THE HYPO-EUCLIDEAN NORM OF AN n−TUPLE OF VECTORS IN INNER PRODUCT SPACES AND APPLICATIONS
S.S. DRAGOMIR
SCHOOL OFCOMPUTERSCIENCE ANDMATHEMATICS
VICTORIAUNIVERSITY
PO BOX14428, MELBOURNECITY
8001, VIC, AUSTRALIA
sever.dragomir@vu.edu.au URL:http://rgmia.vu.edu.au/dragomir Received 19 March, 2007; accepted 28 April, 2007
Communicated by J.M. Rassias
ABSTRACT. The concept of hypo-Euclidean norm for ann−tuple of vectors in inner product spaces is introduced. Its fundamental properties are established. Upper bounds via the Boas- Bellman [1]-[3] and Bombieri [2] type inequalities are provided. Applications forn−tuples of bounded linear operators defined on Hilbert spaces are also given.
Key words and phrases: Inner product spaces, Norms, Bessel’s inequality, Boas-Bellman and Bombieri inequalities, Bounded linear operators, Numerical radius.
2000 Mathematics Subject Classification. Primary 47C05, 47C10; Secondary 47A12.
1. INTRODUCTION
Let (E,k·k) be a normed linear space over the real or complex number field K. On Kn endowed with the canonical linear structure we consider a normk·knand the unit ball
B(k·kn) :={λ= (λ1, . . . , λn)∈Kn| kλkn≤1}. As an example of such norms we should mention the usualp−norms
(1.1) kλkn,p:=
( max{|λ1|, . . . ,|λn|} if p=∞;
(Pn
k=1|λk|p)1p if p∈[1,∞).
The Euclidean norm is obtained forp= 2, i.e., kλkn,2 =
n
X
k=1
|λk|2
!12 .
084-07
It is well known that onEn:=E× · · · ×Eendowed with the canonical linear structure we can define the followingp−norms:
(1.2) kXkn,p:=
( max{kx1k, . . . ,kxnk} if p=∞;
(Pn
k=1kxkkp)1p if p∈[1,∞);
whereX = (x1, . . . , xn)∈En.
For a given normk·knonKnwe define the functionalk·kh,n:En→[0,∞)given by
(1.3) kXkh,n:= sup
(λ1,...,λn)∈B(k·kn)
n
X
j=1
λjxj , whereX = (x1, . . . , xn)∈En.
It is easy to see that:
(i) kXkh,n≥0for anyX ∈En;
(ii) kX+Ykh,n≤ kXkh,n+kYkh,nfor anyX, Y ∈En; (iii) kαXkh,n =|α| kXkh,nfor eachα∈KandX ∈En;
and therefore k·kh,nis a semi-norm on En. This will be called the hypo-semi-norm generated by the normk·knonXn.
We observe thatkXkh,n= 0if and only ifPn
j=1λjxj = 0for any(λ1, . . . , λn)∈B(k·kn). If there exists λ01, . . . , λ0n 6= 0 such that (λ01,0, . . . ,0), (0, λ02, . . . ,0), . . . , (0,0, . . . , λ0n) ∈ B(k·kn)then the semi-norm generated byk·knis a norm onEn.
If byBn,pwithp∈[1,∞]we denote the balls generated by the p−normsk·kn,ponKn,then we can obtain the following hypo-p-norms onXn:
(1.4) kXkh,n,p := sup
(λ1,...,λn)∈Bn,p
n
X
j=1
λjxj , withp∈[1,∞].
Forp= 2,we have the Euclidean ball inKn,which we denote byBn, Bn=
(
λ= (λ1, . . . , λn)∈Kn
n
X
i=1
|λi|2 ≤1 )
that generates the hypo-Euclidean norm onEn,i.e.,
(1.5) kXkh,e := sup
(λ1,...,λn)∈Bn
n
X
j=1
λjxj .
Moreover, ifE =H, H is a Hilbert space overK, then the hypo-Euclidean norm onHnwill be denoted simply by
(1.6) k(x1, . . . , xn)ke := sup
(λ1,...,λn)∈Bn
n
X
j=1
λjxj , and its properties will be extensively studied in the present paper.
Both the notation in (1.6) and the necessity of investigating its main properties are motivated by the recent work of G. Popescu [9] who introduced a similar norm on the Cartesian product of Banach algebraB(H)of all bounded linear operators on H and used it to investigate var- ious properties of n−tuple of operators in Multivariable Operator Theory. The study is also motivated by the fact that the hypo-Euclidean norm is closely related to the quadratic form
Pn
j=1|hx, xji|2 (see the representation Theorem 2.2) that plays a key role in many problems arising in the Theory of Fourier expansions in Hilbert spaces.
The paper is structured as follows: in Section 2 we establish the equivalence of the hypo- Euclidean norm with the usual Euclidean norm on Hn, provide a representation result and obtain some lower bounds for it. In Section 3, on utilising the classical results of Boas-Bellman and Bombieri as well as some recent similar results obtained by the author, we give various upper bounds for the hypo-Euclidean norm. These are complemented in Section 4 with other inequalities betweenp−norms and the hypo-Euclidean norm. Section 5 is devoted to the pre- sentation of some conditional reverse inequalities between the hypo-Euclidean norm and the norm of the sum of the vectors involved. In Section 6, the natural connection between the hypo-Euclidean norm and the operator normk(·, . . . ,·)ke introduced by Popescu in [9] is in- vestigated. A representation result is obtained and some applications for operator inequalities are pointed out. Finally, in Section 7, a new norm for operators is introduced and some natural inequalities are obtained.
2. FUNDAMENTALPROPERTIES
Let(H;h·,·i)be a Hilbert space overKandn ∈ N,n ≥ 1.In the Cartesian productHn :=
H× · · · ×H,for then−tuples of vectorsX = (x1, . . . , xn), Y = (y1, . . . , yn) ∈Hn,we can define the inner producth·,·iby
(2.1) hX, Yi:=
n
X
j=1
hxj, yji, X, Y ∈Hn, which generates the Euclidean normk·k2 onHn,i.e.,
(2.2) kXk2 :=
n
X
j=1
kxjk2
!12
, X ∈Hn.
The following result connects the usual Euclidean norm k·kwith the hypo-Euclidean norm k·ke.
Theorem 2.1. For anyX ∈Hnwe have the inequalities
(2.3) kXk2 ≥ kXke≥ 1
√nkXk2, i.e.,k·k2andk·ke are equivalent norms onHn.
Proof. By the Cauchy-Bunyakovsky-Schwarz inequality we have
(2.4)
n
X
j=1
λjxj
≤
n
X
j=1
|λj|2
!12 n X
j=1
kxjk2
!12
for any(λ1, . . . , λn)∈Kn.Taking the supremum over(λ1, . . . , λn)∈Bnin (2.4) we obtain the first inequality in (2.3).
If byσwe denote the rotation-invariant normalised positive Borel measure on the unit sphere
∂Bn ∂Bn = (λ1, . . . , λn)∈Kn
Pn
i=1|λi|2 = 1
whose existence and properties have been
pointed out in [10], then we can state that Z
∂Bn
|λk|2dσ(λ) = 1 n and (2.5)
Z
∂Bn
λkλjdσ(λ) = 0 if k 6=j, k, j= 1, . . . , n.
Utilising these properties, we have kXk2e = sup
(λ1,...,λn)∈Bn
n
X
k=1
λkxk
2
= sup
(λ1,...,λn)∈Bn
" n X
k,j=1
λkλjhxk, xji
#
≥ Z
∂Bn
" n X
k,j=1
λkλjhxk, xji
#
dσ(λ) =
n
X
k,j=1
Z
∂Bn
λkλjhxk, xji dσ(λ)
= 1 n
n
X
k=1
kxkk2 = 1
nkXk22,
from where we deduce the second inequality in (2.3).
The following representation result for the hypo-Euclidean norm plays a key role in obtaining various bounds for this norm:
Theorem 2.2. For anyX ∈HnwithX = (x1, . . . , xn),we have
(2.6) kXke = sup
kxk=1 n
X
j=1
|hx, xji|2
!12 .
Proof. We use the following well known representation result for scalars:
(2.7)
n
X
j=1
|zj|2 = sup
(λ1,...,λn)∈Bn
n
X
j=1
λjzj
2
, where(z1, . . . , zn)∈Kn.
Utilising this property, we thus have (2.8)
n
X
j=1
|hx, xji|2
!12
= sup
(λ1,...,λn)∈Bn
* x,
n
X
j=1
λjxj +
for anyx∈H.
Now, taking the supremum overkxk= 1in (2.8) we get sup
kxk=1 n
X
j=1
|hx, xji|2
!12
= sup
kxk=1
"
sup
(λ1,...,λn)∈Bn
* x,
n
X
j=1
λjxj +
#
= sup
(λ1,...,λn)∈Bn
"
sup
kxk=1
* x,
n
X
j=1
λjxj +
#
= sup
(λ1,...,λn)∈Bn
n
X
j=1
λjxj ,
since, in any Hilbert space we have thatsupkuk=1|hu, vi|=kvkfor eachv ∈H.
Corollary 2.3. If X = (x1, . . . , xn)is ann−tuple of orthonormal vectors, i.e., we recall that kxkk= 1andhxk, xji= 0fork, j ∈ {1, . . . , n}withk 6=j,thenkXke≤1.
The proof is obvious by Bessel’s inequality.
The next proposition contains two lower bounds for the hypo-Euclidean norm that are some- times better than the one in (2.3), as will be shown by some examples later.
Proposition 2.4. For anyX = (x1, . . . , xn)∈Hn\ {0}we have
(2.9) kXke≥
1 kXk2
Pn
j=1kxjkxj
,
√1 n
Pn j=1xj
.
Proof. By the definition of the hypo-Euclidean norm we have that, if(λ01, . . . , λ0n) ∈ Bn,then obviously
kXke≥
n
X
j=1
λ0jxj . The choice
λ0j := kxjk
kXk2, j ∈ {1, . . . , n},
which satisfies the condition(λ01, . . . , λ0n) ∈ Bn will produce the first inequality while the se- lection
λ0j = 1
√n, j ∈ {1, . . . , n},
will give the second inequality in (2.9).
Remark 2.5. Forn= 2,the hypo-Euclidean norm onH2 k(x, y)ke = sup
(λ,µ)∈B2
kλx+µyk= sup
kzk=1
|hz, xi|2+|hz, yi|212 is bounded below by
B1(x, y) := 1
√2 kxk2+kyk212 , B2(x, y) := kkxkx+kykyk
kxk2+kyk212 and
B3(x, y) := 1
√2kx+yk.
IfH =Cendowed with the canonical inner producthx, yi:=x¯ywherex, y ∈C, then B1(x, y) = 1
√2 |x|2+|y|212 , B2(x, y) = ||x|x+|y|y|
|x|2+|y|212 and
B3(x, y) = 1
√2|x+y|, x, y ∈C.
The plots of the differences D1(x, y) := B1(x, y)−B2(x, y) and D2(x, y) := B1(x, y)− B3(x, y)which are depicted in Figure 2.1 and Figure 2.2, respectively, show that the boundB1 is not always better thanB2orB3.However, since the plot ofD3(x, y) :=B2(x, y)−B3(x, y) (see Figure 2.3) appears to indicate that, at least in the case ofC2,it may be possible that the boundB2is always better thanB3,hence we can ask in general which bound from (2.6) is better for a givenn ≥ 2?This is an open problem that will be left to the interested reader for further investigation.
-10 -5 10
-4 -2
0 0
5 2
y 4
6
x 8
0 10
5 -5 -10
10
Figure 2.1: The behaviour ofD1(x, y)
-10 -5 0 y -4 5
10 -2
5 0
0
-5 10
2
x -10
4 6 8 10
Figure 2.2: The behaviour ofD2(x, y)
-10 -5 0 y 5 10
0
x 5
0.5
0 -5
1
10 -10 1.5
2 2.5
3 3.5
Figure 2.3: The behaviour ofD3(x, y)
3. UPPERBOUNDS VIA THEBOAS-BELLMAN ANDBOMBIERITYPE INEQUALITIES
In 1941, R.P. Boas [3] and in 1944, independently, R. Bellman [1] proved the following generalisation of Bessel’s inequality that can be stated for any family of vectors{y1, . . . , yn}(see also [8, p. 392] or [5, p. 125]):
(3.1)
n
X
j=1
|hx, yji|2 ≤ kxk2
max
1≤j≤nkyjk2+ X
1≤j6=k≤n
|hyk, yji|2
!12
for anyx, y1. . . , ynvectors in the real or complex inner product space(H;h·,·i).This result is known in the literature as the Boas-Bellman inequality.
The following result provides various upper bounds for the hypo-Euclidean norm:
Theorem 3.1. For anyX = (x1, . . . , xn)∈Hn,we have
(3.2) kXk2e ≤
1≤j≤nmax kxjk2+ P
1≤j6=k≤n
|hxk, xji|2
!12 ,
1≤j≤nmax kxjk2+ (n−1) max
1≤j6=k≤n|hxk, xji|;
(3.3) kXk2e ≤
"
1≤j≤nmax kxjk2
n
X
j=1
kxjk2+ max
1≤j6=k≤n{kxjk kxkk} X
1≤j6=k≤n
|hxj, xki|
#12
and
(3.4) kXk4e ≤
1≤j≤nmax kxjk2
n
P
j=1
kxjk2+ (n−1)kXk2e max
1≤j6=k≤n|hxj, xki|, kXk2e max
1≤j≤nkxjk2+ max
1≤j6=k≤n{kxjk kxkk} P
1≤j6=k≤n
|hxj, xki|.
Proof. Taking the supremum over kxk = 1 in (3.1) and utilising the representation (2.6), we deduce the first inequality in (3.2).
In [4], we proved amongst others the following inequalities
(3.5)
n
X
j=1
cjhx, yji
2
≤ kxk2×
1≤j≤nmax |cj|2
n
P
j=1
kyjk2,
n
P
j=1
|cj|2 max
1≤j≤nkyjk2, +kxk2×
1≤j6=k≤nmax {|cjck|} P
1≤j6=k≤n
|hyj, yki|, (n−1)
n
P
j=1
|cj|2 max
1≤j6=k≤n|hyj, yki|, for any y1, . . . , yn, x ∈ H and c1, . . . , cn ∈ K, where (3.5) should be seen as all possible configurations.
The choicecj =hx, yji, j ∈ {1, . . . , n}will produce the following four inequalities:
(3.6)
" n X
j=1
|hx, yji|2
#2
≤ kxk2×
1≤j≤nmax |hx, yji|2
n
P
j=1
kyjk2,
n
P
j=1
|hx, yji|2 max
1≤j≤nkyjk2, +kxk2×
1≤j6=k≤nmax {|hx, yji| |hx, yki|} P
1≤j6=k≤n
|hyj, yki|, (n−1)
n
P
j=1
|hx, yji|2 max
1≤j6=k≤n|hyj, yki|.
Taking the supremum overkxk = 1and utilising the representation (2.6) we easily deduce the
rest of the four inequalities.
A different generalisation of Bessel’s inequality for non-orthogonal vectors is the Bombieri inequality (see [2] or [8, p. 397] and [5, p. 134]):
(3.7)
n
X
j=1
|hx, yji|2 ≤ kxk2 max
1≤j≤n
( n X
k=1
|hyj, yki|
) ,
for any x ∈ H, where y1, . . . , yn are vectors in the real or complex inner product space (H;h·,·i).
Note that, the Bombieri inequality was not stated in the general case of inner product spaces in [2]. However, the inequality presented there easily leads to (3.7) which, apparently, was firstly mentioned as is in [8, p. 394].
On utilising the Bombieri inequality (3.7) and the representation Theorem 2.2, we can state the following simple upper bound for the hypo-Euclidean normk·ke.
Theorem 3.2. For anyX = (x1, . . . , xn)∈Hn,we have
(3.8) kXk2e ≤ max
1≤j≤n
( n X
k=1
|hxj, xki|
) .
In [6] (see also [5, p. 138]), we have established the following norm inequalities:
(3.9)
n
X
j=1
αjzj
2
≤n1p+1t−1
n
X
k=1
|αk|2
n
X
k=1 n
X
j=1
|hzj, zki|q
!uq
1 u
,
where 1p + 1q = 1, 1t +1u = 1and1< p ≤2,1< t≤2andαj ∈C,zj ∈H, j ∈ {1, . . . , n}. An interesting particular case of (3.9) obtained forp=q = 2, t=u= 2is incorporated in
(3.10)
n
X
j=1
αjzj
2
≤
n
X
k=1
|αk|2
n
X
j,k=1
|hzj, zki|2
!12 .
Other similar inequalities for norms are the following ones [6] (see also [5, pp. 139-140]):
(3.11)
n
X
j=1
αjzj
2
≤n1p
n
X
k=1
|αk|2 max
1≤j≤n
" n X
k=1
|hzj, zki|q
#1q
,
provided that1< p≤2and 1p +1q = 1, αj ∈C,zj ∈H, j ∈ {1, . . . , n}.In the particular case p=q = 2,we have
(3.12)
n
X
j=1
αjzj
2
≤√ n
n
X
k=1
|αk|2 max
1≤j≤n
" n X
k=1
|hzj, zki|2
#12 .
Also, if1< m≤2,then [6]:
(3.13)
n
X
j=1
αjzj
2
≤nm1
n
X
k=1
|αk|2 ( n
X
j=1
1≤k≤nmax |hzj, zki|l )1l
,
where m1 + 1l = 1.Form=l= 2,we get
(3.14)
n
X
j=1
αjzj
2
≤√ n
n
X
k=1
|αk|2
" n X
j=1
1≤k≤nmax |hzj, zki|2 #12
.
Finally, we can also state the inequality [6]:
(3.15)
n
X
j=1
αjzj
2
≤n
n
X
k=1
|αk|2 max
1≤j,k≤n|hzj, zki|.
Utilising the above norm-inequalities and the definition of the hypo-Euclidean norm, we can state the following result which provides other upper bounds than the ones outlined in Theorem 3.1 and 3.2:
Theorem 3.3. For anyX = (x1, . . . , xn)∈Hn,we have
(3.16) kXk2e ≤
n1p+1t−1
n
P
k=1 n
P
j=1
|hxj, xki|q
!uq
1 u
where 1p +1q = 1,
1
t +u1 = 1 and 1< p ≤2, 1< t≤2;
n1p max
1≤j≤n
"
n
P
j=1
|hxj, xki|q
#1q
where 1p +1q = 1 and 1< p≤2;
nm1 ( n
P
j=1
1≤k≤nmax |hxj, xki|l )1l
where m1 +1l = 1 and 1< m≤2;
n max
1≤k≤n|hxk, zji|; and, in particular,
(3.17) kXk2e ≤
"
n
P
j,k=1
|hxj, xki|2
#12
;
√n max
1≤j≤n
n P
k=1
|hxj, xki|2 12
;
√n
"
n
P
j=1
1≤k≤nmax
|hxj, xki|2 #12
.
4. VARIOUS INEQUALITIES FOR THE HYPO-EUCLIDEANNORM
For ann−tupleX = (x1, . . . , xn)of vectors inH,we consider the usualp−norms:
kXkp :=
n
X
j=1
kxjkp
!1p , wherep∈[1,∞),and denote withSthe sumPn
j=1xj.
With these notations we can state the following reverse of the inequalitykXk2 ≥ kXke,that has been pointed out in Theorem 2.1.
Theorem 4.1. For anyX = (x1, . . . , xn)∈Hn,we have
(4.1) (0≤)kXk22 − kXk2e ≤ kXk21− kSk2. If
kXk2(2) :=
n
X
j,k=1
xj+xk 2
2
, then also
(0≤)kXk22 − kXk2e ≤ kXk2(2)− kSk2 (4.2)
(≤nkXk22− kSk2).
Proof. We observe, for anyx∈H,that
n
X
j=1
hx, xji
2
=
n
X
j=1
hx, xji
n
X
k=1
hx, xki=
n
X
j=1
hx, xji
n
X
k=1
hx, xki (4.3)
=
n
X
k=1
|hx, xki|2+ X
1≤j6=k≤n
hx, xji hxk, xi
≤
n
X
k=1
|hx, xki|2+
X
1≤j6=k≤n
hx, xji hxk, xi
≤
n
X
k=1
|hx, xki|2+ X
1≤j6=k≤n
|hx, xji| |hxk, xi|. Taking the supremum overkxk= 1,we get
(4.4) sup
kxk=1
n
X
j=1
hx, xji
2
≤ sup
kxk=1 n
X
k=1
|hx, xki|2+ X
1≤j6=k≤n
sup
kxk=1
|hx, xji| · sup
kxk=1
|hxk, xi|.
However,
sup
kxk=1
n
X
j=1
hx, xji
2
= sup
kxk=1
* x,
n
X
j=1
xj +
2
=kSk2,
sup
kxk=1
|hx, xji|=kxjk and sup
kxk=1
|hx, xki|=kxkk forj, k ∈ {1, . . . , n},and by (4.4) we get
kSk2 ≤ kXk2e+ X
1≤j6=k≤n
kxjk kxjk
=kXk2e+
n
X
j,k=1
kxjk kxkk −
n
X
k=1
kxkk2
=kXk2e+kXk21− kXk22, which is clearly equivalent with (4.1).
Further on, we also observe that, for anyx∈H we have the identity:
n
X
j=1
hx, xji
2
= Re
" n X
k,j=1
hx, xji hxk, xi
# (4.5)
=
n
X
k=1
|hx, xki|2+ X
1≤j6=k≤n
Re [hx, xji hxk, xi].
Utilising the elementary inequality for complex numbers
(4.6) Re (u¯v)≤ 1
4|u+v|2, u, v ∈C,
we can state that X
1≤k6=j≤n
Re [hx, xji hxk, xi]≤ 1 4
X
1≤k6=j≤n
|hx, xji+hx, xki|2
= X
1≤k6=j≤n
x,xj+xk 2
2
, and by (4.5) we get
(4.7)
n
X
j=1
hx, xji
2
≤
n
X
k=1
|hx, xki|2+ X
1≤k6=j≤n
x,xj+xk 2
2
for anyx∈H.
Taking the supremum overkxk= 1in (4.7) we deduce
n
X
j=1
xj
2
≤ kXk2e+ X
1≤k6=j≤n
xj +xk 2
2
=kXk2e+
n
X
k,j=1
xj +xk 2
2
−
n
X
k=1
kxkk2 which provides the first inequality in (4.2).
By the convexity ofk·k2 we have
n
X
j,k=1
xj +xk 2
2
≤ 1 2
n
X
j,k=1
kxjk2+kxkk2
=n
n
X
k=1
kxkk2
and the last part of (4.2) is obvious.
Remark 4.2. Forn= 2, X = (x, y)∈H2 we have the upper bounds B1(x, y) :=kxk2+kyk2− kx+yk2
= 2 (kxk kyk −Rehx, yi) and
B2(x, y) :=kxk2+kyk2
for the difference kXk22 − kXk2e, X ∈ H2 as provided by (4.1) and (4.2) respectively. If H = R then B1(x, y) = 2 (|xy| −xy), B2(x, y) = x2 +y2. If we consider the function
∆ (x, y) = B2(x, y)−B1(x, y)then the plot of∆ (x, y)depicted in Figure 4.1 shows that the bounds provided by (4.1) and (4.2) cannot be compared in general, meaning that sometimes the first is better than the second and vice versa.
From a different view-point we can state the following result:
Theorem 4.3. For anyX = (x1, . . . , xn)∈Hn,we have
(4.8) kSk2 ≤ kXke
kXke+
n
X
k=1
kS−xkk2
!12
and
(4.9) kSk2 ≤ kXke
kXke+
1≤k≤nmax kS−xkk2+ X
1≤k6=l≤n
|hS−xk, S−xli|2
!12
1 2
,
-10 -5 0 y 5 -200
10 5 10
0 -5 -10
-100
x 0
100 200
Figure 4.1: The behaviour of∆ (x, y)
respectively.
Proof. Utilising the identity (4.5) above we have
(4.10)
n
X
j=1
hx, xji
2
=
n
X
k=1
|hx, xki|2 + Re
*
x, X
1≤j6=k≤n
hx, xkixj
+
for anyx∈H.
By the Schwarz inequality in the inner product space(H,h·,·i), we have that Re
*
x, X
1≤j6=k≤n
hx, xkixj +
≤ kxk
X
1≤j6=k≤n
hx, xkixj (4.11)
=kxk
n
X
j,k=1
hx, xkixj −
n
X
k=1
hx, xkixk
=kxk
* x,
n
X
k=1
xk + n
X
j=1
xj −
n
X
k=1
hx, xkixk
=kxk
n
X
k=1
hx, xki(S−xk) . Utilising the Cauchy-Bunyakovsky-Schwarz inequality we have
(4.12)
n
X
k=1
hx, xki(S−xk)
≤
n
X
k=1
|hx, xki|2
!12 n X
k=1
kS−xkk2
!12
and then by (4.10) – (4.12) we can state the inequality:
(4.13)
n
X
j=1
hx, xji
2
≤
n
X
k=1
|hx, xki|2
!12
n
X
k=1
|hx, xki|2
!12 +
n
X
k=1
kS−xkk2
!12
for any x ∈ H,kxk = 1. Taking the supremum over kxk = 1 we deduce the desired result (4.8).
Now, following the above argument, we can also state that (4.14)
* x,
n
X
j=1
xj
+
2
≤
n
X
k=1
|hx, xki|2+kxk
n
X
k=1
hx, xki(S−xk) for anyx∈H.
Utilising the inequality (4.15)
n
X
j=1
αjzj
2
≤
n
X
j=1
|αj|2
1≤j≤nmax kzjk2 + X
1≤j6=k≤n
|hzj, zki|2
!12
,
whereαj ∈ C,zj ∈H, j ∈ {1, . . . , n},that has been obtained in [4], see also [5, p. 128], we can state that
(4.16)
n
X
k=1
hx, xki(S−xk)
≤
n
X
k=1
|hx, xki|2
!12
1≤k≤nmax kS−xkk2 + X
1≤k6=l≤n
|hS−xk, S−xli|2
!12
1 2
for anyx∈H.
Now, by the use of (4.14) – (4.16) we deduce the desired result (4.9). The details are omitted.
Remark 4.4. On utilising the inequality:
(4.17)
n
X
j=1
αjzj
2
≤
n
X
j=1
|αj|2
1≤k≤nmax kzkk2+ (n−1) max
1≤k6=l≤n|hzk, zli|
,
whereαj ∈ C, zj ∈ H, j ∈ {1, . . . , n},that has been obtained in [4], (see also [5, p. 130]) in place of (4.15) above, we can state the following inequality for the hypo-Euclidean norm as well:
(4.18) kSk2
≤ kXke
"
kXke+
1≤k≤nmax kS−xkk2+ (n−1) max
1≤k6=l≤n|hS−xk, S−xli|2 12#
for anyX= (x1, . . . , xn)∈Hn.
Other similar results may be stated by making use of the results from [6]. The details are left to the interested reader.
5. REVERSEINEQUALITIES
Before we proceed with establishing some reverse inequalities for the hypo-Euclidean norm, we recall some reverse results of the Cauchy-Bunyakovsky-Schwarz inequality for real or com- plex numbers as follows:
Ifγ,Γ∈K(K=C,R)andαj ∈K,j ∈ {1, . . . , n}with the property that 0≤Re [(Γ−αj) (αj−γ)]¯
(5.1)
= (Re Γ−Reαj) (Reαj −Reγ) + (Im Γ−Imαj) (Imαj−Imγ) or, equivalently,
(5.2)
αj − γ+ Γ 2
≤ 1
2|Γ−γ|
for eachj ∈ {1, . . . , n},then (see for instance [5, p. 9])
(5.3) n
n
X
j=1
|αj|2−
n
X
j=1
αj
2
≤ 1
4 ·n2|Γ−γ|2. In addition, ifRe (Γ¯γ)>0,then (see for example [5, p. 26]):
n
n
X
j=1
|αj|2 ≤ 1 4·
n Re
h Γ + ¯¯ γ Pn j=1αj
io2
Re (Γ¯γ) (5.4)
≤ 1
4· |Γ +γ|2 Re (Γ¯γ)
n
X
j=1
αj
2
and
(5.5) n
n
X
j=1
|αj|2−
n
X
j=1
αj
2
≤ 1
4· |Γ−γ|2 Re (Γ¯γ)
n
X
j=1
αj
2
. Also, ifΓ6=−γ,then (see for instance [5, p. 32]):
(5.6) n
n
X
j=1
|αj|2
!12
−
n
X
j=1
αj
≤ 1
4n·|Γ−γ|2
|Γ +γ|. Finally, from [7] we can also state that
(5.7) n
n
X
j=1
|αj|2−
n
X
j=1
αj
2
≤nh
|Γ +γ| −2p
Re (Γ¯γ)i
n
X
j=1
αj , providedRe (Γ¯γ)>0.
We notice that a simple sufficient condition for (5.1) to hold is that (5.8) Re Γ≥Reαj ≥Reγ and Im Γ≥Imαj ≥Imγ for eachj ∈ {1, . . . , n}.
We can state and prove the following conditional inequalities for the hypo-Euclidean norm k·ke :
Theorem 5.1. Letϕ, φ∈KandX = (x1, . . . , xn)∈Hnsuch that either:
(5.9)
hx, xji − ϕ+φ 2
≤ 1
2|φ−ϕ|
or, equivalently,
(5.10) Re [(φ− hx, xji) (hxj, xi −ϕ)]¯ ≥0 for eachj ∈ {1, . . . , n}and for anyx∈H,kxk= 1.Then
(5.11) kXk2e ≤ 1
n kSk2+ 1
4n|φ−ϕ|2. Moreover, ifRe (φϕ)¯ >0,then
(5.12) kXk2e ≤ 1
4n ·|φ+ϕ|2 Re (φϕ) kSk2 and
(5.13) kXk2e ≤ 1
nkSk2+h
|φ+ϕ| −2p
Re (φϕ)¯ i kSk. Ifφ 6=−ϕ,then
(5.14) kXke ≤ 1
n kSk+ 1
4n· |φ−ϕ|2
|φ+ϕ|, whereS=Pn
j=1xj.
Proof. We only prove the inequality (5.11).
Letx∈H,kxk= 1.Then, on applying the inequality (5.3) forαj =hx, xji, j ∈ {1, . . . , n}
andΓ =φ, γ =ϕ,we can state that (5.15)
n
X
j=1
|hx, xji|2 ≤ 1 n
* x,
n
X
j=1
xj +
2
+ 1
4n|φ−ϕ|2.
Now if in (5.15) we take the supremum over kxk = 1, then we get the desired inequality (5.11).
The other inequalities follow by (5.4), (5.7) and (5.6) respectively. The details are omitted.
Remark 5.2. Due to the fact that
hx, xji −ϕ+φ 2
≤
xj − ϕ+φ 2 ·x
for anyj ∈ {1, . . . , n}andx∈H,kxk= 1,then a sufficient condition for (5.9) to hold is that
xj − ϕ+φ 2 ·x
≤ 1
2|φ−ϕ|
for eachj ∈ {1, . . . , n}andx∈H,kxk= 1.
6. APPLICATIONS FORn−TUPLES OFOPERATORS
In [9], the author has introduced the following norm on the Cartesian product B(n)(H) :=
B(H)× · · · ×B(H),whereB(H)denotes the Banach algebra of all bounded linear operators defined on the complex Hilbert spaceH :
(6.1) k(T1, . . . , Tn)ke := sup
(λ1,...,λn)∈Bn
kλ1T1 +· · ·+λnTnk, where (T1, . . . , Tn) ∈ B(n)(H) and Bn :=
(λ1, . . . , λn)∈Cn
Pn
i=1|λi|2 ≤1 is the Eu- clidean closed ball inCn.It is clear thatk·keis a norm onB(n)(H)and for any(T1, . . . , Tn)∈ B(n)(H)we have
(6.2) k(T1, . . . , Tn)ke =k(T1∗, . . . , Tn∗)ke, whereTi∗ is the adjoint operator ofTi, i∈ {1, . . . , n}.
It has been shown in [9] that the following inequality holds true:
(6.3) 1
√n
n
X
j=1
TjTj∗
1 2
≤ k(T1, . . . , Tn)ke ≤
n
X
j=1
TjTj∗
1 2
for anyn−tuple(T1, . . . , Tn)∈B(n)(H)and the constants √1n and1are best possible.
In the same paper [9] the author has introduced the Euclidean operator radius of ann−tuple of operators(T1, . . . , Tn)by
(6.4) we(T1, . . . , Tn) := sup
kxk=1 n
X
j=1
|hTjx, xi|2
!12
and proved thatwe(·)is a norm onB(n)(H)and satisfies the double inequality:
(6.5) 1
2k(T1, . . . , Tn)ke≤we(T1, . . . , Tn)≤ k(T1, . . . , Tn)ke for eachn−tuple(T1, . . . , Tn)∈B(n)(H).
As pointed out in [9], the Euclidean numerical radius also satisfies the double inequality:
(6.6) 1
2√ n
n
X
j=1
TjTj∗
1 2
≤we(T1, . . . , Tn)≤
n
X
j=1
TjTj∗
1 2
for any(T1, . . . , Tn)∈B(n)(H)and the constants 2√1n and1are best possible.
We are now able to establish the following natural connections that exists between the hypo- Euclidean norm of vectors in a Cartesian product of Hilbert spaces and the norm k·ke for n−tuples of operators in the Banach algebraB(H).
Theorem 6.1. For any(T1, . . . , Tn)∈B(n)(H)we have k(T1, . . . , Tn)ke = sup
kyk=1
k(T1y, . . . , Tny)ke (6.7)
= sup
kyk=1,kxk=1 n
X
j=1
|hTjy, xi|2
!12 .
Proof. By the definition of thek·ke−norm onB(n)(H)and the hypo-Euclidean norm on Hn, we have:
k(T1, . . . , Tn)ke = sup
(λ1,...,λn)∈Bn
"
sup
kyk=1
k(λ1T1+· · ·+λnTn)yk
# (6.8)
= sup
kyk=1
"
sup
(λ1,...,λn)∈Bn
kλ1T1y+· · ·+λnTnyk
#
= sup
kyk=1
k(T1y, . . . , Tny)ke.
Utilising the representation of the hypo-Euclidean norm onHnfrom Theorem 2.2, we have (6.9) k(T1y, . . . , Tny)ke = sup
kxk=1 n
X
j=1
|hTjy, xi|2
!12 .
Making use of (6.8) and (6.9) we deduce the desired equality (6.7).
Remark 6.2. Utilising Theorem 2.1, we have (6.10)
n
X
j=1
kTjyk2
!12
≥ k(T1y, . . . , Tny)ke≥ 1
√n
n
X
j=1
kTjyk2
!12
for anyy∈H,kyk= 1.
Since
n
X
j=1
kTjyk2 =
* n X
j=1
Tj∗Tjy, y +
, kyk= 1
hence, on taking the supremum overkyk= 1in (6.10) and on observing that sup
kyk=1
* n X
j=1
Tj∗Tjy, y +
=w
n
X
j=1
Tj∗Tj
!
=
n
X
j=1
Tj∗Tj
=
n
X
j=1
TjTj∗ , we deduce the inequality (6.3) that has been established in [9] by a different argument.
We observe that, due to the representation Theorem 6.1, some inequalities obtained for the hypo-Euclidean norm can be utilised in obtaining various new inequalities for the operator normk·keby employing a standard approach consisting in taking the supremum overkyk= 1, as described in the above remark.
The following different lower bound for the Euclidean operator normk·kecan be stated:
Proposition 6.3. For any(T1, . . . , Tn)∈B(n)(H),we have (6.11) k(T1, . . . , Tn)ke ≥ 1
√nkT1+· · ·+Tnk. Proof. Utilising Proposition 2.4 and Theorem 6.1 we have:
k(T1, . . . , Tn)ke= sup
kyk=1
k(T1y, . . . , Tny)ke
≥ 1
√n sup
kyk=1
kT1y+· · ·+Tnyk
= 1
√nkT1+· · ·+Tnk
which is the desired inequality (6.11).