Eli Maor, e the story of a number, among references: George F.

Simmons,
Calculus with Analytic Geometry, NY, McGraw Hill, 1985, pp. 734739

How Chebycev polynomials arise

ott
Set y(x) = xn pott
n1 (x) where pn1 is the unique polynomial of degree at
most n 1 solving the minimum problem

min max |xn p(x)|

pPn1 [1,1]

where Pn1 is the set of all polynomials of degree less than or equal to n 1.
Moreover, set = max[1,1] |y(x)|.
Graphical considerations let us find such pott
n1 , for n = 1, 2, 3:
For n = 1 the given problem minpP0 max[1,1] |x p(x)| has the obvious
i
solution pott
0 (x) = 0. Moreover, observe that y(xi ) = (1) , = 1, x1 = 1,
x0 = 1.
1
For n = 2 the problem minpP1 max[1,1] |x2 p(x)| is solved by pott
1 (x) = 2 .
1
i
Moreover, y(xi ) = (1) , = 2 , x2 = 1, x1 = 0, x0 = 1.
3
For n = 3 the solution pott
2 of the problem minpP2 max[1,1] |x p(x)| must
3
be a straight line with positive slope which intersects x in three distinct points
whose abscissas are 2 (1, 0), 1 = 0, 0 = 2 (0, 1), i.e. pott
= x
2
with 0 < < 1. Consider the function g(x) =px3 x in the
interval
[1,
p 1]
p
and notice that g 0 (x) = 3x2 and thus g 0 ( 3 ) = 0, g( 3 ) = 3 3
p
p
( 3 ) = 32 3 . Moreover, g(1) = 1 (1). So, we have to choose
p
(0, 1) so that max{1 , 23 3 } is minimum, i.e. (0, 1) such that
2
3
1= = .
4
3 3

1
3
i
Thus, pott
2 (x) = 4 x. Moreover, observe that y(xi ) = (1) , = 4 , x3 = 1,
1
1
x2 = 2 , x1 = 2 , x0 = 1.
In general, Chebycev-Tonelli theory states that y(x) = xn pott
n1 (x) must
assume the values and alternately in n + 1 points xj of [1, 1], 1
xn < xn1 < . . . < x2 < x1 < x0 1: y(xj ) = (1)j . Obviously y 0 (xi ) = 0,
i = 1, . . . , n 1, whereas y 0 (x0 )y 0 (xn ) 6= 0 since y 0 (x) is a polynomial of degree
n 1. Thus xn = 1, x0 = 1. Consider now the function y(x)2 2 . It is
zero in all the xi and its derivative, 2y(x)y 0 (x), is zero in x1 , x2 , . . . , xn1 . It
follows that y(x)2 2 = c(x2 1)y 0 (x)2 for some real constant c. Noting that
the coefficient of x2n is on the left 1 and on the right cn2 , we conclude that

n2
y 0 (x)2
n
y 0 (x)

p
=
,
.
=

1 x2
2 y(x)2
1 x2
2 y(x)2

The latter equality is solved by y(x) = cos(n arccos x + c), c R. Then the
identity y(1) = implies c = 2k, and thus
2
y(x) = xn pott
n1 (x) = cos(n arccos x), 1 x 0.

Finally, observe that

cos(0 arccos x) = 1|[1,1] =: T0 (x)|[1,1] ,
cos(arccos x) = x|[1,1] =: T1 (x)|[1,1] ,
cos(2 arccos x) = 2 cos(arccos x) cos(arccos x) cos(0 arccos x)
= 2x2 1|[1,1] =: T2 (x)|[1,1] ,
cos((j + 1) arccos x) = 2 cos(arccos x) cos(j arccos x) cos((j 1) arccos x)
= 2xTj (x) Tj1 (x)|[1,1] =: Tj+1 (x)|[1,1] .
Thus, =
result:

1
2n1

because Tn (x) = 2n1 xn + . So, we have the important

y(x) = xn pott
n1 (x) =

1
2n1

cos(n arccos x) =

1
2n1

Tn (x), 1 x2 0.

Let us see two examples. The already studied specific case n = 3 is now immediately obtained:
1
3
1
3
3
y(x) = x3 pott
2 (x) = 4 cos(3 arccos x) = 4 (4x 3x) = x 4 x,
1
3
3
3
3
y(xj ) = (1)j 4 , pott
2 (x) = x (x 4 x) = 4 x.

The cases n > 3 are analogously easily solved. In particular, for n = 4 we have
1
1
1
4
2
4
2
y(x) = x4 pott
3 (x) = 8 cos(4 arccos x) = 8 (8x 8x + 1) = x x + 8 ,
1
1
j1
ott
4
4
2
2
y(xj ) = (1) 8 , p3 (x) = x (x x + 8 ) = x 8 .

Deflation
Le A be a n n matrix. Denote by i , i = 1, . . . , n, the eigenvalues of A and
by yi the corresponding eigenvectors. So, we have Ayi = i yi , i = 1, . . . , n.
Assume that 1 , y1 are given and that 1 6= 0. Choose w Cn such that

w y1 6= 0 (given y1 choose w not orthogonal to y1 ) and set

W =A

1
y1 w .
w y1

It is known that the eigenvalues of W are

0, 2 , . . . , j , . . . , n
i.e. they are the same of A except 1 which is replaced with 0. Let us prove
this fact. Consider a matrix S whose first column is y1 and whose remaining
columns x2 , . . . , xn are chosen such that S is non singular. Observe that
S 1 AS = S 1 [Ay1 Ax2 Axn ] = [1 e1 S 1 Ax2 S 1 Axn ].
So, if we call B the (n 1) (n 1) lower right submatrix of S 1 AS, then
pA () = ( 1 )pB (). But we also have
S 1 W S

= S 1 AS S 1 w1y1 y1 w S


1 c T
=
w1y1 e1 [w y1 w x2 w xn ]
0
B

 

1 d T
1 c T

=
 0 TB
 0 O
0 c dT
=
,
0
B
2

and thus the identity pW () = pB (), from which the thesis.

Let w1 , w2 , . . ., wj , . . ., wn be the corresponding eigenvectors (W w1 = 0,
W wj = j wj j = 2, . . . , n). Is it possible to obtain the wj from the yj ?
First observe that
Ay1 = 1 y1 W y1 = 0 : w1 = y1 .

(a)

Then, for j = 2, . . . , n,
W yj = Ayj

1
w yj
y1 w yj = j yj 1 y1 .

w y1
w y1

(1)

If we impose yj = wj + cy1 , j = 2, . . . , n, then (1) becomes,

w w

= j wj + cj y1 1 w y1j y1 c1 y1

W wj + cW y1

w w

= j wj + y1 [cj 1 w y1j 1 c]

So, if j 6= 1 and
wj = y j

1 w wj
y1 ,
j 1 w y 1

(2)

then W wj = j wj . If, moreover, j 6= 0, then w yj = w wj +

w yj = w

j
wj j
1

w wj =

j 1
j w y j .

for all j {2 . . . n} | j 6= 1 , 0 :
Ayj = j yj
w y
W (yj 1j w y1j y1 ) = j (yj 1j

So, by (2),

w yj
w y1 y1 )

: wj = y j

j 1 w wj

1 w y j
j w y 1 y 1 .

(b)

Note that a formula for yj in terms of wj holds: see (2).

As regards the case j = 1 , it is simple to show that
for all j {2 . . . n} | j = 1 :
Ayj = j yj
w y
w y
W (yj w y1j y1 ) = j (yj w y1j y1 ) : wj = yj

w yj
w y1 y1 .

(c)

w y

Note that the vectors yj w y1j y1 are orthogonal to w. Is it possible to find

from (c) an expression of yj in terms of wj ?
It remains the case j = 0: find ? in
for all j {2 . . . n} | j = 0 :
Ayj = j yj = 0 W (?) = j (?) = 0 : wj =?
(yj = wj

w wj
w y1 y1

w yj = 0) . . .

(d?)

Choices of w. Since y1 y1 6= 0 one can set w = y1 . In this way, if A is

hermitian also W is hermitian. . . . . If i is such that (y1 )i 6= 0 then eTi Ay1 =
1 (y1 )i 6= 0. So one can set w = eTi A = row i of A. In this way the row i of W
is null and therefore we can introduce a matrix of order n 1 whose eigenvalues
are 2 , . . ., n (the unknown eigenvalues of A).
Exercise on deflation
3

The matrix

G=

1
4
3
4
11
16

1
4
1
8
1
4

1
2
1
8
1
16

satisfies the identity Ge = e, e = [1 1 1]T . So, G has the eigenvalue 1 with corresponding eigenvector e. Moreover, since (G) kGk = 1, all the eigenvalues
of G have modulus less than or equal to 1.
Let 1, 2 , 3 be the eigenvalues of G. It is known that the matrix
W =G

1
1
ew = G T
eeT G = G eeTi G
w e
ei Ge i

for any i = 1, 2, 3 has 0, 2 , 3 as eigenvalues. For i = 1 we obtain

0
0
0
W = 21 81 38 ,
7
7
0 16
16

7
.
thus the remaining eigenvalues of G are 81 and 16
1
7
Now observe that 1, 2 = 8 , 3 = 16 are eigenvalues also of GT . In
particular, there exists p such that GT p = p, but p has to be computed. The
following inverse power iterations

v0 , kv0 k1 = 1, ak = (GT (1 + )I)1 vk ,

vk+1 = ak /kak k1 , . . .

1+1
generate vk convergent to p, kpk1 = 1, with a convergence rate O( 1++
1 ).
8

One eigenvalue at a time with power iterations

Assume A diagonalizable with eigenvalues j such that |1 | > |k |, k =
2, . . . , n. Let v 6= 0 be a vector. Then
Ak v =

j A k xj =

j kj xj ,

Thus

X kj
1 k
A v = 1 x1 +
j k xj .
k
1
1
j6=1

X kj
1 k

z
A
v
=

z
x
+
j k z xj ,
1
1
k1
1
j6=1
1

z Ak+1 v = 1 z x1 +
k+1

X
j6=1

k+1
j
k+1
1

z xj ,

z Ak+1 v
1 , k .
z Ak v
So, if an eigenvalue dominates the other eigenvalues, then such eigenvalue can
be approximated better and better by computing the quantities:
Av,

z Av
z A2 v
z A3 v
, A2 v = A(Av),
, A3 v = A(A2 v), 2 , . . .

z v
z Av
z A v

It is clear that each new approximation requires a multiplication Aw.

Positive definite matrices and the choice w = y1

Let A be a positive definite n n matrix and let j , yj be such that Ayj =
j yj . Assume that 0 < n < n1 < < 2 < 1 . Then compute 1 via
power iterations, and y1 from a weak approximation 1 of 1 via inverse power
iterations, both applied to A. Then the eigenvalues of
A

1
y1 y1 are 0, n , n1 , , 2 .
ky1 k2

Compute 2 via power iterations, and y2 from a weak approximation 2 of 2 via

inverse power iterations, both applied to A ky11k2 y1 y1 . Then the eigenvalues
of
1
2
(A
y1 y1 )
y2 y2 are 0, 0, n , , 3 .
ky1 k2
ky2 k2
...
( (A

2
n
1
y1 y1 )
y2 y2 )
yn yn are 0, 0, . . . , 0.
ky1 k2
ky2 k2
kyn k2

Pn

It follows that A = j=1 kyjjk2 yj yj = QDQ , Q = [ ky11 k2 y1 ky12 k2 y2 kyn1 k2 yn ].

Note that the matrix Q is unitary (eigenvectors corresponding to distinct eigenvalues of a hermitian matrix must be orthogonal).
The QR method for 2 2 matrices
Set

x y
z w

z
.
x2 +z 2

Then

A=

x, y, w, z R.

Choose , R such that 2 + 2 = 1 and [Q1 A]21 = 0, where




Q1 =
,

i.e. =

x
,
x2 +z 2

"
x2 + z 2
Q1 A =
0

xy+zw

x2 +z 2
zy+xw

x2 +z 2

Now define the matrix B = RQT1 :

"
x + z(xy+zw)
x2 +z 2
B=
z(xwzy)

=: R.

x(xy+zw)
x2 +z 2
x(xwzy)
x2 +z 2

z +

x2 +z 2

Note that B = Q1 AQT1 , QT1 = Q1

1 , so that B has the same eigenvalues of A.
(Moreover, B is real symmetric if A is real symmetric).
So, by setting x0 = x, y0 = y, w0 = w, z0 = z we can define the four
sequences
z(xk yk +zk wk )
,
x2k +zk2
z(xk wk zk yk )
, wk+1
x2k +zk2

xk+1 = xk +

zk+1 =
k = 0, 1, 2, . . . ,

xk (xk yk +zk wk )
,
x2k +zk2
xk (xk wk zk yk )
,
x2k +zk2

yk+1 = zk +
=

which satisfy (by the theory on QR method) the properties:

zk 0,

xk , wk eigenvalues of A,

k +

provided the eigenvalues of A are distinct in modulus (try to prove this asser3
tion). For example, if x = w = 2 and y = z = 1, then x1 = 14
5 , y1 = 5 ,
6
3
122
9
42
9
w1 = 5 , z1 = 5 , x2 = 41 , y2 = 41 , w2 = 41 , z2 = 41 , . . .. It is clear that
xk and wk tend to 3 and 1, the eigenvalues of A.
Some results on matrix algebras
Given a n n matrix X, set
KX = {A : AX XA = 0},

P(X) = {p(X) : p polynomials}.

Note that P(X) KX , and

P(X) = KX iff dim P(X) = dim KX = n.
Let Z denote the n n shift-forward matrix, i.e. [Z]ij = 1 if i = j + 1, and
[Z]ij = 0 otherwise. Note that
KZ = P(Z) = {lower triangular Toeplitz matrices},
KZ T = P(Z T ) = {upper triangular Toeplitz matrices},
KZ T +en eT1 = P(Z T + en eT1 ) = { circulant matrices},
KZ T +Z = P(Z T + Z) = { matrices},
{symmetric circulant matrices} = P(Z T + Z + en eT1 + e1 eTn )
KZ T +Z+en eT1 +e1 eTn = {A + JB : A, B circulant matrices }
(eTi J = eni+1 i = 1, . . . , n, J =counteridentity).
Set X = Z+Z T . Then the condition AX = XA, A = (aij )ni,j=1 , is equivalent
to the n2 conditions:
ai,j1 + ai,j+1 = ai1,j + ai+1,j ,

1 i, j n,

ai,0 = ai,n+1 = a0,j = an+1,j = 0. Thus a generic matrix of has the form (in
the case n = 5):

a
b
c
d
e
b a+c
b+d
c+e d

c b+d a+c +e b+d c .

d c+e
b+d
a+c b
e
d
c
b
a
q
ij
j
2
Since XS = SD, Sij = n+1
sin n+1
(S 2 = I), D = diag (2 cos n+1
), and ma-

trices from are determined from their first row zT , we have the representation:
(z) = Sd(Sz)d(Se1 )1 S
( (z) = matrix of whose first row is zT ).
Given a generic non singular matrix M , we have the representation
{M d(z)M 1 : z Cn } = {M d(z)d(M T v)1 M 1 : z Cn }
for any vector v such that (M T v)j 6= 0, j (note that vT M d(z)d(M T v)1 M 1 =
zT ). For M =Fourier, sine matrices, one can choose v = e1 (so circulants and
6

 matrices are determined by their first row). But there are significant matrices
M (associated to fast discrete transforms) for which v cannot be chosen equal
to e1 (i.e. matrices diagonalized by M are not determined by their first row).
An example of matrix algebra which is not commutative is L = {A + JB :
A, B circulants}. The best approximation (in the Frobenius norm) in L of a
given matrix A, call it LA , is well defined. It is known that LA is hermitian
any time A is hermitian. But it is not known if (in case A hermitian) z Az > 0
z 6= 0 implies z LA z > 0 z 6= 0.
Assume {tk }+
k=0 , tk R, such that
+
X
k=0

|tk | < +.

(1)

P+
Set t() = k= t|k| eik , tmin = min t(), tmax = max t(). Then the eigenvalues of T (n) = (t|ij| )nj,j=1 are in the interval [tmin , tmax ] for all n (proof omitted).
Let CT (n) be the best circulant approximation of T (n) . Since
1
CT (n) = F diag ((F T (n) F )ii )F , Fij = n(i1)(j1) , n = ei2/n ,
n
we have
tmin min (T (n) ) min (CT (n) ), max (CT (n) ) max (T (n) ) tmax .
In particular, if
tmin > 0,

(2)

then the T (n) and the CT (n) are positive definite, and 2 (CT (n) ) 2 (T (n) )
(n)
(n)
tmax
T
and j are the eigenvalues, respectmin ; moreover, if En En = CT (n) , and j
tively, of I En1 T (n) EnT and CT (n) T (n) in nondecreasing order, then
1
1
1
1
(n)
(n)
(n)
(n)
(n)
| |
| | |j |
| |
| |
tmax j
max (CT (n) ) j
min (CT (n) ) j
tmin j
(2.5)
(apply the Courant-Fisher minimax characterization of the eigenvalues of a real
symmetric matrix to I En1 T (n) EnT ).

Theorem. If (1) holds, then the eigenvalues of CT (n) T (n) are clustered
(n)
around 0. If (1) and (2) hold, then the eigenvalues of I CT1
are clustered
(n) T
around 0.
Proof. For the sake of simplicity, set T = T (n) . Fix a number N , n > 2N ,
and let W (N ) and E (N ) be the n n matrices defined by

[CT T ]ij i, j n N
(N )
[W
]ij =
0
otherwise
and
CT T = E (N ) + W (N ) .

(3)

Note that [CT ]1j = ((n j + 1)tj1 + (j 1)tnj+1 )/n, j = 1, . . . , n, and thus,
for i, j = 1, . . . , n, we have
[CT T ]ij =

s|ij| |i j|
, sk = tk tnk .
n
7

Now observe that the rank of E (N ) is less than or equal to 2N , so E (N ) has at

least n 2N null eigenvalues. Also observe that CT T , E (N ) and W (N ) are
all real symmetric matrices. In the following we prove that, for any fixed > 0,
there exist N and 2N such that
kW (N ) k1 < n > .

(4)

As a consequence of this fact and of the identity (3) for N = N , we shall

have that for all n > at least n 2N eigenvalues of CT T are in (, ).
Moreover, if tmin > 0, then, by (2.5), we shall also obtain the clustering around
0 of the eigenvalues of I CT1 T .
So, let us prove (4). First we have
kW (N ) k1

nN 1
n1
N
X
2 X
2X
j|sj | 2
|tj | +
j|tj |.
n j=1
n j=1

(5)

j=N +1

P+
Then, for any > 0 choose N such that 2 j=N +1 |tj | < 2 and set N = N
in (5) and in the previous arguments. If , 2N , is such that, n > ,
2 PN
1 Pn1

j=1 j|tj | tends to 0 if (1) holds), then by

j=1 j|tj | < 2 (the sequence n
n
(5) we have the thesis (4).
Stai usando il seguente algoritmo (il primo a p.18 dellarticolo) che calcola
direttamente una successione di e vettori xk convergente a x tale che p =
1
kxk1 x ? Se non lo stai usando, allora leggilo attentamente ed implementalo
accuratamente, rispondendomi alle domande che troverai.

xk+1 = xk + F (I nd(F c))1 F (v AT xk )

1
Fs,j = n(s1)(j1) , s, j = 1, . . . , n,
n

z1

z2

d(z) =
..

.
zn

n = ei2/n

c = [c0 c1 cn1 ]T , c0 = s0 /n = 0, ci = (si + sn+i )/n, i = 1, . . . , n 1

s1 =

n1
X
i=1

[P ]i,i+1 , s1 =

n1
X

[P ]i+1,i , . . .

i=1

Quindi, ogni volta che n e una potenza di 2: calcolo dei cj , j = 0, . . . , n 1

(c0 = 0); calcolo di F c; calcolodei F c (il vettore coniugato di F c); calcolo della
matrice diagonale D = (I nd(F c))1 . Poi, per ogni k = 0, 1, . . ., calcolo di
xk+1 = xk + F DF (v (I P T )xk )
(scegliendo x0 = v = [1/n 1/n]T ).
Nota che esiste una matrice di permutazione Q tale che F = QF , F = QF ,
hai usato questo fatto per calcolare xk+1 ? Quindi F z e semplicemente una
permutazione di F z (e viceversa); la FFT che hai tu calcola F z o F z ?
8

I vettori xk dovrebbero convergere a un vettore x che una volta normalizzato

1
dovrebbe coincidere con il vettore page-rank p, cioe p = kxk
x.
1
Mi scrivi dettagliatamente i tre criteri di arresto che usi? Quello per potenze
dovrebbe differire da quelli usati per RE e RE precondizionato perche i vettori
generati dal metodo delle potenze sono gia normalizzati.
y 0 (t) =

1
, y(0) = 1 (y(t) = 1 t)
2y(t)

1
1
1 t = if f t = 1
p
p
Integrate in [0, 1 p1 ] the Cauchy problem to obtain an approximation of

1 .
p

p = 3: Eulero for h = 13 , two steps; for h = 16 , four steps.

(xi + h) = (xi ) + hf (xi , (xi )) = (xi ) h

1
2(xi )

1 1
1
5
1
=1 =
(0 + ) = (0)
3
3 2(0)
6
6
1
1 1
19
5 1
1 1
= =
( + ) = ( )
3 3
3
3 2( 31 )
6 5
30
Idem, implicit Euler: h =

1
3

not ok; h =

1
6

ok?.

(xi + h) = (xi ) + hf (xi + h, (xi + h)) = (xi ) h

(xi + h)2 (xi + h)(xi ) + h

1
2(xi + h)

1
=0
2

p
1
(xi + h) = ((xi ) (xi )2 2h)
2
r
r

2
2
1
3+1
1
1
1
1
1

(0 + ) = ((0) (0)2 ) = (1 1 ) =
3
2
3
2
3
2 2 3 2 3
r
1
1
2
1 1
1
( + ) = (( ) ( )2 )
3 3
2
3
3
3
not real!
The given matrix is non negative and stochastic by columns
3 2 (1 a b) b a = ( 1)(2 + (a + b) + a)
Eigenvalues:
p
(a + b)2 4a
a+b
1,

2
2
We know that their absolute value is less than or equal to 1. Question: when is
it equal to 1?
Assume they are real. Then question becomes:
p
(a + b)2 4a
a+b

+
=1
2
2
9

p
(a + b)2 4a
a+b

= 1

2
2
Assume they are not real. Then they can be rewritten as follows:
p
4a (a + b)2
a+b

i
2
2
Thus, question becomes:
(a + b)2
4a (a + b)2
+
=1
4
4
equality which is satisfied iff a = 1
An equivalent definition of Bernoulli polynomials
The degree n Bernoulli polynomial Bn (x) is uniquely determined by the
conditions
Z
Bn (x + 1) Bn (x) = nxn1 ,

Bn (x) dx = 0.

(1)

Note that the first condition in (1) implies:

Z

t+1

Bn (x + 1)dx

t+2
t+1

Bn (y)dy

t+1

Bn (x)dx = n
t

t+1
t

xn
n

t+1

Bn (x)dx = (t + 1)n tn .

By writing the latter identity for t = 0, 1, . . . , x 1, taking into account the

second condition in (1), and summing, we obtain:
Z

x+1
x

Bn (y)dy = xn , x R.

(2)

So, (1) implies (2). Of course, (2) implies the second condition in (1) (choose
x = 0). It can be shown that (2) implies also that Bn must be a polynomial of
degree at least n and must satisfy the first condition in (1).
Assume that we know that (2) implies that Bn must be a polynomial. Let us
show that then its degree is at least n. If, on the contrary, Bn (y) = a0 y n1 + . . .
R x+1
= an0 [(x + 1)n xn ] + . . . is a degree n 1
then x Bn (y)dy = [ an0 y n + . . .]x+1
x
polynomial, and thus cannot be equal to xn .
Finally, the fact that (2) implies the first condition in (1) can be shown by
deriving (2) with respect to x, and remembering the rule:
d
dx

g(x)
f (x)

h(x, y)dy = h(x, g(x))g 0 (x) h(x, f (x))f 0 (x) +

10

g(x)
f (x)

h(x, y) dy.
x