Strong Converse Rates for QuantumCommunication
Marco Tomamichel1, Mark M. Wilde2, Andreas Winter3
1School of Physics, The University of Sydney2Center for Computation and Technology, Dept. of Physics & Astronomy,
Louisiana State University3ICREA & Fısica Teorica, Informacio i Fenomens Quantics,
Universitat Autonoma de Barcelona
ISIT 2015, Hong Kong
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Information Theory
• Information theory is concerned with establishing fundamentallimits for information processing tasks.
• The most precise modeling of physical information carriers(known today) uses the formalism of quantum mechanics.
Source: Wikipedia
• Many fundamental questions remain unanswered, for example:
1 No single-letter formula for point-to-point channel capacity.2 Strong converse for memoryless channels unknown.
2 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Primer: States
classical world quantum world
alphabet: X = {0, 1} Hilbert space: X = C2 — “qubit”orthonormal basis: |0〉, |1〉 ∈ X.
pmf: PX(x) = px quantum state: ρX =
(p0 00 p1
)generally: ρX ∈ S(X), linear operator on Xwith ρX ≥ 0 and tr(ρX) = 1.
t t
t t
t
|1〉
|0〉
t t|−〉 |+〉
|±〉 = 1√2(|0〉 ± |1〉)
3 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Primer: Correlation vs. Entanglement
• Relative entropy: D(ρ‖σ) = tr(ρ(log ρ− log σ)
).
• Entropy: H(A)ρ = −D(ρA‖1A) = − tr(ρ log ρ).
• Cond. entropy: H(A|B)ρ = −D(ρAB‖1A ⊗ ρB) = H(AB)ρ −H(B)ρ.
classical world quantum world
correlated pmf: entangled quantum state:
PXX′(x, x′) =
{px x = x′
0 else|φ〉XX′ =
√p0|0〉 ⊗ |0〉+
√p1|1〉 ⊗ |1〉
φXX′ = |φ〉〈φ|XX′ =
p0 0 0
√p0p1
0 0 0 00 0 0 0√p0p1 0 0 p1
not extremal extremal (pure, rank one)
entropy: H(X)P H(X)φ = H(X)Pcond. entropy: H(X|X ′)P = 0 H(X|X ′)φ = −H(X)P
4 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Primer: Correlation vs. Entanglement
• Relative entropy: D(ρ‖σ) = tr(ρ(log ρ− log σ)
).
• Entropy: H(A)ρ = −D(ρA‖1A) = − tr(ρ log ρ).
• Cond. entropy: H(A|B)ρ = −D(ρAB‖1A ⊗ ρB) = H(AB)ρ −H(B)ρ.
classical world quantum world
correlated pmf: entangled quantum state:
PXX′(x, x′) =
{px x = x′
0 else|φ〉XX′ =
√p0|0〉 ⊗ |0〉+
√p1|1〉 ⊗ |1〉
φXX′ = |φ〉〈φ|XX′ =
p0 0 0
√p0p1
0 0 0 00 0 0 0√p0p1 0 0 p1
not extremal extremal (pure, rank one) not extremal (mixed)
entropy: H(X)P H(X)φ = H(X)Pcond. entropy: H(X|X ′)P = 0 H(X|X ′)φ = −H(X)P
5 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Primer: Correlation vs. Entanglement
• Relative entropy: D(ρ‖σ) = tr(ρ(log ρ− log σ)
).
• Entropy: H(A)ρ = −D(ρA‖1A) = − tr(ρ log ρ).
• Cond. entropy: H(A|B)ρ = −D(ρAB‖1A ⊗ ρB) = H(AB)ρ −H(B)ρ.
classical world quantum world
correlated pmf: entangled quantum state:
PXX′(x, x′) =
{px x = x′
0 else|φ〉XX′ =
√p0|0〉 ⊗ |0〉+
√p1|1〉 ⊗ |1〉
φXX′ = |φ〉〈φ|XX′ =
p0 0 0
√p0p1
0 0 0 00 0 0 0√p0p1 0 0 p1
not extremal extremal (pure, rank one)
entropy: H(X)P H(X)φ = H(X)Pcond. entropy: H(X|X ′)P = 0 H(X|X ′)φ = −H(X)P
6 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Coding: Channels
• Quantum channel: completely positive trace-preservinglinear map N ≡ NA→B from (states on) A to (states on) B.
NA B
Assume A and B are finite-dimensional.
• The channel is memoryless:
NAn Bn
NA2 B2
NA1 B1
≡ N⊗nAn Bn
7 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Coding: Encoder and Decoder
• Entanglement transmission code (for N⊗n):
Cn = {dn, En,Dn) .
1 code size dn:• M , M ′, M ′′ of dimension dn.• maximally entangled state
|φ〉MM′ =1√dn
dn∑i=1
|i〉M⊗|i〉M′ .
2 encoder En: quantum channelfrom M ′ to An.
3 decoder Dn: quantum channelfrom Bn to M ′′.
|φ〉
M ′
M
8 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Coding: Encoder and Decoder
• Entanglement transmission code (for N⊗n):
Cn = {dn, En,Dn) .
1 code size dn:• M , M ′, M ′′ of dimension dn.• maximally entangled state
|φ〉MM′ =1√dn
dn∑i=1
|i〉M⊗|i〉M′ .
2 encoder En: quantum channelfrom M ′ to An.
3 decoder Dn: quantum channelfrom Bn to M ′′.
M ′
En
An
A2
A1
9 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Coding: Encoder and Decoder
• Entanglement transmission code (for N⊗n):
Cn = {dn, En,Dn) .
1 code size dn:• M , M ′, M ′′ of dimension dn.• maximally entangled state
|φ〉MM′ =1√dn
dn∑i=1
|i〉M⊗|i〉M′ .
2 encoder En: quantum channelfrom M ′ to An.
3 decoder Dn: quantum channelfrom Bn to M ′′.
M ′′
Dn
Bn
B2
B1
10 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Coding: Entanglement Fidelity
N
N
N|φ〉
M ′
M
M ′′
En Dn
An
A2
A1
Bn
B2
B1
• Fidelity with maximally entangled state:
F (Cn, N⊗n) = tr(
(Dn ◦ N⊗n ◦ En)(φMM ′)φMM ′′).
• Corresponds to Pr[M = M ′′] classically.
11 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Capacity
• A triple (R,n, ε) is achievable on N if ∃ Cn with
1
nlog dn ≥ R, and F (CnN⊗n) ≥ 1− ε .
• Boundary of (non-asymptotic) achievable region:
R(n; ε,N ) := max{R : (R,n, ε) is achievable on N
}.
• The quantum capacity, Q(N ), is the rate at which qubits canbe transmitted with fidelity approaching one asymptotically.
Qε(N ) := limn→∞
R(n; ε,N ), ε ∈ (0, 1)
Q(N ) := limε→0
Qε(N ) .
12 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Quantum Capacity Theorem
• Barnum, Nielsen and Schumacher (1996-2000) as well asLloyd, Shor and Devetak (1997-2005) established
Q(N ) = lim`→∞
1
`Ic
(N⊗`
), where
Ic(N ) = maxρA{−H(A|B)ω} , ωAB = NA′→B(ψρA′A)
• This result is unsatisfactory for many reasons:
1 It is not a single-letter formula.2 The limit `→∞ is necessary in general (Cubitt et al.’14).3 It cannot be calculated except for e.g. degradable channels
which satisfy Ic(N⊗n
)= nIc(N ).
4 We do not know anything about Qε(N ).
13 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Strong Converse Property
• A quantum channel satisfies the strong converse property iff
Qε(N ) = Q(N ) for all ε ∈ (0, 1) .
0.40 0.45 0.50 0.55 0.60 0.650.0
0.2
0.4
0.6
0.8
1.0
n→∞
rate, R
toleratederror,ε
HHYachievable region
capacity ?
• A strong converse rate is a uniform upper bound on Qε.
14 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
State of the Art
• Until this work, the strong converse property could onlybe established for some channels with trivial capacity.
• Morgan and Winter showed that degradable quantumchannels satisfy a “pretty strong” converse:
Qε(N ) = Q(N ) for all ε ∈(
0,1
2
)
(Extending their proof to all ε ∈ (0, 1) appears difficult.)
• Strong converse rates are known, for example theentanglement-assisted capacity established via channelsimulation by Bennett et al. and Berta et al.
• However, they are not tight except for trivial channels.
A lot of (fundamental) work still needs to be done!
15 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Result 1: Rains Entropy is Strong Converse Rate
• The Rains relative entropy of the channel is defined as
R(N ) := maxρA
minσAB∈Rains(A:B)
D(NA′→B(ψρA′A)
∥∥σAB).
• A state σAB ∈ Rains(A : B) (cf. Rains’99) satisfies
tr(φABσAB
)≤ 1
d∀ maximally entangled φAB.
Result
For any channel N , communication at a rate exceeding R(N )implies (exponentially) vanishing fidelity.
• Key Idea: Consider correlations σAB that are useless forquantum communication. Classically:
C(W ) = maxPX
minQX ,QY
D(PX ×WY |X
∥∥QX ×QY).
16 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Step 1: Arimoto-Type (One-Shot) Converse Bound
• Consider C = {d, E ,D} for N with F (C,N ) ≥ 1− ε.
• Test if a state is φMM ′′ , or not:
T (·) = p|0〉〈0|+ (1− p)|1〉〈1|, p = tr(φMM ′′ ·).
• Let ρAM = E(φMM ′). Due to data-processing, we have
Dα(N (ρAM )‖σBM ) ≥ Dα
(T ◦ D ◦ N (ρAM )‖T ◦ D(σRB)
)
≥ log d+α
α− 1log(1− ε),
for Renyi divergence with α > 1
• Sandwiched Renyi divergence (Lennert et al., Wilde et al.’13):
Dα(ρ‖σ) =1
α− 1log tr
((σ
1−α2α ρσ
1−α2α
)α).
17 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Step 2: Asymptotics
• Minimizing σAB ∈ Rains(A : B) and optimizing over codes:
Lemma
We have the following one-shot converse:
R(1; ε,N ) ≤ maxρA
minσAB
Dα(NA′→B(ψρAA′)‖σAB) +α log 1
1−εα− 1
• This yields an upper bound on the ε-capacity:
Qε(N ) ≤ lim supn→∞
1
nmaxρAn
minσAnBn
Dα
(N⊗n(ψρAnA′n)
∥∥σAnBn)
︸ ︷︷ ︸Rα(N⊗n)
.
• It remains to show that Rα(N ) satisfies an asymptoticsub-additivity property, i.e. Rα(N⊗n) ≤ nRα(N ) + o(n).
18 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Step 3: Covariant Channels and Permutations
• Covariance group of the channel N : Group G with unitaryrepresentations UA and VB such that
NA→B(UA(g)( · )U †A(g)) = VB(g)NA→B( · )V †B(g) ∀g ∈ G
• We show that:
Rα(N ) = maxρA
minσAB
Dα
(NA′→B(ψρAA′)
∥∥σAB)
where ρA = UA(g)ρAU†A(g), i.e. ρA is invariant under G.
• Covariance group of N⊗n always contains permutations Sn.
• Thus, we can restrict the optimization in Rα(N⊗n) topermutation invariant states ρAn .
19 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Step 4: Asymptotic Sub-Additivity
• Employ the fact that ψρAnA′n is in the symmetric subspace:
ψρAA′ ≤ PsymmAnRn ≤ n|A|
2
∫dµ(θ) θ⊗nAR .
• The quantum way to restrict to product states in the converse.
• This allows us to show (skipping a few technical steps) that
Rα(N⊗n) ≤ nRα(N ) +O(log(n)).
• Hence, Qε(N ) ≤ Rα(N ) for all α > 1.
• And, thus, by continuity as α→ 1, we find Qε(N ) ≤ R(N ).
• A more detailed analysis reveals that the fidelity convergesexponentially fast to 0 for any d > R(N ).
20 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Result 2: Dephasing Channels Satisfy Strong Converse
• For all quantum channels we thus have
Ic(N ) ≤ Q(N ) ≤ Qε(N ) ≤ R(N )
for all ε ∈ (0, 1).
Result
For generalized dephasing channels Z, we have Ic(Z) = R(Z).
• The inequalities collapse and Qε(Z) = Q(Z).
• Includes qubit dephasing channel:
Zλ : ρ 7→ (1− λ)ρ+ λZ ρZ .(a cc† b
)7→(
a (1− 2λ)c(1− 2λ)c† b
)
21 / 22
Introduction Coding for Quantum Channels Main Result and Proof Dephasing Channels Conclusion
Summary and Outlook
• Our strong converse also holds with one-way classicalcommunication assistance. However, it is unclear if it alsoholds for two-way assistance.
• For more details on this result, see:
arX
iv:1
406.
2946
v2 [
quan
t-ph]
22
Nov
201
4
1
Strong Converse Rates for Quantum CommunicationMarco Tomamichel∗, Mark M. Wilde†, Andreas Winter‡
Abstract
We revisit a fundamental open problem in quantum information theory, namely whether it is possible totransmit quantum information at a rate exceeding the channel capacity if we allow for a non-vanishing probabilityof decoding error. Here we establish that the Rains information of any quantum channel is a strong converse ratefor quantum communication: For any code with a rate exceeding the Rains information of the channel, we showthat the fidelity vanishes exponentially fast as the number of channel uses increases. This remains true even if weconsider codes that perform classical post-processing on the transmitted quantum data.
Our result has several applications. Most importantly, for generalized dephasing channels we show that theRains information is also achievable, and thereby establish the strong converse property for quantum communicationover such channels. This for the first time conclusively settles the strong converse question for a class of quantumchannels that have a non-trivial quantum capacity. Moreover, we show that the classical post-processing assistedquantum capacity of the quantum binary erasure channel satisfies the strong converse property.
I. INTRODUCTION
The quantum capacity of a quantum channel N , denoted Q(N ), is defined as the maximum rate (inqubits per channel use) at which it is possible to transmit quantum information over many memoryless usesof the channel, such that a receiver can recover the state with a fidelity that asymptotically converges toone as we increase the number of channel uses. (We will formally introduce capacity in Section II-B.) Thequestion of determining the quantum capacity was set out by Shor in his seminal paper on quantum errorcorrection [1]. Since then, a number of works established a “multi-letter” upper bound on the quantumchannel capacity in terms of the coherent information [2]–[4], and the coherent information lower boundon quantum capacity was demonstrated by a sequence of works [5]–[7] which are often said to bear“increasing standards of rigor.”1 In more detail, the work in [5]–[7] showed that the following coherentinformation of the channel N is an achievable rate for quantum communication:
Ic (N ) := maxφRA
I(R⟩B)ρ, where ρRB = NA→B (φRA) (1)
and the optimization is over all pure bipartite states φRA. Here, the coherent information of a bipartite stateρRB is defined as I(R⟩B)ρ := H(ρB)−H(ρRB), with the von Neumann entropies H(ρ) := −Tr{ρ log ρ}.2From the above result, we can also conclude that the rate Ic
!N ⊗l
"/l is achievable for any positive integer l,
simply by applying the formula in (1) to the “superchannel” N ⊗l and normalizing. By a limiting argument,we find that the regularized coherent information liml→∞ Ic
!N ⊗l
"/l is also achievable, and Refs. [2]–[4]
established that this regularized coherent information is also an upper bound on quantum capacity. Thisestablishes that
Q(N ) = liml→∞
Ic(N ⊗l)
l. (2)
Clearly, the regularized coherent information is not a tractable characterization of quantum capacity. Butthe later work of Devetak and Shor proved that the “single-letter” coherent information formula in (1) is
∗ School of Physics, The University of Sydney, Sydney, NSW 2006, Australia. (Email: [email protected]).† Hearne Institute for Theoretical Physics, Department of Physics and Astronomy, Center for Computation and Technology, Louisiana
State University, Baton Rouge, Louisiana 70803, USA.‡ ICREA & Fısica Teorica, Informacio i Fenomens Quantics, Universitat Autonoma de Barcelona, ES-08193 Bellaterra (Barcelona), Spain.1However, see the later works [8] and [9], which set [5] and [6], respectively, on a firm foundation.2All logarithms in this paper are taken base two.
arXiv:1406.2
946
• The techniques also allow to establish a one-shot“meta-converse” in terms of binary quantum hypothesistesting (see arXiv: 1504.04617).
22 / 22
Top Related