another look at some new cauchy–schwarz type inner ... · pdf fileanother...
TRANSCRIPT
Applied Mathematics and Computation 231 (2014) 463–477
Contents lists available at ScienceDirect
Applied Mathematics and Computation
journal homepage: www.elsevier .com/ locate/amc
Another look at some new Cauchy–Schwarz type inner productinequalities
0096-3003/$ - see front matter � 2013 Elsevier Inc. All rights reserved.http://dx.doi.org/10.1016/j.amc.2013.11.047
⇑ Corresponding author at: University of Pittsburgh, Department of Mathematics, Pittsburgh, PA 15260, USA.E-mail addresses: [email protected], [email protected] (C. Lupu), [email protected] (D. Schwarz).
Cezar Lupu a,b,⇑, Dan Schwarz c
a University of Pittsburgh, Department of Mathematics, Pittsburgh, PA 15260, USAb University of Craiova, Department of Mathematics, Str. Alexandru Ioan Cuza 13, RO–200585 Craiova, Romaniac Intuitext Softwin Group, D. Pompeiu St. 10A, RO–020337 Bucharest, Romania
a r t i c l e i n f o
Keywords:Inner product spaceCauchy–Schwarz inequalityGram determinantBuzano’s inequalityIntegral inequalitiesDiscrete inequalitiesLagrange multipliers
a b s t r a c t
In this paper we take a refreshing look at a Cauchy–Schwarz type inner product inequali-ties. We also provide a Cauchy–Schwarz based proof and a refinement of Buzano’s inequal-ity in inner product spaces. Some elementary applications such as trace inequalities forunitary matrices and discrete or integral inequalities are given.
� 2013 Elsevier Inc. All rights reserved.
1. Introduction
The theory of inner product inequalities plays a central role in many branches of mathematics like Linear Operators, Par-tial Differential Equations, Approximation & Optimization Theory, Information Theory or Statistics and many other fields.Such inequalities were the pioneering work of mathematicians like Cauchy, Minkovski, Holder, Hilbert, Hardy, Kantorovichand many others.
On the other hand, it is worth mentioning the contributions of Bellman, Boas, van der Corput, Ostrowski, Selberg, Enfloand Bombieri who have applied successfully in obtaining applications for Fourier and Mellin transforms, oscillatory integrals,approximation of polynomials or large sieve. All these results as well as their applications are covered in the monograph [4].
In [2], Buzano gave the following extension of the celebrated Cauchy–Schwarz’s inequality in a real or complex innerproduct space ðH; h�; �iÞ,
jha; cihc; bij 6 12ðjjajjjjbjj þ jha; bijÞ � jjcjj2 ð1Þ
for all a; b; c 2 H.Clearly, when a ¼ b, the above inequality becomes the celebrated Cauchy–Schwarz’s inequality,
jha; cij2 6 jjajj2jjcjj2 ð2Þ
for all a; c 2 H.As pointed out in [5], the original proof of Buzano is a little bit complicated. Buzano’s inequality also mentioned in [8]
as an interesting generalization of the Cauchy–Schwarz’s inequality and its proof requires some facts about orthogonal
464 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
decomposition of a complete inner product space. Moreover, Fuji and Kubo [5] gave a simple proof by using the orthogonalprojection on a subspace of an inner product space H and Cauchy–Schwarz’s inequality.
The equality case in (1) holds if
c ¼a ajjajj þ
ha;bijha;bij � b
jjbjj
� �; when ha; bi – 0;
a ajjajj þ b b
jjbjj
� �; when ha; bi ¼ 0;
8><>:
where a; b 2 K 2 fR;Cg. A detailed proof of Buzano’s inequality can also be found in [15], pp. 60–61.For real inner product spaces, Richard [14] obtained the following stronger inequality
ha; cihc; bi � 12ha; bijjcjj2
�������� 6 1
2jjajjjjbjjjjcjj2; ð3Þ
a; b; c 2 H.Dragomir [4] showed that the above inequality is true with coefficients 1
jaj instead of 12, where a is a non-zero number with
j1� aj ¼ 1. Also, Dragomir [4] obtained a refinement of Buzano’s inequality,
jha; cihc; bij 6 jha; cihc; bi � 12ha; bijjcjj2j þ 1
2ha; bijjcjj2 6 1
2ðjjajjjjbjj þ jha; bijÞÞjjcjj2; ð4Þ
a; b; c 2 H.In [6], Gavrea generalized Buzano’s inequality to nonegative real exponents. In fact, he proved that
jha; xija � jhx; bijb 6 ta21t
b22jjxjj
aþb;
where t1 and t2 are expressed in terms of the constants a and b and vectors a; b.
2. A Cauchy–Schwarz setting to prove Buzano’s inequality
In this section we prove Buzano’s inequality using a Krein type inequality which is deduced only from Cauchy–Schwarz’sinequality. As mentioned in [9], for any given two vectors a; b 2 Cn � f0g, one can define
cos Uab ¼Reha; bijjajjjjbjj
and
cos Wab ¼jha; bijjjajjjjbjj :
In 1969, Krein proved the following inequality,
Uac 6 Uab þUbc
for any a; b; c 2 Cn.A proof of the above result can be found in [7,9]. Moreover, the main idea from the proof of Krein’s inequality is hidden
in the following Krein-type inequality,
Theorem 2.1. For any vectors a; b; c 2 ðH; h�; �iÞ we have the following inequality
jjajj2jhb; cij2 þ jjbjj2jhc; aij2 þ jjcjj2jha; bij2 6 jjajj2jjbjj2jjcjj2 þ 2jha; bihb; cihc; aij:
First proof. We will give a proof using only Cauchy–Schwarz’s inequality. Without loss of generality, we assume jjcjj– 0,otherwise the inequality is trivial. Now, for any complex scalar k, by the Cauchy–Schwarz’s inequality, we have
jja� kbjj2jjcjj2 P jha� kb; cij2
for all a; b; c 2 H. By expanding, this inequality is successively equivalent to
ðjjajj2 � 2kjha; bij þ k2jjbjj2Þjjcjj2 P jhc; aij2 � 2kjhc; aijjhb; cij þ k2jhc; aij2;
ðjjbjj2jjcjj2 � jhb; cij2Þk2 þ 2kðjha; bijjjcjj2 � jhc; aijjhb; cijÞ þ jjajj2jjcjj2 � jhc; aij2 P 0
for all complex scalars k. The last inequality can be viewed as a quadratic polynomial in k with its dominant coefficient non-negative (by Cauchy–Schwarz’s inequality), and thus, the discriminant D must be negative,
Dk ¼ 4ðjha; bijjjcjj2 � jhc; aijjhb; cijÞ2� 4ðjjbjj2jjcjj2 � jhb; cij2Þðjjajj2jjcjj2 � jhc; aij2Þ 6 0;
which is equivalent to
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 465
jha; bij2jjcjj4 � 2jha; bijjhb; cijjhc; aijjjcjj2 þ jhc; aij2jhb; cij2 6 jjajj2jjbjj2jjcjj4 � jhc; aij2jjbjj2jjcjj2
and the desired inequality follows.Second proof. Consider the following matrix
Dða; bÞ ¼jjcjj2 jhc; aji
jhb; cij jha; bij
24
35:
We will prove that
jdet Dða; bÞj2 6 jdet Dða; aÞj � jdet Dðb; bÞj:
Denote PðtÞ ¼ detðDða; aÞÞt2 � 2 detðDða; bÞÞt þ detðDðb; bÞÞ. A simple calculation yields
PðtÞ ¼ det Dðat � b; at � bÞ ¼ detjjcjj2 hta� b; ci
hc; ta� bi hta� b; ta� bi
" # !¼ jjcjj2jjat � bjj2 � jhat � b; cij2 P 0;
by Cauchy–Schwarz’s inequality. This implies that the quadratic polynomial is nonnegative and thus, its discriminant mustbe negative,
DPðtÞ ¼ jdet Dða; bÞj2 � jdet Dða; aÞj � jdet Dðb; bÞj 6 0;
which is equivalent to
ðjjcjj2jha; bij � jhc; aijjhc; bijÞ26 ðjjcjj2jjajj2 � jhc; aij2Þðjjcjj2jjbjj2 � jhb; cij2Þ
and by expanding our inequality follows. h
Remark. One can consider the Gram matrix
Gða; b; cÞ ¼
jjajj2 jha; bij jha; cij
jha; bij jjbjj2 jhb; cij
jhc; aij jhb; cij jjcjj2
26664
37775:
It is well-known that Gða; b; cÞ is positive semidefinite and thus det Gða; b; cÞ is nonnegative, and finally this implies
det Gða; b; cÞ ¼ jjajj2jjbjj2jjcjj2 þ 2jha; bijjhb; cijjhc; aij � ðjjajj2jhb; cij2 þ jjbjj2jhc; aij2 þ jjcjj2jha; bij2Þ
and the inequality follows. This idea appears also in [1].As a consequence we have the following interesting.
Corollary 2.2. For any vectors a; b; c 2 ðH; h�; �iÞ we have the following inequality
jjajj2jjbjj2 � jha; bij2 P jjajj2jhb;1ij2 þ jjbjj2jha;1ij2 � 2jha;1ijjhb;1ijjha; bij:
Proof. Take c to be the unit vector in Theorem 1.2. h
The above corollary served as application of several refinements of Heisenberg type inequalities related to uncertaintyprinciple in the paper [13].
As a consequence of Theorem 1.2 we have the general integral version.
Corollary 2.3. Let ðH;R;lÞ be a positive measure space and let f ; g;h 2 L2ðH;R;lÞ the Hilbert space of complex-valued 2� lintegrable functions defined on H. Then
ZHf 2ðtÞdlðtÞ
� � ZH
gðtÞhðtÞdlðtÞ����
����� �2
þZ
Hg2ðtÞdlðtÞ
� � ZH
hðtÞf ðtÞdlðtÞ����
����� �2
þZ
Hh2ðtÞdlðtÞ
� � ZH
f ðtÞgðtÞdlðtÞ����
����� �2
6
ZH
f 2ðtÞdlðtÞ� � Z
Hg2ðtÞdlðtÞ
� � ZH
g2ðtÞdlðtÞ� �
þ 2Z
Hf ðtÞgðtÞdlðtÞ
� � ZH
f ðtÞgðtÞdlðtÞ� � Z
Hf ðtÞgðtÞdlðtÞ
� ���������:
It might be useful to see that out of Theorem 3.2 one can get the following discrete version.
Corollary 2.4. If pi P 0; i ¼ 1;2; . . . ;n and ai; bi; ci are complex numbers, then
466 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
Xn
i¼1
pijaij2 ! Xn
i¼1
pibici
����������
!2
þXn
i¼1
pijbij2 ! Xn
i¼1
piciai
����������
!2
þXn
i¼1
pijaij2 ! Xn
i¼1
pibici
����������
!2
6
Xn
i¼1
pijaij2 ! Xn
i¼1
pijbij2 ! Xn
i¼1
pijcij2 !
þ 2Xn
i¼1
piaibi
����������Xn
i¼1
pibici
����������Xn
i¼1
piciai
����������:
In the case when H is the space of real-valued continuous functions defined on the interval ½a; b� endowed with the innerproduct hf ; gi :¼
R ba f ðxÞgðxÞdx, we have.
Corollary 2.5. Let f ; g;h be continuous real-valued functions on ½a; b�. Then, we have the following inequality
Z b
af 2ðtÞdt
! Z b
agðtÞhðtÞdt
!2
þZ b
ag2ðtÞdt
! Z b
ahðtÞf ðtÞdt
!2
þZ b
ah2ðtÞdt
! Z b
af ðtÞgðtÞdt
!2
6
Z b
af 2ðtÞdt
! Z b
ag2ðtÞdt
! Z b
ah2ðtÞdt
!þ 2
Z b
af ðtÞgðtÞdt
����������Z b
agðtÞhðtÞdt
����������Z b
ahðtÞf ðtÞdt
����������
!:
The discrete real version of Corollary 2.4 is given by.
Corollary 2.6. Given real numbers ai; bi; ci;1 6 i 6 n, then
Xn
i¼1
a2i
! Xn
i¼1
bici
!2
þXn
i¼1
b2i
! Xn
i¼1
aici
!2
þXn
i¼1
a2i
! Xn
i¼1
bici
!2
6
Xn
i¼1
a2i
! Xn
i¼1
b2i
! Xn
i¼1
c2i
!þ 2
Xn
i¼1
aibi
����������Xn
i¼1
bici
����������Xn
i¼1
ciai
����������:
In other words, if f ; g;h are random variables of the space L2ðX;lÞ, where l is a probability measure. Define the follwow-ing covariance between f ; g and h by
covhðf ; gÞ :¼ Eðf gÞEðh2Þ � EðfhÞEðghÞ;
where Eðf Þ ¼R
X fdl is the expectation of f. The variance of f is given by
varhðf Þ ¼ covhðf ; f Þ ¼ Eðjf j2ÞEðh2Þ � jEðfhÞj2:
Thus, after some computations, Theorem 1.2 can be rewritten as the following variance–covariance inequality:
jcovhðf ; gÞj2 6 varhðf ÞvarhðgÞ:
In the case when h ¼ 1 we obtain the classical variance–covariance inequality which is used in Statistics.Now, we are able to prove Buzano’s inequality which is given by.
Theorem 2.7. For any a; b; c 2 ðH; h�; �iÞ we have the inequality
jha; cihb; cij 6 12ðjjajjjjbjj þ jha; bijÞjjcjj2:
Proof. Theorem 2.1 can be rewritten as
jha; bij2
jjajj2jjbjj2þ jhb; cij
2
jjbjj2jjcjj2þ jhc; aij
2
jjcjj2jjajj26 1þ 2jha; bijjhb; cijjhc; aij
jjajj2jjbjj2jjcjj2;
which is equivalent to
ajjajj ;
bjjbjj
� ���������2
þ bjjbjj ;
cjjcjj
� ���������2
þ cjjcjj ;
ajjajj
� ���������2
6 1þ 2ajjajj ;
bjjbjj
� ��������� bjjbjj ;
cjjcjj
� ��������� cjjcjj ;
ajjajj
� ���������:
Denote u ¼ ajjajj ;v ¼ b
jjbjj and x ¼ cjjcjj. Our inequality rewrites as
jhu; vij2 þ jhx;uij2 þ jhx;vij2 6 1þ 2jhu; vihx; uihx; vij;
which is equivalent further with
jhx;uij2 þ jhx;vij2 6 1þ jhu; vij � jhu;vijð1þ jhu;vij � 2jhx;uijjhx;vijÞ:
By the arithmetic–geometric mean, we obtain
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 467
jhx;uij2 þ jhx;vij2 P 2jhx;uijjhx;vij
and we have
2jhx;uijjhx; vij 6 1þ jhu; vij � jhu;vijð1þ jhu; vij � 2jhx;uijjhx;vijÞ;
or
ð1� jhu; vijÞð1þ jhu; vij � 2jhx;uijjhx;vijÞP 0:Then either jhu;vij ¼ 1 which by Cauchy–Schwarz’s inequality, we have 2 ¼ 1þ jhu;vijP 2jhx;uijjhx;vij or 1 > jhu;vijwhich implies
1þ jhu;vij � 2jhx;uijjhx; vijÞP 0;
or
jhx;uijjhx;vijÞ 6 12ð1þ jhu; vijÞ
which is exactly what we needed to prove. h
Remark. In fact, one can prove even more:
jhx;uij2 þ jhx;vij2 6 1þ jhu; vij:
Indeed, at some point in the proof of Theorem 2.2 we derived the following inequality
jhx;uij2 þ jhx;vij2 6 1þ jhu; vij � jhu;vijð1þ jhu; vij � 2jhx;uijjhx;vijÞ:
By Theorem 2.2 we have
1þ jhu;vij � jhu;vijð1þ jhu;vij � 2jhx;uijjhx; vijÞ 6 1þ jhu; vij:
Putting u ¼ ajjajj ; v ¼ b
jjbjj and x ¼ cjjcjj we obtain the following inequality
ðjjbjjjha; cijÞ2 þ ðjjajjjhb; cijÞ2 6 jjajjjjbjjðjjajjjjbjj þ jha; bijÞjjcjj2
for all a; b; c 2 H.Other proofs and applications of this last inequality are given in the section that will follow and extensions and general-
izations to different settings will be given elsewhere.
3. Main result
Now, we are ready to state and prove the main result of this paper. But first, the following result will be crucial in provingour refinement of Buzano’s inequality.
Theorem 3.1. Given vectors u;v 2 H, with jjujj ¼ jjv jj ¼ 1 (thus u;v 2 Sn�1), then
supjjxjj¼1ðjhu; xij2 þ jhv ; xij2Þ ¼ 1þ jhu;vij:
First proof. (Trigonometric proof) Consider the unique vectorial representation x ¼ x0 þ x00 with x0 2 spanðu;vÞ andx00 ? spanðu; vÞ. Then 1 ¼ jjxjj2 ¼ hx0 þ x00; x0 þ x00i ¼ jjx0jj2 þ 2hx0; x00i þ jjx00jj2 ¼ jjx0jj2 þ jjx00jj2 (Pythagoras’ relation), sojjx0jj 6 1. Now, jhu; xij ¼ jhu; x0ij 6 jhu; yij where y ¼ 0 if x0 ¼ 0 and y ¼ x0
jjx0 jj if x0 – 0 (thus jjyjj ¼ 1). Similarlyjhv; xij ¼ jhv ; x0ij 6 jhv; yij. The maximum asked is therefore obtained when x 2 spanðu;vÞ.
Thus our vectorial problem has been transferred in the 2-dimensional space, with unit vectors u;v and x. The endpoints ofthe vectors ±u and ±v partition the circle S1 into four arcs, each of measure at most p (and possibly 0, when u = ±v); the end-point of x will fall into one of them. Let x be the measure of that arc, and a; b, with aþ b ¼ x, the measures of the arcs be-tween the endpoint of x and the ends of that arc. Then hu; xi2 þ hv ; xi2 ¼ cos2 aþ cos2 b, by the well-known geometricinterpretation of the dot-product (independently of the dimension of the space). But thenhu; xi2 þ hv ; xi2 ¼ cos2 aþ cos2 b ¼ 1þ 1
2 ðcos 2aþ cos 2bÞ ¼ 1þ cosðaþ bÞ cosða� bÞ 6 1þ j cos xj ¼ 1þ jhu;vij, with equal-ity in the obvious cases:
� when a ¼ b ¼ x=2, therefore when jhu; xij ¼ jhv ; xij, therefore x ¼ ð�u� vÞ=jj � u� v jj, where the signs are such that0 6 x < p=2;� when x ¼ p=2, therefore when hu;vi ¼ 0, i.e., u ? v , for all x 2 spanðu;vÞ.
Second Proof. (Quadratic forms) For jjxjj ¼ jjujj ¼ jjv jj ¼ 1 we have
0 6 jjkxþ luþ mv jj2 ¼ hkxþ luþ mv; kxþ luþ mvi ¼ k2 þ l2 þ m2 þ 2klhx;ui þ 2kmhx;vi þ 2lmhu;vi;
468 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
a quadratic form which takes non-negative values for any real parameters k;l; m, thus corresponding to a positive semide-fined matrix
1 jhx;uij jhx;vijjhx;uij 1 jhu;vijjhx;vij jhu;vij 1
264
375
The principal minors of order 1 are thus non-negative, which yields the semi-positivity of the norm; the principal minors oforder 2 are non-negative, which yields the Cauchy–Schwarz’s inequality, e.g., 1 P jhu;vij2; finally, also the determinant ofthe matrix is non-negative
D ¼ 1� ðjhu; vij2 þ jhx; uij2 þ jhx;vij2Þ þ 2jhu;vihx;uihx;vijP 0;
which can be arranged as jhx;uij2 þ jhx;vij2 6 1þ jhu;vij � jhu; vijð1þ jhu;vij � 2jhx;uihx;vijÞ. But jhx;uij2 þ jhx;vij2 P2jhx;uihx;vij, which plugged into the relation yields that ð1� jhu;vijÞð1þ jhu;vij � 2jhx;uihx; vijÞP 0. Then either1 ¼ jhu;vij, when 1þ jhu;vij ¼ 2 P 2jhx;uihx;vij (by Cauchy–Schwarz), or 1 > jhu;vij, which implies 1þ jhu;vijP2jhx;uihx;vij. Thus always hx;ui2 þ hx;vi2 6 1þ jhu;vij.
Third proof. (Lagrange multipliers) Define
Lðx; kÞ ¼ hu; xi2 þ hv; xi2 � kðjjxjj2 � 1Þ
and consider the system
@L@xi¼ 2uihu; xi þ 2v ihv ; xi � 2kxi ¼ 0
for 1 6 i 6 n, and
@L@k¼ jjxjj2 � 1 ¼ 0:
Now
0 ¼ 12
Xn
i¼1
xi@L@xi¼ hu; xi
Xn
i¼1
uixi þ hv ; xiXn
i¼1
v ixi � kXn
i¼1
x2i ¼ hu; xi
2 þ hv ; xi2 � kjjxjj2 ¼ hu; xi2 þ hv; xi2 � k;
so k needs be, computed at the critical point(s), the very expression we are considering (of no relevance, in fact). On the otherhand,
0 ¼ 12
Xn
i¼1
ui@L@xi¼ hu; xi
Xn
i¼1
u2i þ hv ; xi
Xn
i¼1
v iui � kXn
i¼1
xiui ¼ hu; xijjujj2 þ hv ; xihu;vi � khu; xi
¼ hu; xi þ hu; vihv ; xi � khu; xi
and similarly for v, thus reaching the system of two equations (in the variables hu; xi and hv ; xi)
ð1� kÞhu; xi þ hu; vihv; xi ¼ 0;hu;vihu; xi þ ð1� kÞhv; xi ¼ 0:
The determinant D of the matrix of this particular system is D ¼ ð1� kÞ2 � hu;vi2, and, if not null, the only solution is thetrivial hu; xi ¼ hv ; xi ¼ 0, when our expression reaches an evident minimum, equal to zero (therefore since then k ¼ 0, thismeans hu;vi – � 1, which translates in u – � v , and x ? spanðu;vÞ). We are therefore interested in D ¼ 0 (for all other crit-ical points), leading to k ¼ 1� hu;vi, thus k ¼ 1þ jhu;vij at a maximum point, and k ¼ 1� jhu;vij at a critical point. There aresome particular cases.
When u ¼ �v , thus hu;vi ¼ �1, the situation is simple. We have a maximum of 2 when x ¼ �u, and a minimum of 0 whenx ? u.
When u ? v , thus hu;vi ¼ 0, then k (thus the expression) is 1 at the maximum points, for all x 2 spanðu;vÞ, and k (thus theexpression) is 0 at the minimum points, for all x ? spanðu;vÞ. h.
The main result of this paper is given by
Theorem 3.2. Let ðH; h�; �iÞ be a real or complex inner product space and a; b; c 2 H. Then
ðjjajjjhb; cijÞ2 þ ðjjbjjjha; cijÞ2 6 jjajjjjbjjðjjajjjjbjj þ jha; bijÞjjcjj2:
Proof. For any not-null a; b; c, let us take u ¼ ajjajj ;v ¼ b
jjbjj ; x ¼ cjjcjj in Theorem 2.1 and, therefore hu; xi2 ¼ 1
jjajj�jjcjj2ha; ci2;
hv ; xi2 ¼ 1jjbjj�jjcjj2
hb; ci2; hu;vi ¼ ha;bijjajj�jjbjj, so the proven relation becomes
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 469
1jjajj jha; cij
2 þ 1jjbjj jhb; cij
26 1þ jha; bij
jjajj � jjbjj
� �jjcjj2;
equivalent with the required inequality, also true for a; b or c equal to zero. h
Remark. We can continue by Arithmetic–Geometric mean to obtain
jhu; xi � hv ; xij 6 12jhu; xij2 þ jhv; xij2� �
612
1þ jhu; vijð Þ:
A simple computation now yields
jha; ci � hb; cij 6 12ðjjajj � jjbjj þ jha; bijÞjjcjj2;
also true for a; b or c equal to zero, which is exactly Buzano’s inequality.As a consequence of Theorem 3.2 we have the general integral version.
Corollary 3.3. Let ðH;R;lÞ be a positive measure space and let f ; g;h 2 L2ðH;R;lÞ the Hilbert space of complex-valued 2� lintegrable functions defined on H. Then
ZHf 2ðtÞdlðtÞ
� � ZH
gðtÞhðtÞdlðtÞ����
����� �2
þZ
Hg2ðtÞdlðtÞ
� � ZH
hðtÞf ðtÞdlðtÞ����
����� �2
6
ZH
f 2ðtÞdlðtÞ� �1=2 Z
Hg2ðtÞdlðtÞ
� �1=2
�Z
Hf 2ðtÞdlðtÞ
� �1=2 ZH
g2ðtÞdlðtÞ� �1=2
þZ
Hf ðtÞgðtÞdlðtÞ
��������
!ZH
h2ðtÞdlðtÞ:
It might be useful to see that out of Theorem 3.2 one can get the following discrete version.
Corollary 3.4. If pi P 0; i ¼ 1;2; . . . ;n and ai; bi; ci are complex numbers, then
Xn
i¼1
pijaij2 ! Xn
i¼1
pibici
����������
!2
þXn
i¼1
pijbij2 ! Xn
i¼1
piciai
����������
!2
6
Xn
i¼1
pijaij2 !1=2 Xn
i¼1
pijbij2 !1=2 Xn
i¼1
pijcij2 !
�Xn
i¼1
pijaij2 !1=2 Xn
i¼1
pijbij2 !1=2
þXn
i¼1
piaibi
����������
0@
1A:
In the case when H is the space of real-valued continuous functions defined on the interval ½a; b� endowed with the innerproduct hf ; gi :¼
R ba f ðxÞgðxÞdx, we have.
Corollary 3.5. Let f ; g;h be continuous real-valued functions on ½a; b�. Then, we have the following inequality
Z b
af 2ðtÞdt
! Z b
agðtÞhðtÞdt
!2
þZ b
ag2ðtÞdt
! Z b
ahðtÞf ðtÞdt
!2
6
Z b
af 2ðtÞdt
!1=2 Z b
ag2ðtÞdt
!1=2 Z b
ah2ðtÞdt �
Z b
af 2ðtÞdt
!1=2 Z b
ag2ðtÞdt
!1=2
þZ b
af ðtÞgðtÞdt
����������
0@
1A:
The discrete real version of Corollary 2.4 is given by.
Corollary 3.6. Given real numbers ai; bi; ci;1 6 i 6 n, then
Xn
i¼1
a2i
! Xn
i¼1
bici
!2
þXn
i¼1
b2i
! Xn
i¼1
aici
!2
6
Xn
i¼1
a2i
! Xn
i¼1
b2i
! Xn
i¼1
c2i
!þXn
i¼1
aibi
����������Xn
i¼1
a2i
!1=2 Xn
i¼1
b2i
!1=2 Xn
i¼1
c2i
!:
4. Some elementary applications
In [16] Wang and Zhang proved an inequality for unitary matrices. Now, we are able to give other trace inequalities forunitary matrices. A similar inequality was obtained by Zhang [17] but with a slightly different method.
Theorem 4.1. For a complex n� n matrix X denote mðXÞ the algebraic mean of its eigenvalues. For any two unitary n� n complexmatrices U;V, we have
470 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
jmðUÞj2 þ jmðVÞj2 þ jmðUVÞj2 6 1þ 2jmðUÞjjmðVÞjjmðUVÞj
and
jmðUÞj2 þ jmðVÞj2 6 1þ jmðUVÞj:
Proof. Let c ¼ 1ffiffinp I; a ¼ 1ffiffi
np U and b ¼ 1ffiffi
np V where U and V are unitary matrices. Clearly, jjajj ¼ jjbjj ¼ jjcjj ¼ 1 and moreover
ha; bi ¼ 1n
trðUV�Þ ¼ 1n
trðUVÞ;
hb; ci ¼ 1n
trðUÞ; hc; ai ¼ 1n
trðVÞ:
Now, by Theorems 2.1 and 3.2 we have
1n2 ðjtrðUÞj2 þ jtrðVÞj2 þ jtrðUVÞj2Þ 6 1þ 2
n3 ðjtrðUÞjjtrðVÞjjtrðUVÞjÞ
and
1njtrðUÞj
� �2
þ 1njtrðVÞj
� �2
6 1þ 1njtrðUVÞj:
Taking into account that mðXÞ ¼ 1n trðXÞ, we obtain exactly our inequalities. h
Remark. In general, for the Hilbert–Schmidt inner product, hX;Yi :¼ trðX�YÞ, for any complex square matrices A;B and C, wehave the following inequalities
trðAA�ÞjtrðBC�Þj2 þ trðBB�ÞjtrðCA�Þj2 þ trðCC�ÞjtrðCA�Þj2 6 trðAA�ÞtrðBB�ÞtrðCC�Þ þ 2jtrðAB�ÞtrðBC�ÞtrðCA�Þj
and
trðAA�Þ � jtrðB�CÞj2 þ trðBB�Þ � jtrðC�AÞj2 6 tr1=2ðAA�Þ � tr1=2ðBB�ÞtrðCC�Þ tr1=2ðAA�Þ � tr1=2ðBB�Þ þ jtrðA�BÞj� �
:
Theorem 4.2. For reals ai; bi; ci; i ¼ 1;2; . . . ;n such thatPn
i¼1aibi ¼ 0, we have
Xn
i¼1
a2i
! Xn
i¼1
c2i
! ! Xn
i¼1
b2i
! Xn
i¼1
c2i
! !P 4
Xn
i¼1
aici
!2 Xn
i¼1
bici
!2
and
Xn
i¼1
a2i
! Xn
i¼1
c2i
! ! Xn
i¼1
ciai
!2
þXn
i¼1
b2i
! Xn
i¼1
c2i
! ! Xn
i¼1
bici
!2
P 4Xn
i¼1
aici
!2 Xn
i¼1
bici
!2
:
Proof. By Arithmetic–Geometric mean inequality, we have
Xn
i¼1
a2i
! Xn
i¼1
bici
!2
þXn
i¼1
b2i
! Xn
i¼1
ciai
!2
P 2Xn
i¼1
bici
���������� �Xn
i¼1
ciai
���������� �
Xn
i¼1
a2i
!1=2
�Xn
i¼1
b2i
!1=2
:
By Corollary 3.6 and taking into account thatPn
i¼1aibi ¼ 0, we obtain
Xn
i¼1
a2i
! Xn
i¼1
b2i
! Xn
i¼1
c2i
!P 2
Xn
i¼1
bici
����������Xn
i¼1
ciai
���������� �
Xn
i¼1
a2i
!1=2 Xn
i¼1
b2i
!1=2
;
which is equivalent with the desired inequality. As for the second inequality, applying again arithmetic–geometric meaninequality, we have
Xn
i¼1
a2i
! Xn
i¼1
c2i
! Xn
i¼1
aici
!2
þXn
i¼1
b2i
! Xn
i¼1
c2i
! Xn
i¼1
bici
!2
P 2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i¼1
a2i
! Xn
i¼1
b2i
! Xn
i¼1
c2i
!2vuut Xn
i¼1
ciai
! Xn
i¼1
bici
!
and by using the first inequality we finally obtain the second one. h
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 471
Theorem 4.3. For a matrix X with real entries, let sðXÞ be the sum of its entries. If A;B and C are n� n real matrices, then
sðCCTÞðsðAATÞsðBBTÞ � sðABTÞsðBATÞÞP sðAATÞsðBCTÞsðCBTÞ þ sðBBTÞsðACTÞsðCT AÞ � sðACTÞsðCBTÞsðBATÞ
� sðBCTÞsðCATÞsðABTÞ
and
sðAT AÞsðBT CÞsðBCTÞ þ sðBT BÞsðCT AÞsðAT CÞ 66 s1=2ðAT AÞs1=2ðBT BÞsðCT CÞðs1=2ðAT AÞs1=2ðBT BÞ þ sðAT BÞÞ:
Proof. For a square matrix A 2MnðRÞ; A ¼ ðaijÞ16i;j6n we consider sðAÞ ¼P
16i;j6naij, the sum of all the elements of the matrix.We denote AT the transpose of the matrix A.
We want to express sðABTÞ in terms of the elements of A and B. We have
A � BT ¼
a11 a12 . . . a1n
a21 a22 . . . a2n
..
. ... . .
. ...
an1 an2 . . . ann
0BBBB@
1CCCCA �
b11 b21 . . . bn1
b12 b22 . . . bn2
..
. ... . .
. ...
b1n b2n . . . bnn
0BBBB@
1CCCCA:
If we look at the element a1i, this element appears in the product matrix multiplied only by the elements of the i-th line of BT ,and this means that the sum of all the terms that contain a1i is a1iðb1i þ b2i þ . . .þ bniÞ. We repeat the same argument for allthe elements for the i-th column of A, and we get that the sum of all the terms which contain one of them isða1i þ a2i þ . . .þ aniÞðb1i þ b2i þ . . .þ bniÞ. Writing this for all the columns of A, we get that
sðABTÞ ¼Xn
i¼1
ða1i þ a2i þ . . .þ aniÞðb1i þ b2i þ . . .þ bniÞ:
Thus, we have xi ¼Pn
i¼1a1i; yi ¼Pn
i¼1b1i and zi ¼Pn
i¼1c1i.With xi and yi; zi considered above, we have
Xn
i¼1
x2i ¼ sðAATÞ;
Xn
i¼1
y2i ¼ sðBBTÞ
andPn
i¼1z2i ¼ sðCCTÞ. On the other hand,Pn
i¼1xiyi ¼ sðABTÞ ¼ sðAT BÞ;Pn
i¼1yizi ¼ sðBCTÞ ¼ sðCT BÞ andPn
i¼1zixi ¼ sðACTÞ ¼ sðCT AÞ. Now, by Corollaries 2.6 and 3.6 weobtain the desired inequalities. h
Remark. If we put C ¼ I in the first inequality we recover [12].
Theorem 4.4. If m; n are positive integers and A; B and C are m� n real matrices, then
ðdetðABTÞÞ2
detðCCTÞ þ ðdetðBCTÞÞ2 detðAATÞ þ ðdetðCATÞÞ2
detðBBTÞ
6 detðAATÞdetðBBTÞdetðCCTÞ þ 2jdetðABTÞdetðBCTÞdetðCT AÞj:
and
detðAT AÞdetðBT CÞdetðBCTÞ þ detðBT BÞdetðCT AÞdetðAT CÞ 6 ðdetðAT AÞÞ1=2ðdetðBT BÞÞ1=2
detðCT CÞ
� ððdetðAT AÞÞ1=2ðdetðBT BÞÞ1=2 þ detðAT BÞÞ:
Proof. If m > n then ABT;BCT ;CAT and AAT
; BBT ;CCT are m�m matrices with rank at most m, so our inequality is obvious. So,we may assume wothout loss of generality that m 6 n. Consider S ¼ ðj1; j2; . . . ; jmÞ : 1 6 j1 < � � � < jm 6 nf g. Now, for a m by nmatrix X with real entries and w ¼ ðj1; j2; . . . ; jmÞ 2 S denote by Xw the m�m matirx whose columns are columns j1; j2; . . . ; jm
of X. For X;Y two m� n matrices with real entries, by the Cauchy-Binet formula, we have
detðXYTÞ ¼Xw2S
det Xw det Yw;
n� �
n� �
with respect to the inner product h�; �i : R m � R m ! R. Now, if a ¼ ðdet AwÞw2S ; b ¼ ðdet BwÞw2S and c ¼ ðdet CwÞw2S ,then by applying Theorems 2.1 and 3.2, we obtain
472 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
Xw2S
det Aw det Bw
!2 Xw2Sðdet CwÞ2
!þ
Xw2S
det Bw det Cw
!2 Xw2Sðdet AwÞ2
!þ
Xw2S
det Cw det Aw
!2 Xw2Sðdet BwÞ2
!
6
Xw2Sðdet CwÞ2
! Xw2Sðdet CwÞ2
! Xw2Sðdet CwÞ2
!þ 2
Xw2S
det Aw det Bw
! Xw2S
det Aw det Bw
! Xw2S
det Aw det Bw
!����������
and
Xw2Sðdet AwÞ2
! Xw2S
det Bw det Cw
!2
þXw2Sðdet BwÞ2
! Xw2S
det Cw det Aw
!2
6
Xw2Sðdet AwÞ2
!����������
1=2 Xw2Sðdet BwÞ2
!����������1=2
�Xw2Sðdet CwÞ2
!�
Xw2Sðdet AwÞ2
!����������
1=2 Xw2Sðdet BwÞ
!����������þ
Xw2S
det Aw det Bw
!����������
0@
1A
and the conclusion follows immediately. h
An interesting particular case was one of the hardest problems given at Romanian selection test for International Math-ematical Olympiad in 2007.
Theorem 4.5. For n P 2 positive integer, ai; bi reals, i ¼ 1;2; . . . ;n, such that
Xn
i¼1
a2i ¼
Xn
i¼1
b2i ¼ 1;
Xn
i¼1
aibi ¼ 0;
we have
Xn
i¼1
ai
!2
þXn
i¼1
bi
!2
6 n:
Proof. Take ci ¼ 1; i ¼ 1;2; . . . ; n in Corollary 3.6. h
The integral analog of Theorem 4.2 is given by.
Theorem 4.6. Let f ; g;h be contionuous real-valued functions on ½0;1� satisfying the conditionR½0;1� fg ¼ 0. Then, we have
Z½0;1�
f 2Z½0;1�
g2Z½0;1�
h2
!2
P 4Z½0;1�
hf
!2 Z½0;1�
gh
!2
and
Z½0;1�
f 2Z½0;1�
h2
! Z½0;1�
hf
!2
þZ½0;1�
g2Z½0;1�
h2
! Z½0;1�
gh
!2
P 4Z½0;1�
hf
!2 Z½0;1�
gh
!2
:
Proof. Let us consider by Cð½0;1�Þ be the space of continuous functions endowed with the inner product
hf ; gi :¼Z½0;1�
f ðxÞgðxÞdx
and the associated norm jjf jj ¼R½0;1� f
2ðxÞdx� �1=2
. Our inequalities can be rewritten as
jjf jj2jjhjj2 � jjgjj2jjhjj2 P 4jhg;hij2 � jhh; f ij2 ð5Þ
and
jjf jj2jjhjj2jhf ;hij2 þ jjgjj2jjhjj2jhg; hij2 P 4jhg;hij2 � jhh; f ij2; ð6Þ
by the Corollary 3.4, we have
ðjjf jjhg;hiÞ2 þ ðjjgjjhf ;hiÞ2 6 jjf jjjjgjjðjjf jjjjgjj þ jhf ; gijÞjjhjj2
for any f ; g;h 2 Cð½0;1�Þ. On the other hand, by Arithmetic–Geometric mean inequality we obtain
ðjjf jjhg;hiÞ2 þ ðjjgjjhf ;hiÞ2 P 2jjf jjjjgjj � jhg; hijjhh; f ij:
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 473
Combining this inequality with the one obtained from the Corollary 2.5 and taking into account that hf ; gi ¼ 0, we get
jjf jj2jjgjj2jjhjj2 P 2jjf jjjjgjj � jhg;hijjhh; f ij;
which is finally equivalent to
jjf jj2jjgjj2jjhjj4 P 4jhg;hij2jhh; f ij2
and thus the first inequality is proved. For the second inequality, we shall use the first inequality. Hence, again by Arithme-tic–Geometric mean inequality we have
jjf jj2jjhjj2jhf ;hij2 þ jjgjj2jjhjj2jhg; hij2 P 2jjf jjjjhjj � jhh; f ij � jjgjjjjhjj � jhg;hij:
Now, we only need to see that
2jjf jjjjhjj � jhh; f ij � jjgjjjjhjj � jhg; hijP 4jhg;hij2jhh; f ij2:
This last inequality finally reduces to
jjf jjjjgjjjjhjj2 P 2jhg; hijjhh; f ij
and this exactly inequality (5) written in the form
jjf jj2jjgjj2jjhjj4 P 4jhg;hij2jhh; f ij2: �
Remark. For h 1, we obtain [10].
Theorem 4.7. [[11]] Let n be a positive integer, and write a vector x 2 Rn as ðx1; x2; . . . ; xnÞ. For x; y; a; b 2 Rn let
½x; y�a;b ¼X
16i;j6n
xiyj minðai; bjÞ:
Then, for x; y; z; a; b; c 2 Rnwith nonnegative entries,
½x; x�a;a � ½y; z�2b;c þ ½y; y�b;b � ½z; x�
2c;a þ ½z; z�c;c � ½x; y�
2a;b 6 ½x; x�a;a � ½y; y�b;b � ½z; z�c;c þ 2½x; y�a;b � ½y; z�b;c � ½z; x�c;a:
and
½x; x�a;a � ½y; z�2b;c þ ½y; y�b;b � ½z; x�
2c;a 6 ½x; x�
1=2a;a � ½y; y�
1=2b;b � ½z; z�c;c � ½x; x�
1=2a;a � ½y; y�
1=2b;b þ ½x; y�a;b
� �:
Proof. Our inequalities can be restated as
X16i;j6n
xixj minðai; ajÞ !
�X
16i;j6n
yizj minðbi; cjÞ !2
þX
16i;j6n
yiyj minðbi; bjÞ !
�X
16i;j6n
zixj minðci; ajÞ !2
þX
16i;j6n
zizj minðci; cjÞ !
�X
16i;j6n
yixj minðbi; cjÞ !2
6
X16i;j6n
xixj minðai; ajÞ !
�X
16i;j6n
yiyj minðbi; bjÞ !
�X
16i;j6n
zizj minðci; cjÞ !
þ 2X
16i;j6n
xiyj minðai; bjÞ !
�X
16i;j6n
yizj minðbi; cjÞ !
�X
16i;j6n
zixj minðci; ajÞ !
and
X16i;j6n
xixj minðai; ajÞ !
�X
16i;j6n
yizj minðbi; cjÞ !2
þX
16i;j6n
yiyj minðbi; bjÞ !
�X
16i;j6n
zixj minðci; ajÞ !2
6
X16i;j6n
xixj minðai; ajÞ !1=2 X
16i;j6n
yiyj minðbi; bjÞ !1=2
�X
16i;j6n
xixj minðai; ajÞ !1=2 X
16i;j6n
yiyj minðbi; bjÞ !1=2
0@
þX
16i;j6n
xiyj minðai; bjÞ!�
X16i;j6n
zizj minðci; cjÞ !
:
Let vA be the characteristic function of an arbitrary set A. Consider the functions f ; g;h : ½0;1Þ ! R defined by
f ðxÞ ¼Xn
i¼1
xiv½0;ai �ðxÞ;
474 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
gðxÞ ¼Xn
i¼1
yiv½0;bi �ðxÞ;
hðxÞ ¼Xn
i¼1
ziv½0;ci �ðxÞ:
We have
Z 10f 2ðxÞdx ¼
X16i;j6n
xixj
Z 1
0v½0;ai �ðxÞv½0;bi �ðxÞdx ¼
X16i;j6n
xixj minðai; ajÞ:
Analogously, we obtain
Z 10g2ðxÞdx ¼
X16i;j6n
yiyj minðbi; bjÞ;Z 1
0h2ðxÞdx ¼
X16i;j6n
zizj minðci; cjÞ;
as well as
Z 10f ðxÞgðxÞdx ¼
X16i;j6n
xiyj minðai; bjÞ;
Z 1
0gðxÞhðxÞdx ¼
X16i;j6n
yizj minðbi; cjÞ
and
Z 10gðxÞhðxÞdx ¼
X16i;j6n
zixj minðci; ajÞ:
Finally, by applying Theorem 2.1 and Corollary 3.3 with the inner product given by hf ; gi :¼R1
0 f ðxÞgðxÞdx, with f ; gdefined as above, we derive the inequalities
Z 10f 2ðxÞdx
Z 1
0gðxÞhðxÞdx
� �2
þZ 1
0g2ðxÞdx
Z 1
0hðxÞf ðxÞdx
� �2
þZ 1
0h2ðxÞdx
Z 1
0f ðxÞgðxÞdx
� �2
6
Z 1
0f 2ðxÞdx
Z 1
0g2ðxÞdx
Z 1
0h2ðxÞdxþ 2
Z 1
0hðxÞf ðxÞdx
Z 1
0hðxÞf ðxÞdx
Z 1
0hðxÞf ðxÞdx
��������
and
Z 10f 2ðxÞdx
Z 1
0gðxÞhðxÞdx
� �2
þZ 1
0g2ðxÞdx
Z 1
0hðxÞf ðxÞdx
� �2
6
Z 1
0f 2ðxÞdx
� �1=2 Z 1
0g2ðxÞdx
� �1=2 Z 1
0h2ðxÞdx
� ��
Z 1
0f 2ðxÞdx
� �1=2 Z 1
0g2ðxÞdx
� �1=2
þZ 1
0f ðxÞgðxÞdx
��������
!
and, now the conclusion is obvious. h
Remark. These kind of min–max inequalities for vectors arise in coding of messages. A list of such inequalities is provided in[3].
Theorem 4.8. For a matrix A 2 MnðRÞ, let fA :MnðRÞ ! R,
fAðXÞ ¼Xn
i;j¼1
bijxij;
where B ¼ ½bij�16i;j6n ¼ AAT . Then, we have
fAðXXTÞ � f 2A ðYZTÞ þ fAðYYTÞ � f 2
A ðZXTÞ þ fAðZZTÞ � f 2A ðXYTÞ 6 fAðXXTÞ � fAðYYTÞ � fAðZZTÞ þ 2f AðXYTÞ � fAðYZTÞ � fAðZXTÞ
and
fAðXXTÞ � f 2A ðYZTÞ þ fAðYYTÞ � f 2
A ðZXTÞ 6 f 1=2A ðXXTÞ � f 1=2
A ðYYTÞ � fAðZZTÞ f 1=2A ðXXTÞ � f 1=2
A ðYYTÞ þ fAðXYTÞ� �
:
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 475
Proof. We define the inner product
hX;Yi :¼ fAðXYTÞ ¼X
16i;j;k6n
bi;jxi;kyj;k ¼Xn
k¼1
columnkðXÞT BcolumnkðYÞ:
Moreover, since B is positive semidefinite, we have
fAðXXTÞ ¼Xn
k¼1
columnkðXÞT BcolumnkðXÞ
and clearly, fAðXXTÞ ¼ 0 implies that columnkðXÞ ¼ 0 for every k, that is X ¼ 0. Now, by Theorems 2.1 and 3.2, we derive ourinequalities. h
Theorem 4.9. Let A ¼ ½aij�16i;j6n 2 MnðRÞ, be a symmetric matrix with aii >Pn
j¼1;j–ijaijj and xi; yi; zi; i ¼ 1;2; . . . ;n are nonnegativereal numbers. Then, we have
X16i;j6n
aijxixj
! X16i;j6n
aijyizj
!2
þX
16i;j6n
aijyiyj
! X16i;j6n
aijzixj
!2
þX
16i;j6n
aijzizj
! X16i;j6n
aijxiyj
!2
6
X16i;j6n
aijxixj
! X16i;j6n
aijyiyj
! X16i;j6n
aijzizj
!þ 2
X16i;j6n
aijxiyj
����������X
16i;j6n
aijyizj
����������X
16i;j6n
aijzixj
����������
and
X16i;j6n
aijxixj
! X16i;j6n
aijyizj
!2
þX
16i;j6n
aijyiyj
! X16i;j6n
aijzixj
!2
6
X16i;j6n
aijxixj
!1=2 X16i;j6n
aijyiyj
!1=2 X16i;j6n
aijzizj
!�
X16i;j6n
aijxixj
!1=2 X16i;j6n
aijyiyj
!1=2
þX
16i;j6n
aijxiyj
����������
0@
1A:
Proof. For vectors x; y 2 Rn, define the function f : Rn � Rn ! R,
f ðx; yÞ :¼X
16i;j6n
aijxiyj:
Clearly, f ðax; byÞ ¼ a � bf ðx; yÞ; f ðx; yÞ ¼ f ðy; xÞ and f ðxþ y; zÞ ¼ f ðx; zÞ þ f ðy; zÞ. The only property that we need to verify isf ðx; xÞ > 0 for x – 0. We have
f ðx; xÞ ¼X
16i;j6n
aijxiyj ¼Xn
i¼1
aiix2i þ 2
Xi<j
aijxixj >Xn
i¼1
jaijj� �
x2ij þ 2
Xi<j
aijxixj ¼Xi<j
jaijjðx2i þ x2
j Þ þ 2Xi<j
aijxixj
P 2Xi<j
jaijjjxijjxjj þ 2Xi<j
aijxixj ¼ 2Xi<j
ðjaijjjxijjxjj þ aijxixjÞP 0:
Thus, f ðx; yÞ is an inner product on Rn � Rn. Now, by applying Theorems 2.1 and 3.2, we obtain our result. h
Remark. More generally, one can prove that if A is a positive definite symmetric matrix with real entries, then the functionf : Rn � Rn ! R,
f ðx; yÞ ¼X
16i;j6n
aijxiyj;
defines an inner product on Rn. Thus, Theorem 4.8 holds in a slightly general setting.One can consider some classical examples of positive definite and symmetric matrices. For instance, the Hilbert matrix
A1 ¼ ðaijÞ16i;j6n ¼ 1iþj�1
h i16i;j6n
which is positive definite. This is equivalent with showing that for positive xi; i ¼ 1;2; . . . ;n
we haveP
16i;j6nxixj
iþj�1 > 0. For instance, one ca easily prove that
X16i;j6n
xixj
iþ j� 1¼Z 1
0
Xn
i¼1
xiti�1
!2
dt > 0;
476 C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477
when f1ðtÞ ¼Pn
i¼1xiti�1� �2
is not the zero function. The same for Cauchy matrix, A2 ¼ ðaijÞ16i;j6n ¼ 1iþj
h i16i;j6n
, one can provethat
X16i;j6n
xixj
iþ j¼Z 1
0
Xn
i¼1
xie�it
!2
dt > 0;
when f2ðtÞ ¼Pn
i¼1xie�it� �2 is not the zero function.
On the other hand, on the axis of real numbers we consider the intervals Ii; i ¼ 1;2; . . . ;n and denote by aij be the length of
the interval Ii \ Ij, for i; j ¼ 1;2; . . . ; n. One can prove that the matrix A3 ¼ ðaijÞ16i;j6n ¼ lengthðIi \ IjÞ� �
16i;j6n is positive definite.
Indeed, denote I ¼Sn
i¼1Ii and for J I consider the characteristic function of the set J;vJ : I ! f0;1g. If J is an interval, then the
length of J is given by lengthðJÞ ¼R
I vJðxÞdx ¼R
I v2J dx. On the other hand,
aij ¼Z
IvIiðxÞvIj
ðxÞdx
and thus,
X16i;j6n
aijxixj ¼Z
I
X16i;j6n
vIiðxÞvIj
ðxÞxixj
!dx ¼
ZI
Xn
i¼1
xivIiðxÞ
!2
dx > 0
for positive xi; i ¼ 1;2; . . . ;n.The same idea works when A4 ¼ ðaijÞ16i;j6n ¼ areaðDi \ DjÞ, where Di; i ¼ 1;2; . . . ;n are discs in the Euclidian plane. Indeed,
we have the characteristic function of the set Di;vDi: R2 ! f0;1g, and moreover
areaðDiÞ ¼Z Z
R2vDiðx; yÞdxdy ¼
Z ZR2
v2Diðx; yÞdxdy
and
areaðDi \ DjÞ ¼Z Z
R2vDiðx; yÞvDj
ðx; yÞdxdy:
Now, we have
X16i;j6n
aijxixj ¼Z Z
R2
X16i;j6n
xivDiðx; yÞxjvDj
ðx; yÞdxdy ¼Z Z
R2
Xn
i¼1
xivDiðx; yÞ
!2
dxdy > 0:
Last but not least, we end the note with an elementary inequality with its remniscents in the theory of improper integrals.
Theorem 4.10. For all nonnegative reals a; b; c we have the inequalities
ab
ðaþ bÞ2þ bc
ðbþ cÞ2þ ca
ðc þ aÞ26
14þ 4abcðaþ bÞðbþ cÞðc þ aÞ
and
bc
ðbþ cÞ2þ ca
ðc þ aÞ26
18abc
þ 14cðaþ bÞ :
Proof. The key is the following identity:
1aþ b
¼Z 1
0e�ðaþbÞxdx:
Define the inner product hf ; gi :¼R1
0 f ðxÞgðxÞ, where f ðxÞ ¼ e�ax and gðxÞ ¼ e�bx. A short computation shows thatR10 e�2axdx ¼ 1
2a. So by Theorem 2.1 we have
1
ðbþ cÞ2aþ 1
ðc þ aÞ2bþ 1
ðaþ bÞ2c6
18abc
þ 2ðaþ bÞðbþ cÞðc þ aÞ ;
which is equivalent to the first inequality.
C. Lupu, D. Schwarz / Applied Mathematics and Computation 231 (2014) 463–477 477
Now, by Corolarry 3.3, we obtain
1
aðbþ cÞ2� 1
bðc þ aÞ26
1
4abc2
12abþ 1
aþ b
� �;
which is our second inequality. h
Acknowledgements
The first author would like to thank Politehnica University of Bucharest for hospitality during May–June 2012 when partof this work was initiated. The authors would also like to thank Beniamin Bogos�el for crucial observations which improvedthe initial version of this paper.
References
[1] M. Arav, F. Hall, Z. Li, A Cauchy–Schwarz inequality for triples of vectors, Math. Inequal. Appl. 11 (2008) 629–634.[2] M.L. Buzano, Generalizzatione della diseguagliazza di Cauchy–Schwarz, Rend. Sem. Mat. Univ. Politech. Torino 31 (1971/73) 405–409.[3] J.E. Cohen, J.H.B. Kemperman, G. Zbaganu, Elementary inequalities that involve two nonnegative vectors or functions, Proc. Nat. Acad. Sci. 101 (2004)
15018–15022.[4] S.S. Dragomir, Advances in Inequalities of the Schwarz, Triangle and Heisenberg type in Inner Product Spaces, Nova Science Publishers, New York, 2007.[5] M. Fuji, F. Kubo, Buzano’s inequality and bounds for roots of algebraic equations, Proc. Am. Math. Soc. 117 (1993) 359–361.[6] I. Gavrea, An extension of Buzano’s inequality in inner product spaces, Math. Inequal. Appl. 10 (2007) 281–285.[7] K.E. Gustafson, D.K.M. Rao, Numerical Range, Springer Verlag, 1997.[8] V.I. Istratescu, Inner Product Spaces Structures, D. Riedel Publishing, Dordrecht Holland, 1987.[9] M. Lin, Remarks on Krein’s inequality, Math. Intelligencer 34 (2012) 3–4.
[10] C. Lupu, T. Lupu, Problem 11360, Am. Math. Mon. 115 (2008) 248.[11] C. Lupu, D. Schwarz, Problem 11632, Am. Math. Mon. 119 (2012) 365–366.[12] B. Bogosel, C. Lupu, Problem 11496, Am. Math. Mon. 117 (2010) 370.[13] G. Mingzhe, On Heisenberg inequality, J. Math. Anal. Appl. 234 (1999) 727–734.[14] U. Richard, Sur des inegalites du type Wirtinger et leurs application aux equations differentiales ordinaires, Colloq. Anal. Rio Janeiro (1972) 233–244.[15] J.M. Steele, The Cauchy–Schwarz Master Class, Cambridge University Press, USA, 2004.[16] B. Wang, F. Zhang, A trace inequality for unitary matrices, Am. Math. Mon. 101 (1994) 453–455.[17] F. Zhang, Positivity of matrices with generalized matrix functions, Acta Math. Sinica 28 (2012) 1779–1786.