Beruflich Dokumente
Kultur Dokumente
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
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.
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
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
that satisfies
a
+ 1 b ,
U
U =
(II.3)
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
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)
INEQUALITY
f (S) := g 1 (S) g g 1 (S) .
(III.2)
N
X
n=0
pn |nihn| .
(III.5)
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
(III.10)
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)
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
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) =
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)
zN = 0 .
(III.21)
(N )
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 )
(N )
zn+1
and hence
the hypothesis
(N )
zn
(N )
zn1
(N )
(N )
0,
(III.26)
zn
(N )
zn1
0,
(III.28)
<1,
(III.29)
where
n
= min n N : n + 2 > eS
(III.30)
=0.
(III.24)
.
(N )
(N )
< z0 1
ln z0 .
(N )
(III.27)
Lemma 9. We have
(N )
lim sup zn
. . . zN 1 > zN
(N )
(N )
(N )
. We will
(N )
zn
(N )
ln zn
lim sup zn
0.
(N )
ln z0
(N )
ln z0
(N )
1 z0
z0
zn
(N )
+ n ln
.
(N )
ln z0
(N )
n = 0, . . . , N 1 ,
(N )
+n ln
(N )
0 < zn(N ) 1
(N )
z0
(N )
(N )
p=p
pn
z0
(III.18)
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 ) =
1.
(III.31)
(N )
n = 0, . . . , n
.
(III.32)
pn
qn(N ) = Pn +1
k=0
(N )
pk
n = 0, . . . , n
+1.
(III.33)
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 )
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 )
(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 .
(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 )
= 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
(III.43)
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
=
(n + 2) (zn zn+1 ) zn 1
+
.
ln zn
ln zn
(III.47)
n=0
(III.48)
(n + 2) (zn zn+1 )
=0.
ln zn
(III.49)
,
ln z
ln z0
ln z
(III.50)
(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 )
p
.
.
.
z
z
p(N
=
p
n
0
0
n1
0
0
(III.57)
(N )
p(N
.
(III.52)
n
2
PNk (Nk )
ThePsums
are then dominated for any k N
n=0
n pn
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)
(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)
OF
T HEOREM 1
n=0
pn |nihn| ,
p0 p1 . . . 0 .
(IV.1)
X
=
pn |nihn| ,
(IV.2)
n=0
(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
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 .
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)
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
N
X
qn(N ) ln qn(N )
n=0
= S ( (
)) ,
(IV.15)
d
0 (t) = f (0 (t)) ,
dt
where f is defined by (III.2).
(IV.19)
f (S) =
(1 +
g 1 (S)) ln
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)
THE DEGRADED
BROADCAST CHANNEL
G AUSSIAN
(V.1)
(V.3)
(V.4)
A (n)
S j
(n)
qj
n RB
(n)
pi
A (n)
S ij
(V.6)
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)
i=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 =
THINNING
rn|k pk ,
(VI.2)
k=0
(VI.3)
X
X
pn |nihn| =
[T (p)]n |nihn| .
(VI.4)
n=0
n=0
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)
X
X
c
xn
=,
(A.3)
2n
n=n
n=0
0
10
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).