1

Gaussian States Minimize the Output Entropy of

the One-Mode Quantum Attenuator

arXiv:1605.00441v2 [quant-ph] 21 Dec 2016

Giacomo De Palma, Dario Trevisan, Vittorio Giovannetti

AbstractWe prove that Gaussian thermal input states minimize the output von Neumann entropy of the one-mode Gaussian quantum-limited attenuator for fixed input entropy. The
Gaussian quantum-limited attenuator models the attenuation of
an electromagnetic signal in the quantum regime. The Shannon
entropy of an attenuated real-valued classical signal is a simple
function of the entropy of the original signal. A striking consequence of energy quantization is that the output von Neumann
entropy of the quantum-limited attenuator is no more a function
of the input entropy alone. The proof starts from the majorization
result of De Palma et al., IEEE Trans. Inf. Theory 62, 2895
(2016), and is based on a new isoperimetric inequality. Our result
implies that geometric input probability distributions minimize
the output Shannon entropy of the thinning for fixed input
entropy. Moreover, our result opens the way to the multimode
generalization, that permits to determine both the triple tradeoff region of the Gaussian quantum-limited attenuator and the
classical capacity region of the Gaussian degraded quantum
broadcast channel.
Index TermsGaussian quantum channels, Gaussian quantum attenuator, thinning, von Neumann entropy, isoperimetric
inequality.

I. I NTRODUCTION

OST communication schemes encode the information

into pulses of electromagnetic radiation, that is transmitted through metal wires, optical fibers or free space, and
is unavoidably affected by signal attenuation. The maximum
achievable communication rate of a channel depends on the
minimum noise achievable at its output. A continuous classical
signal can be modeled by a real random variable
X. Signal
attenuation corresponds to a rescaling X 7 X, where
0 1 is the attenuation coefficient (the power of
the signal is proportional to X 2 and gets rescaled by ).
The noise of a real random variable is quantified by its
Shannon differential entropy H [1]. The Shannon entropy of
the rescaled signal is a simple function of the entropy of the
original signal [1]:



H
X = H (X) + ln .
(I.1)
G. De Palma is with QMATH, Department of Mathematical Sciences,
University of Copenhagen, Universitetsparken 5, 2100 Copenhagen, Denmark;
NEST, Scuola Normale Superiore, Istituto Nanoscienze-CNR and INFN, I56126 Pisa, Italy.
D. Trevisan is with Universit`a degli Studi di Pisa, I-56126 Pisa, Italy.
V. Giovannetti is with NEST, Scuola Normale Superiore and Istituto
Nanoscienze-CNR, I-56126 Pisa, Italy.
This paper was presented at the 11th Conference on the Theory of Quantum
Computation, Communication, and Cryptography, Berlin (Germany), September 2016; at the Beyond IID in Information Theory 4 Workshop, Barcelona
(Spain), July 2016; at the Quantum Roundabout Conference, Nottingham
(United Kingdom), July 2016; and at the 13th Central European Quantum
Information Processing Workshop, Valtice (Czech Republic), June 2016.

This property is ubiquitous in classical information theory. For

example, it lies at the basis of the proof of the Entropy Power
Inequality [1][7].
Since the energy carried by an electromagnetic pulse is
quantized, quantum effects must be taken into account [8].
They become relevant for low-intensity signals, such as for
satellite communications, where the receiver can be reached
by only few photons for each bit of information [9]. In the
quantum regime the role of the classical Shannon entropy is
played by the von Neumann entropy [10], [11] and signal
attenuation is modeled by the Gaussian quantum-limited attenuator [11][15].
A striking consequence of the quantization of the energy
is that the output entropy of the quantum-limited attenuator
is not a function of the input entropy alone. A fundamental
problem in quantum communication is then determining the
minimum output entropy of the attenuator for fixed input
entropy. Gaussian thermal input states have been conjectured
to achieve this minimum output entropy [16][21]. The first
attempt of a proof has been the quantum Entropy Power
Inequality (qEPI) [22][25], that provides the lower bound
 



S ( (
)) n ln e S()/n
1 +1
(I.2)

to the output entropy of the n-mode quantum-limited attenuator in terms of the entropy of the input state . However,
the qEPI (I.2) is not saturated by thermal Gaussian states, and
thus it is not sufficient to prove their conjectured optimality.
Here we prove that Gaussian thermal input states minimize
the output entropy of the one-mode quantum-limited attenuator
for fixed input entropy (Theorem 1). The proof starts from
a recent majorization result on one-mode Gaussian quantum
channels [26], that reduces the problem to input states diagonal
in the Fock basis. The key point of the proof is a new
isoperimetric inequality (Theorem 2), that provides a lower
bound to the derivative of the output entropy of the attenuator
with respect to the attenuation coefficient.
The restriction of the one-mode quantum-limited attenuator
to input states diagonal in the Fock basis is the map acting
on discrete classical probability distributions on N known in
the probability literature under the name of thinning [26].
The thinning has been introduced by Renyi [27] as a discrete
analogue of the rescaling of a continuous real random variable.
The thinning has been involved with this role in discrete
versions of the central limit theorem [28][30] and of the
Entropy Power Inequality [31], [32]. Most of these results
require the ad hoc hypothesis of the ultra log-concavity (ULC)
of the input state. In particular, the Restricted Thinned Entropy
Power Inequality [32] states that the Poisson input probability

distribution minimizes the output Shannon entropy of the

thinning among all the ULC input probability distributions
with a given Shannon entropy. We prove (Theorem 23) that
the geometric distribution minimizes the output entropy of the
thinning among all the input probability distributions with a
given entropy, without the ad hoc ULC constraint.
Theorem 1 constitutes a strong evidence for the validity
of the conjecture in the multimode scenario, whose proof
could exploit a multimode generalization of the isoperimetric
inequality (III.1). The multimode generalization of Theorem
1 is necessary for the proof of the converse theorems for
the achievable rates in two communication scenarios. The
first is the triple trade-off coding for simultaneous classical
and quantum communication and entanglement sharing, or
simultaneous public and private communication and secret key
distribution through the Gaussian quantum-limited attenuator
[20], [21], [33]. The second is the transmission of classical
information to two receivers through the Gaussian degraded
quantum broadcast channel [16], [17]. The multimode generalization of Theorem 1 would imply the optimality of Gaussian
encodings in both scenarios.
The paper is structured as follows. In Section II we define
the Gaussian quantum attenuator and state the main result
(Theorem 1). Section III presents the isoperimetric inequality
(Theorem 2). Theorem 1 is proved in Section IV. Section V
discusses the relation with the Gaussian degraded broadcast
channel. Section VI links the Theorems 1 and 2 to the thinning
operation, and the conclusions are in Section VII. Finally,
Appendix A contains the proof of some auxiliary lemmas.
II. S ETUP

AND MAIN RESULT

We consider the Hilbert space of one harmonic oscillator, or

one mode of electromagnetic radiation. Its ladder
operator
a



satisfies the canonical commutation relation a
, a
= I, and
=a
its Hamiltonian N
a
counts the number of excitations, or
photons. The state annihilated by a
is the vacuum |0i, from
which the Fock states are built:

n
a

|ni = n|ni . (II.1)

|ni = |0i , hm|ni = mn , N
n!
The quantum-limited attenuator of transmissivity 0
1 mixes the input state with the vacuum state of
an ancillary quantum system B through a beamsplitter of
transmissivity . The beamsplitter is implemented by the
unitary operator



= exp a
(II.2)
b arccos ,
U
b a

that satisfies

a
+ 1 b ,
U
U =

(II.3)

where b is the ladder operator of the ancilla system B (see

Section 1.4.2 of [34]). We then have
h
i
(
.
|0iB h0|) U
(II.4)
(
) = TrB U

For input states diagonal in the Fock basis (II.1), i.e. with
definite photon number, lets each photon be transmitted
with probability and reflected or absorbed with probability

1 (see Section VI), hence the name quantum-limited

attenuator.
The Gaussian thermal state with average energy E 0 is

n

i
h
X
E
1

|nihn| , Tr N
E = E ,

E =
E+1 E+1
n=0
(II.5)
with von Neumann entropy
S (
E ) = (E + 1) ln (E + 1) E ln E := g(E) ,

(II.6)

and corresponds to a geometric probability distribution of the

energy. The quantum-limited attenuator sends thermal states
into themselves, i.e. (
E ) =
E , hence


S ( (
E )) = g(E) = g g 1 (S (
E )) .
(II.7)
Our main result is:

Theorem 1. Gaussian thermal input states (II.5) minimize the

output entropy of the quantum-limited attenuator among all
the input states with a given entropy, i.e. for any input state
and any 0 1


S ( (
)) g g 1 (S (
)) .
(II.8)
III. I SOPERIMETRIC

INEQUALITY

The main step to prove Theorem 1 is the proof of its

infinitesimal version. It states that Gaussian states minimize
the derivative of the output entropy of the quantum-limited
attenuator with respect to the attenuation parameter for fixed
entropy:
Theorem 2 (Isoperimetric inequality). For any quantum state
with finite support


d
S (et (
))
:= F (
) f (S (
)) ,
(III.1)
dt
t=0
where

f (S) := g 1 (S) g g 1 (S) .

(III.2)

Proof: The starting point of the proof is the recent result

of Ref. [26], that links the constrained minimum output entropy
conjecture to the notions of passive states. The passive states of
a quantum system [35][40] minimize the average energy for
a given spectrum. They are diagonal in the energy eigenbasis,
and their eigenvalues decrease as the energy increases. The
passive rearrangement of a quantum state is the only
passive state with the same spectrum of . The result is the
following:
Theorem 3. The passive rearrangement of the input does not
increase the output entropy, i.e. for any quantum state and
any 0 1



(III.3)
S ( (
)) S .
Proof: Follows from Theorem 34 and Remark 20 of [26].

In the following, we will show that this result reduces the

proof to the set of passive states. On this set, we will use the
Karush-Kuhn-Tucker (KKT) necessary conditions [41] for the

maximization of the right-hand side of (III.1) for fixed entropy,

and prove that in the limit of infinite support the maximizer
tends to a thermal Gaussian state.
Let us fix S (
) = S.
If S = 0, from the positivity of the entropy we have for any
quantum state F (
) 0 = f (0), and the inequality (III.1)
is proven.
We can then suppose S > 0. Taking the derivative of (III.3)
with respect to t for t = 0 we get


(III.4)
F (
) F ,
hence it is sufficient to prove Theorem 2 for passive states with
finite support.
Let us fix N N, and consider a quantum state with
entropy S of the form
=

N
X

n=0

pn |nihn| .

(III.5)

Let DN be the set of decreasing probability distributions

on {0, . . . , N } with Shannon entropy S. We recall that the
Shannon entropy of p coincides with the von Neumann entropy
of . The state in (III.5) is passive if p DN .
Lemma 4. DN is compact.
Proof: The set of decreasing probability distributions on
{0, . . . , N } is a closed bounded subset of RN +1 , hence it is
compact. The Shannon entropy H is continuous on this set.
DN is the counterimage of the point S, hence it is closed.
Since DN is contained in a compact set, it is compact, too.
Definition 5 (Connected support). A probability distribution
p on {0, . . . , N } has connected support iff pn > 0 for n =
0, . . . , N , and pN +1 = . . . = pN = 0, where 0 N N
can depend on p (N = N means pn > 0 for any n). We call
PN the set of probability distributions on {0, . . . , N } with
connected support and Shannon entropy S.
We relax the passivity hypothesis, and consider all the states
as in (III.5) with p PN . We notice that any decreasing p
has connected support, i.e. DN PN .
From Equations (VIII.5), (VIII.6) and Theorem 56 of Ref.
[26], we have for any t 0
et (
) =

N
X

n=0

pn (t) |nihn| ,

(III.6)

F (
) =

N
X

pn (0) (ln pn

+ 1) =

n=0

N
X

n pn ln

n=1

pn1
. (III.9)
pn

Let FN be the sup of F (p) for p PN , where with a bit of

abuse of notation we have defined F (p) = F (
) for any as
in (III.5). From (III.4), FN is also the sup of F (p) for p DN .
From Lemma 4 DN is compact. Since F is continuous on DN ,
the sup is achieved in a point p(N ) DN . This point satisfies
the Karush-Kuhn-Tucker (KKT) necessary conditions [41] for
the maximization of F with the entropy constraint. The proof
then comes from
Lemma 6. There is a subsequence {Nk }kN such that
lim FNk = f (S) .

(III.10)

Proof: The point p(N ) is the maximum of F for p PN .

The constraints read
p0 , . . . , pN 0 ,

N
X

pn = 1 ,

n=0

N
X

pn ln pn = S .

n=0

(III.11)
p(N ) must then satisfy the associated KKT necessary conditions [41]. We build the functional
F (p) = F (p) N
Let N be such that
(N )

p0

N
X

p n + N

n=0

(N )

N
X

pn ln pn .

(III.12)

n=0

(N )

(N )

. . . pN > pN +1 = . . . = pN

=0.

(III.13)

Remark 7. We must have N 1.

(N )
(N )
Indeed, if N = 0 we must have p0 = 1 and p1 =
(N )
. . . = pN = 0, hence S = 0, contradicting the hypothesis
S > 0.
The KKT stationarity condition for n = 0, . . . , N reads

(N )
(N )
pn+1
p


n
+
(n
+
1)
F
= n ln n1
(N )
(N )
pn p=p(N )
pn
pn
)
N + N ln p(N
+ N = 0 . (III.14)
n
If N < N , p(N ) satisfies the KKT dual feasibility condition
(N )
associated to pN +1 . To avoid the singularity of the logarithm
in 0, we make the variable change
y = pN +1 ln pN +1 ,

pN +1 = (y) ,

(III.15)

where satisfies

where
N  
X

kn
k nt
pn (t) =
pk
1 et
e
n

(III.7)

k=n

satisfies pn (0) = (n + 1) pn+1 n pn for n = 0, . . . , N , and

we have set for simplicity pN +1 = 0.
Since pN +1 = . . . = pN = 0, from (III.7) we get
pN +1 (t) = . . . = pN (t) = 0 for any t 0. We then have

S (et (
)) =

and

N
X

n=0

pn (t) ln pn (t) ,

(III.8)

1
.
(III.16)
e
Since (0) = 0, the point pN +1 = 0 corresponds to y = 0.
Differentiating (III.16) with respect to x, we get
(x ln x) = x

0x

1
1
0<x< ,
(III.17)
1 + ln x
e
and taking the limit for x 0 we get that (y) is continuous
in y = 0 with (0) = 0.
For hypothesis p(N ) PN PN +1 PN . Then, p(N ) is
a maximum point for F (p) also if we restrict to p PN +1 .
(x ln x) =

We can then consider the restriction of the functional F on

PN +1 :

N
X

F (p) =

n pn ln

n=1

pn1
+ (N + 1) (y) ln pN
pn

N
X

+ (N + 1) y N

+ N

N
X

n=0

pn N (y)

n=0

(N )

(III.19)

(III.20)

Condition (III.13) implies

(N )

zN = 0 .
(III.21)

(N )

From Remark 7 N 1, hence z0 > 0.

Taking the difference of (III.14) for two consecutive values
of n we get for n = 0, . . . , N 1
(N )

(n + 2) zn+1 = (n + 2) zn(N ) + 1 zn(N )

+ (1

N ) ln zn(N )

(N )

(N )

Moreover, zn

(N )
zn
(N )
zn1

. (III.22)

(III.23)

(N )

z0

(N )
ln z0
(N )

zn

(N )

(N )
ln zn

(N )

(III.25)
(N )

(N )

and hence (1 N ) ln zn zn 1. Since zn1 zn ,

from (III.22) we have


(N )
(n + 2) zn+1 zn(N )
= 1 zn(N ) + (1 N ) ln zn(N )
+n ln
(N )
zn . However,
(N )
zN = 0.

(N )
zn+1

and hence
the hypothesis

(N )
zn
(N )
zn1

(N )

(N )

0,

(III.26)

this is in contradiction with

zn

(N )

zn1

0,

(III.28)

<1,

(III.29)

where


n
= min n N : n + 2 > eS

(III.30)

does not depend on N .

(N )
Proof: We recall that zn 1 for any n and N , hence
(N )

=0.
(III.24)

.
(N )
(N )
< z0 1
ln z0 .

(N )

(III.27)

Lemma 9. We have

We will prove by induction on n that the sequence zn is

(N )
increasing in n. The inductive hypothesis is 0 < z0 . . .
(N )
zn 1, true for n = 0. Since the function (z 1)/ ln z is
strictly increasing for 0 z 1, we have
1 N <

(N )

lim sup zn

. . . zN 1 > zN

Proof: Let us suppose 1 N

(N )

(N )

(N )

. We will

(N )
zn

(N )
ln zn

lim sup zn

0.

(N )
ln z0

(N )

ln z0

and hence zn+1 zn . Since n + 1 < N , we also have

(N )
zn+1 > 0, and the claim is proven.
From Definition 5 and Remark 7 we have 1 N N . Let
us suppose 1 N < N . Then, the sequence p(N ) satisfies
the KKT dual feasibility condition (III.19), and N N 1.
From (III.23) we get N 1 0, hence N 0, in
contradiction with N 1. We must then have N = N .

is decreasing in n and N = N , i.e.

(N )

1 z0

z0

zn

(N )

+ n ln

.

and hence (1 N ) ln zn zn 1. Since zn1 zn ,

from (III.22) we have


(N )
(n + 2) zn+1 zn(N )
= 1 zn(N ) + (1 N ) ln zn(N )

Lemma 8. We must have

1 N

(N )
ln z0

(N )

n = 0, . . . , N 1 ,

(N )

+n ln

(N )

0 < zn(N ) 1

(N )

z0

(N )

where we have used that (0) = 0.

We define for n = 0, . . . , N
(N )

(N )

p=p

pn

z0

(III.18)

The KKT dual feasibility condition is then



F
= N + 1 N 0 ,
y
(N )
pn+1

is decreasing
prove by induction on n that the sequence
(N )
in n. The inductive hypothesis is now 1 z0
...
(N )
(N )
zn > 0, true for n = 0. If n + 1 = N , since zN = 0
there is nothing to prove. We can then suppose n + 1 < N .
We have
1 N

pn ln pn N y .

zn(N ) =

We must then have 1N

1.

(III.31)

(N )

Let us suppose that lim supN zn = 1. Then, there is a

(N )
subsequence {Nk }kN such that limk zn k = 1. Since
(N )
zn is decreasing in n for any N , we also have
lim zn(Nk ) = 1

n = 0, . . . , n
.

(III.32)

Let us define for any N the probability distribution q (N )

Dn +1 as
(N )

pn
qn(N ) = Pn +1
k=0

(N )

pk

n = 0, . . . , n
+1.

(III.33)

From (III.32) we get for n = 0, . . . , n

+1
(Nk )

lim

qn

(Nk )

q0

(Nk )

= lim z0
k

(N )

k
. . . zn1
=1.

(III.34)

For any k
n
+1
X

n=0

qn(Nk ) = 1 .

(III.35)

(N )

Dividing both members of (III.35) by q0 k and taking the

(N )
limit k we get limk q0 k = 1/ (
n + 2), hence
(Nk )
limk qn
= 1/ (
n + 2) for n = 0, . . . , n
+ 1, and


n + 2) > S .
(III.36)
lim H q (Nk ) = ln (
k




However, from Lemma 25 we have H q (N ) H p(N ) =
S.
Corollary 10. There exists 0 z < 1 (that does not depend
(N )
.
on N ) such that zn z for any N n
Lemma 11. The sequence {N }N N is bounded.
Proof: An upper bound for N is provided by (III.23).
Let us then prove a lower bound.
(N )
For any N n
+ 1 we must have zn +1 0. The recursive
equation (III.22) for n = n
gives
(N )

0 (
n + 2) zn +1
(N )

= (
n + 1) zn

(N )

+ 1 + (1 N ) ln zn

(N )

(N )

+n
ln

(N )

zn

.
(N )
zn 1
(III.37)

(N )

Since zn is decreasing in n, we have zn zn 1 . Recalling

from (III.23) that 1 N 0, and from Corollary 10 that
(N )
zn z < 1, (III.37) implies
0

(N )

(N )

(
n + 1) zn + 1 + (1 N ) ln zn
(
n + 1) z + 1 + (1 N ) ln z ,

(III.38)

hence 1 N ((
n + 1) z + 1)/ ln z < .
The sequence {N }N N has then a converging subsequence
{Nk }kN with limk
o
n
n Nko= .
(N )
(N )
are conand p0
Since the sequences z0
N N
N N
strained between 0 and 1, we can also assume
(Nk )

lim z0

(Nk )

= z0 ,

lim p0

= p0 .

Taking the limit of (III.23) we get

z0 1
1
0.
ln z0

(III.39)

(III.40)

(N )
limk zn k

Lemma 12.
= zn for any n N, where the zn
are either all 0 or all strictly positive, and in the latter case
they satisfy for any n in N the recursive relation (III.22) with
N replaced by :
(n + 2) zn+1

= (n + 2) zn + 1 zn + (1 ) ln zn
zn
+n ln
.
(III.41)
zn1

Proof: If z0 = 0, since
for any n in N

(N )
zn

is decreasing in n we have

(Nk )

lim sup zn(Nk ) lim sup z0

k

= z0 = 0 ,

(III.42)

(N )

hence limk zn k = 0.
Let us now suppose z0 > 0, and proceed by induction on
n. From the inductive hypothesis, we can suppose
(Nk )

z0 = lim z0
k

. . . lim zn(Nk ) = zn > 0 .

k

(III.43)

Then, taking the limit in (III.22) we get

(N )

k
zn+1 = lim zn+1

zn
1 zn + (1 ) ln zn + n ln zn1

. (III.44)
n+2
If zn+1 > 0, the claim is proven. Let us then suppose zn+1 =
0. From (III.22) we get then
= zn +

1 + (n + 2 ) ln 0 (n + 1) ln zn
,
n+3
(III.45)
that implies n + 2 2. However, (III.40) implies 1.
(N )

k
0 lim zn+2
=

Lemma 13. There exists 0 z < 1 such that zn =

(N )
limk zn k = z for any n N.
Proof: If z0 = 0, Lemma 12 implies the claim with z = 0.
Let us then suppose z0 > 0.
If z0 = 1, we can prove by induction on n that zn = 1 for
any n N. The claim is true for n = 0. From the inductive
hypothesis we can suppose z0 = . . . = zn = 1. The relation
(III.41) implies then zn+1 = 1.
However, from Lemma 12 and Corollary 10 we must have
(N )
zn = limk zn k nz < 1.oThen, it must be 0 < z0 < 1.
(N )
Since the sequence zn
is decreasing for any N ,
nN
also the sequence {zn }nN is decreasing. Since it is also
positive, it has a limit limn zn = inf nN zn = z, that
satisfies 0 z z0 < 1. Since zn zn1 z0 < 1, (III.41)
implies
(n + 2) (zn zn+1 ) + 1 zn + (1 ) ln zn 0 , (III.46)
hence
1

(n + 2) (zn zn+1 ) zn 1
+
.
ln zn
ln zn

(III.47)

The sequence {zn zn+1 }nN is positive and satisfies

n=0

(zn zn+1 ) = z0 z < .

(III.48)

Then, from Lemma 26 lim inf n (n + 2) (zn zn+1 ) = 0,

and since ln zn ln z0 > 0, also
lim inf
n

(n + 2) (zn zn+1 )
=0.
ln zn

(III.49)

Then, taking the lim inf of (III.47) for n we get 1

(z 1)/ ln z. Combining with (III.40) and recalling that
z z0 we get
z1
z1
z0 1
1

,
ln z
ln z0
ln z

(III.50)

that implies z = z0 . Since zn is decreasing and z = inf nN zn ,

we have z0 = z zn z0 for any n, hence zn = z.
(N )

Lemma 14. limk pn k = p0 z n for any n N.

Proof: The claim is true for n = 0. The inductive

(N )
hypothesis is limk pn k = p0 z n for n = 0, . . . , n. We
(Nk )
(N ) (N )
then have limk pn+1
= limk pn k zn k = p0 z n+1 ,
where we have used the inductive hypothesis and Lemma 13.

(N )

Lemma 15. p0 = 1 z, hence limk pn k = (1 z) z n

for any n N.
PN
(N )
= 1 for any N N.
Proof: We have
n=0 pn
(N )
Moreover, since zn is decreasing in n, we also have
n

(N )
(N )
(N )
(N ) (N )
)
z
.
(III.51)

p
.
.
.
z
z
p(N
=
p
n
0
0
n1
0
0

Then, since DN DN +1 for any N , FN is increasing in

N , and for any p PN
F (p) FN sup FN = lim FN = lim FNk = f (S).
N N

(III.57)

(N )

Since limk z0 k = z < 1, for sufficiently large k we have

(N )
(N )
z0 k (1 + z) /2, and since p0 1,

n
1+z
k)

p(N
.
(III.52)
n
2
PNk (Nk )
ThePsums
are then dominated for any k N
n=0

n pn

< , and from the dominated convergence

by n=0 1+z
2
theorem we have
1

lim

Nk
X

n=0

k)
p(N
=
n

n=0

k)
lim p(N
= p0
n

zn

n=0

p0
,
(III.53)
1z
where we have used Lemma 14.



Lemma 16. z = g 1 (S) g 1 (S) + 1 .
Proof: The function x ln x is increasing for 0 x
1/e. Let us choose n0 such that ( (1 + z)/ 2)n0 1/e. RecallPNk
(N )
(N )
ing (III.52), the sums n=n
pn
k ln pn k are dominated
0
P
1+z
1+z n
ln 2 < . For any N
for any k N by n=n0 n 2
PN
(N )
(N )
we have S = n=0 pn ln pn . Then, from the dominated
convergence theorem and Lemma 15 we have
=

S=
=

k)
lim p(N
ln pn(Nk )
n

n=0

n=0

z
,
1z
(III.54)

where we have used the definition of g (II.6). Finally, the claim

follows solving (III.54) with respect to z.


It is convenient to rewrite FNk = F p(Nk ) as

(N )
zn k

NX
k 1

k ) (Nk )
(n + 1) p(N
zn ln zn(Nk ) .
n

(III.55)

n=0

Since
1, each term of the sum is positive. Since
x ln x 1/e for 0P x 1, and
recalling (III.52), the

1+z n
sum is dominated by n=0 n+1
< . We then have
e
2
from the dominated convergence theorem, recalling Lemmas
15 and 13,

X
k ) (Nk )
lim FNk =
(n + 1) lim p(N
zn ln zn(Nk ) =
n
k

n=0

n=0

(n + 1) (1 z) z n+1 ln z =


1
= g (S) ln 1 +

1
g 1 (S)

z ln z
=
z1

= f (S) , (III.56)

where we have used Lemma 16 and the definitions of f (III.2)

and g (II.6).

OF

T HEOREM 1

The idea for the proof is integrating the infinitesimal version

(III.1).
From Theorem 3, it is sufficient to prove Theorem 1 for
passive states, i.e. states of the form
=

n=0

pn |nihn| ,

p0 p1 . . . 0 .

(IV.1)

Lemma 17. If Theorem 1 holds for any passive state with

finite support, then it holds for any passive state.
Proof: Let be a passive state. If S ( (
)) = , there
is nothing to prove. We can then suppose S ( (
)) < .
We can associate to the probability distribution p on N
such that

X
=
pn |nihn| ,
(IV.2)
n=0

). Let us define for any

satisfying n=0 pn ln pn = S (
N N the quantum state
N =
where sN =

(1 z) z n (ln (1 z) + n ln z) = g

FNk =

IV. P ROOF

PN

n=0

N
X
pn
|nihn| ,
s
n=0 N

(IV.3)

pn . We have

X
1 sN X
pn +
pn ,
k
N k1 =
sN n=0

(IV.4)

n=N +1

where kk1 denotes the trace norm [10], [11]. Since

lim sN = 1 ,

pn = 1 ,

(IV.5)

n=0

we have limN k
N k1 = 0. Since is continuous in
the trace norm, we also have
lim k (
N ) (
)k1 = 0 .

(IV.6)

Moreover,
lim S (
N ) = lim

N
X
pn
ln pn
ln sN
s
n=0 N

= S (
) .

(IV.7)
Notice that (IV.7) holds also if S (
) = .
Let us now define the probability distribution q on N as
(
) =

n=0

satisfying

S ( (
)) =

qn |nihn| ,

(IV.8)

(IV.9)

n=0

qn ln qn .

From [42], Section IV.B, or [26], Equation (II.12), the channel

sends the set of states supported on the span of the first
N + 1 Fock states into itself. Then, for any N N there is a
probability distribution q (N ) on {0, . . . , N } such that
(
N ) =

N
X

n=0

qn(N ) |nihn| .

(IV.10)

lim

= qn .

(IV.11)

P
Since is trace preserving, we have n=0 qn = 1, hence
limn qn = 0. Then, there is n0 N (that does not depend
on N ) such that for any n n0 we have qn p0 /e.
Since sN N and the channel is positive, we have
sN (
N ) (
). Then, for any n n0
qn(N )

qn
qn
1

,
sN
p0
e

(IV.12)

where we have used that sN p0 > 0. Since the function

x ln x is increasing for 0 x 1/e, the sums

N
X

qn(N ) ln qn(N )

(IV.13)

n=n0

are dominated by

X
qn ln p0 qn ln qn
ln p0 + S ( (
))

<,
p0
p0
n=n0
(IV.14)
where we have used (IV.9). Then, from the dominated convergence theorem we have

lim S ( (
N ))

= lim

n=0

(0) = 0 (0) ,

(IV.18)

and

From (IV.6) we get for any n N

qn(N )

for S 0, and differentiable for S > 0. We define for any

t 0 the functions


(t) = S (et (
)) ,
0 (t) = g et g 1 (S (
)) .
(IV.17)
It is easy to show that

N
X

qn(N ) ln qn(N )

n=0

lim qn(N ) ln qn(N )

= S ( (
)) ,

(IV.15)

where we have also used (IV.11).

If Theorem 1 holds for passive states with finite support, for
any N in N we have


S ( (
N )) g g 1 (S (
N )) .
(IV.16)
Then, the claim follows taking the limit N .

From Lemma 17, we can suppose to be a passive state

with finite support.
Lemma 18. The quantum-limited attenuator satisfies the
composition rule = .
Proof: Follows from Lemma 13 of [26].
The function g(x) defined in (II.6) is differentiable for x >
0, and continuous and strictly increasing for x 0, and its
image is the whole interval [0, ). Then, its inverse g 1 (S) is
defined for any S 0, it is continuous and strictly increasing

d
0 (t) = f (0 (t)) ,
dt
where f is defined by (III.2).

(IV.19)

Lemma 19. f is differentiable for any S 0.

Proof: We have
1

f (S) =
(1 +

g 1 (S)) ln

hence limS0 f (S) = 1.


1+

1
g1 (S)

 1,

(IV.20)

Since the quantum-limited attenuator sends the set of passive states with finite support into itself (see Equation (II.12)
of [26]), we can replace et (
) in equation (III.1), and
from Theorem 2 and Lemma 18 we get
d
(t) f ((t)) .
dt
The claim then follows from

(IV.21)

Theorem 20 (Comparison theorem for first-order ordinary

differential equations). Let , 0 : [0, ) [0, ) be
differentiable functions satisfying (IV.18), (IV.19) and (IV.21)
with f : [0, ) R differentiable. Then, (t) 0 (t) for
any t 0.
Proof: See e.g. Theorem 2.2.2 of [43].
V. R ELATION WITH

THE DEGRADED
BROADCAST CHANNEL

G AUSSIAN

The quantum degraded Gaussian broadcast channel [16],

[17] maps a state A of the quantum system A to a state
A B of the joint quantum system A B with
(
,
A |0iB h0|) U
A B = U

(V.1)

is the unitary operator defined in (II.2), and 1/2

where U
1. It can be understood as follows. A encodes the
information into the state of the electromagnetic radiation A ,
and sends it through a beamsplitter of transmissivity . A and
B receive the transmitted and the reflected part of the signal,
respectively, whose joint state is A B .
The rate pair (RA , RB ) is achievable if for any n N A
can send to A and B with n uses of the channel any couple
(n)
(n)
of messages chosen from sets IA and IB with




 (n) 
 (n) 
(V.2)
IB  enRB
IA  enRA ,

with vanishing maximum error probability in the limit n

(see [44], Section II, and [17], Sections II and III; see also
[45], Definition 1). The closure of the set of the achievable
rate pairs constitutes the capacity region of the channel.

Let E > 0 be the maximum average energy per copy of the

input states. Ref. [17] first proves in Section IV that superposition coding allows to achieve with the quantum degraded
Gaussian broadcast channel (V.1) any rate pair (RA , RB )
satisfying
RA g ( E) ,
RB g ((1 ) E) g ((1 ) E)

(V.3)
(V.4)

for some 0 1. For the converse, Ref. [17] proves

an outer bound for the achievable rate pairs in Eqs. (22)
and (23). Any achievable rate pair (RA , RB ) must satisfy
the following property. For any n
exist an
n N there must o
A(n)
(n) (n)
on n
ensemble of encoding pure states pi qj , ij
copies of the quantum system A such that
n RA

 
A (n)
S j

(n)

qj

n RB

(n)

pi

A (n)

S ij

(V.6)

where n , n 0 for n , and



A B (n)
n ,
n A(n) (|0iB h0|)n U
=U
ij
ij

X (n) B (n)
B (n)
=
,
j
pi ij

(V.7)
(V.8)

(n)

(n)

qj

B (n)

(V.9)

Using the outer bounds (V.5), (V.6) and assuming the multimode version of the inequality (II.8), Ref. [17] proves that any
achievable rate pair (RA , RB ) must satisfy (V.3) and (V.4)
for some 0 1. Eqs. (V.3) and (V.4) then describe the
capacity region of the quantum degraded Gaussian broadcast
channel.
One may ask whether the one-mode inequality (II.8) implies
the outer bounds (V.3), (V.4) in the setting where the sender
A cannot entangle the input state among successive uses of
A(n)
the channel, i.e. when the pure states ij are product states.
This would be the case if the bounds (V.5), (V.6) were additive,
i.e. if they did not require the regularization over n. In this
case determining them for n = 1 would be sufficient. The
answer is negative.
the subadditivity of the entropy

 Indeed,
B (n)
in (V.6) goes in the wrong direction.
for the terms S j
B (n)

were product states, but from

Additivity would hold if j
(V.8) in general this is not the case.
VI. T HE

i=1

where the {Bn }nN+ are independent Bernoulli variables with

parameter , i.e. each Bi is one with probability , and zero
with probability 1 .

From a physical point of view, the thinning can be understood as follows: each incoming photon has probability of
being transmitted, and 1 of being reflected or absorbed.
Let N be the random variable associated to the number of
incoming photons, and {pn }nN its probability distribution,
i.e. pn is the probability that N = n (i.e. that n photons are
sent). Then, T (p) is the probability distribution of the number
of transmitted photons. It is easy to show that
[T (p)]n =

+ n n , (V.5)

 
 X

B (n)
(n)
(n)
+ n n ,
qj S j
S B

B =

Definition 21 (Thinning). Let N be a random variable with

values in N. The thinning with parameter 0 1 is defined
as
N
X
Bi ,
(VI.1)
T (N ) =

THINNING

The thinning [27] is the map acting on classical probability

distributions on the set of natural numbers that is the discrete
analogue of the continuous rescaling operation on positive real
numbers.

rn|k pk ,

(VI.2)

k=0

where the transition probabilities rn|k are given by

 
k n
rn|k =
(1 )kn ,
n

(VI.3)

and vanish for k < n.

The thinning coincides with the restriction of the attenuator
to input states diagonal in the Fock basis:
Theorem 22. Let and T be the quantum-limited attenuator and the thinning of parameter 0 1, respectively.
Then for any probability distribution p on N
!

X
X

pn |nihn| =
[T (p)]n |nihn| .
(VI.4)
n=0

n=0

Proof: See Theorem 56 of [26].

Thanks to Theorem 22, our main results Theorems 1 and 2

apply also to the thinning:
Theorem 23. Geometric input probability distributions minimize the output Shannon entropy of the thinning for fixed input
entropy, i.e. for any probability distribution p on N and any
0 1 we have


H (T (p)) g g 1 (H(p)) ,
(VI.5)
where g has been defined in (II.6).

Theorem 24. For any probability distribution p on N



d
H (Tet (p))
f (H(p)) ,
(VI.6)
dt
t=0

where f has been defined in (III.2).

VII. C ONCLUSION
We have proved that Gaussian thermal input states minimize
the output von Neumann entropy of the Gaussian quantumlimited attenuator for fixed input entropy (Theorem 1). The
proof is based on a new isoperimetric inequality (Theorem

2). Theorem 1 implies that geometric input probability distributions minimize the output Shannon entropy of the thinning
for fixed input entropy (Theorem 23). Its multimode extension
would permit to determine both the triple trade-off region of
the Gaussian quantum-limited attenuator [20], [21], [33] and
the classical capacity region of the Gaussian quantum degraded
broadcast channel [16], [17]. The proof of Theorem 1 for
the multimode attenuator would follow from the multimode
generalization of the isoperimetric inequality (III.1). However,
our proof of (III.1) relies on the majorization result of Ref.
[26], that does not hold for the multimode attenuator (see [46],
Section IV.A).
A PPENDIX A
AUXILIARY L EMMAS
Lemma 25. Let us choose a probability distribution p DN ,
fix 0 N N , and define the probability distribution q
DN as
pn
qn = PN
,
n = 0, . . . , N .
(A.1)
k=0 pk
Then, H(q) H(p).
Proof: We have for n = 0, . . . , N
Pn
n
n
X
X
pk

pk ,
qk = Pk=0

N
l=0 pl
k=0
k=0

(A.2)

Then, q majorizes p and the claim follows from Remark 20 of

[26].
Lemma 26. Let {xn }nN be a positive sequence with finite
sum. Then lim inf n n xn = 0.
Proof: Let us suppose lim inf n n xn = c > 0. Then,
there exists n0 N such that n xn c/2 for any n n0 .
Then,

X
X
c
xn
=,
(A.3)
2n
n=n
n=0
0

contradicting the hypothesis.

ACKNOWLEDGMENT
The Authors thank Luigi Ambrosio for comments and
fruitful discussions.
R EFERENCES
[1] T. Cover and J. Thomas, Elements of Information Theory, ser. A WileyInterscience publication. Wiley, 2006.
[2] A. Dembo, T. M. Cover, and J. Thomas, Information theoretic inequalities, Information Theory, IEEE Transactions on, vol. 37, no. 6, pp.
15011518, 1991.
[3] R. Gardner, The brunn-minkowski inequality, Bulletin of the American
Mathematical Society, vol. 39, no. 3, pp. 355405, 2002.
[4] C. E. Shannon, A mathematical theory of communication, ACM
SIGMOBILE Mobile Computing and Communications Review, vol. 5,
no. 1, pp. 355, 2001.
[5] A. Stam, Some inequalities satisfied by the quantities of information of
fisher and shannon, Information and Control, vol. 2, no. 2, pp. 101112,
1959.
[6] S. Verdu and D. Guo, A simple proof of the entropy-power inequality,
IEEE Transactions on Information Theory, vol. 52, no. 5, pp. 2165
2166, 2006.

[7] O. Rioul, Information theoretic proofs of entropy power inequalities,

Information Theory, IEEE Transactions on, vol. 57, no. 1, pp. 3355,
2011.
[8] J. P. Gordon, Quantum effects in communications systems, Proceedings of the IRE, vol. 50, no. 9, pp. 18981908, 1962.
[9] J. Chen, J. L. Habif, Z. Dutton, R. Lazarus, and S. Guha, Optical
codeword demodulation with error rates below the standard quantum
limit using a conditional nulling receiver, Nature Photonics, vol. 6,
no. 6, pp. 374379, 2012.
[10] M. Wilde, Quantum Information Theory. Cambridge University Press,
2013.
[11] A. S. Holevo, Quantum Systems, Channels, Information: A Mathematical Introduction, ser. De Gruyter Studies in Mathematical Physics. De
Gruyter, 2013.
[12] V. W. Chan, Free-space optical communications, Lightwave Technology, Journal of, vol. 24, no. 12, pp. 47504762, 2006.
[13] S. L. Braunstein and P. Van Loock, Quantum information with continuous variables, Reviews of Modern Physics, vol. 77, no. 2, p. 513,
2005.
[14] C. Weedbrook, S. Pirandola, R. Garcia-Patron, N. J. Cerf, T. C. Ralph,
J. H. Shapiro, and S. Lloyd, Gaussian quantum information, Reviews
of Modern Physics, vol. 84, no. 2, p. 621, 2012.
[15] A. S. Holevo, Gaussian optimizers and the additivity problem in
quantum information theory, Uspekhi Matematicheskikh Nauk, vol. 70,
no. 2, pp. 141180, 2015.
[16] S. Guha and J. H. Shapiro, Classical information capacity of the
bosonic broadcast channel, in Information Theory, 2007. ISIT 2007.
IEEE International Symposium on. IEEE, 2007, pp. 18961900.
[17] S. Guha, J. H. Shapiro, and B. I. Erkmen, Classical capacity of bosonic
broadcast communication and a minimum output entropy conjecture,
Physical Review A, vol. 76, no. 3, p. 032303, 2007.
[18] S. Guha, B. Erkmen, and J. H. Shapiro, The entropy photon-number inequality and its consequences, in Information Theory and Applications
Workshop, 2008. IEEE, 2008, pp. 128130.
[19] S. Guha, J. H. Shapiro, and B. Erkmen, Capacity of the bosonic wiretap
channel and the entropy photon-number inequality, in Information
Theory, 2008. ISIT 2008. IEEE International Symposium on. IEEE,
2008, pp. 9195.
[20] M. M. Wilde, P. Hayden, and S. Guha, Information trade-offs for optical
quantum communication, Physical Review Letters, vol. 108, no. 14, p.
140501, 2012.
[21] , Quantum trade-off coding for bosonic communication, Physical
Review A, vol. 86, no. 6, p. 062306, 2012.
[22] R. Konig and G. Smith, Limits on classical communication from
quantum entropy power inequalities, Nature Photonics, vol. 7, no. 2,
pp. 142146, 2013.
[23] R. Konig and G. Smith, The entropy power inequality for quantum
systems, Information Theory, IEEE Transactions on, vol. 60, no. 3, pp.
15361548, 2014.
[24] G. De Palma, A. Mari, and V. Giovannetti, A generalization
of the entropy power inequality to bosonic quantum systems,
Nature Photonics, vol. 8, no. 12, pp. 958964, 2014. [Online]. Available:
http://www.nature.com/nphoton/journal/v8/n12/full/nphoton.2014.252.html
[25] G. De Palma, A. Mari, S. Lloyd, and V. Giovannetti,
Multimode quantum entropy power inequality, Physical Review
A, vol. 91, no. 3, p. 032320, 2015. [Online]. Available:
http://journals.aps.org/pra/abstract/10.1103/PhysRevA.91.032320
[26] G. De Palma, D. Trevisan, and V. Giovannetti, Passive
states optimize the output of bosonic gaussian quantum
channels, IEEE Transactions on Information Theory, vol. 62,
no. 5, pp. 28952906, May 2016. [Online]. Available:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=7442587
[27] A. Renyi, A characterization of poisson processes, Magyar Tud. Akad.
Mat. Kutato Int. Kozl, vol. 1, pp. 519527, 1956.
[28] P. Harremoes, O. Johnson, and I. Kontoyiannis, Thinning and the
law of small numbers, in Information Theory, 2007. ISIT 2007. IEEE
International Symposium on. IEEE, 2007, pp. 14911495.
[29] Y. Yu, Monotonic convergence in an information-theoretic law of small
numbers, Information Theory, IEEE Transactions on, vol. 55, no. 12,
pp. 54125422, 2009.
[30] P. Harremoes, O. Johnson, and I. Kontoyiannis, Thinning, entropy, and
the law of thin numbers, Information Theory, IEEE Transactions on,
vol. 56, no. 9, pp. 42284244, 2010.
[31] Y. Yu and O. Johnson, Concavity of entropy under thinning, in
Information Theory, 2009. ISIT 2009. IEEE International Symposium
on. IEEE, 2009, pp. 144148.

10

[32] O. Johnson and Y. Yu, Monotonicity, thinning, and discrete versions of

the entropy power inequality, Information Theory, IEEE Transactions
on, vol. 56, no. 11, pp. 53875395, 2010.
[33] M. M. Wilde and M.-H. Hsieh, Public and private resource trade-offs
for a quantum channel, Quantum Information Processing, vol. 11, no. 6,
pp. 14651501, 2012.
[34] A. Ferraro, S. Olivares, and M. G. Paris, Gaussian states in continuous
variable quantum information, arXiv preprint quant-ph/0503237, 2005.
[35] W. Pusz and S. Woronowicz, Passive states and kms states for general
quantum systems, Communications in Mathematical Physics, vol. 58,
no. 3, pp. 273290, 1978.
[36] A. Lenard, Thermodynamical proof of the gibbs formula for elementary
quantum systems, Journal of Statistical Physics, vol. 19, no. 6, pp. 575
586, 1978.
[37] J. Gorecki and W. Pusz, Passive states for finite classical systems,
Letters in Mathematical Physics, vol. 4, no. 6, pp. 433443, 1980.
[38] S. Vinjanampathy and J. Anders, Quantum thermodynamics, Contemporary Physics, vol. 0, no. 0, pp. 135, 2016.
[39] J. Goold, M. Huber, A. Riera, L. del Rio, and P. Skrzypczyk, The
role of quantum information in thermodynamics a topical review,
Journal of Physics A: Mathematical and Theoretical, vol. 49, no. 14, p.
143001, 2016.
[40] F. Binder, S. Vinjanampathy, K. Modi, and J. Goold, Quantum thermodynamics of general quantum processes, Physical Review E, vol. 91,
no. 3, p. 032119, 2015.
[41] H. W. Kuhn and A. W. Tucker, Nonlinear programming, in Proceedings of the Second Berkeley Symposium on Mathematical Statistics and
Probability. Berkeley, Calif.: University of California Press, 1951, pp.
481492.
[42] J. S. Ivan, K. K. Sabapathy, and R. Simon, Operator-sum representation
for bosonic gaussian channels, Physical Review A, vol. 84, no. 4, p.
042311, 2011.
[43] W. Ames and B. Pachpatte, Inequalities for Differential and Integral
Equations, ser. Mathematics in Science and Technology.
Elsevier
Science, 1997.
[44] J. Yard, P. Hayden, and I. Devetak, Quantum broadcast channels,
Information Theory, IEEE Transactions on, vol. 57, no. 10, pp. 7147
7162, 2011.
[45] I. Savov and M. M. Wilde, Classical codes for quantum broadcast
channels, Information Theory, IEEE Transactions on, vol. 61, no. 12,
pp. 70177028, 2015.
[46] G. De Palma, A. Mari, S. Lloyd, and V. Giovannetti, Passive states
as optimal inputs for single-jump lossy quantum channels, Physical
Review A, vol. 93, no. 6, p. 062328, 2016.

Giacomo De Palma was born in Lanciano (CH), Italy, on March 15, 1990.
He received the B.S. degree in Physics and the M.S. degree in Physics from
the University of Pisa, Pisa (PI), Italy, in 2011 and 2013, respectively. He
also received the Diploma di Licenza in Physics and the Ph.D. degree in
Physics from Scuola Normale Superiore, Pisa (PI), Italy, in 2014 and 2016,
respectively.
He is currently a postdoc at the University of Copenhagen, Copenhagen,
Denmark.
His research interests include quantum information, quantum statistical
mechanics and quantum thermodynamics. He is author of eleven scientific
papers published in peer-reviewed journals.

Dario Trevisan was born in the Province of Venice, Italy. He received the
B.S. degree in mathematics from the University of Pisa, Pisa, Italy, in 2009,
the M.S. degree in mathematics from the University of Pisa, in 2011, and the
Ph.D. degree in mathematics from the Scuola Normale Superiore, Pisa, Italy,
in 2014.
He is currently Assistant Professor at the University of Pisa.
Dr. Trevisan is a member of the GNAMPA group of the Italian National
Institute for Higher Mathematics (INdAM).

Vittorio Giovannetti was born in Castelnuovo di Garfagnana (LU) Italy, on

April 1, 1970. He received the M.S. degree in Physics from the University of
Pisa and PhD degree in theoretical Physics from the University of Perugia.
He is currently Associate Professor at the Scuola Normale Superiore of
Pisa.