M athematical

I nequalities
& A pplications
Volume 14, Number 2 (2011), 247–270

REAL AND COMPLEX OPERATOR NORMS

BETWEEN QUASI–BANACH L p − Lq SPACES

L ECH M ALIGRANDA AND NATALIA S ABOUROVA

(Communicated by N. Elezović)

Abstract. Relations between the norms of an operator and its complexification as a mapping
from L p to Lq has been recognized as a serious problem in analysis after the publication of
Marcel Riesz’s work on convexity and bilinear forms in 1926. We summarize here what it is
known about these relations in the case of normed Legesgue spaces and investigate the quasi-
normed case, i. e. we consider all 0 < p,q  ∞ . In particular, in the lower triangle, that is, for
0 < p  q  ∞ these norms are the same. In the upper triangle and the normed case, that is, when
1  q < p  ∞ the norm of the complexification of a real operator is obviously not bigger√ than 2
times its real norm. In 1977 Krivine proved that the constant 2 can be replaced by 2 . On the
other hand, it was suspected that in the case of quasi-normed Lebesgue spaces ( 0 < q < p  ∞ )
the corresponding constant could be arbitrarily large, but as we will see this is not the case. More
precisely, we prove that this constant for quasi-normed Lebesgue spaces is between 1 and 2 .
Some additional properties and estimates of this constant with some results about the relation
between complex and real norms of operators, including those between two-dimensional Orlicz
spaces are presented in the first four chapters. Finally, in Chapter 5, we use the results on the
estimates of the norms in the proof of the real Riesz-Thorin interpolation theorem valid in the
first quadrant.

1. On the norm of operator between real and complex Lebesgue spaces

Any bounded linear operator T : L p (μ ) → Lq (ν ) between real (quasi-)normed

Lebesgue spaces (that is, when 0 < p, q  ∞) has its natural extension to the complex-
valued functions (also called complexification) TC : LCp (μ ) → LqC (ν ) given by
TC ( f + ig) = T ( f ) + i T (g).
Operator TC is linear and bounded. Moreover,

TC ( f + ig)q =  |T f + iT g|q =  (T f )2 + (T g)2 q
 |T f | + |T g|q  max{1, 21/q−1} [T f q + T gq ]
 max{1, 21/q−1}T L p →Lq [ f  p + g p]

 max{2, 21/q}T L p →Lq  f 2 + g2  p
= max{2, 21/q}T L p →Lq  | f + ig| p
= max{2, 21/q}T L p →Lq  f + ig p,
Mathematics subject classification (2010): Primary 46E30, 47A30; secondary 446B20, 46B25.
Keywords and phrases: Lebesgue spaces, Orlicz spaces, quasi-normed spaces, linear bounded oper-
ators, complexification, Marcinkiewicz-Zygmund vector-valued estimates, complex and real Riesz-Thorin
interpolation theorem.

c  , Zagreb 247
Paper MIA-14-21
248 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

that is,
TC L p →Lq  C p,q T L p →Lq , where C p,q  max{2, 21/q }. (1)
C C

Using these estimates we can see that the constant C p,q is the best constant in the
inequality  
 (T f )2 + (T g)2 q  C p,q T L p →Lq  f 2 + g2  p (2)
for all f , g ∈ L p (μ ). In the case of normed Lebesgue spaces (that is, when 1  p, q  ∞)
we immediately obtain √ from the relation (1) that C p,q  2 and an even more precise
estimate C p,q  C∞,1 = 2 was provided by Krivine in his remarkable paper [20]. On
the other hand, relying on (1), it seems reasonable to suggest, that the constant C p,q can
be very large when q is small, i.e. when 0 < q < 1 . However, it is not the case and we
present here a simple proof of the statement (see Theorem 1)

1  C p,q  2 for all p, q ∈ (0, ∞]. (3)

To obtain the upper estimate in (3) we need the following lemma, the first part of which
was already mentioned in the classical book of Zygmund [44, p. 181].

L EMMA 1. (a) Let 0 < p  ∞. Then, for all a, b ∈ R

 1
1/p
|a cos 2π t + b sin 2π t| p dt = d p (a2 + b2 )1/2 , (4)
0

 1/p
where d p = 01 | cos 2π t| p dt .
(b) The function d : (0, ∞] → (0, ∞) given by d(p) := d p is strictly increasing and
d1 = π2 , d2 = √12 , d∞ = 1 . Moreover, d0 := lim p→0+ d p = 12 and so d((0, ∞]) = ( 12 , 1].

Proof. If we divide both sides of the equality (4) by (a2 + b2 )1/2 (we can assume
that a2 + b2 > 0 , since otherwise the equality obviously takes place) and note that
a b
1 cos 2π t + 1 sin 2π t = cos θ cos 2π t ± sin θ sin 2π t
(a2 + b2 ) 2 (a2 + b2 ) 2
= cos(2π t ∓ θ ),

then by 2π -periodicity of cosine we obtain

 1  2 π ∓θ
ds
| cos(2π t ∓ θ )| p dt = | cos s| p
0 ∓θ 2π
 2π  1
ds
= | cos s| p = | cos 2π t| p dt,
0 2π 0

that gives the required equality. Note that the constant d p can be written in terms of
 1/p  1/p
Γ( p+1 )
gamma function as follows d p = 01 | cos 2π t| p dt = √ 2p+2 .
π Γ( 2 )
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 249

(b) The property that d(p) is increasing in p can be easily obtained by the Hölder-
Rogers inequality and since the equality in the Hölder-Rogers inequality applied to our
case does not hold for any p = q we have even that d(p) is strictly increasing.
The values of d(p) at 1, 2 and infinity can be straightforwardly calculated and it
only remains to show the equality d0 = 1/2 .
By the F. Riesz lemma ([29]; see also Rudin [32, p. 71]) we have
 1
1/p  1

p
d0 = lim | cos 2π t| dt = exp ln | cos 2π t| dt .
p→0+ 0 0

  1/4  1/4
Since 01 ln | cos 2π t|dt = 4 0 ln(cos 2π t) dt , then denoting I1 := 0 ln(cos 2π t) dt
we obtain   1/4
1/4
π 
I1 = ln cos − 2π t dt = ln(sin 2π t) dt
0 2 0
and so
 1/4  1/4
sin 4π t
2I1 = ln(sin 2π t cos 2π t) dt = ln dt
0 0 2
 1/4  1/2
1 1 1
= ln(sin 4π t) dt − ln 2 = ln(sin 2π t) dt − ln 2
0 4 2 0 4
1
= I1 − ln 2.
4
Thus, I1 = − 14 ln 2 and d0 = exp(4I1 ) = exp(− ln 2) = 12 .
 π /2
Note, that I1 = 21π 0 ln cost dt = 21π (− π2 ln 2) = − 14 ln 2 can also be proved by
applying Taylor series expansion of the logarithmic function (see Russell [33]). 
Using Lemma 1 we now obtain the early mentioned statement about the constant
C p,q for 0 < p, q  ∞ or what is the same statement, Theorem 1, on the operator es-
timate. Note, Lemma 1 as the key ingredient in the proof of this theorem for values
p = q  1 was firstly used by Zygmund [44] and for 1  p, q  ∞ by Verbickiı̆-Sereda
[41], Verbickiı̆ [40]. Instead of using Lemma 1 Marcinkiewicz-Zygmund in an ear-
lier paper from 1939 [25] based their proof of the restricted statement on the operator
estimate for 0 < p  q  ∞ on utilizing Gaussian variables.

T HEOREM 1. Let T : L p (μ ) → Lq (ν ) be an arbitrary bounded linear operator

between real Lebesgue spaces, where μ and ν are arbitrary positive σ -finite mea-
sures.
If 0 < p  q  ∞, then
TC  p,q = T  p,q . (5)

If 0 < q  p  ∞, then

dp
TC  p,q  T  p,q . (6)
dq
250 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

In particular, in the relation TC  p,q  C p,q T  p,q we always have 1  C p,q  2 , and
if, in addition, 0 < p  q  ∞, then C p,q = 1 .

Proof. First, we show that for all f , g ∈ L p (μ ) and 0 < p, q  ∞ we have

 1 1/q
q
 f cos 2π t + g sin 2π t p dt  max{d p, dq } f + ig p. (7)
0

If 0 < p  q < ∞, then from the integral Minkowski inequality for  · Lq/p [0,1]
(here we use the assumption that q/p  1 ) and Lemma 1 it follows
 1
 f cos 2π t + g sin 2π tqp dt
0
 1  q/p
= | f (x) cos 2π t + g(x) sin 2π t| p d μ (x) dt
0 Ω
 q/p
 

=  | f (x) cos 2π t + g(x) sin 2π t| d μ (x)
p
 q/p
Ω L [0,1]
 q/p
 | f (x) cos 2π t + g(x) sin 2π t| p Lq/p [0,1] d μ (x)
Ω
   p/q q/p
1
q
= | f (x) cos 2π t + g(x) sin 2π t| dt d μ (x)
Ω 0
 q/p
p
= dqq | f (x) + ig(x)| d μ (x) = dqq  f + igqp .
Ω

Note, that for p = q the equality holds in each step of this derivation. A simple modi-
fication of this proof justifies the same statement for q = ∞ (recall d∞ = 1 ).
If 0 < q  p  ∞, then making use of the Hölder-Rogers inequality with p/q  1
and just proved relation we obtain
 1 1/q  1 1/p
q p
 f cos 2π t + g sin 2π t p dt   f cos 2π t + g sin 2π t p dt
0 0
= d p  f + ig p .
Now, from Lemma 1 and the estimate (7), it follows that
 
 
TC ( f + ig)q = 
 (T f )2 + (T g)2 

q
 1
1/q
1
= T ( f cos2π t + g sin 2π t)qq dt
dq 0
 1/q
T  p,q 1
  f cos 2π t + g sin 2π tqp dt
dq 0
d p dq
 max{ , } T  p,q  f + ig p .
dq dq
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 251

Thus, if 0 < p  q  ∞, then TC  p,q  T  p,q and since the reverse inequality ob-
viously holds, then TC  p,q = T  p,q for all such p and q . On the other hand, if
d
0 < q  p  ∞, then TC  p,q  dqp T  p,q . In particular, from Lemma 1(b) it follows
that  
dp d∞ d∞
C p,q  max 1,  = = 2.  (8)
dq lim p→0+ d p d0
Using Theorem 1, Lemma 1(b) and Krivine theorem we can now easily write the
following estimates of the constant C p,q , which is dependent on the points (1/p, 1/q)
as the elements of some regions in the first quadrant Q++ = [0, ∞) × [0, ∞) (see Fig. 1):

A : If 0 < p  q  ∞, then C p,q = 1.

√
B : If 1  q < p  ∞, then C p,q  C∞,1 = 2 (Krivine [20])
C : If 0 < q < p < 1, then C p,q  d p /dq  d1 /d0 = 4/π .
√
D : If 0 < q  1 < p  2, then C p,q  d p /dq  d2 /d0 = 2.
E : If 0 < q < 1, 2  p  ∞, then C p,q  d p /dq  d∞ /d0 = 2.

1
q

√ 4
2  2  π

1
√
 2 1

1
0 1 1 p
2

Fig. 1. The estimates of the constant Cp,q

Let us mention that for the case 1  p  q < ∞ and n -dimensional spaces, i. e.,
q
for T : lnp → ln the result that C p,q (dim n) = 1 was already obtained by Taylor [37] in
1958 by using differential calculus. He showed that the norm of T is attained on the
real vector both in the real and in the complex case for q  p  1 .
In the case of operators between two-dimensional Banach spaces T : l2p → l2q the
constant C p,q (dim 2) = 1 if and only if either 1  p  2 or 2  q  ∞ (see Verbickiı̆-
Sereda [41, Th. 2]). These conditions are equivalent to three cases, namely either
1  p  q  ∞ or 1  q  p  2 or 2  q  p  ∞. Thus C p,q (dim 2) > 1 if and only
if 1  q < 2 < p  ∞. On the other hand, if at least one of the spaces is quasi-Banach,
namely, when 0 < q  p  2 we have C p,q (dim 2) = 1 (see Sabourova [34, Prop. 11]).
252 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

For three dimensional spaces already M. Riesz, in his paper on convexity and
bilinear forms [30], found an operator with different real and complex norms (detailed
proof of the Riesz example can be found in Gasch-Maligranda [16, pp. 112–113] and
Vogt [42, pp. 6–8]). More precisely, if p > q  1 then there exist ε = ε p,q > 0 such
111
that Tε : l3p → l3q satisfies Tε  p,q < (Tε )C  p,q , where Tε = I − ε 111 . The same
111
operator is working even for p > q > 0 .

1 1
q q

1 >1
1 1
1
>1

1 1
2 1 2 1
1

1 1
0 1 1 p 0 1 1 p
2 2
(a) C(l2p , l2q ) (b) C(l3p , l3q )
Fig. 2. Constant Cp,q for two and three dimensional spaces

2. On the norm of operator between some two-dimensional Orlicz spaces in the

real and in the complex case

Having the equality TC L p →L p = T L p →L p for all 0 < p  ∞ it is an interesting

question whether similar equality is always preserved for the operators between the
same spaces but different from L p . We give a negative answer to this question by
showing the existence of operators between two-dimensional symmetric spaces and
two-dimensional Orlicz spaces with different real and complex norms. The following
example of such operator between two-dimensional symmetric spaces for p = 1 was
firstly given by Sokolovski [35, Th. 1] and Gasch-Maligranda [16, Ex. 7].

E XAMPLE 1. For 1  p < 2 let us consider real two-dimensional symmetric spaces

X p = (R2 ,  · 1p) equipped with the norms
1
(|x1 | p + |x2 | p )1/p + 2 2p max{|x1 |, |x2 |}
x1p = 1 , x = (x1 , x2 ) ∈ R2 . (9)
1+2 2p

For the operator T : X p → X p given by the formula T (x, y) = (x + y, x − y) we have

1
TC  > T  = 21− 2p .
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 253

In fact, using the generalized real Clarkson inequality twice (cf. [24], Th. 2.1), we
find the real norm of this operator:
1
(|x1 + x2 | p + |x1 − x2 | p )1/p + 2 2p max{|x1 + x2 |, |x1 − x2|}
T (x1 , x2 )1p = 1
1 + 2 2p
1 1
2 max{|x1 |, |x2 |} + 2 2p 21− p (|x1 | p + |x2 | p )1/p
 1
1 + 2 2p
1
1 (|x1 | p + |x2 | p )1/p + 2 2p max{|x1 |, |x2 |}
= 21− 2p 1
1 + 2 2p
1
= 21− 2p (x1 , x2 )1p .
1
The equality holds for x1 = x2 = 1 and therefore T  = 21− 2p . To show that the com-
1
plex norm of this operator is strictly bigger than 21− 2p we consider two cases. Firstly,
if p = 1 , then taking z1 = 1, z2 = √12 + √12 i we obtain
√ √ √
TC (z1 , z2 )11 (1 + 2)(2 + 2)1/2 + (2 − 2)1/2 √
= √ > 2 = T .
(z1 , z2 )11 2+ 2

Secondly, in the case when 1 < p < 2 we consider the function

   p p
1 1
 1

p
TC (1, eiθ )1 (2+2 cos θ ) 2 +(2−2 cos θ ) 2 +2 2p max (2 ± 2 cos θ ) 2
p
ψ (θ ) = 1
= 1 1
(1, eiθ ) p 2 p +2 2p
√  1  
2 p p p 1 1
= 1 1 ϕ (θ ), ϕ (θ )= (1+ cos θ ) 2 +(1− cos θ ) 2 +2 2p max (1 ± cos θ ) 2 .

2 p +2 2p
For 0 < θ  π /2 we have

sin θ p p 1 p p
ϕ  (θ ) = − [(1 + cos θ ) 2 + (1 − cos θ ) 2 ] p −1 [(1 + cos θ ) 2 −1 − (1 − cos θ ) 2 −1 ]
2
1 sin θ
−2 2p (1 + cos θ )−1/2 ,
2
where

sin θ 21/2−p/2 1 θ θ
ϕ  (0+ ) = lim+ = √ lim sin p−1 cos = 0.
θ →0 2 (1 − cos θ )1−p/2 2 θ →0+ 2 2

Moreover,

1 1 3 1 1 − (2 − p) cos2 θ2
ϕ  (0+ ) = − √ − 2 2p − 2 + p+1 lim 1− p
= +∞.
2 2 2 2 θ →0+ (1 − cos θ ) 2
254 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

√ 1 1
Therefore there exists θ0 ∈ (0, π2 ) such that ϕ (θ0 ) > ϕ (0) = 2 + 2 2 + 2p and conse-
quently
√
2 √ 1 1
2 + 2p
1
ψ (θ0 ) > ψ (0) = 1 1 2 + 2 = 21− 2p .
2 p + 2 2p
Thus, we showed that TC  > T  in both cases.

E XAMPLE 2. For 1  p < 2 let us consider the p -convexification X (p) of the

space X = X1 given by the norm of x = (x1 , x2 ) ∈ R2
√ 1/p
1/p |x1 | p + |x2 | p + 2 max{|x1 |, |x2 |} p
xX (p) = |x| p X = √ .
1+ 2
1 1

For the operator T = 1 −1 we can prove similarly as in Example 1 that
1 1
T X (p) →X (p) = 21− 2p and TC  > 21− 2p .

ϕ
T HEOREM 2. There exists a real two-dimensional Orlicz space l2 such that for
ϕ ϕ
the linear operator T : l2 → l2 given by T (x, y) = (x + y, x − y) we have TC  >
ϕ
T  , where l2 is equipped with the Luxemburg-Nakano norm. Moreover, for the same
ϕ∗ ϕ∗
operator T : l2 → l2 , where ϕ ∗ is a complementary function to ϕ and the Orlicz
ϕ∗
space l2 is equipped with the Orlicz norm we have TC  > T  .

Proof. Let us consider two-dimensional space X = X1 from Example 1. Clearly,

the unit ball defined by the norm (9) is a compact symmetric convex non-empty subset
of R2 . Using the result of Grza̧ślewicz (cf. [17]) saying that every compact symmetric
ϕ
convex subset of R2 with non-empty interior is a unit ball of some Orlicz space l2 we
can construct a two-dimensional Orlicz space isometric to X .
ϕ
Let us take real two-dimensional Orlicz space l2 , generated by the convex Orlicz
function on [0, ∞) given by

√u , if 0  u  √1 ,
ϕ (u) = 2 2 (10)
u(1 + √12 ) − √12 , if u  √12 .

ϕ
It is simple to check that the Luxemburg-Nakano norm on l2 which is given by
 x   x 
 1  2
(x1 , x2 )ϕ = inf{k > 0 : ϕ   + ϕ    1} (11)
k k
coincides with the original norm (9) on X1 . Indeed, without loss of generality, let
us fix 0  x1  x2 . Then, (x1 , x2 )11 = x2 + 1+x1√2 . To calculate (11) we consider
the following three cases obtained naturally from the representation of ϕ , namely
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 255

(x1 , x2 )ϕ = min{inf A, inf B, infC} , where

 x x x + x   
√ 1 2 1 2 1 √
A = k > 0 : k  2x1 and ϕ +ϕ = 1+ √ − 2  1 ,
k k k 2
 √ 
√ √ x x x1 x2 1+ 2 1
1 2
B = k > 0 : 2x1  k  2x2 and ϕ +ϕ =√ + √ −√  1 ,
k k 2k k 2 2
 x x x + x 
√ 1 2 1 2
C = k > 0 : k  2x2 and ϕ +ϕ = √ 1 .
k k k 2
√ √ √
The set A consists only of 2x1 since x1 +x k
2 1+ √
 1+1/ 2
2
= 2 and to satisfy
+x2
x1√
√ +x2
x1√
√
 k  2x 1  we should have x 1 = x 2 and therefore k = 2x1 . Since
2 √ 2 √ √ x1√
infC  2x2 and in the set B we have 2x1  k  2x2 and k  x2 + 1+ 2 , and also
√ √ √
since 2x1  x2 + 1+x1√2  2x2 and x1 + 1+x1√2 = 2x1 it follows that (x1 , x2 )ϕ =
inf B = x2 + 1+x1√2 , and the equality (x1 , x2 )ϕ = (x1 , x2 )11 is proved.
From Example 1 we have that TC  > T  for the Orlicz space with the Luxemburg-
Nakano norm. To show that the same is true for the Orlicz norm we need some duality
results which we present below. 
From the theory of Orlicz spaces it is well-known that the dual norm to the Lu-
xemburg-Nakano  · ϕ is an Orlicz norm  · oϕ (cf. Maligranda [23]). Let us then
calculate it.

P ROPOSITION 1. The dual space X ∗ = (X ∗ ,  · ∗) to X = X1 has the norm

 
∗ |x1 y1 + x2y2 | |y1 | + |y2 |
y = sup = max √ , |y1 |, |y2 | .
x=0 x 2

Proof. Without loss of generality we can assume that 0  y2  y1 . Then

 
∗ x1 y1 + x2 y2 y1 + xy2 y1 + y2
y = sup √ = sup x = max √ , y1 .
x1 x2 0 x1 +x2 +
√ 2x1 0x1 1 + 1+√2 2
1+ 2

We can also calculate this dual norm in another way using the theory of Orlicz spaces
(cf. Maligranda [23]). The complementary function ϕ ∗ to ϕ is
ϕ ∗ (v) = sup [uv − ϕ (u)]
u>0
 
 √  √ √
= max sup√
uv − u/ 2 , sup√ uv − u(1 + 1/ 2) + 1/ 2 ,
0<u1/ 2 u1/ 2

and so ⎧
⎨ 0,
⎪ if 0  v  √12 ,
ϕ ∗ (v) = √v − 1
if √12  v  1 + √12 ,
2, (12)
⎪
⎩ ∞,
2
if v  1 + √12 ,
256 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

The dual norm to the Luxemburg-Nakano norm is the Orlicz norm, which can be written
in the Amemiya form as follows

1 + ϕ ∗(k|y1 |) + ϕ ∗ (k|y2 |)
y∗ϕ = yOA
ϕ ∗ = inf := N.
k>0 k
√
Assume that |y1 |  |y2 | and consider three cases. If k max(|y1 |, |y2 |)  1/ 2 , then

1 √
N = inf = 2 max(|y1 |, |y2 |).
k>0 k
√ √ √
If 1/ 2  k|y1 |  1 + 1/ 2 and k|y2 |  1/ 2 , then
√  
1 + k|y1|/ 2 − 1/2 1 |y1 |
N = inf = inf +√
k>0 k k>0 2k 2
 
|y1 | |y1 | |y2 | |y1 |
= max √ +√ ,√ +√
2(1 + 1/ 2) 2 2 2
 
|y1 | + |y2|
= max |y1 |, √ .
2
√ √ √ √
If 1/ 2  k|y1 |  1 + 1/ 2 and 1/ 2  k|y2 |  1 + 1/ 2 , then
√ √
1 + k|y1|/ 2 − 1/2 + k|y2|/ 2 − 1/2 |y1 | + |y2|
N = inf = √ .
k>0 k 2
Finally,
  
√ |y1 | + |y2|
N = min 2 max(|y1 |, |y2 |), max |y1 |, √
2
 
|y1 | + |y2 |
= max |y1 |, √ . 
2
 1
To finish the proof of Theorem 2 note that the operator T = 11 −1 in the dual
√
space has the norm T  = 2 that can be obtained similarly as in Example 1 (with
p = 1 ) or it follows from the fact that T is self-adjoint, that is, T ∗ = T . Moreover,

TC X ∗ →X ∗ = TC X→X > T X→X = T X ∗ →X ∗ .

ϕ∗
Therefore, for the operator T between two-dimensional Orlicz space l2 with the
Orlicz-Amemiya norm we also have TC  > T  .

R EMARK 1. In Theorem 2 we showed one example of Orlicz function (Orlicz

space), but we can also construct a number of examples of two-dimensional Orlicz
spaces having the same property as those in Theorem 2 by considering family of Or-
licz functions ϕ p given by ϕ p (u) = ϕ (u p ) with ϕ from (10) and 1  p < 2 . The
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 257
ϕ
two-dimensional real Orlicz spaces l2 p with the Luxemburg-Nakano
 1 norms are p -
ϕ
convexifications of l2 and for 1  p < 2 the operator T = 11 −1 gives TC  > T 
 ∗  ϕ ∗
ϕ
by Example 2. The dual space l2 p ,  · ϕ p = l2 p ,  · oϕ p∗ , where
⎧ p √ 1
⎪ v 2 p−1 p
⎪
⎪
⎪ p ( p ) , if 0  v  1
1− 2p
,
⎪
⎨ v 2 √
1 p p(1+ 2)
ϕ p∗ (v) = 1 − 2, if 1
1− 2p
v 1
1− 2p
,
⎪
⎪ 2 2p
 √ 1 2 2
⎪
⎪ p
√
⎪
⎩ vp 2√ p−1
+ √12 , if v  p(1+ 2)
,
p(1+ 2) 1
1− 2p
2

and  
|y1 | + |y2 |  p  √ 1 p1
y0ϕ p∗ = max 1 , |y1 | + |y2 | p (1 + 2) p−1 .
2 2p

3. Vector-valued constants and the monotonicity property

More general estimates than those given by the inequality (2) are called vector-
valued estimates or Marcinkiewicz-Zygmund estimates and they are defined as follows:
(n)
for 0 < p, q, r  ∞ and natural number n  2 let K p,q (r) be the best constant C  1 in
the inequality
n n
( ∑ |T fk |r )1/r q  CT L p →Lq ( ∑ | fk |r )1/r  p (13)
k=1 k=1

for all f1 , f2 , . . . , fn ∈ L p (μ ) and any bounded linear operator T : L p (μ ) → Lq (ν ) be-

tween real (quasi-) normed Lebesgue spaces with arbitrary σ -finite measures μ and
(2)
ν . Of course, K p,q (2) = C p,q .
(n) (n)
Properties of the constants K p,q (r) and K p,q (r) = sup K p,q (r) for 1  p, q, r 
n2
∞ were investigated by Marcinkiewicz-Zygmund [25], Herz [18], Krivine [21], [22],
Gasch-Maligranda [16], Defant-Floret [12], Vogt [42] and Defant-Junge [13].
The assertions that K p,p (2) = 1 and K p,p (r) = 1 , where 0 < p < r  2 , were
proved by Marcinkiewicz-Zygmund with the help of Gaussian and r -stable Gaussian
variables [25, Thm 1 and Thm 2] (see also Edwards-Gaudry [14, pp. 203–204] and
Andersen [1]). Some similar results on the tensor products of operators from L p to
Lq , from which one can obtain the equality K p,q (2) = 1 for 1  p  q  ∞, were
established by Beckner [2], Figiel-Iwaniec-Pełczyński [15] and Rosenthal-Szarek [31]
(cf. also Defant-Floret [12, p. 87], Reinov [28]).
(n)
Besides obvious fact that the constant K p,q (r) is increasing in n it is also in-
(n)
creasing in p and decreasing in q . These properties of K p,q (r) for 1  p, q, r  ∞
follow directly, for example, from their tensor product description (see Krivine [21],
[22] for r = 2 and Gasch-Maligranda [16] for 1  r  ∞). However, for the values
0 < p, q, r  ∞ it is a problem if such a description is possible and therefore another
(n)
method of the proof of the monotonicity of K p,q (r) is needed. One such simple proof
258 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

for the case when 1  p, q  ∞ and r = 2 was obtained by Vogt in [42, Thm 1.9]. It
(n)
turned out that this proof also provides the monotonicity property of K p,q (r) for p, q
in the entire first quadrant and 0 < r  ∞. It is a weighty reason to present it here. The
idea of the proof is based on the generalized Hölder-Rogers inequality and the equality
case in this inequality.
(n)
T HEOREM 3. Let 0 < r  ∞. If 0 < p  s  ∞ and 0 < t  q  ∞, then K p,q (r) 
(n)
Ks,t (r).

Proof. To prove the statement of this theorem is equivalent to prove that for T ∈
L (L p (μ ), Lq (ν )) and f1 , ..., fn ∈ L p (μ )
n n
( ∑ |T fk |r )1/r q  Ks,t (r)T ( ∑ | fk |r )1/r  p .
(n)

k=1 k=1

Let us denote F := (∑nk=1 | fk |r )1/r ∈ L p (μ ) and G := (∑nk=1 |T fk |r )1/r ∈ Lq (ν ).

If q < ∞, then from the equality in the generalized Hölder-Rogers inequality it
follows that for any G ∈ Lq (ν ) there exists ψ ∈ Lα (ν ), where 1t = 1q + α1 , such that

|G| q/α
ψ α = 1 and ψ Gt = Gq . We can take, for example, ψ := G q
.

If q = ∞ (and α = t ), then for any G ∈ L∞ (ν ) and any ε > 0 there exists ψ ∈

Lα (ν ) such that ψ α = 1 and ψ Gt  (1 − ε )G∞ . In fact, let A = {x ∈ Ω2 :
|G(x)| > G∞ (1 − ε )} . Then ν (A) > 0 and therefore since a σ -finite measure ν
has the finite subset property it follows that there exists a set B ⊂ A, B ∈ Σ2 such that
χB
0 < ν (B) < ∞. If we take ψ := ν (B) 1/α , then ψ α = 1 and

 1/α
1 α
ψ Gt = ψ Gα = |G(x)| d ν  (1 − ε )G∞.
ν (B)1/α B

Moreover, from the generalized Hölder-Rogers inequality, in both above cases, it fol-
lows that the norm of the multiplication linear operator Mψ : Lq (ν ) → Lt (ν ) is Mψ q→t
 ψ α = 1 .
At the same time, for any F ∈ L p (μ ), 0 < p < ∞ there exists ϕ ∈ Lβ (μ ), where
1 1 1 F
p = s + β , such that  ϕ s = F p and ϕ β = 1 . We can take, for example, ϕ :=
 1−p/s
|F|
F p . By the generalized Hölder-Rogers inequality we have that the norm of
the multiplication operator Mϕ : Ls (μ ) → L p (μ ) is Mϕ s→p  ϕ β = 1 .
Now, let consider the factorization of operator T : Ls (μ ) → Lt (ν ) in the form
T = Mψ ◦ T ◦ Mϕ as it is shown on Fig. 3. Set gk := ϕ1 fk . By convention 00 = 0 and if
ϕ (x) = 0 , then F(x) = 0 and therefore each fk (x) = 0 . Thus, clearly Mϕ gk = fk and
consequently
1/r 1/r
n n
∑ |T gk |r = ∑ |ψ T fk |r = |ψ |G
k=1 k=1
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 259

T -
F ∈ L p (μ ) Lq (ν )  G
6
Mϕ Mψ

?
F
ϕ ∈ Ls ( μ ) - Lt (ν )  ψ G
T
Fig. 3. Operator diagram

1
and (∑nk=1 |gk |r )1/r = |ϕ | F . Putting all together, in the case when 0 < q < ∞, we obtain

n
Gq = ψ Gt = ( ∑ |T gk |r )1/r t
k=1
n
 Ks,t (r) T s→t ( ∑ |gk |r )1/r s
(n)

k=1
(n) 1
 Ks,t (r) Mψ q→t T  p→q Mϕ s→p  Fs
ϕ
(n) 1 (n)
 Ks,t (r) T  p→q  Fs = Ks,t (r) T  p→q F p ,
ϕ
(n) (n)
whence K p,q (r)  Ks,t (r). If q = ∞, then we similarly can obtain that
(n)
(1 − ε )G∞  Ks,t (r) T  p→∞ F p ,
(n) (n)
and so (1 − ε ) K p,∞ (r)  Ks,t (r). Since ε > 0 was taken arbitrary we have the required
(n) (n)
estimate K p,∞ (r)  Ks,t (r) and the proof is complete. 

(2)
R EMARK 2. Theorem 3 together with the fact that C p,p = K p,p (2) = 1 which
(2)
simply follows from the Fubini theorem and Lemma 1 also shows that C p,q = K p,q (2) =
1 for 1  p  q  ∞ (cf. Theorem 1 and also Verbickiı̆-Sereda [41], Verbickiı̆ [40]).
(n)
Moreover, the same Theorem 3 and the equality K p,p (2) = 1 , which can be easily
obtained by using the relation for the spherical coordinates (see Stechkin [36]) and the
(n)
Fubini theorem, proves the more general equality K p,q (2) = 1 for 0 < p  q  ∞.

(n)
R EMARK 3. For any natural n  2 we have K p,q (r) = 1 if either r = 2 and 0 <
p  q < ∞ or 0 < p < r < 2 and 0 < p  q < ∞ or 1 < p  q < ∞ and min(p, 2) 
r  max(p, 2).
In fact, by Theorem 3, Marcinkiewicz-Zygmund and Herz results we obtain
(n) (n)
K p,q (r)  K p,p (r)  K p,p (r) = 1.
260 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

The last equality we are getting from the Marcinkiewicz-Zygmund result, which tells
us that K p,p (r) = 1 if either r = 2 and 0 < p < ∞ or 0 < p < r < 2 (cf. [25] or [16]).
The case when 1 < p < ∞ and min(p, 2)  r  max(p, 2), which gives that K p,p (r) = 1
was proved by Hertz (cf. [18] or [16]). There are some more cases where we can have
the equality K p,p (r) = 1 and this was investigated by Gasch-Maligranda [16, Th. 2],
but this is not our main here therefore we don’t describe all of them.

R EMARK 4. We can see from the proof that the monotonicity property, as in The-
(n)
orem 3, will also work for constant K p,q (r, μ , ν ) with two fixed σ -finite measures μ
and ν .

4. On one exact description in the upper triangle

(n)
We have discussed so far the exact description of the constant K p,q (r) for 0 <
p, q, r  ∞ only in the lower triangle. In regard to such a description in the upper
triangle Verbickiı̆ [40, Th. 3] proved that C2,q = dd2q for 1  q < 2 and later on Defant
dp
[11] obtained the equality C p,q = dq for 1  q  p  2 . Of course, from this description
d q
and the duality principle we have C p,q = Cq ,p = d p for 2 < q  p  ∞. Note, that in
d q dp
the latter case d p < dq (see Verbickiı̆ [40, Remark I]).
In this chapter we extend the result of Defant [11] obtained for 1  q  p  2 to
the larger set in the first quadrant. In fact, the proof of Defant is also working in the
case 0 < q  p  2 but it is necessary to control carefully each step of the proof. It
is interesting to note that the method of the proof establishes the connection between
(n)
the constant K p,q (r) and the theories of p -summing operators, (p, q)-mixing operators
and the probability theory in Banach spaces.

dp
T HEOREM 4. If 0 < q  p  2 , then C p,q = dq .

Proof. The proof is based on the following arguments. First is the Maurey theory
of (p, q)-mixing operators [26] (see also [27, Chapter 20] and [10]). More precisely,
on the estimating of q -summing norm via a quotient space. Secondly, on the Lévy
theorem that states the fundamental stability property of r -stable random variables.
At last, on the expression of p -summing norm of the identity map on ln2 in terms of
Gaussian measures and on Theorem 1. All these statements can equally well be applied
even when some of the positive indices p and q is less than one and their proofs can
be found in the books of Pietsch [27] and Defant-Floret [12] (see also Carl-Defant [10]
and Maurey [26]).
In what follows assume that E is a Banach space, BE is its unit ball and E ∗ is its
dual space. We use the standard notations from [27] and [12]. By definition, a bounded
linear operator T : E → F is absolutely p -summing ( 0 < p < ∞, F Banach space) if
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 261

1
q
dp
dq
1 √
2 dp
dq

1
2
d q
1
d p

1
0 1 1 p
2

Fig. 4. The best constant in the upper triangle

there exists a constant C > 0 such that for all x1 , ..., xn ∈ E and n = 1, 2, ... we have
n
( ∑ T (xi )Fp )1/p  C ω p (xi ),
i=1

1
where ω p (xi ) = sup{(∑ni=1 | xi , a| p ) p : a ∈ BE ∗ } . The infimum over all possible C
is denoted by π p (T ). Also, a bounded linear operator T : E → F is (p, q)-mixing
( 0 < q  p < ∞, F Banach space) if there exists C > 0 such that for all x1 , ..., xn ∈ E
and the functionals b1 , .., bm ∈ F ∗ , m, n = 1, 2, ... we have
⎛  qp ⎞ 1q
n m
⎝∑ ∑ | T xi , bk | p ⎠  C ωq (xi )l p(bk ),
i=1 k=1

where l p (bk ) = (∑m p 1/p . The infimum over all possible C is the norm of the
k=1 bk  )
operator T which is denoted by μ p,q (T ). By [27, Sec. 20.2.1], for 0 < q  p  ∞ the
product S ◦ T of an absolutely p -summing operator S and the (p, q)-mixing operator
T is an absolutely q -summing operator, moreover,

πq (S ◦ T )  μ p,q (T )π p (S).

The Lévy theorem states: for 0 < q < p  2 there is a constant a p,q such that for every
n ∈ N there is a probability measure μ pn on Rn satisfying
 1/q
1
| x, ω |q μ pn (d ω ) = x p
a p,q Rn

for all x ∈ Rn or in other words if {ωk } is an p -stable sequence for any finite sequence
of real numbers {xk } , then ∑nk=1 xk ωk has the same distribution as (∑nk=1 |xk | p )1/p ω1 .
Using the Gauss-Khinchine equality and the Grothendieck-Pietsch domination theorem
262 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

the p -summing norm of the identity map on ln2 can be calculated (Gordon result – see,
for example, Pietsch [27, p. 300] for 0 < p < ∞ and Defant-Floret [12, p. 137] for
1  p < ∞):
(n) (1)
π p (idRn2 ) = c2,p /c2,p ,
 1/p
(n)  p 1/p (n) √ Γ( p+n 2 )
where c2,p = Kn ω 2 γn (d ω ) and c2,p,R = 2 Γ n . Thus, taking the
(2)
expression for d p from Lemma 1 we have
 ⎛√ ⎞1/p
π Γ p+22
π p (idR2 ) = ⎝  ⎠ = 1/d p.
p+1
2
Γ 2

Now, clearly
dp πq (idR2 )
= 2
 μ p,q (idR2 ). (14)
dq π p (idR2 ) 2
2

(2)
To show that μ p,q (idR2 )  K p,q (2), let us consider the operator F n : lnp → Lq (μ pn ) given
2
1
by the formula F n (x) = a p,q ·, x. Then, for x1 , ..., xn ∈ R2 and b1 , .., bm ∈ R2
⎛ q/p ⎞1/q
n m
⎝∑ ∑ | bk , xi | p ⎠
i=1 k=1

1/q
n m
1
=
a p,q ∑ n
|∑ ωk bk , xi |q μ pn (d ω )
i=1 R k=1
 1/q
n m
1
=
a p,q Rn i=1
∑ | ∑ ωk bk , xi |q μ pn (d ω )
k=1
 1/q
m n m m
1
= ∑ ωk bk q2 ∑ |  ∑ ωk bk −q
2 ∑ ωk bk , xi |q μ pn (d ω )
a p,q Rn k=1 i=1 k=1 k=1
 1/q 1/q
m n
1
 ∑ ωk bk q2 μ pn (d ω ) sup ∑ | y, xi |q
a p,q Rn k=1 y2 1 i=1
1/p 1/q
m n
 FCn  ∑ p
bk 2 sup ∑ | y, xi | q
k=1 y2 1 i=1
1/p 1/q
m n
∑ ∑ | y, xi |
(n)
 K p,q (2)F n  bk 2p sup q
.
k=1 y2 1 i=1

Taking into account that F n is an isometry, that is F n  = 1 and by the definition

(2)
of (p, q)-mixing norm we get μ p,q (idR2 )  K p,q (2). Putting this inequality together
2
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 263

(2) (2)
with the relation (14) we conclude that K p,q (2)  d p /dq . Since K p,q (2)  d p /dq for
(2)
0 < q  p  2 holds by Theorem 1 we obtain equality K p,q (2) = d p /dq . By the same
 p+1 1/p
(n) (n) (n) (n) Γ( 2 )Γ( 2n )
proof we even get that K p,q (2) = d p /dq , where d p = √ p+n . 
π Γ( 2 )

(2) (2)
Observe that K p,q (2) = d p /dq > 1 for 0 < q < p  2 (and K p,q (2) = dq /d p > 1
for 2  q < p  ∞, cf. Defant [11]) and because of the strictly increasing property
(2)
of the constant d p and the monotonicity property of K p,q (2) established in Theorem
(2)
3 we obtain a strong inequality C p,q = K p,q (2) = d p /dq > 1 for the extended region
0 < q < p  ∞.

In the following two examples we make use of the Fourier transform and first
produce a simpler, in comparison with the construction used in Theorem 4, operator for
(2)
which K2,q (2) = d2 /dq , where 0 < q  2 . Secondly, we show that there exists a couple
(2) √
of p and q connected by 0 < q < p  ∞ and p  1 such that K p,q (2)  π2 > 2 .

E XAMPLE 3. Let 0 < q  2 . Consider an operator T : l22 → Lq [0, 1] that maps

the basis vectors e1 and e2 into values T (ei ) = αi , where α1 (t) = cos 2π t and α2 (t) =
sin 2π t. Then,
T l 2 →Lq [0,1] = dq and TC l 2 →Lq [0,1] = d2 . (15)
2 2

(2)
Thus, K2,q (2)  TC /T  = d2 /dq . On the other hand, Theorem 1 gives for 0 < q  2
(2)
the upper bound K2,q (2)  d2 /dq and hence
(2) d2
K2,q (2) = . (16)
dq

Proof. To show (15) we note, that for arbitrary f ∈ l22 there exist c1 , c2 ∈ R such
that f = c1 e1 + c2 e2 . Then, by Lemma 1 we have
T f q = c1 Te1 + c2 Te2 q = c1 cos 2π t + c2 sin 2π tq
 1/2
= dq c21 + c22 = dq  f 2 .
Hence, T l 2 →Lq [0,1] = dq . Let show that TC l 2 →Lq [0,1] = d2 . For this purpose let us
2 2
take arbitrary f , g ∈ l22 , where f = a1 e1 + a2 e2 and g = b1 e1 + b2 e2 . Using the Hölder-
Rogers inequality with 2/q  1 we obtain
1
T ( f + ig)q = (∫ [(a1 cos 2π t + a2 sin 2π t)2 + (b1 cos 2π t + b2 sin 2π t)2 ]q/2 dt)1/q
0
1
 (∫ (a1 cos 2π t + a2 sin 2π t)2 + (b1 cos 2π t + b2 sin 2π t)2 dt)1/2
0
 1/2
= d22 (a21 + a22) + d22 (b21 + b22)
% &1/2
= d2 (a21 + b21 ) + (a22 + b22) = d2  f + ig2
264 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

and therefore T ( f + ig)q  d√2  f + ig2 . √

The equality holds when f = e1 and g = e2
since in this case  f + ig2 = 2 , d2 = 1/ 2 and T ( f + ig)q = 1 . 
If, in the above example, we took the Rademacher functions instead of sin and
cos, then the relation between complex and real operator norms would be equal to one
(the proof is by the Riesz-Thorin interpolation theorem). In other words, the operator
which maps an element from l22 into the linear combination of Rademacher functions
(2)
is not enough to attain the best constant K2,q (2), for this purpose we need, for example,
the orthogonal set of trigonometric functions.

E XAMPLE 4. If 0 < q < p  ∞ and p  1 , then there exists an operator F ∈

L (L p (0, 1), Lq (0, 1)) such as

d22
FC  p,q  F p,q.
d p d q
√
For p = ∞ and q → 0+ by continuity we get FC  p,q/F p,q → π /2 > 2.

Proof. The idea of this proof is due to Verbickiı̆ [40] but we extend it to the case
when 0 < q < 1 . Consider the operator F : L p (0, 1) → Lq (0, 1) defined by

F f (x) = a( f ) cos 2π x + b( f ) sin2π x, (17)

 1  1
with a( f ) = 2 f (t) cos 2π t dt, b( f ) = 2 f (t) sin 2π t dt and f ∈ L p (0, 1).
0 0
Applying Lemma 1 twice we get
 1/2 dq
F f q = dq a2 ( f ) + b2 ( f ) = F f 2 .
d2
Consequently,
dq
F p,q =
F p,2 . (18)
d2
To find F p,2 we note that by duality F p,2 = F ∗ 2,p and F ∗ is defined by

F ∗ ϕ (x) = a(ϕ ) cos 2π x + b(ϕ ) sin2π x,

1 1
where as before a(ϕ ) = 2 0 ϕ (t) cos 2π t dt and b(ϕ ) = 2 0 ϕ (t) sin 2π t dt. Applying
Lemma 1 it yields
F ∗ ϕ  p = d p [a2 (ϕ ) + b2(ϕ )]1/2 .
By the Parseval identity the norm F p ,2 is attained on the function having form
ϕ (x) = a(ϕ ) cos 2π x + b(ϕ ) sin2π x. This function is in L2 (0, 1) and has the norm that
can be derived by using Lemma 1 again
 1 1/2
2
ϕ 2 = |a(ϕ ) cos 2π t + b(ϕ ) sin 2π t| dt
0
% &1/2
= d2 a2 (ϕ ) + b2(ϕ ) .
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 265

Then,
F p,2 = F ∗ 2,p = F ∗ ϕ  p /ϕ 2 = d p /d2 .
Substituting this expression into the equality (18) we obtain
d p d q
F p,q = . (19)
d22
Let FC be the natural complexification of the operator F , then by (17) we find

FC (e2π ix ) = F(cos 2π x) + iF(sin 2π x) = cos 2π x + i sin2π x = e2π ix ,

that implies
FC (e2π ix )q e2π ix q
FC  p,q  2 π ix
= 2π ix = 1.
e  p e  p
Therefore,
d22
FC  p,q  F p,q.
d p d q
If we take p = ∞ and q → 0+ , then for the operator F given by (17) we have
FC  p,q π √
→ > 2. 
F p,q 2
π
P ROBLEM 1. Is it true that C p,q  2 for 0 < p, q  ∞?

5. The complex and the real Riesz-Thorin interpolation theorem in the first
quadrant

Let 0 < p, q  ∞. A linear operator T defined for all simple complex-valued

functions f on (Ω1 , μ ) is of strong type (p, q) if T f is ν -measurable and

T f q  M f  p

with constant M > 0 independent of f . The smallest constant M , called the norm of
T , we denote by T  p,q .
The complex Riesz-Thorin interpolation theorem, proved by Marcel Riesz in 1926
and Thorin in 1939, holds with constant 1 (cf. [30], [38], [39] and also [5]): let (Ω1 , μ )
and (Ω2 , ν ) be two measure spaces with σ -finite measures and T be a linear operator
defined for all simple complex-valued functions f on (Ω1 , μ ). Let 1  p0 , p1 , q0 , q1 
∞, 0  θ  1 and
     
1 1 1 1 1 1
, = (1 − θ ) , +θ ,
p q p0 q0 p1 q1
If T is simultaneously of strong types (p0 , q0 ) and (p1 , q1 ), then T is also of strong
type (p, q) and its norm satisfies
θ θ
T  p,q  T 1−
p0 ,q0 T  p1 ,q1 .
266 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

Geometrically we can say that the set of points

  
1 1
, ∈ [0, 1] × [0, 1] : T is of strong type (p, q)
p q

is a convex subset of [0, 1] × [0, 1].

In 1950 Calderón-Zygmund [8, Th. E] (cf. also [9, Th. A1 ]) extended the complex
Riesz-Thorin convexity theorem to the strip [0, 1]× [0, ∞) and their proof works equally
well for the entire first quadrant [0, ∞) × [0, ∞). This result seems to be a complete
version of the complex Riesz-Thorin interpolation theorem.

T HEOREM 5. (Complex Riesz-Thorin interpolation theorem in the first quadrant).

Let (Ω1 , μ ) and (Ω2 , ν ) be two measure spaces with σ -finite measures and T be a
linear operator defined for all simple complex-valued functions f on (Ω1 , μ ). Let
0 < p0 , p1 , q0 , q1  ∞, 0 < θ < 1 and
     
1 1 1 1 1 1
, = (1 − θ ) , +θ , (20)
p q p0 q0 p1 q1

If T is simultaneously of strong types (p0 , q0 ) and (p1 , q1 ), then T is also of strong

type (p, q) and its strong type norm satisfies
θ θ
T  p,q  T 1−
p0 ,q0 T  p1 ,q1 . (21)

Note that in the case of quasi-normed Lebesgue spaces, that is, when 0 < p, q  ∞
it can happen that the operators T ∈ L (L p (μ ), Lq (ν )) can be only trivial ones. For
example, if T : L p (μ ) → Lq (ν ) is a bounded linear operator, where μ is a non-atomic
measure and 0 < p < min{1, q} , then T = 0 (see e.g. Brudnyı̆-Krugljak [6, p. 89] and
Maligranda [23, p. 150]). Of course, if the measure μ is atomic Theorem 5 is no longer
trivial (cf. Bennett [3], [4]).
In the real case, as it was already noticed by Riesz [30], the estimate (21) can be
not true in the uppertriangle (cf. Riesz [30, pp. 493–494] √ and [6, pp. 16–17]).
√ For the
operator T = 11 −1 1 we have T ∞,1 = 2, T 2,2 = 2 and T 4,4/3 = 3. Taking
θ = 1/2 in (20) we obtain ( 14 , 34 ) = (1 − 12 ) (0, 1) + 12 ( 12 , 12 ) and
√ 1/2 1/2 √
T 4, 4 = 3 > T ∞,1 T 2,2 = 21/2 ( 2)1/2 = 23/4 .
3

Using the relation between the norms of an operator and its complexification, that
were considered in the previous chapters of this paper, we obtain the real case of the
Riesz-Thorin theorem from its complex version with, in general, some additional con-
stant on the right hand side of the estimate (21), that is,
θ θ
T  p,q  CT 1−
p0 ,q0 T  p1 ,q1 . (22)

The real Riesz-Thorin theorem for the normed Lebesgue spaces has, obviously,
constant 2 in the estimate (22) but, as it follows from the Krivine result this estimate
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 267
√
is even true with constant 2 . More precisely, in the real case we have two different
cases:
(a) If we are in the lower triangle of the square [0, 1] × [0, 1], that is 1  pi  qi 
∞, i = 0, 1 , then estimate (22) holds with constant C = 1 .
(b) If we are in the upper triangle of the square [0, 1] × [0, 1], that is 1  q√i < pi 
∞, i = 0, 1 , then estimate (22) holds with constant C  C1− θ θ
p0 ,q0 C p1 ,q1  C∞,1 = 2 .
Brudnyı̆-Krugljak [6, p. 87] asked the question if in the real case we still have con-
stant 1 in the estimate (22), if the conditions 1  pi  qi  ∞, i = 0, 1 will be replaced
by the weaker one 1  p  q  ∞. The answer  1.1 is “NO” and the counterexample was
given by Vogt [43]: for the operator T = 11 −1 we have T ∞,1 = 2.1, T 3/2,3 =
3
7/3 and T 2,2 = 1.5 . If we take θ = 3/4 , then (1 − 34 ) (0, 1) + 34 ( 23 , 13 ) = ( 12 , 12 )
and  3/4 1−3/4 3/4
T 2,2 = 1.5 > 2.11/4 3 7/3 = T ∞,1 T 3/2,3 .
In the real case for the quasi-normed Lebesgue spaces, that is, when 0 < pi , qi  ∞, i =
0, 1 Krasnosel’skiı̆-Zabreı̆ko-Pustyl’nik-Sobolevskiı̆ [19, Th. 2.4] just mentioned about
some constant C = C(q0 , q1 , θ ) in (22), but gave no information about it. Burenkov [7,
pp. 76–77] was more precise (using the relation (1) between real norm and its complex-
ification) and obtained that
(1−θ ) max{1, q1 }+θ max{1, q1 }
C  (max{2, 21/q0 })1−θ (max{2, 21/q1 })θ = 2 0 1

in the estimate (22).

The last estimate suggests that the constant can be large, but this is not the case.
Using our Theorem 1 for any bounded linear operator between quasi-normed Lebesgue
spaces and taking a “good” estimate of the complex norm of this operator by the norm
of its real part we obtain the following result:

T HEOREM 6. The real Riesz-Thorin interpolation theorem in the entire first quad-
rant holds with the estimate (22), where the constant C  C1− θ θ
p0 ,q0 C p1 ,q1 and more pre-
cisely:
A : C = 1 if 0 < pi  qi  ∞, i = 0, 1.
√
B : C  C∞,1 = 2 if 1  qi < pi  ∞, i = 0, 1.
C : C  C1,0+  4/π if 0 < qi < pi < 1, i = 0, 1.
√
D : C  C2,0+ = 2 if 0 < qi  1 < pi  2, i = 0, 1.
E : C  C∞,0+  2 if 0 < qi < 1, 2  pi  ∞, i = 0, 1.
 
1 1
F : If one point , is in one of the above regions and the other
p0 q0
 
1 1
one , is in the second one, then the constant C is estimated
p1 q1
by the larger of the corresponding estimates.
268 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

1
q

E D C

1
B
A

1
0 1 1 p
2

Fig. 5. The areas in Theorem 6

From the above discussion we have that in the complex Riesz-Thorin interpolation
theorem the estimate (22) holds with constant C = 1 in the entire first quadrant Q++ =
[0, ∞) × [0, ∞) while in the real case the estimate (22) holds with constant C = 1 only
in the lower triangle of the square Q++ , i.e., for 0 < pi  qi  ∞, i = 0, 1 . In the upper
triangle of the square Q++ the estimate (22) holds with constant 2 .

Acknowledgement. We would like to thank referee for some useful corrections in

this paper.

REFERENCES

[1] K. F. A NDERSEN, Inequalities for scalar-valued linear operators that extend to their vector-valued
analogues, J. Math. Anal. Appl., 77, 1 (1980) 264–269.
[2] W. B ECKNER, Inequalities in Fourier analysis, Ann. of Math. (2), 102, 1 (1975) 159–182.
[3] G. B ENNETT, An extension of the Riesz-Thorin theorem, In: Banach Spaces of Analytic Functions
(Proc. Pelczynski Conf., Kent State Univ., Kent, Ohio, 1976), Lecture Notes in Math., Vol. 604,
Springer, Berlin 1977, 1–11.
[4] G. B ENNETT, Lectures on matrix transformations of l p spaces, Notes in Banach Spaces, Univ. Texas
Press, Austin 1980, 39–80.
[5] J. B ERGH AND J. L ÖFSTR ÖM, Interpolation Spaces. An Introduction, Springer, Berlin-New York,
1976.
[6] Y U . A. B RUDNY Ĭ AND N. YA . K RUGLJAK, Interpolation Functors and Interpolation Spaces I, North-
Holland, Amsterdam 1991.
[7] V. I. B URENKOV, Functional Spaces. Basic integral inequalities connected with L p -spaces, Izd. Univ.
Druzhby Narodov, Moscow 1989 (in Russian).
[8] A. P. C ALDER ÓN AND A. Z YGMUND, On the theorem of Hausdorff-Young and its applications,
In: Contributions to Fourier Analysis, Annals of Math. Studies, No. 25, Princeton University Press,
Princeton, N. J. 1950, 166–188.
[9] A. P. C ALDER ÓN AND A. Z YGMUND, A note on the interpolation of linear operators, Studia Math.,
12 (1951), 194–204.
[10] B. C ARL AND A. D EFANT, Tensor products and Grothendieck type inequalities of operators in L p -
spaces, Trans. Amer. Math. Soc., 331, 1 (1992), 55–76.
[11] A. D EFANT, Best constants for the norm of the complexification of operators between L p -spaces, In:
Functional Analysis (Proc. of the Essen Conference, Nov. 24 - 30, 1991, Essen, Germany), New York,
Dekker. Lect. Notes Pure Appl. Math., 150 (1993), 173–180.
 R EAL AND COMPLEX OPERATOR NORMS BETWEEN QUASI -BANACH L p − Lq SPACES 269

[12] A. D EFANT AND K. F LORET, Tensor Norms and Operator Ideals, North-Holland Mathematics Stud-
ies 176, North-Holland, Amsterdam 1993.
[13] A. D EFANT AND M. J UNGE, Best constants and asymptotics of Marcinkiewicz-Zygmund inequalities,
Stud. Math., 125, 3 (1997), 271–287.
[14] R. E. E DWARDS AND G. I. G AUDRY, Littlewood-Paley and Multiplier Theory, Springer-Verlag,
Berlin-Heidelberg-New York, 1977.
[15] T. F IGIEL , T. I WANIEC AND A. P EŁCZY ŃSKI , Computing norms and critical exponents of some
operators in L p -spaces, Studia Math., 79, 3 (1984), 227–274.
[16] J. G ASCH AND L. M ALIGRANDA, On vector-valued inequalities of the Marcinkiewicz-Zygmund, Herz
and Krivine type, Math. Nachr., 167 (1994), 95–129.
[17] R. G RZA̧ŚLEWICZ, Finite dimensional Orlicz spaces, Bull. Polish Acad. Sci. Math., 33, 5-6 (1985),
277–283.
[18] C. H ERZ, The theory of p -spaces with an application to convolution operators, Trans. Amer. Math.
Soc., 154 (1971), 69–82.
[19] M. A. K RASNOSEL’ SKI Ĭ , P. P. Z ABRE ĬKO , E. I. P USTYL’ NIK AND P. E. S OBOLEVSKI Ĭ , Integral
Operators in Spaces of Summable Functions, Nauka, Moscow, 1966 (in Russian).
[20] J. L. K RIVINE, Sur la complexification des opérateurs de L∞ dans L1 , C. R. Acad. Sci. Paris, 284
(1977), 377–379.
[21] J. L. K RIVINE, Constantes de Grothendieck et fonctions de type positif sur les sphères, Séminaire sur
la Géométrie des Espaces de Banach (1977–1978), École Polytech., Palaiseau, 1978, Exp. No. 1-2, 17
pp.
[22] J. L. K RIVINE, Constantes de Grothendieck et fonctions de type positif sur les sphères, Adv. Mat., 31
(1979), 16–30.
[23] L. M ALIGRANDA, Orlicz Spaces and Interpolation, Seminars in Math. 5, Univ. of Campinas, Camp-
inas SP, Brazil 1989.
[24] L. M ALIGRANDA AND N. S ABOUROVA, On Clarkson’s inequality in the real case, Math. Nachr.,
280, 12 (2007), 1363–1375.
[25] J. M ARCINKIEWICZ AND A. Z YGMUND, Quelquels inéqualités les opérations linéaires, Fund. Math.,
32 (1939), 115–121; Reprinted in: J. Marcinkiewicz, Collected Pappers, PWN, Warszawa 1964, 541–
546.
[26] B. M AUREY, Théorémes de factorisation pour les opérateurs linéaires á valeurs dans les espaces L p ,
Astérisque, 11 (1974), 1–163.
[27] A. P IETSCH, Operator Ideals, North-Holland, Amsterdam, New York, Oxford 1980.
[28] O. I. R EINOV, On the tensor products of operators in Lebesgue spaces, J. Math. Sci. (New York), 102,
5 (2000), 4487–4507.
[29] F. R IESZ, Sur les valeurs moyennes des fonctions, J. London Math. Soc., 5 (1930), 120–121.
[30] M. R IESZ, Sur les maxima des formes bilinéaires et sur les fonctionelles linéaires, Acta Math., 49
(1926), 465–497.
[31] H. P. ROSENTHAL AND S. J. S ZAREK, On tensor products of operators from L p to Lq , Functional
Analysis (Austin, TX, 1987/1989), Lecture Notes in Math. 1470, Springer, Berlin 1991, 108–132.
[32] W. RUDIN, Real and Complex Analysis, 3. ed., McGraw-Hill, New York 1987.
[33] D. C. RUSSELL, Another Eulerian-type proof, Amer. Math. Monthly, 64 (1991), 349.
[34] N. S ABOUROVA, Real and Complex Operator Norms, Licentiate Thesis, Luleå University of Technol-
ogy 2007, v+69 pp.
[35] Y U . S OKOLOVSKI , The norm of a complex extension of a linear operator, Mat. Issled., 54 (1980),
152–154 (in Russian).
[36] S. B. S TECHKIN, On best lacunary systems of functions, Izv. Akad. Nauk SSSR Ser. Mat., 25 (1961),
357–366 (in Russian).
[37] A. E. TAYLOR, On norm of a real linear transformation in Minkowski space, Enseign. Math. II Sér, 4
(1958), 101–107.
[38] G. O. T HORIN, An extension of a convexity theorem due to M. Riesz, Kungl. Fysiografiska Sällskapets
i Lund Förhandlingar, No. 8, 14 (1938), 166–170; Medd. Lunds Univ. Mat. Sem., 4 (1939), 1–5.
[39] G. O. T HORIN, Convexity theorems, Thesis, University of Lund (= Medd. Lunds Mat. Sem., 9 (1948),
1–58), Uppsala 1948, 1–58.
[40] I. E. V ERBICKI Ĭ , Some relations between the norm of an operator and that of its complex extension,
Mat. Issled., 42 (1976), 3–12 (in Russian).
270 L ECH M ALIGRANDA AND N ATALIA S ABOUROVA

[41] I. E. V ERBICKI Ĭ AND P. P. S EREDA, On norm of the complexification of an operator, Mat. Issled.,
10, 3(37) (1975), 201–206 (in Russian).
[42] H. V OGT, Kompexifizierung von Operatoren zwischen L p -Räumen, Diplomarbeit, Carl von Ossietzky
Universität Oldenburg, Oldenburg 1995, 61 pp.
[43] H. V OGT, On the constant in real Riesz-Thorin interpolation, Arch. Math., 71 (1998), 112–114.
[44] A. Z YGMUND, Trigonometric Series, Cambridge, University Press, Cambridge 1968.

(Received June 15, 2009) Lech Maligranda

Department of Mathematics
Luleå University of Technology
SE-971 87 Luleå
Sweden
e-mail: [email protected]
Natalia Sabourova
Department of Civil, Mining and Environmental Engineering
Luleå University of Technology
SE-971 87 Luleå
Sweden
e-mail: [email protected]

Mathematical Inequalities & Applications

www.ele-math.com
[email protected]