classical-quantum channels
Omar Fawzi, Patrick Hayden, Pranab Sen, Ivan Savov†, Mark Wilde
McGill University School of Computer Science
September 19, 2012
x1
Tx1
x2
Tx2
y Rx
(a) MAC ≡ p(y|x1, x2)
x1
Tx1
x2
Tx2
y1 Rx1
y2 Rx2
(b) IC ≡ p(y1, y2|x1, x2)
x
Tx
y1 Rx1
y2 Rx2 (c) BC ≡ p(y1, y2|x)
x
Tx
y1 Re
x1
y Rx (d) RC ≡ p(y1, y|x, x1)
Classical-quantum channel
(X, NX→B(x) ≡ ρBx, HB) ⇔ Tx x NX→B ρBx Rx
I Input: x ∈ X (finite set).
I Outputs: conditional density matrices {ρBx} ∈ HB, hv|ρBx|vi ≥ 0, ∀|vi ∈ HB, (ρBx)† =ρBx, Tr[ρBx] = 1.
Each output state can be decomposed as follows:
ρBx =X
y
λρx;y|eρx;yiheρx;y| =X
y
pY |X(y|x)|eρx;yiheρx;y|,
where we identify eigenvalues ofρxwith a conditional probability dist.
B
ya ∈ Y ⇐⇒ |viB ∈ HB
symbol from a finite set vector in a Hilbert space
Y ≡ pY ∈ P(Y) ⇐⇒ ρB ≡ ρB ∈ D(HB)
probability distribution density matrix ≡ quantum state pY(y) ≥ 0, ∀y ∈ Y hv|ρB|vi ≥ 0, ∀|vi ∈ HB
P
ypY(y) = 1 Tr[ρB] = 1, (ρB)†= ρB pY |X ⇐⇒ {ρBx}, x ∈ X
conditional probability distribution conditional states
≡ classical-classical channel ≡ classical-quantum channel pXY(x, y) ≡ pX(x)pY |X(y|x) ⇐⇒ θXB≡P
xpX(x) |xihx|X⊗ ρBx
joint input-output distribution joint input-output state 1n
yn∈Tδ(n)(Y |xn)o ⇐⇒ Πxn≡ ΠBρxnn,δ indicator function for the conditionally typical
conditionally typical set projector for the state ρBxnn
Quantum multi-user scenarios
I Classical multi-user decoding is based on the notion of jointly typical sequences: (xn1(m1), xn2(m2), yn) ∈ Jδn(X1, X2, Y ).
I Equivalently, we can use the conditionally-typical sets: T (Y ), T (Y |xn1), T (Y |xn2), T (Y |xn1, xn2):
1{yn∈T (Y )}1{yn∈T (Y |xn1(m1))} 1{yn∈T (Y |xn2(m2))} 1{yn∈T (Y |xn1(m1),xn2(m2))}
I The quantum equivalents of 1{yn ∈T(Y |xn
1 (m1))}and
1{yn ∈T(Y |xn2 (m2))}are the conditionally typical projectors Πxn1(m1)
and Πxn2(m2), which in generaldo not commute:
Πxn
1(m1)Πxn
2(m2)6= Πxn2(m2)Πxn 1(m1).
I Quantum multi-user decodingrequires new techniques.
Outline
I Introduce basic notions:
I Classical-quantum communication model
I Conditionally typical projectors
I Show the proof of simultaneous decoder for the quantum multiple access cannel
I Briefly discuss other problems in network information theory:
I Quantum interference channels
I Quantum broadcast channels
I Quantum relay channels
Ivan Savov QNIT
x1
Tx1
x2
Tx2
ρBx1,x2 Rx
(a) QMAC ≡˘ρBx1,x2¯
x1
Tx1
x2
Tx2
ρB1 x1,x2 Rx1
ρB2 x1,x2 Rx2
(b) QIC ≡ n
ρBx11,xB22
o
Tx x
ρB1x Rx1
ρB2x Rx2 n B1B2o
Tx x
ρB1 x,x1
Re x1
ρBx,x1 Rx
n B1Bo
Classical-quantum channel coding
Consider first thepoint-to-pointcommunication scenario.
WANTED Noiseless classical communication:
[c → c]
Ivan Savov QNIT
Classical-quantum channel coding
AVAILABLE Noisy classical-quantum channel:
(X, NX→B(x) ≡ ρBx, HB)
Tx x NX→B ρBx Rx
Classical-quantum channel coding
We are allowed to use the channel many times in parallel:
x1 ρB1
x1 N X→B
x2 ρB2
x2 N X→B
x3 ρB3
x3 N X→B
x4 ρB4
x4 N X→B
Ivan Savov QNIT
Classical-quantum channel coding
We are allowed to use the channel many times in parallel:
x1
m E
∈ [1 : 2nR]
ρB1 x1 N X→B
x2 ρB2
x2 N X→B
x3 ρB3
x3 N X→B
x4 ρB4
x4 N X→B
{ΛM0} M0
∈ [1 : 2nR]
Decoding POVM : {Λm} X
Λm=I and Λm≥ 0, ∀m.
Classical-quantum channel coding
m
∈ [1 : 2nR]
E Xn
∈ Xn
ρBXnn
∈ HBn
N⊗n
{ΛM0} M0
∈ [1 : 2nR]
p¯e≡ 1 2nR
X
m∈[1:2nR]
Pr{M06= m | m is sent} ≤ .
Ivan Savov QNIT
Classical-quantum channel coding
m
∈ [1 : 2nR]
E Xn
∈ Xn
ρBXnn
∈ HBn
N⊗n
{ΛM0} M0
∈ [1 : 2nR]
p¯e≡ 1 2nR
X
m∈[1:2nR]
Trn
I − ΛBmn ρBxnn(m)
o≤ .
Classical-quantum code
A rateR is achievable over the c-q channel NX→Bif there exists a codebookxn(m), m ∈ [1 : 2nR]and a corresponding decoding measurement POVM {Λm}, m ∈ [1 : 2nR]such that the average probability of error is bounded from above by epsilon: ¯pe≤ .
⇔
n · NX→B (1−)−→ nR · [c → c].
Ivan Savov QNIT
Classical decoding: conditionally typical sets
Consider an input distributionpX(x) and the channel pY |X(y|x).
For any input codewordxn, we define the conditionally typical set:
Tδ(n)(Y |xn) ≡
yn ∈ Yn:
−logpYn|Xn(yn|xn)
n − H(Y |X)
≤ δ
.
IDEA: An input stringxnis passed through the channel is likely to result in a conditionally typical output string.
Quantum decoding: conditionally typical subspaces
We define thexn-conditionally typical projector as follows:
ΠnρB
xn,δ= X
yn∈Tδ(n)(Y |xn)
|eρxn;yniheρxn;yn|,
where:
I The sum is over the set of conditionally typical strings of eigenvalue:
Tδ(n)(Y |xn) ≡
yn:
−logpYn|Xn(yn|xn)
n − H(Y |X)
≤ δ
, withpYn|Xn(yn|xn) =Qn
i=1pY |X(yi|xi).
I The states |eρxn;yni are built from tensor products of eigenvectors for the individual signal states:
|eρxn;yni = |eρx1;y1i ⊗ |eρx2;y2i ⊗ · · · ⊗ |eρxn;yni.
Ivan Savov QNIT
Conditionally typical sets and subspaces
1n
yn∈Tδ(n)(Y |xn)o
Xn
Tδ(n)(X)
Yn
Tδ(n)(Y |xn) xn
XnE X
yn ∈Yn
p`yn|Xn´ 1
yn ∈T(n) δ (Y |Xn)
ff≥ 1 − ,
X
yn ∈Yn
1
yn ∈T(n) δ (Y |xn)
ff ≤ 2n[H(Y |X)+δ]
Πxn≡ ΠBρn
xn,δ Xn
Tδ(n)(X) xn
HBn
Πρxn
XEn Trh
ρBXnΠnρB Xn,δ
i≥ 1 − ,
Trh ΠnρB
xn,δ
i≤ 2n[H(B|X)ρ+δ], ΠnρB
Xn,δρBXnΠnρB
Xn,δ≤ 2−n[H(B|X)ρ−δ]ΠnρB Xn,δ
Quantum multiple access channel
Ivan Savov QNIT
Quantum multiple access channel
X1×X2, NX1X2→B(x1, x2) ≡ρBx1,x2, HB
⇔
x1
Tx1
x2
Tx2
ρBx1,x2 Rx
Communication task:
n · NX1X2→B (1−)−→ nR1· [c1→ c] + nR2· [c2→ c]
IDEA: Tradeoff between the ratesR1andR2for the two senders.
QMAC capacity region
Theorem 10 in [Winter2001]: The capacity region for the
classical-quantum multiple access channel (X1× X2, ρBx1,x2, HB)is:
CMAC(N ) ≡ [
pX1,pX2
(R1, R2) ∈ R2+
R1 ≤ I(X1;B|X2)θ R2 ≤ I(X2;B|X1)θ R1+R2 ≤ I(X1X2;B)θ
,
where
θX1X2B ≡ X
x1,x2
pX1(x1)pX2(x2) |x1ihx1|X1⊗ |x2ihx2|X2⊗ ρBx1,x2.
Ivan Savov QNIT
Winter’s achievability proof: sequential decoding
Coding for the corners:
Two codesCαand Cβwith rates:
αp: (I(X1;B), I(X2;B|X1) ) βp: (I(X1;B|X2), I(X2;B) ).
Time sharing:
Intermediate points can be switching between the two codes, e.g., use Cα70% of the time, and Cβ30% of the time.
Simultaneous decoding
New theorem: Any rate pair (R1, R2) ∈CMAC(N )can be achieved usingsimultaneous decoding. [FHSSW11]
Simultaneous decoding allows us to use asingle codeto achieve any rate pair in the QMAC capacity region.
Ivan Savov QNIT
Proof sketch
We want to show that ∃ codebooksxn1(m1),xn2(m2)and a decoding POVM {Λm1,m2} such that the average error probability is small:
pe≡ 1
|M1||M2| X
m1,m2
Tr(I − Λm1,m2)ρxn1(m1),xn2(m2) ≤ .
We will show that the expectation ofpefor random codebooks X1n(m1),X2n(m2)is small:
Xn1E,X2n{ pe} ≤ .
This implies the existence of a good code.
Random codebook construction
Codebook construction: Randomly and independently generate 2nRi sequencesxni(mi),mi∈1 : 2nRi, according toQ pn Xi. Transmission: When the message pair (m1, m2)is transmitted, the channel output will be
ρBmn1,m2 ≡ ρBxnn
1(m1),xn2(m2).
Define the conditionally typical projector for this output state:
Πnm1,m2 ≡ ΠBρBnn xn1 (m1),xn
2 (m2)
.
Ivan Savov QNIT
POVM construction
Define the averaged output states:
ρ¯m1 ≡ E
X2nρxn1(m1),Xn2, ρ¯m2 ≡ E
X1nρX1n,xn2(m2), ρ ≡¯¯ E
Xn1,X2nρX1n,X2n, and the corresponding conditionally typical projectors:
Πnm1, Πnm2, Πnρ¯¯. POVM construction: Construct the “projector sandwich”
Pm1,m2≡ Πnρ¯¯ Πnm1 Πnm1,m2 Πnm1Πnρ¯¯, and normalize these to form a valid POVM:
Λm1,m2≡
X
m01,m02
Pm01,m02
−12
Pm1,m2
X
m01,m02
Pm01,m02
−12
.
Proof (1): State smoothing trick
Recall the definition of the average probability of error:
pe≤ 1
|M1||M2| X
m1,m2
Tr[(I − Λm1,m2)ρm1,m2].
We make a substitution of the output stateρm1,m2 with a Πnm2-smoothed version:
ρ˜m1,m2 ≡ Πnm2ρm1,m2Πnm2.
This substitution introduces a“smoothing penalty”term:
pe≤ 1
|M1||M2| X
m1,m2
"
Tr[(I − Λm1,m2) ˜ρm1,m2] + k˜ρm1,m2− ρm1,m2k1
# .
Ivan Savov QNIT
Proof (2): Hayashi-Nagaoka inequality
Next we use the Hayashi-Nagaoka operator inequality [HN03]:
I − (S + T )−12S (S + T )−12 ≤ 2 (I − S) + 4T, to obtain:
Tr[(I − Λm1,m2) ˜ρm1,m2] ≤
2Tr[(I − Pm1,m2) ˜ρm1,m2] + 4 X (m01,m02)6=(m1,m2)
TrPm01,m02ρ˜m1,m2 .
Proof (3): Epsilon terms
First term:
X1nE,X2n2Tr[(I − Pm1,m2) ˜ρm1,m2] ≤ 2( + 6√
).
Smoothing-penalty:
EXn1,X2nk˜ρm1,m2− ρm1,m2k1= EX1n,X2nkΠnm2ρm1,m2Πnm2 − ρm1,m2k1
≤ 2√
.
Therefore we are left with:
EXn1,X2n{pe} ≤
"
2( + 6√
) + 4X (m01,m02)6=(m1,m2)
TrPm01,m02ρ˜m1,m2 + 2√
# .
Ivan Savov QNIT
Proof (4): Split into three error terms
Split the “wrong message” error term into three parts:
X (m01,m02)6=(m1,m2)
TrPm01,m02ρ˜m1,m2 =
= X
m016=m1
TrPm01,m2ρ˜m1,m2
(E1)
+ X
m026=m2
TrPm1,m02ρ˜m1,m2
(E2)
+ X
m016=m1,m026=m2
TrPm01,m02ρ˜m1,m2 . (E12)
We will attack each of these in turn.
Xn1E,X2n{(E1)} = E
X1n,X2n
X
m016=m1
TrPm01,m2ρ˜m1,m2
= X
m016=m1
XE2n
Tr
XE1nPm01,m2
Πnm2E
X1n{ρm1,m2} Πnm2
=¬ X
m016=m1
X1nEX2nTr Pm01,m2Πnm2ρ¯m2Πnm2
≤ 2−n[H(B|X2)−δ] X
m016=m1
Xn1EX2nTr Pm01,m2Πnm2
≤ 2−n[H(B|X2)−δ]X
m016=m1
Trh Πnm0
1,m2
i
≤ 2® −n[H(B|X2)−δ] X
m016=m1
2n[H(B|X1X2)+δ]
≤ |M1| 2−n[I(X1;B|X2)−2δ]. (1)
X1nE,X2n
n(E2)o
= E
X1n,X2n
8
<
: X
m026=m2
Trh
Pm1,m02ρ˜m1,m2
i 9
=
;
= X
m026=m2 XE1n
( Tr
"
XE2n n
Πnρ,δ¯ Πnm1Πnm1,m0
2 Πnm1Πnρ,δ¯
o
XE2n
{˜ρm1,m2}
#)
≤¬ 2n[H(B|X1X2)+δ] X
m026=m2 XE1n
( Tr
"
XEn2 n
Πnρ,δ¯ Πnm1ρBm1,m0
2Πnm1Πnρ,δ¯ o
XE2n{˜ρm1,m2}
#)
= 2n[H(B|X1X2)+δ] X
m026=m2 XE1n
( Tr
"
Πnρ,δ¯ Πnm1 E
X2n
n
ρBm1,m02o
Πnm1Πnρ,δ¯ E
Xn2
{˜ρm1,m2}
#)
= 2n[H(B|X1X2)+δ] X
m026=m2 XE1n
( Tr
"
Πnρ,δ¯ Πnm1ρ¯m1Πnm1Πnρ,δ¯ E
X2n
{˜ρm1,m2}
#)
≤ 2 n[H(B|X1X2)+δ]2−n[H(B|X1)−δ] X
m026=m2 XEn1
( Tr
"
Πnρ,δ¯ Πnm1Πnρ,δ¯ E
Xn2{˜ρm1,m2}
#)
®
X1nE,X2n
n(E12)o
= E
X1n,Xn2
<
:
X
m016=m1,m026=m2
Trh
Pm01,m02ρ˜m1,m2
i=
;
=¬ X
m016=m1,m026=m2
Tr
"
X1nE,X2n
n
Pm01,m02o
Xn1E,X2n
{˜ρm1,m2}
#
≤ X
m016=m1,m026=m2
Tr
"
X1nE,X2n
n Pm0
1,m02
oρ¯¯⊗n
#
= X
m016=m1,m026=m2 X1nE,X2n
nTrh
Πnm01Πnm01,m02Πnm01 Πnρ,δ¯¯ ρ¯¯⊗nΠnρ,δ¯¯
io
≤®2−n[H(B)−δ] X
m016=m1,m026=m2 X1nE,X2n
n Trh
Πnm01 Πnm01,m02Πnm01Πnρ,δ¯¯
io
≤ 2¯ −n[H(B)−δ] X
m016=m1,m026=m2
Trh Πnm1,m0
2
i
≤ 2° −n[H(B)−δ]2n[H(B|X1X2)+δ] X
m016=m1,m026=m2
1
≤ 2−n[I(X1X2;B)−2δ]|M1||M2|. (3)
I If we want the error terms (E1), (E2) and (E12) to be small, we have to choose the size of the message sets appropriately.
I For example, if we choose the rateR1=I (X1;B|X2) − 3δ we will have:
n→∞lim E
X1n,X2n{(E1)} = lim
n→∞|M1| 2−n[I(X1;B|X2)−2δ]
= lim
n→∞2nR1 2−n[I(X1;B|X2)−2δ]
= lim
n→∞2n[I(X1;B|X2)−3δ]2−n[I(X1;B|X2)−2δ]
= lim
n→∞2−nδ= 0. Thus, any rate pair (R1, R2)which satisfies:
R1 ≤ I(X1;B|X2)θ
R2 ≤ I(X2;B|X1)θ
R1+R2 ≤ I(X1X2;B)θ
Summary of new techniques
For two-sender QMAC:
I Projector sandwich: Pm1,m2 ≡ Πnρ,δ¯ Πnm1 Πnm1,m2Πnm1 Πnρ,δ¯ .
I Smoothing trick: ˜ρm1,m2 ≡ Πnm2 ρm1,m2 Πnm2. Three-sender simultaneous decoding conjecture:
There exists a simultaneous decoding POVM Λm1,m2,m3 for the three-sender quantum multiple access channel.
Result would follow if we hadsimultaneous smoothing(Omar’s talk):
I Detection POVM:
Λm1,m2,m2 ≡ (P Π...)−1/2Πm1,m2,m3(P Π...)−1/2.
I Use smooth state ˜ρm1,m2,m3 which corresponds to channel outputρm1,m2,m3 smoothed by Πnm1,Πnm2,Πnm3,
Πnm1,m2,Πnm1,m3,Πnm2,m3 and Πρ,δ¯¯ .
Ivan Savov QNIT
Quantum interference channel
Quantum interference channel
(X1× X2, NX1X2→B1B2(x1, x2) ≡ρBx11,xB22, HB1⊗ HB2),
x1
Tx1
x2
Tx2
ρB1 x1,x2 Rx1
ρB2 x1,x2 Rx2
n · NX1X2→B1B2 (1−)−→ nR1· [c1→ c1] + nR2· [c2→ c2]
IDEA: Decoding the interference signals instead of treating them as noise.
IDEA: Use QMAC results.
Ivan Savov QNIT
Special case: IC with very strong interference
Theorem 5.1: The channel’s capacity region is given by:
S
pQX1X2(R1, R2) ∈ R2+such that:
R1 ≤ I (X1;B1|X2Q)θ, R2 ≤ I (X2;B2|X1Q)θ whereθQX1X2B is the following state:
X
x1,x2,q
pQ(q)pX1|Q(x1|q) pX2|Q(x2|q) |qihq|Q⊗|x1ihx1|X1⊗|x2ihx2|X2⊗ρBx1,x2.
IDEA: Decoding the interfering signal before own signal.
Special case: IC with strong interference
Theorem 5.2: The channel’s capacity region is given by:
S
pQX1X2(R1, R2) ∈ R2+such that:
R1 ≤ I(X1;B1|X2Q)θ, R2 ≤ I(X2;B2|X1Q)θ, R1+R2 ≤ min I(X1X2;B1|Q)θ
I(X1X2;B2|Q)θ
whereθQX1X2B = X
x1,x2,q
pQ(q)pX1|Q(x1|q) pX2|Q(x2|q) |qihq|Q⊗|x1ihx1|X1⊗|x2ihx2|X2⊗ρBx1,x2.
IDEA: Receivers use QMAC simultaneous decoding.
Ivan Savov QNIT
IDEA: Partialinterference cancellation.
I Split msg. into “personal” and “common” partsm1= (m1p, m1c).
I Generate separate random codebooks for each message U1n(m1p)andW1n(m1c), and mix them: X1i=f1(U1i, W1i).
X1
X2
B1
B2 W1
W2 U1
U2 f1
f2
Ŵ1Û1 Ŵ2 Ŵ2Û2 Ŵ1 ρBx11,xB22
I Receiver 1 will decodem1p,m1c,m2candignoresm2p.
I Achieving the rates of the Han-Kobayashi rate region RHK(N ) requires thethree-sender quantum simultaneous decoder.
I Workaround: Can show that the quantum Chong-Motani-Garg rate region R (N )is achievable [Sen12, Sav12]. Note that:
Quantum broadcast channel
Ivan Savov QNIT
(X, NX→B1B2 ≡ ρBx1B2, HB1B2)
Tx x
ρBx1
Rx1
ρBx2
Rx2
n·NX→B1B2 (1−)−→ nR1·[c → c]X→B1
| {z }
for Rx1 only
+nR·[c → cc]X→B1B2
| {z }
common message
+nR2·[c → c]X→B2
| {z }
for Rx2 only
Tx x
ρBx1
Rx1
ρBx2
Rx2
n · NX→B1B2 (1−)−→ nR ·[c → cc]X→B1B2
| {z }
common message
+ nR1· [c → c]X→B1
| {z }
superimposed message
Theorem 6.1: A rate pair (R, R1)is achievable for the quantum broadcast channel {ρBx1B2} if it satisfies the following inequalities:
R1≤ I(X; B1|W )θ, R1+R ≤ I(X; B1)θ,
R ≤ I(W ; B2)θ,
where the above information quantities are with respect to a state θW XB1B2 =P
w,xpW(w)pX|W(x|w) |wihw|W ⊗ |xihx|X⊗ ρBx1B2. IDEA: Simultaneous decoding used by Receiver 1.
Quantum relay channel
Quantum relay channel
X × X1, NXX1→B1B(x, x1) ≡ρBx,x1B1, HB1⊗ HB
Source x
ρB1 x,x1
Relay x1
ρBx,x1 Destination
n · NXX1→B1B (1−)−→ nR · [c → c]X→B
IDEA: Combine the information from the Source and the Relay.
IDEA: Transmission of messages can happen during multiple blocks.
Ivan Savov QNIT
Conclusion
Summary
I Showed that we can adapt some classical multiuser coding techniques to the setting of c-q channels.
I Discussed the obstacles that stand in the way of a general theory of a multi-user classical-quantum communication.
I Described new constructions and ad hoc tricks which can be applied for the case of two senders.
I The general problem withn ≥ 3 senders will require different techniques.
Ivan Savov QNIT
Further research
I Simultaneous decoding of three-sender QMAC?
I Simultaneous smoothing approach? (Omar’s talk)
I Auxiliary “tag space” approach of Pranab Sen.
I Extend other problems in network information theory to the classical-quantum setting:
I Joint source-channel coding problem. See M. M. Wilde and I.
Savov. Joint source-channel coding for a quantum multiple access channel. February 2012. arXiv:1202.3467.
I Entanglement-assisted communication?
1. O. Fawzi, P. Hayden, I. Savov, P. Sen, and M. M. Wilde.
Classical communication over a quantum interference channel.
IEEE Transactions on Information Theory, 58(6):3670-3691, June 2012. arXiv:1102.2624.
2. I. Savov and M. M. Wilde. Classical codes for quantum broadcast channels.
IEEE International Symposium on Information Theory, 2012.
arXiv:1111.3645.
3. I. Savov, M. M. Wilde, and M. Vu. Partial decode-forward for quantum relay channels.
IEEE International Symposium on Information Theory, 2012.
arXiv:1201.0011.
4. I. Savov. Network information theory for classical-quantum channels.
PhD Thesis. McGill University. July 2012. arXiv:1208.4188.