2 )2
+ (c01 c03 + c05 c07 + + c02n3 c02n1 )2
2
2
n1
n1
X
X
c02j +
c02j+1 .
=
j=0
(74)
j=0
It was shewn in (70) that, for any n, the eigenvalue bound (67) can also be
attained with = 1 by complex B0 .
3.2. Alternating circulant matrix with defective eigenvectors. As a real
example with defective eigenvectors, consider
1
0 3
2
2 1
0
3
B =
(75)
3
2
1
0
0
3 2 1
with n = 2, so that the values of are 1 and 1. For = 1 the pair of eigenvalues
must be both real or both imaginary or both zero: in this case (78) they prove to
be both zero.
For this real alternating circulant matrix B, n = 2 is even and all inequalities
(71) do hold, since 1 = c0 0 c2 = 3, and c1 = 0 0 2 = c3 . Accordingly,
210
GARRY J. TEE
,
with
orthogonal eigenvectors
giving real symmetric H = 24 2
4
v(1) =
2
4 + 20
v(2) =
2
4 20
(76)
2
2
4 + 20
, w(2) = 4 20
w(1) =
(77)
2
2
4 20
4 + 20
2 2
, with the single eigenvector
2 2
(3)
2
2
(78)
0 1
2 f
2 f
2 2
,
=
0 0
2 g
2 g
2 2
(79)
where f and g are the elements of the generalized eigenvector with double zero
eigenvalue. This matrix equation is equivalent to 2f + 2g = 2. Hence, f is an
arbitrary parameter and g = 1 + f .
Therefore, the Jordan canonical form BW = W (cf. (46)) has the matrix of
3 eigenvectors and 1 generalized eigenvector (with arbitrary parameter f )
f
2
2 2
20 4 20 4 2 1 + f
,
W =
(80)
2
2 2
f
20 + 4
20 + 4 2 1 + f
and the eigenvalues as the diagonal elements of
20
20
=
.
1
0
(81)
20
4
W =
20 + 4
2
20 4
2
20 + 4
2
2
2
2
1/2
1/2
,
1/2
1/2
211
(82)
then column 4 is also orthogonal to column 3, and hence the entire matrix (82) is
then orthogonal.
Acknowledgments
I thank David Brillinger, who read a draft version (in 2002) and informed me of
Bernard Friedmans paper (Friedman, 1961).
References
[1] Chong-Yun Chao, A note on block circulant matrices, Kyungpook Mathematical Journal, 14 (1974), 97100.
[2] P.R. Davis, Circulant Matrices, John Wiley, New York, 1979.
[3] R.W. Farebrother, S.T. Jensen and G.P.H. Styan, Sir Thomas Muir and
NineteenthCentury Books on Determinants, IMAGE, 28 (2002, 615.
[4] B. Friedman, Eigenvalues of composite matrices, Proceedings of the Cambridge
Philosophical Society, 57 (1961), 3749.
[5] T. Muir, The Theory Of Determinants In The Historical Order of Development
(4 volumes), Macmillan & Co., London, (1906, 1911, 1920, 1923).
[6] T. Muir, A Treatise On The Theory Of Determinants, (revised and enlarged
by William H. Metzler), Albany NY, Privately published, 4+766pp, 1930.
[7] T. Muir, (1933). A Treatise On the Theory of Determinants, (revised and
enlarged by William H. Metzler), Longmans, Green & Co; London, 4+766pp.
(not seen). Reprint edition, Dover, New York, 1960 (not seen).
[8] G.J. Tee (1963). A novel finite-difference approximation to the biharmonic
operator, The Computer Journal, 6(2), pp. 177192. MR 29 #6646.
[9] G.E. Trapp, Inverses of circulant matrices and block circulant matrices,
Kyungpook Mathematical Journal, 13 (1973), 1120.
[10] R. Turcajova, Numerical condition of discrete wavelet transforms, SIAM J.
Matrix Anal. Appl., 18 (1997), 981999.
[11] S.J. Watson, (2001). NA Digest (e-mail) v. 01, #45, 2001122.
Garry Tee
Department of Mathematics
University of Auckland
Private Bag 92019
Auckland
NEW ZEALAND
tee@math.auckland.ac.nz