1

Quantum Data Processing

Rudolf Ahlswede and Peter Löber

Abstract
We prove a data processing inequality for quantum communication channels, which states that processing
a received quantum state may never increase the mutual information between input and output states.

I. Introduction
Our starting point was the question whether Holevo’s upper bound could be viewed as
arXiv:quant-ph/9907081v1 26 Jul 1999

a special case of a more general theorem, the quantum data processing inequality (that
should hold). An inequality of this kind states for an information transfer via (quantum)
communication channels that if a received state is “processed”, this may never increase
the mutual information between input and output states.
We soon realized that a quantum data processing inequality is a corollary of Uhlmann’s
Monotonicity Theorem (see Sec. 5) which is a generalization of a theorem of Lieb ([7]).
Like in the proof of the classical data processing inequality (which plays around with con-
ditional mutual information1 ), convexity properties play a central role. Here, we will need
such properties “for operators” (see Sec. 3).

II. Pick Functions

This section introduces the notion of Pick functions. We will see in the next section that
Pick functions are “operator monotone” (cf. Cor. 3.4 for a rigorous formulation), a fact
that is very useful because often it’s quite easy to decide whether a function is a Pick func-
tion or not. The theory we introduce here is developed in great detail (and with complete
proofs) in [4].

Definition 2.1: H ± , {x + iy ∈ C|y ≷ 0} denote the two half spaces, respectively.

P , {ϕ : H + → H + analytic } denotes the set of Pick functions.

Remark 2.2: P is a convex cone, and if f, g ∈ P then g ◦ f ∈ P , too.

ExampleP2.3: The function ϕ(z) , z µ with 0 < µ ≤ 1 is in P . A function ψ(z) ,

αz + β + m γi
i=1 δi −z with α, γi > 0 and β, δi ∈ R is in P , too.

The next theorem shows that the latter examples give essentially (i.e. up to limits) all
Pick functions:

Theorem 2.4: Every Pick function ϕ ∈ P has a (unique) representation

1
Z
x
ϕ(z) , αz + β + ( − 2 )dµ(x) (1)
x−z x +1
with α ≥ 0, β ∈ R and µ a positive Borel measure on R for which (x2 + 1)−1 dµ(x) < ∞.
R

The authors are with Fakultät für Mathematik, Universität Bielefeld, Postfach 100131, 33501 Bielefeld.
E-mail: ahlswede@mathematik.uni-bielefeld.de, loeber@mathematik.uni-bielefeld.de
1
cf. [2], p. 55 or [1], p. 32
2

This theorem is from [4], p. 20 ff. Its proof transforms the Pick function with an ap-
propriate Möbius transformation (and its inverse) to a function which maps the unit disc
into itself and which has a positive real part. This real part is a positive harmonic function.

Definition 2.5: For an open interval (a, b) ⊆ R let C(a, b) , H + ∪ H − ∪ (a, b) and

P (a, b) , {ϕ : C(a, b) → C : ϕ|H + ∈ P ∧ ϕ(·) = ϕ(·)} .

Remark 2.6: P (a, b) is a convex cone. Moreover, ϕ ∈ P (a, b) implies that ϕ|(a,b) is a
monotonically increasing real function. (As φ maps H + into H + and H − into H − it has
d
to be real on the real axis. Let now ϕ = u + iv and z = x + iy. By definition dy v(x) ≥ 0,
d
and the Cauchy-Riemann differential equations imply that dx u(x) ≥ 0, too.)

Example 2.7: The

Pmfunction ϕ(z) , z µ with 0 < µ ≤ 1 is in P (0, ∞). A function
ψ(z) , αz + β + i=1 δiγ−z
i
with α, γi > 0, β ∈ R and δi ∈ R\(a, b) is in P (a, b), too.

Remark 2.8: Let ϕ ∈ P a Pick function and µ its corresponding Borel measure (cf. (1)).
Let (a, b) ∈ R an interval. Then:

ϕ ∈ P (a, b) ⇔ µ((a, b)) = 0 .

This remark is again from [4] (p. 26), like the next example (p. 27):
√ R0 √
x
Example 2.9: z= √1 + ( 1 − x
) dx
2 −∞ x−z x2 +1 π

III. Operator Monotonicity

The first part of this section introduces the notion of operator monotonicity. The results
are taken from [4], pp. 67ff. We show that Pick functions are operator monotone.
In the sequel we consider operator convexity properties like it was done in [5], pp. 230ff.
It will be important for the next sections that root functions are operator concave (cf. Ex-
ample 3.9). Here, all Hilbert spaces (usually denoted by H, etc.) are supposed to be finite
dimensional.

Definition 3.1: Let H a finite dimensional (complex) Hilbert space.

• L(H) denotes the algebra of linear operators on H.
• L(H)s.a. denotes the real vector space of self-adjoint operators on H.
• An operator A ∈ L(H) is called positive if < v|Av > ≥ 0 ∀ v ∈ H.
• L(H)+ denotes the convex cone of positive operators on H. (It holds L(H)+ ⊂ L(H)s.a. .)
• For operators A, B ∈ L(H) we write A ≤ B if B − A is positive.

Definition 3.2: Let H a finite dimensional (complex) Hilbert space, I ⊆ R an interval,

f : I → R aP function, and A ∈ L(H)s.a. with all its eigenvalues in I.
n
• For A = a |a >< ai |, where (|ai >)1≤i≤n denotes an ON basis of H, we define
Pn i=1 i i
f (A) , i=1 f (ai )|ai >< ai |. (Clearly, this is independent of the chosen basis.) In matrix
 3

notation we have
a1 0 f (a1 ) 0
   

A= .. f
 7→ f (A) =  .. .
. .
0 an 0 f (an )

• f is called operator monotone of order n (“f ∈ Pn (I)”) if

B≤C =⇒ f (B) ≤ f (C) ∀ B, C ∈ L(H)s.a. .

• f is called operator monotone if f ∈ Pn (I) ∀ n ∈ N.

Lemma 3.3: • Pn (I) is a convex cone.

• If f ∈ Pn (I), g ∈ Pn (J) with im(f ) ⊆ J then g ◦ f ∈ Pn (I).
• Pn (I) is closed (in the topology of pointwise convergence).
• Pn+1 (I) ⊆ Pn (I).
• For α > 0, β ∈ R it holds that x 7→ αx + β is operator monotone.
• x 7→ − x1 is operator monotone on (0, ∞).

Proof: All assertions but the last one are obvious.

So, for A ∈ L(H)+ strictly positive (0 is no eigenvalue) and v, w ∈ H it holds that
1 1 1 1 1 1
| < v|w > |2 = | < A− 2 v|A 2 w > |2 ≤ < A− 2 v|A− 2 v >< A 2 w|A 2 w >
= < v|A−1v >< w|Aw > ,

with equality (e.g.) for w = A−1 v. (This is Cauchy-Schwarz-Buniakowski Inequality!)

Therefore:
| < v|w > |2
< v|A−1 v > = max ,
w6=0 < w|Aw >

and this immediately implies that for B, C ∈ L(H)s.a. :

B≤C ⇔ B −1 ≥ C −1 ⇔ −B −1 ≤ −C −1 .

Corollary 3.4: Let ϕ ∈ P (a, b). Then, ϕ|(a,b) is operator monotone.

Example 3.5: f (x) , xµ with 0 ≤ µ ≤ 1 is operator monotone.

This is not (!) true if µ > 1. For (e.g.) µ = 2:
       
2 2 1 −1 3 1 3.1 0
+ = ≤ but
2 2 −1 1 1 3 0 3.1
 2    2  
3 1 10 6 3.1 0 9.61 0
= and = .
1 3 6 10 0 3.1 0 9.61

Theorem 3.6: Let f ≥ 0 a continuous and operator monotone function on [0, ∞), and
let x ∈ L(H)+ and a ∈ L(H) with ||a||op ≤ 1. Then:

f (a∗ xa) ≥ a∗ f (x)a . (2)

4

1 1
Proof: Let b , (1 − aa∗ ) 2 and c , (1 − a∗ a) 2 , and define operators
   
x 0 a b
X, , U,
0 0 c −a∗

on H ⊕ H. Given now ε > 0 let λ large enough such that:

 ∗   ∗ 
a xa + ε1 0 a xa a∗ xb
Y , ≥ = U ∗ XU .
0 λ1 bxa bxb
   
∗ ∗ ∗ f (x) 0 a∗ f (x)a a∗ f (x)b
=⇒ f (Y ) ≥ F (U XU) = U f (X)U ≥ U U=
0 0 bf (x)a bf (x)b
=⇒ a∗ f (x)a ≤ f (a∗ xa + ε1) −→ f (a∗ xa)
ε→0

Remark 3.7: We call functions f that fulfill (2) operator concave. Indeed, a non-negative
continuous function on [0, ∞) that is operator concave is also operator monotone (cf. [5],
p. 232). Furthermore, operator concave functions f P fulfill Jensen’s Inequality:
+ ∗
P ∗
∀ x
P ∗ 1 , . . . , x n ∈ L(H) and a1 , . . . , an ∈ L(H) with i ai ai ≤ 1 it holds: f ( i ai xi ai ) ≥
a
i i f (x )a
i i .

Proof:

a∗i xi ai 0 · · · 0
 P 

x1 0

a1
  i
0
..  , A ,  ... 0  ⇒ A∗ XA = 
 
X , . .. 
 . 0 
0 xn an
0

Therefore:
 P ∗  P ∗
f ( i ai xi ai ) 0 · · · 0 i ai f (xi )ai 0 · · · 0
 
 0  0
 = f (A∗ XA) ≥ A∗ f (X)A = 
 
 .
.. .
..

 0  (2)
 0 
0 0

Corollary 3.8: Let f ≥ 0 a continuous and operator monotone function on [0, ∞), and let
x ∈ L(H)+ and a : H′ → H a (lin.) contraction (||a||op ≤ 1). It holds: f (a∗ xa) ≥ a∗ f (x)a.

Example 3.9: For 0 ≤ µ ≤ 1 we have (a∗ xa)µ ≥ a∗ xµ a (with notation from Cor. 3.8).

IV. Finite Quantum Systems and Physical Maps

In this section we introduce the notion of finite quantum systems and the maps of such
systems that are considered to be in accordance with quantum physics’ laws.

Definition 4.1: A finite quantum system A is a finite dimensional C ∗ -algebra, i.e. a self-
adjoint subalgebra (with identity) of some L(H), where dim(H) < ∞.
 5

A physical state (on this system) is an element A ∈ A+ , L(H)+ ∩A for which tr (A) = 1.

a1 0
 

Remark 4.2: Let A = L(H) and A a physical state on A. Then A =  ..

. 
0 an
for an ONB of A-eigenvectors (by abuse of notation), and A may be interpreted as a
probability distribution on these basis vectors.

Definition 4.3: A C-linear map α∗ : A → B is positive if α∗ (A+ ) ⊆ B+ . It is called

completely positive if all maps of the form
α∗ ⊗ 1 : A ⊗ C → B ⊗ C
X
ai ⊗ ci 7→ α∗ (ai ) ⊗ ci
i

are positive. A physical map of finite quantum systems A, B is a trace preserving and
completely positive C-linear map α∗ : A → B.

Fortunately, the completely positive maps have a nice representation by Stinespring’s The-
orem (see [3], p. 137 or [9] for a proof):

Theorem 4.4: Let α∗ : A → L(H) a completely positive map of finite quantum systems.
Then
α∗ (A) = V ∗ ρ(A)V (∀ A ∈ A)
for some representation ρ of A on a finite dimensional Hilbert space K and a linear map
V : H → K. Here, the representation ρ is an (algebra-)homomorphism with ρ(A∗ ) = ρ(A)∗
for all A ∈ A, and ρ(1) = 1.

Remark 4.5: A physical map α∗ : A → B maps physical states on A on physical states on

B. There are 3 equivalent ways to describe a physical map of finite (!) quantum systems:
(1) α∗ : A → B
A 7→ α∗ (A)
(2) α : A∗ → B ∗
tr A (A · ) 7→ tr B (α∗ (A) · )
(3) α∗ : B → A
. . . such that α(λ) = λ ◦ α∗ .
A C-linear map α∗ : A → B is a physical map if α∗ is completely positive and unity pre-
serving. This implies that β , α∗ is a Schwarz map, i.e. β(x∗ x) ≥ β(x)∗ β(x) ∀ x ∈ B.
(This can be deduced from Stinespring’s Theorem.)
If A ∈ A+ is a physical state we call tr (A · ) ∈ A∗ a physical state, too.

V. Uhlmann’s Monotonicity Theorem

Uhlmann’s Monotonicity Theorem is our key tool to prove the quantum version of a data
processing inequality. Its proof (the proof of the following lemma) uses the operator con-
cavity of the root functions (see Ex. 3.9). For this section we closely followed [8], p. 18 ff.
6

Lemma 5.1: Let β = α∗ : A1 → A2 a physical map of finite quantum systems, S1 , T1 ∈

A+
1 and S2 , T2 ∈ A+ +
2 with the Ti invertible (i = 1, 2). If for all a ∈ A1 it holds

tr (S2 β(a)) ≤ tr (S1 a) and tr (T2 β(a)) ≤ tr (T1 a),

then
tr (β(x)∗ S2t β(x)T21−t ) ≤ tr (x∗ S1t xT11−t ) ∀ 0 ≤ t ≤ 1, x ∈ A1 .

Example 5.2: tr (S2t T21−t ) ≤ tr (S1t T11−t ).

Proof: We remark that A1 and A2 are Hilbert spaces with inner product < a, b > =
tr (a∗ b).
1 1
Define a linear map V : A1 → A2 by aT12 7→ β(a)T22 for all a ∈ A1 .
V is a contraction, in fact:
1 1
||β(a)T22 ||2 = tr (T2 β(a)∗ β(a)) ≤ tr (T2 β(a∗ a)) ≤ tr (T1 a∗ a) = ||aT12 ||2 .
1
− 12
For i = 1, 2 define ∆i ∈ L(Ai )+ by ∆i (aTi2 ) , Si aTi for all a ∈ Ai .
1 1 1
− 21 1
− 12
(It is positive because < aTi |∆i |aTi > = < aTi |Si aTi
2 2 2
> = tr (Ti2 a∗ Si aTi ) =
tr (a∗ Si a) ≥ 0.)
1 1
−t
It holds ∆ti (aTi2 ) = Sit aTi2 (t ≥ 0), and V ∗ ∆2 V ≤ ∆1 :
1 1 1 1 1
− 21
< aT12 |V ∗ ∆2 V |aT12 > = < β(a)T22 |∆2 |β(a)T22 > = < β(a)T22 |S2 β(a)T2 >
1 1
= tr (β(α)∗ S2 β(a)) ≤ tr (S2 β(αα∗ )) ≤ tr (S1 αα∗ ) = < aT12 |∆1 |aT12 > .
So2 , V ∗ ∆t2 V ≤ (V ∗ ∆2 V )t ≤ ∆t1 , and
1 1 1 1
−t
tr (T22 β(x)∗ S2t β(x)T22 ) = < xT12 |V ∗ ∆t2 V |xT12 >
1 1 1 1
−t
≤ < xT12 |∆t1 |xT12 > = tr (T12 x∗ S1t xT12 ) .

Definition 5.3: Given a physical state A ∈ A+ we denote in the sequel its support,
i.e. the projector on the space spanned by the eigenvectors corresponding to non-zero
eigenvalues, by supp A.
Two physical states A, B ∈ A+ have divergence
(
tr (A(log A − log B)), if supp A ≤ supp B
D(A||B) ,
∞, otherwise.

Equivalently, physical states ω = tr (A · ), ϕ = tr (B · ) ∈ A∗ have divergence D(ω||ϕ) ,

D(A||B). Sometimes, divergence is called relative entropy, too.

We are ready for Uhlmann’s Monotonicity Theorem ([10]):

2
We use Ex. 3.9 (like we promised above).
 7

Theorem 5.4: Let β = α∗ : A1 → A2 a physical map of finite quantum systems and

ω, ϕ ∈ A∗2 physical states. Then:

D(ω ◦ β||ϕ ◦ β) ≤ D(ω||ϕ) .

Proof: Let ω = tr (S2 · ), ϕ = tr (T2 · ), ω ◦ β = tr (S1 · ) and ϕ ◦ β = tr (T1 · ),

where w.l.o.g. supp T2 = 1.
As tr (S2 β(·)) = ω ◦ β = tr (S1 · ) and tr (T2 β(·)) = ϕ ◦ β = tr (T1 · ) Lemma 5.1 implies
that:
tr (S2µ T21−µ ) ≤ tr (S1µ T11−µ ) (∀ 0 ≤ µ ≤ 1).
1− tr (S2µ T21−µ ) 1− tr (S1µ T11−µ )
Consequently, 1−µ
≥ 1−µ
, and by the limit µ → 1:

D(ω||ϕ) = tr (S2µ T21−µ )′ |µ→1 ≥ tr (S1µ T11−µ )′ |µ→1 = D(ω ◦ β||ϕ ◦ β) .

VI. The Quantum Data Processing Inequality

In this final section we derive the Quantum Data Processing Inequality from Uhlmann’s
Monotonicity Theorem. We start with a modified formulation of the latter:

Corollary 6.1: Let α∗ : A → B a physical map of finite quantum systems and A1 , A2 ∈

A+ physical states. Then:

D(α∗ (A1 )||α∗ (A2 )) ≤ D(A1 ||A2 ).

Proof: This is only a question of notation:

D(A1 ||A2 ) = D( tr (A1 · )|| tr (A2 · )), and

D(α∗ (A1 )||α∗ (A2 )) = D( tr (α∗ (A1 ) · )|| tr (α∗ (A2 ) · )).

To reduce the claim to Theorem 5.4 notice that tr (α∗ (Ai ) · ) = α( tr (Ai · )) = tr (Ai α∗ (·)).

Definition 6.2: Let W∗ : A → B a quantum channel,

P i.e. a physical map of finite quantum
+
systems. Let A ∈ A a physical state with A = a a|a >< a| its spectral decomposition,
P let B , W∗ (A) the corresponding output (physical) state. Furthermore, let (A, B) ,
and
a a|a >< a| ⊗ W∗ (|a >< a|) the joint state. The mutual information is given by

I(A; W∗ ) , H(A) + H(B) − H(A, B)

where H(X) , − tr (X log X) denotes the Shannon-von Neumann entropy.

Here is the Quantum Data Processing Inequality:

Corollary 6.3: Let W∗ : A1 → A2 and D∗ : A2 → A3 physical maps of finite quantum

systems and A ∈ A+ 1 a physical state. Then:

I(A; D∗ ◦ W∗ ) ≤ I(A; W∗ ) .
8

Proof: Because of I(A; W∗ ) = D((A, B)||A ⊗ B) this is just a consequence of

D((1 ⊗ D∗ )(A, B)||(1 ⊗ D∗ )(A ⊗ B)) ≤ D((A, B)||A ⊗ B) .

As special case we get Holevo’s Upper Bound ([6]):

Corollary 6.4: Let W∗ : A1 → A2 a quantum channel and D∗ a measurement on its

output space, i.e. a physical map from A2 into some commutative C ∗ -algebra A3 . Then
I(A; D∗ ◦ W∗ ) ≤ I(A; W∗ ).

We remind the reader that Holevo’s result is from 1973 whereas Uhlmann’s Monotonicity
Theorem is from 1977. Both use analytical considerations for the proofs. There is also an
“elementary” proof of Holevo’s Upper Bound, using only information theoretical consid-
erations, in [11].

VII. Acknowledgments
We thank Andreas Winter for fruitful discussions about quantum information theory.

References
[1] Th. M. Cover, J. A. Thomas, “Elements of Information Theory,” Wiley-Interscience, New York, 1991.
[2] I. Csiszár, J. Körner, “Information Theory: Coding Theorems for Discrete Memoryless Systems”, Academic
Press, London, 1981.
[3] E. B. Davis, “Quantum Theory of Open Systems”, Academic Press, London, 1976.
[4] W. F. Donoghue (Jr.), “Monotone Matrix Functions and Analytic Continuation”, Springer, Berlin, 1970.
[5] F. Hansen, G. K. Pederson, “Jensen’s Inequality and Löwner’s Theorem”, Math. Ann., vol. 258, pp. 229-241,
1982.
[6] A. S. Holevo, “Bounds for the Quantity of Information Transmitted by a Quantum Communication Channel”,
Probl. Peredachi Inform., vol. 9, no. 3, pp. 3-11, 1973. (Engl. transl.: Probl. of Inf. Transm., vol. 9, no. 3,
pp. 177-183, 1973).
[7] E. H. Lieb, “Convex Trace Functions and the Wigner-Yanase-Dyson Conjecture”, Advan. Math., vol. 11,
pp. 267-288, 1973.
[8] M. Ohya, D. Petz, “Quantum Entropy and Its Use”, (Texts and Monographs in Physics), Springer, Berlin,
1993.
[9] W. F. Stinespring, “Positive Functions on C ∗ -algebras”, Proc. Am. Math. Soc., vol. 6, pp. 211-216, 1955.
[10] A. Uhlmann, “Relative Entropy and the Wigner-Yanase-Dyson-Lieb Concavity in an Interpolation Theory”,
Commun. math. Phys., vol. 54, pp. 21-32, 1977.
[11] A. Winter, “Coding Theorem and Strong Converse for Quantum Channels”, to appear in: IEEE Trans. In-
form. Theory, vol. 45, no. 7, 1999.