THE NONCOMMUTATIVE ℓ1 − ℓ2 INEQUALITY FOR HILBERT C*-MODULES AND

THE EXACT CONSTANT

K. MAHESH KRISHNA AND P. SAM JOHNSON
Department of Mathematical and Computational Sciences
National Institute of Technology Karnataka (NITK), Surathkal
Mangaluru 575 025, India
Emails: kmaheshak@gmail.com, sam@nitk.edu.in
Date: October 7, 2020

Abstract: Let A be a unital C*-algebra. Then the theory of Hilbert C*-modules tells that
arXiv:2010.02549v1 [math.OA] 6 Oct 2020

n n
! 12
X 1 √ X
(ai a∗i ) 2 ≤ n ai a∗i , ∀n ∈ N, ∀a1 , . . . , an ∈ A.
i=1 i=1

By modifications of arguments of Botelho-Andrade, Casazza, Cheng, and Tran given in 2019, for certain
tuple x = (a1 , . . . , an ) ∈ An , we give a method to compute a positive element cx in the C*-algebra A
such that the equality
n n
! 12
X 1 √ X
(ai a∗i ) 2 = cx n ai a∗i .
i=1 i=1

holds. We give an application for the integral of G. G. Kasparov. We also derive the formula for the
exact constant for the continuous ℓ1 − ℓ2 inequality.
Keywords: C*-algebra, Hilbert C*-module, Hilbert space.
Mathematics Subject Classification (2020): 46L05, 46L08, 46C05.

1. Introduction

Let K = C or R and x ∈ Kn . Universally known ℓ1 − ℓ2 inequality for Hilbert spaces states that
√
kxk1 ≤ nkxk2 . In 2019, Botelho-Andrade, Casazza, Cheng, and Tran [1] gave a characterization which
√
allows to compute a constant cx , for a given x such that kxk1 = cx nkxk2 . First we recall this result.
Definition 1.1. [1] A vector x = √1 (c1 , . . . , cn ) ∈ Kn is said to be a constant modulus vector if |ci | = 1,
n
for all i = 1, . . . , n.
Theorem 1.2. [1] Let x = (a1 , . . . , an ) ∈ Kn . The following are equivalent.
(i) We have
 cx  √
kxk1 = 1 − nkxk2 .
2
(ii) We have
n 2
X |ai | 1
−√ = cx .
i=1
kxk 2 n
x √
(iii) The infimum of the distance from kxk2 to the constant modulus vector is cx .
In particular,
√ cx  √ √
r
 cx s
kxk1 ≤ skxk2 ⇐⇒ 1 − n ≤ s ⇐⇒ 1 − ≤ .
2 2 n
Theorem 1.2 says that as long as we have equality connecting one-norm and two-norm, the constant can
be determined using two-norm and the dimension of space. Further, it also helps to find the distance
1
2

x
between kxk2 to certain types of vectors (constant modulus vectors). This result found uses in nonlinear
diffusion and diffusion state distances [2, 6]. A variation of Theorem 1.2 which concerns subspaces is the
following.

Theorem 1.3. [1] Let W be a subspace of Kn and let P : Kn → W be onto orthogonal projection. Then
the following are equivalent.
cx

√
(i) For every unit vector x ∈ W , kxk1 ≤ 1 − 2 n.
(ii) The distance of any unit vector in W to any constant modulus vector x ∈ W is greater than or equal
√
to cx .
cx
(iii) For every constant modulus vector x ∈ W , kP xk2 ≤ 1 − 2 .

We organized this paper as follows. In Section 2, we obtain a result (Theorem 2.1), which is similar to
first two implications of Theorem 1.2, in the context of Hilbert C*-modules. A partial result is obtained
(Proposition 2.4) which corresponds to (iii) in Theorem 1.2. In Section 3, we derive results which are
similar to Theorems 1.2 and 1.3, namely Theorems 3.2 and 3.4, respectively, for the function space L2 (X)
whenever µ(X) < ∞.

2. The noncommutative ℓ1 − ℓ2 inequality for Hilbert C*-modules and the exact

constant

Let A be a unital C*-algebra. Then the space An becomes (left) Hilbert C*-module over the C*-algebra
A w.r.t. the inner product
n
X
hx, yi := ai b∗i , ∀x = (a1 , . . . , an ), y = (b1 , . . . , bn ) ∈ An
i=1

and the norm

1
n 2
1
X
kxk := khx, xik = 2 ai a∗i , ∀x = (a1 , . . . , an ) ∈ An
i=1

(see [4, 8] for Hilbert C*-modules). Let a1 , . . . , an ∈ A and let

1 1
x = ((a1 a∗1 ) 2 , . . . , (an a∗n ) 2 ), y = (1, . . . , 1) ∈ An .

By applying the Cauchy-Schwarz inequality in Hilbert C*-modules (Proposition 1.1 in [4]) for this pair
we get
n
!2 n
1
X X
(ai a∗i ) 2 ≤n ai a∗i , ∀n ∈ N, ∀a1 , . . . , an ∈ A.
i=1 i=1

By taking C*-algebraic square root (see Theorem 1.4.11 in [5])

n n
! 12
X
∗ 12 √ X
∗
(1) (ai ai ) ≤ n ai ai , ∀n ∈ N, ∀a1 , . . . , an ∈ A.
i=1 i=1

We call the Inequality (1) as the noncommutative ℓ1 − ℓ2 inequality for Hilbert C*-modules. A standard
result in C*-algebra is that an element a ∈ A is positive if and only if a = bb∗ for some b ∈ A. Thus
 3

Inequality (1) can also be written as

n n
! 21
X √ X
ai a∗i ≤ n (ai a∗i )2 , ∀n ∈ N, ∀a1 , . . . , an ∈ A.
i=1 i=1

Note that Inequality (1) is the ℓ1 − ℓ2 inequality for Hilbert spaces whenever the C*-algebra is the field
of scalars.

Theorem 2.1. Let x = (a1 , . . . , an ) ∈ An be such that hx, xi is invertible. The following are equivalent.
(i) We have
n
! n
X 1 1 1
X 1 √ 1 1
(ai a∗i ) 2 hx, xi 2 + hx, xi 2 (ai a∗i ) 2 = nhx, xi 2 (2 − cx )hx, xi 2 .
i=1 i=1

(ii) We have
n   ∗
X −1
∗ 12 1 −1
∗ 12 1
hx, xi (ai ai ) − √
2 hx, xi (ai ai ) − √
2 = cx .
i=1
n n

Proof. We make expansion and see

n   ∗
X 1 1 1 1
hx, xi 2 (ai a∗i ) 2 − √ hx, xi 2 (ai a∗i ) 2 − √
−1 −1

i=1
n n
n 1 1
!
hx, xi 2 (ai a∗i ) 2 (ai a∗i ) 2 hx, xi 2
−1 −1
X 1 1 1
(ai a∗i ) 2 (ai a∗i ) 2 hx, xi 2
−1 −1
= hx, xi 2 + − √ − √
i=1
n n n
n
! −1 n n
! −1
−1 X
∗ −1 hx, xi 2 X ∗ 12
X
∗ 21 hx, xi 2
= hx, xi 2 ai ai hx, xi + 1 − √
2 (ai ai ) − (ai ai ) √
i=1
n i=1 i=1
n
−1 n n
! −1
hx, xi 2 X 1
X 1 hx, xi 2
=2− √ (ai a∗i ) 2 − (ai a∗i ) 2 √ = cx
n i=1 i=1
n

if and only if
−1 n n
! −1
hx, xi 2 X 1
X 1 hx, xi 2
√ (ai a∗i ) 2 + (ai a∗i ) 2 √ = 2 − cx
n i=1 i=1
n

if and only if

n
! n
X 1 1 1
X 1 √ 1 1
(ai a∗i ) 2 hx, xi 2 + hx, xi 2 (ai a∗i ) 2 = nhx, xi 2 (2 − cx )hx, xi 2 .
i=1 i=1

A particular case of Theorem 2.1 which is very similar to Theorem 1.2 is the following.
Pn ∗ 12
Corollary 2.2. Let x = (a1 , . . . , an ) ∈ An be such that hx, xi is invertible and commutes with i=1 (ai ai ) .
The following are equivalent.
(i) We have
n
X 1
 cx  √ 1 √ 1
 cx 
(ai a∗i ) 2 = 1 − nhx, xi 2 = nhx, xi 2 1 − .
i=1
2 2
4

(ii) We have
n   ∗
X −1
∗ 1 1 −1
∗ 1 1
hx, xi 2 (ai ai ) 2 − √ hx, xi 2 (ai ai ) 2 − √ = cx .
i=1
n n

In particular,
n
√ cx  √ √
r
X 1 1
 cx s
(ai a∗i ) 2 ≤ shx, xi 2 ⇐⇒ 1 − n ≤ s ⇐⇒ 1 − ≤ .
i=1
2 2 n

We next derive a result which gives one sided implication in Theorem 1.2. For this, we need to generalize
Definition 1.1.

Definition 2.3. A vector x = √1 (c1 , . . . , cn ) ∈ An is said to be a constant modulus vector if ci c∗i = 1,

n
for all i = 1, . . . , n.

Recall that an element a in a unital C*-algebra A is said to be an isometry if a∗ a = 1. Thus a vector is

a constant modulus vector is adjoint of each of its coordinates is an isometry upto scalar.

Proposition 2.4. Let x = (a1 , . . . , an ) ∈ An be such that ai a∗i is invertible for each i. Define
−1 n n
! −1
hx, xi 2 X ∗ 21
X
∗ 21 hx, xi 2
cx := 2− √ (ai ai ) − (ai ai ) √ .
n i=1 i=1
n
−1
Then the infimum of the distance from hx, xi 2 x to the constant modulus vector is less than or equal to
p
kcx k.

Proof. Note that the condition ai a∗i is invertible for each i implies that hx, xi is invertible. Now consider
√1 ((a1 a∗ ) 2 a1 , . . . , (an a∗ ) 2 an ),
−1 −1
the vector n 1 n which is unit modulus. Using the definition of infimum and
by an expansion we get

 
−1 1 n
inf khx, xi x − yk : y = √ (c1 , . . . , cn ) ∈ A is constant modulus vector
2
n
∗ 12
 
n  
 X ci ci 
: ci ∈ A, ci c∗i = 1, i = 1, . . . , n
−1 −1
= inf hx, xi 2 ai − √ hx, xi 2 ai − √

i=1
n n 
! !∗ 1
n 2
(ai a∗i ) 2 ai (ai a∗i ) 2 ai
−1 −1
X −1 −1
≤ hx, xi 2 ai − √ hx, xi 2 ai − √
i=1
n n
! 1
n −1 n n −1 2
X
∗−1 hx, xi−12 X
∗ 21
X
∗ 2 1 hx, xi 2
≤ hx, xi ai ai hx, xi − √
2 2 (ai ai ) − (ai ai ) √ +1
i=1
n i=1 i=1
n
! 1
−1 n n −1 2
hx, xi 2 X ∗ 12
X
∗ 21 hx, xi 2 1
= 2− √ (ai ai ) − (ai ai ) √ = kcx k 2 .
n i=1 i=1
n

Proposition 2.4 and Theorem 1.2 lead to the following question: Does converse of Proposition 2.4 hold?
We see that when n = 1, cx = 0 and hence converse holds. It is not known that for n ≥ 2. Next we
derive a result which concerns the ℓ1 − ℓ2 inequality for submodules of Hilbert C*-modules.
 5

Proposition 2.5. Let N be a submodule of An and x ∈ N be a vector such that hx, xi = 1. If the
distance of x to the constant modulus vector is greater than or equal to cx , then
1
n 2
2 X 1
cx ≤ 2 − √ (ai a∗i ) 2 .
n i=1

Proof. By doing a similar calculation as in the proof of Proposition 2.4 we get that
 
1 n
cx ≤ inf kx − yk : y = √ (c1 , . . . , cn ) ∈ A is a constant modulus vector
n
∗ 12
 
n   
 X ci ci 
= inf ai − √ ai − √ : ci ∈ A, ci c∗i = 1, i = 1, . . . , n

i=1
n n 
1
n 2
2 X 1
≤ 2− √ (ai a∗i ) 2 .
n i=1

Again a look at Proposition 2.5 and Theorem 2.4 in [1] which lead to the following question: Does con-
verse of Proposition 2.5 hold?

In the spirit of Theorem 3.1 in [1], we next give an application of the previous theorem. For this we need
some concepts.
Let G be a compact Lie group and µ be the left Haar measure on G such that µ(G) = 1 (see [9]). If
f, g : G → A are continuous functions, then we define
Z
hf, gi := f (x)g(x)∗ dµ(x),
G

where the integral is in the sense of G. G. Kasparov (see [3, 7]). Now we can state the result.

Theorem 2.6. Let G be a compact Lie group, µ(G) = 1, f : G → A be continuous, f (x) ≥ 0, ∀x ∈ G

and hf, f i = 1. The following are equivalent.
(i) We have G f (x) dµ(x) = 1 − 2c .
R

(ii) We have hf − 1, f − 1i = c.
R
Proof. Consider 4 = hf − 1, f − 1i + hf + 1, f + 1i = hf − 1, f − 1i + 1 + 1 + 2 G f (x) dµ(x) = hf − 1, f −
R R
1i + 2 + 2 G f (x) dµ(x) which implies hf − 1, f − 1i = 2 − 2 G f (x) dµ(x). Conclusion follows by taking
R
c = hf − 1, f − 1i = 2 − 2 G f (x) dµ(x). 

3. Exact constant for the continuous ℓ1 − ℓ2 inequality

Let X be a measure space with finite measure. Continuous Cauchy-Schwarz inequality tells that kf k1 ≤
p
µ(X)kf k2 . Given f ∈ L2 (X), we now derive a method for the exact constant in the equality kf k1 =
p
cf µ(X)kf k2 . For this, we reform the Definition 1.1.

Definition 3.1. A function f ∈ L2 (X) is said to be a constant modulus function if |f (x)| = √ 1 , ∀x ∈

µ(X)
X.

Definition 3.1 says that a function is constant modulus function if its image lies in the circle of radius
√ 1 , centered at origin.
µ(X)
6

Theorem 3.2. For f ∈ L2 (X), the following are equivalent.

(i) We have
 cf  p
kf k1 = 1 − µ(X)kf k2 .
2
(ii) We have
2
|f (x)| 1
Z
−p dµ(x) = cf .
X kf k2 µ(X)
f √
(iii) The infimum of the distance from kf k2 to the constant modulus function is cf .

In particular,
√ cf  p √ cf s
 r
kf k1 ≤ skf k2 ⇐⇒ 1 − µ(X) ≤ s ⇐⇒ 1 − ≤ .
2 2 µ(X)

Proof. (i) ⇐⇒ (ii) Starting from the integral in (ii) we see that

2
|f (x)| 1 1 1
Z Z Z
2
−p dµ(x) = |f (x)| dµ(x) + dµ(x)
X kf k2 µ(X) kf k22 X µ(X) X
1
Z
−2 p |f (x)| dµ(x)
kf k2 µ(X) X
!
1
Z
=2 1− p |f (x)| dµ(x)
kf k2 µ(X) X
= cf

if and only if
1
Z
cf
p |f (x)| dµ(x) = 1 −
kf k2 µ(X) X 2
if and only if
Z  cf  p
|f (x)| dµ(x) = 1 − kf k2 µ(X).
X 2
 7

(i) ⇐⇒ (iii) This follows from the calculation

 
f
inf − g : g ∈ L2 (X) is constant modulus function
kf k2 2
 ! 21 
 Z f (x) 2 
= inf − g(x) dµ(x) : g ∈ L2 (X) is a constant modulus function
 X kf k2 

 Z 2 Z ! 21
f (x) 2
Z
= inf dµ(x) + |g(x)|2 dµ(x) − Re f (x)g(x) dµ(x) :
X kf k2 X kf k2 X

g ∈ L2 (X) is a constant modulus function
 Z  21 
2
= inf 1+1− Re f (x)g(x) dµ(x) : g ∈ L2 (X) is a constant modulus function
kf k2 X
! 12
2
Z
= 2− p |f (x)| dµ(x) .
µ(X)kf k2 X


To obtain further results we need a result whose proof will follow from the routine argument using Hilbert
projection theorem.

Theorem 3.3. Let K be a closed subspace of a Hilbert space H and let P : H → K be onto orthogonal
Ph
projection. Then for each h ∈ H with P h 6= 0, kP hk is the closest unit vector in K to h.

We now use Theorem 3.3 to obtain relations between closed subspaces of L2 (X) and continuouss ℓ1 − ℓ2
inequality.

Theorem 3.4. Let W be a closed subspace of L2 (X) and let P : L2 (X) → W be onto orthogonal
projection. Then the following are equivalent.
cf
p
(i) For every unit vector f ∈ W , kf k1 ≤ 1 − 2 µ(X).
(ii) The distance of any unit vector in W to any constant modulus function f ∈ W is greater than or
√
equal to cf .
cf
(iii) For every constant modulus function f ∈ W , kP f k2 ≤ 1 − 2 .

Proof. (i) ⇐⇒ (ii) We do a similar calculation as in the proof of Theorem 3.2 and get

inf kf − gk2 : g ∈ L2 (X) is a constant modulus function


(Z  21 )
2 2
= inf |f (x) − g(x)| dµ(x) : g ∈ L (X) is a constant modulus function
X
 Z Z Z  12
2 2
= inf |g(x)| dµ(x) − 2Re
|f (x)| dµ(x) + f (x)g(x) dµ(x) :
X X X

2
g ∈ L (X) is a constant modulus function
 Z  21 
2
= inf 1 + 1 − 2Re f (x)g(x) dµ(x) : g ∈ L (X) is a constant modulus function
X
! 21
2
Z
= 2− p |f (x)| dµ(x) .
µ(X) X
8

Therefore

! 21
√ 2
Z
cf ≤ 2− p |f (x)| dµ(x)
µ(X) X

if and only if
Z  cf  p
kf k1 = |f (x)| dµ(x) ≤ 1 − µ(X).
X 2
(ii) ⇐⇒ (iii) Let f ∈ W be constant modulus function. In view of Theorem 3.3 we calculate
2  2
P 2f
      
Pf Pf Pf P f
−f =1+1− , f − f, =2− , f − f,
kP f k kP f k kP f k kP f k kP f k
= 2 − 2kP f k.

Therefore
Pf cf
cf ≤ −f if and only if kP f k ≤ 1 − .
kP f k 2


4. Acknowledgements

First author thanks National Institute of Technology (NITK) Surathkal for financial assistance.

References
[1] Sara Botelho-Andrade, Peter G. Casazza, Desai Cheng, and Tin T. Tran. The exact constant for the ℓ1 − ℓ2 norm
inequality. Math. Inequal. Appl., 22(1):59–64, 2019.
[2] Lenore Cowen, Kapil Devkota, Xiaozhe Hu, James M. Murphy, and Kaiyi Wu. Diffusion state distances: Multitemporal
analysis, fast algorithms, and applications to biological networks. arXiv:2003.03616v1 [stat.ML] 7 March 2020.
[3] G. G. Kasparov. Topological invariants of elliptic operators. I. K-homology. Izv. Akad. Nauk SSSR Ser. Mat., 39(4):796–
838, 1975.
[4] E. C. Lance. Hilbert C ∗ -modules: A toolkit for operator algebraists, volume 210 of London Mathematical Society Lecture
Note Series. Cambridge University Press, Cambridge, 1995.
[5] Huaxin Lin. An introduction to the classification of amenable C ∗ -algebras. World Scientific Publishing Co., Inc., River
Edge, NJ, 2001.
[6] Mauro Maggioni and James M. Murphy. Learning by unsupervised nonlinear diffusion. J. Mach. Learn. Res., 20:Paper
No. 160, 56, 2019.
[7] V. M. Manuilov and E. V. Troitsky. Hilbert C ∗ -modules, volume 226 of Translations of Mathematical Monographs.
American Mathematical Society, Providence, RI, 2005.
[8] William L. Paschke. Inner product modules over B ∗ -algebras. Trans. Amer. Math. Soc., 182:443–468, 1973.
[9] Mark R. Sepanski. Compact Lie groups, volume 235 of Graduate Texts in Mathematics. Springer, New York, 2007.