Construction πA Lattices: AReviewandRecent Results · 2019. 12. 20. · • Secrecy - He & Yener,...
Transcript of Construction πA Lattices: AReviewandRecent Results · 2019. 12. 20. · • Secrecy - He & Yener,...
Construction πA Lattices: A Review and Recent
Results
Yu-Chih (Jerry) Huang
Department of Communication Engineering
National Taipei University
Joint work with Krishna Narayanan @ Texas A&M University
1 / 38
Lattice codes are everywhere
(Nested) Lattice codes have many applications in IT
• See paper “Lattices are Everywhere” by Zamir
• Single user Gaussian channel - Erez & Zamir
• Coding with side information (Wyner-Ziv and Costa) - Zamir, Erez & Shamai
• Physical layer network coding - Wilson et al, Nam et al
• Secrecy - He & Yener, Ling et al
• Interference alignment - Sridharan, Jafaraian, Vishwanath & Jafar andOrdentlich, Erez, & Nazer
• Dirty multiple access channel - Philosof, Khisti, Erez & Zamir
• Compute-and-forward - Nazer & Gastpar
Most of these results are based on Construction A lattices
2 / 38
Lattice codes are everywhere
(Nested) Lattice codes have many applications in IT
• See paper “Lattices are Everywhere” by Zamir
• Single user Gaussian channel - Erez & Zamir
• Coding with side information (Wyner-Ziv and Costa) - Zamir, Erez & Shamai
• Physical layer network coding - Wilson et al, Nam et al
• Secrecy - He & Yener, Ling et al
• Interference alignment - Sridharan, Jafaraian, Vishwanath & Jafar andOrdentlich, Erez, & Nazer
• Dirty multiple access channel - Philosof, Khisti, Erez & Zamir
• Compute-and-forward - Nazer & Gastpar
Most of these results are based on Construction A lattices
Generate asymptotically good lattices; but comes with large decoding complexity
2 / 38
Lattice codes are everywhere
(Nested) Lattice codes have many applications in IT
• See paper “Lattices are Everywhere” by Zamir
• Single user Gaussian channel - Erez & Zamir
• Coding with side information (Wyner-Ziv and Costa) - Zamir, Erez & Shamai
• Physical layer network coding - Wilson et al, Nam et al
• Secrecy - He & Yener, Ling et al
• Interference alignment - Sridharan, Jafaraian, Vishwanath & Jafar andOrdentlich, Erez, & Nazer
• Dirty multiple access channel - Philosof, Khisti, Erez & Zamir
• Compute-and-forward - Nazer & Gastpar
Most of these results are based on Construction A lattices
Generate asymptotically good lattices; but comes with large decoding complexity
Main theme: Construction that generates good lattices with low complexity
2 / 38
My view of these constructions
After Alister’s talk on Monday, I was like
3 / 38
My view of these constructions
After Alister’s talk on Monday, I was like
3 / 38
My view of these constructions
Compute-and-forward
Nazer-Gastpar 09
Algebraic approach
Feng-Silva-Kschichang 11
Multistage CFHuang-Narayanan 14
Elementary divisorWang-Burr 15
Structure theorem forfinitely generated modules
Structure theorem forfinitely generated modules
Chinese
remainder
theorem
Chinese
remainder
theorem
(rings) (modules)
Construction πA is special case of EDC where we can show interesting things
3 / 38
Lattices
n-dimensional lattice Λn: A discrete subgroup of Rn
• Can be expressed by generator matrix G as
Λn = {Gz : z ∈ Zn}
• Closed underAddition: λ1, λ2 ∈ Λn implies λ1 + λ2 ∈ Λn
Reflection: λ1 ∈ Λn implies −λ1 ∈ Λn
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
Figure: Rectangular (Z2) lattice, Gaussianintegers Z[i]
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
Figure: Hexagonal (A2) lattice, Eisensteinintegers Z[ω] 4 / 38
Lattices
• Lattice quantizer: For x ∈ Rn, QΛ(x) , argmin
λ∈Λ‖x− λ‖2
• Fundamental Voronoi region: V(Λ) , {x : QΛ(x) = 0}• Modulo operation: For x ∈ R
n, x mod Λ , x−QΛ(x)
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
5 / 38
Lattices
• Lattice quantizer: For x ∈ Rn, QΛ(x) , argmin
λ∈Λ‖x− λ‖2
• Fundamental Voronoi region: V(Λ) , {x : QΛ(x) = 0}• Modulo operation: For x ∈ R
n, x mod Λ , x−QΛ(x)
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
5 / 38
Lattices
• Lattice quantizer: For x ∈ Rn, QΛ(x) , argmin
λ∈Λ‖x− λ‖2
• Fundamental Voronoi region: V(Λ) , {x : QΛ(x) = 0}• Modulo operation: For x ∈ R
n, x mod Λ , x−QΛ(x)
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
5 / 38
Lattices
• Lattice quantizer: For x ∈ Rn, QΛ(x) , argmin
λ∈Λ‖x− λ‖2
• Fundamental Voronoi region: V(Λ) , {x : QΛ(x) = 0}• Modulo operation: For x ∈ R
n, x mod Λ , x−QΛ(x)
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
5 / 38
Goodness for MSE quantization
• Let U ∼Uniform(V)
• Second moment per dim associated with Λ
σ2(Λ) ,1
nE‖U‖2 =
1
n
∫
V‖x‖2dx
Vol(Λ)
• Normalized second moment (NSM) of Λ
G(Λ) ,σ2(Λ)
Vol(Λ)2
n
>1
2πe
• Note that rB has G(rB) → 1/(2πe)
• A seq of Λn is good for MSE quantization if has G(Λ) → 1/(2πe)
• Related to performance of lattice quantizer at high resolution
6 / 38
Goodness for channel coding
rpack
V(Λ)
rz =√
nσ2z
reff
• Consider using Λ as our transmitted constellation, no power constraint, overAWGN channel
y = λ+ z, z ∼ N(0, σ2zI)
• Use lattice decoding, i.e., decoding to QΛ(y)
• From LLN, z lies inside rzB w.h.p.
• A seq of lattices is good for channel coding if pe → 0 whenever reff > rz
7 / 38
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
8 / 38
Construction A Lattices Review
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
9 / 38
Construction A Lattices Review
N -Dim Lattice Using Construction A (Leech-Sloane 71)
• C linear code over Fp
• M : Fp → Z natural mapping
• Λ = M(C) + pZN
C Mw ∈ F
mp
+
v ∈ pZN
λ ∈ Λ
x ∈ (Z/pZ)Nc ∈ FNp
0 2 4 6 8 10 12
0
2
4
6
8
10
12
10 / 38
Construction A Lattices Review
N -Dim Lattice Using Construction A (Leech-Sloane 71)
• C linear code over Fp
• M : Fp → Z natural mapping
• Λ = M(C) + pZN
• λ ∈ Λ iff λ mod pZN ∈ C C Mw ∈ F
mp
+
v ∈ pZN
λ ∈ Λ
x ∈ (Z/pZ)Nc ∈ FNp
0 2 4 6 8 10 12
0
2
4
6
8
10
12
0 5 10 15 20 250
5
10
15
20
25
10 / 38
Construction A Lattices Review
Construction A: What makes this a Lattice?
Construction A: Λ = M(C) + pZN
• λ1 = M(c1) + pk1 and λ2 = M(c2) + pk2
• λ1 + λ2 = M(c1) +M(c2) + pk1 + pk2
• It becomes M(c1 ⊕ c2) + pk3
• Thus, (λ1 + λ2) mod pZN = M(c1 ⊕ c2)
0 5 10 15 20 250
5
10
15
20
25
11 / 38
Construction A Lattices Review
Construction A: What makes this a Lattice?
Construction A: Λ = M(C) + pZN
• λ1 = M(c1) + pk1 and λ2 = M(c2) + pk2
• λ1 + λ2 = M(c1) +M(c2) + pk1 + pk2
• It becomes M(c1 ⊕ c2) + pk3
• Thus, (λ1 + λ2) mod pZN = M(c1 ⊕ c2)
0 5 10 15 20 250
5
10
15
20
25
11 / 38
Construction A Lattices Review
Construction A: What makes this a Lattice?
Construction A: Λ = M(C) + pZN
• λ1 = M(c1) + pk1 and λ2 = M(c2) + pk2
• λ1 + λ2 = M(c1) +M(c2) + pk1 + pk2
• It becomes M(c1 ⊕ c2) + pk3
• Thus, (λ1 + λ2) mod pZN = M(c1 ⊕ c2)
0 5 10 15 20 250
5
10
15
20
25
Natural mapping and mod p preserve ring structures between Z and Fp
11 / 38
Construction A Lattices Review
Why Construction A with natural mapping would work?
Λ = M(C) + pZN
• M−1((λ1 + λ2) mod pZ) = c1 ⊕ c2
• This would work if M−1 ◦ mod pZ is ring homomorphism
12 / 38
Construction A Lattices Review
Why Construction A with natural mapping would work?
Λ = M(C) + pZN
• M−1((λ1 + λ2) mod pZ) = c1 ⊕ c2
• This would work if M−1 ◦ mod pZ is ring homomorphism
Quotient ring
• pZ is an ideal in Z
• Coset decomposition Z/pZ results in a quotient ring
• For prime p, pZ is maximal ideal (since Z is PID)
• Z/pZ ∼= Fp
1 201211 4 531098
0 1 2 3-4 -3 -2 -1 4 5-5 6 7 8 9 10-9 -8 -7 -6-10
6 7 9 108
11 12 13
12 0111 20 4 53 6 7
-12-11-13
Z/13Z
12 / 38
Construction A Lattices Review
Generalizations
Construction A over Z: M natural mapping
• Natural mapping happens to be ring isomorphism
• mod p is canonical ring homomorphism
• ϕ , M−1 ◦ mod p is ring homo
13 / 38
Construction A Lattices Review
Generalizations
Construction A over Z: M natural mapping
• Natural mapping happens to be ring isomorphism
• mod p is canonical ring homomorphism
• ϕ , M−1 ◦ mod p is ring homo
• This is all the math I am going to use
Generalizations:
13 / 38
Construction A Lattices Review
Generalizations
Construction A over Z: M natural mapping
• Natural mapping happens to be ring isomorphism
• mod p is canonical ring homomorphism
• ϕ , M−1 ◦ mod p is ring homo
• This is all the math I am going to use
Generalizations:
• Construction A over Z[ω]Increase average computation rates for compute-and-forward (CF)Construction and application to CF: T-IT 15
• Go beyond PID, Construction A over OK
Consider only rings of imaginary quadratic integersPropose adaptive CF where TX adaptively work with best OK
ISIT 15 and will be submitted to T-IT soon
• Construction πA
Can be used to decrease decoding complexityNaturally suited for broadcasting with receiver side informationITW 13, ISIT 14, T-IT submitted 15, under revision 16
13 / 38
Construction πA Multilevel Lattices
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
14 / 38
Construction πA Multilevel Lattices
Motivation: Problem with Construction A
λ = M(c) + p · k where c ∈ C and k ∈ ZN
0 5 10 15 20 250
5
10
15
20
25
• k is unprotected; p has to be large to have a good lattice
• Complexity depends on decoding the linear code over Fp
• E.g., simulation results by di Pietro, Boutros, Zemor, Brunel with F11 and F41
15 / 38
Construction πA Multilevel Lattices
Main result in this part
Construction πA based on the Chinese Remainder Theorem
• p does not have to be prime - can be replaced by p1p2 . . . pL
• Instead of working over Fp, we can work over Fp1× Fp2
× . . .× Fpl
• Ex: q = 210 with just codes over F2,F3,F5,F7
• Show existence of sequence of lattices that are optimal
In short: New construction of lattices that preserve algebraic structures andgoodness with substantially lower complexity
16 / 38
Construction πA Multilevel Lattices
Chinese Remainder Theorem for Commutative Rings
Let p1, . . . , pL be distinct primes, q = p1 · p2 . . . pL, and ql = q/pl
Z/qZ ∼= Zp1× Zp2
× . . .× ZpL∼= Fp1
× Fp2× . . .× FpL
An isomorphism:
M(v1, . . . , vL) = s1q1v1 + . . .+ sLqLv
L mod qZ,
where s1, . . . , sL are sols to Bezout’s identity s1q1 + . . .+ sLqL = 1.
Example 1
17 / 38
Construction πA Multilevel Lattices
Chinese Remainder Theorem for Commutative Rings
Let p1, . . . , pL be distinct primes, q = p1 · p2 . . . pL, and ql = q/pl
Z/qZ ∼= Zp1× Zp2
× . . .× ZpL∼= Fp1
× Fp2× . . .× FpL
An isomorphism:
M(v1, . . . , vL) = s1q1v1 + . . .+ sLqLv
L mod qZ,
where s1, . . . , sL are sols to Bezout’s identity s1q1 + . . .+ sLqL = 1.
Example 1
Consider Z/6Z ∼= F2 × F3. An isomorphism: M = 3v1 − 2v2 mod 6Z
1,0 1,20,10,21,10,0
1 2 3 50 4......
1,00,21,10,0
7 8 96
17 / 38
Construction πA Multilevel Lattices
Construction πA (previously called product construction)
Let p1, . . . , pL be distinct primes and q = p1 p2 . . . pL.
Z/qZ ∼= Fp1× . . .× FpL
CL
C1
M
Fp1
FpL
Λ∗
+
ΠLl=1
plZN
Λ...
Product construction - L levels1 Choose (C1, . . . , CL) independently (no nesting) where Cl is over Fpl
18 / 38
Construction πA Multilevel Lattices
Construction πA (previously called product construction)
Let p1, . . . , pL be distinct primes and q = p1 p2 . . . pL.
Z/qZ ∼= Fp1× . . .× FpL
CL
C1
M
Fp1
FpL
Λ∗
+
ΠLl=1
plZN
Λ...
Product construction - L levels1 Choose (C1, . . . , CL) independently (no nesting) where Cl is over Fpl
2 Λ∗ , M(C1, . . . , CL) where M is ring isomorphism
18 / 38
Construction πA Multilevel Lattices
Construction πA (previously called product construction)
Let p1, . . . , pL be distinct primes and q = p1 p2 . . . pL.
Z/qZ ∼= Fp1× . . .× FpL
CL
C1
M
Fp1
FpL
Λ∗
+
ΠLl=1
plZN
Λ...
Product construction - L levels1 Choose (C1, . . . , CL) independently (no nesting) where Cl is over Fpl
2 Λ∗ , M(C1, . . . , CL) where M is ring isomorphism3 Λ , Λ∗ + qZN
4 λ ∈ Λ iff ϕ(λ) = (c1, . . . , cL) where ϕ , M−1 ◦ mod qZ
Works for other rings such as Z[i] and Z[ω] (and OK in general)
18 / 38
Construction πA Multilevel Lattices
Construction πA (previously called product construction)
Let p1, . . . , pL be distinct primes and q = p1 p2 . . . pL.
Z/qZ ∼= Fp1× . . .× FpL
CL
C1
M
Fp1
FpL
Λ∗
+
ΠLl=1
plZN
Λ...
Product construction - L levels1 Choose (C1, . . . , CL) independently (no nesting) where Cl is over Fpl
2 Λ∗ , M(C1, . . . , CL) where M is ring isomorphism3 Λ , Λ∗ + qZN
4 λ ∈ Λ iff ϕ(λ) = (c1, . . . , cL) where ϕ , M−1 ◦ mod qZ
Works for other rings such as Z[i] and Z[ω] (and OK in general)
Reduce to Construction A when L = 118 / 38
Construction πA Multilevel Lattices
Connection to Construction A with Coding over Ring
Construction A with Coding over Zq
• C linear code over Zq with a generator matrix G
• M : Zq → Z natural mapping
• Λ = M(C) + qZN = C + qZN
Construction πA is in fact a special case of this construction:
• q = p1 · p2 . . . · pL
• For z ∈ Zq, z = M(z1, . . . , zL) where zl ∈ Fplif z mod pl = zl
• G mod pl generates Cl for l ∈ {1, . . . , L}
This is an interesting special case that has connection to multilevelcoding/multistage decoding
19 / 38
Construction πA Multilevel Lattices
Theorem 2
Exist Construction πA lattices that are good for channel coding under multistage
ML decoding
Proof.
• Follow the steps by Forney-Trott-Chung
• Modulo-qZN channel is symmetric (regular)
• Random multilevel linear codes achieve modulo-qZN channel capacity
• Let q = p1p2 . . . pL tend to ∞
Theorem 3
Exist Construction πA lattices that are good for MSE quantization
Proof.
• Follows the steps by Ordentlich-Erez
• Random multilevel linear codes induce uniform distribution over RN
• Let q = p1p2 . . . pL tend to ∞20 / 38
Construction πA Multilevel Lattices
Power Constrained AWGN Channel: y = x+ z
Generalize Ordentlich-Erez’s construction to multilevel lattices:
Clc = {Gl
c ⊙wl|wl ∈ Fml
cpl }, Cl
f = {Glf ⊙wl|wl ∈ F
mlf
pl },
where Glf =
[
Glc Gl
]
,
Λf , γq−1M(C1f , . . . , C
Lf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
Lc ) + γZN ,
• Clearly, Clc ⊂ Cl
f ; thus, Λc ⊂ Λf
• C = Λf ∩ Vol(VΛc) with R =
∑Ll=1
mlf−ml
c
Nlog(pl)
• Choose Λc good for MSE quantization and Λf good for coding
• Achieve AWGN capacity under multistage decoding
21 / 38
Construction πA Multilevel Lattices
Low-complexity decoders
Serial modulo decoder (SMD):• Stage 1: Form estimate of c1 from
y1 = y mod p1Z
=(
M(c1, . . . , cL) + ΠLl=1plζ + z
)
mod p1Z
=(
c1 + z mod p1Z)
mod p1Z from CRT
• Stage s:Subtract all the contribution from the previous decoded stages to get
M(0, . . . , 0, cs, . . . , cL) + ΠLl=1plζ + z
Divide it by Πs−1
l=1pl to get
M(cs, . . . , cL) + ΠLl=splζ + z Construction πA with L− s+ 1 levels
where z , z/Πs−1
l=1pl
Form estimate of cs from
ys =
(
M(cs, . . . , cL) + ΠLl=splζ + z
)
mod psZ
= (cs + z mod psZ) mod psZ from CRT
22 / 38
Construction πA Multilevel Lattices
Low-complexity decoders
Parallel modulo decoder (PMD):
• For s ∈ {1, . . . , L}, simultaneously form
ys = y mod psZ
=(
M(c1, . . . , cL) + ΠLl=1plζ + z
)
mod psZ
= (cs + z mod psZ) mod psZ from CRT
• Form estimate of cs from ys
• More loss but substantially lower latency
23 / 38
Construction πA Multilevel Lattices
Extensions
Construction πD lattices:
• CRT only requires relatively prime rather than primes
• Allow all nature numbers: EX 12 = 4 · 3 hence Z/12Z ∼= Z4 × F3
• Coding over rings for those levels which do not happen to be fields
• Construction D is a special case with only 1 level Feng-Silva-Kschischang
• Can only show goodness for channel coding so far
Multilevel lattices over algebraic integers:
• Every OK forms a Dedekind domain
• Let I be ideal s.t. I = ΠLl=1pl
• CRT: OK/I ∼= OK/p1 × . . .×OK/pL ∼= Fpf11
× . . .× FpfLL
24 / 38
Application 1: Multistage Compute-and-Forward
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
25 / 38
Application 1: Multistage Compute-and-Forward
Lattice codes and a modern view of interference
Nazer-Gastpar, Compute-and-forward: Harnessing interference through structuralcodes, T-IT 11
DM
uM
yM
zm
zM
+
z1
+
.
.
.
+
y1
ymxk
xK
h11
hMK
hm1
x1
w1
S1 D1
Sk Dm...
SK
u1
um
.
.
....
hmk
C
b b
b b b
b b
c1
wkC
b b
b b b
b b
ck
wKC
b b
b b b
b b
cK
· · ·
· · ·
• Source (Sk): Has message wk over Fp, where p is prime.
• Destination (Dm): ym =∑K
k=1 hmkxk + zm. No CSIT, only CSIR
• Recover um = ⊕Kk=1bmkwk where bmk ∈ Fp
• A building block of a large network
26 / 38
Application 1: Multistage Compute-and-Forward
The Compute-and-Forward Paradigm
Theorem 4
Nazer-Gastpar For channel vector hm ∈ RK and integer vector am ∈ Z
K , the
following computation rate is achievable at Dm
R(hm, am) =1
2log+
(
1 + P‖hm‖2
‖am‖2 + P (‖hm‖2‖am‖2 − (hTmam)2)
)
How: To exploit the structural gains offered by the channel!
• Channel: ym =∑K
k=1 hmkxk + zm
• Sk: Use Construction A lattice code to match the channel structures tocertain extent
• Dm: Directly decode to a linear integer combination of codewords
• e.g.∑K
k=1 amkxk where amk ∈ Z
• Map this combination back to um = ⊕Kk=1bmk ⊙wk
27 / 38
Application 1: Multistage Compute-and-Forward
Multistage Compute-and-Forward
Theorem 5
Same computation rate can be achieved with multistage decoding.
Proof.
• Split message into w1k × . . .×wL
k over Fp1× . . .× FpL
• Use the proposed multilevel lattices
• By CRT, uniquely represent amk = M(b1mk, . . . , bLmk) + qξ
K∑
k=1
amkxk =
K∑
k=1
[M(b1mk, . . . , bLmk) + qξ] · [M(c1k, . . . , c
Lk ) + qζ]
= M(
⊕Kk=1b
1mk ⊙ c1k, . . . ,⊕
Kk=1b
Lmk ⊙ cLk
)
+ qη
• Decoding can be done level by level without losing optimality
Substantially reduce decoding complexity28 / 38
Application 1: Multistage Compute-and-Forward
Example of Multistage CF
DM
uM
yM
zm
zM
+
z1
+
.
.
.
+
y1
ymxk
xK
h11
hMK
hm1
x1
w1
S1 D1
Sk Dm...
SK
u1
um
.
.
....
hmk
C
b b
b b b
b b
c1
wkC
b b
b b b
b b
ck
wKC
b b
b b b
b b
cK
y = x1 + 5x2 + z
• Consider Z/6Z ∼= F2 × F3, same isomorphism
0 ↔ (0, 0), 1 ↔ (1, 1), 2 ↔ (0, 2),
3 ↔ (1, 0), 4 ↔ (0, 1), 5 ↔ (1, 2),
29 / 38
Application 1: Multistage Compute-and-Forward
Example of Multistage CF
C2
C1
M
c11= 0
c21= 2
F2
F3
1,0 1,20,10,21,10,0
1 2 3 50 4...
1,00,21,10,0
7 8 96
0
2
x1 = M(0, 2) = 2
C2
C1
M
c12= 1
c22= 1
F2
F3
1,0 1,20,10,21,10,0
1 2 3 50 4...
1,00,21,10,0
7 8 96
1
1
x2 = M(1, 1) = 1
y = h1x1 + h2x2 = 7 = M(1, 1) + 6
h1 = 1 = M(1, 1)
h2 = 5 = M(1, 2)
1,0 1,20,10,21,10,0
1 2 3 50 4...
1,00,21,10,0
7 8 96
C2
C1
M
c11⊕ c1
2= 1
c21⊕ 2c2
2= 1
F2
1
1
30 / 38
Application 2: Lattice Index Coding
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
31 / 38
Application 2: Lattice Index Coding
Broadcast channel with message side information
Natarajan-Hong-Viterbo T-IT 15
{w1, w2, . . . , wK}
noise
user 1
wS1
{w1, w2, . . . , wK}
...
noise
user L
wSL
{w1, w2, . . . , wK}
• Sender has independent messages {w1, w2, . . . , wK}.• Each receiver requests all the messages• Receiver l has wSl
a set of side info described by index set Sl
• For eg, S1 = {1, 2}, then wS1= {w1, w2}
• Noisy network yl = x+ zl where E[x2] ≤ P , and zl ∼ i.i.d. N (0, ρ2l )
32 / 38
Application 2: Lattice Index Coding
Broadcast channel with message side information
Natarajan-Hong-Viterbo T-IT 15
{w1, w2, . . . , wK}
noise
user 1
wS1
{w1, w2, . . . , wK}
...
noise
user L
wSL
{w1, w2, . . . , wK}
• Sender has independent messages {w1, w2, . . . , wK}.• Each receiver requests all the messages• Receiver l has wSl
a set of side info described by index set Sl
• For eg, S1 = {1, 2}, then wS1= {w1, w2}
• Noisy network yl = x+ zl where E[x2] ≤ P , and zl ∼ i.i.d. N (0, ρ2l )
Noisy broadcasting problem with receiver side information
32 / 38
Application 2: Lattice Index Coding
Capacity region and capacity-achieving codes
Capacity region (Tuncel T-IT 06): For every l ∈ {1, . . . , L},
1
2log2
(
1 +P
ρ2l
)
> H(w1, . . . , wK |wSl) =
K∑
k=1
Rk −RSl
• RSl=
∑
k∈SlRk
• Slepian-Wolf coding: Random codebooks + random binning + typicality
33 / 38
Application 2: Lattice Index Coding
Capacity region and capacity-achieving codes
Capacity region (Tuncel T-IT 06): For every l ∈ {1, . . . , L},
1
2log2
(
1 +P
ρ2l
)
> H(w1, . . . , wK |wSl) =
K∑
k=1
Rk −RSl
• RSl=
∑
k∈SlRk
• Slepian-Wolf coding: Random codebooks + random binning + typicality
A good code should translate every bit of side info into 6 dB SNR reduction
33 / 38
Application 2: Lattice Index Coding
Capacity region and capacity-achieving codes
Capacity region (Tuncel T-IT 06): For every l ∈ {1, . . . , L},
1
2log2
(
1 +P
ρ2l
)
> H(w1, . . . , wK |wSl) =
K∑
k=1
Rk −RSl
• RSl=
∑
k∈SlRk
• Slepian-Wolf coding: Random codebooks + random binning + typicality
A good code should translate every bit of side info into 6 dB SNR reduction
Lattice index codes by Natarajan-Hong-Viterbo T-IT 15
• Uniform side info gain of 6 dB by exploiting algebraic structure of CRT
• Extension to general ring of algebraic integers, Huang T-IT submitted 15
• Obtain diversity gains on top of side information gains
CRT seems to provide a right structure for this problem
33 / 38
Application 2: Lattice Index Coding
Capacity-achieving lattice index codes
• Let wk ∈ Fpk, k ∈ {1, . . . ,K}
• Encode the messages by C = Λf ∩ Vol(VΛc)
Λf , γq−1M(C1f , . . . , C
kf , . . . , C
Kf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
kc , . . . , C
Kc ) + γZN ,
• Here, M(v1, . . . , vL) , q1v1 + . . .+ qLv
L mod qZ
• Receiver l sees a codebook with messages in Sl fixed
Example 6 (3-User Case. S1 = {1}, S2 = {2, 3}, S3 = {1, 3})
Λf = γq−1M(C1f , C
2f , C
3f ) + γZN
Λc = γq−1M(C1c , C
2c , C
3c ) + γZN
34 / 38
Application 2: Lattice Index Coding
Capacity-achieving lattice index codes
• Let wk ∈ Fpk, k ∈ {1, . . . ,K}
• Encode the messages by C = Λf ∩ Vol(VΛc)
Λf , γq−1M(C1f , . . . , C
kf , . . . , C
Kf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
kc , . . . , C
Kc ) + γZN ,
• Here, M(v1, . . . , vL) , q1v1 + . . .+ qLv
L mod qZ
• Receiver l sees a codebook with messages in Sl fixed
Example 6 (3-User Case. S1 = {1}, S2 = {2, 3}, S3 = {1, 3})
Λf = γq−1M(C1f , C
2f , C
3f ) + γZN
Λc = γq−1M(C1c , C
2c , C
3c ) + γZN
R2 +R3 ≤ 12 log(1 + P/ρ21) if Λf with C1
f fixed is good for coding
34 / 38
Application 2: Lattice Index Coding
Capacity-achieving lattice index codes
• Let wk ∈ Fpk, k ∈ {1, . . . ,K}
• Encode the messages by C = Λf ∩ Vol(VΛc)
Λf , γq−1M(C1f , . . . , C
kf , . . . , C
Kf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
kc , . . . , C
Kc ) + γZN ,
• Here, M(v1, . . . , vL) , q1v1 + . . .+ qLv
L mod qZ
• Receiver l sees a codebook with messages in Sl fixed
Example 6 (3-User Case. S1 = {1}, S2 = {2, 3}, S3 = {1, 3})
Λf = γq−1M(C1f , C
2f , C
3f ) + γZN
Λc = γq−1M(C1c , C
2c , C
3c ) + γZN
R1 ≤ 12 log(1 + P/ρ22) if Λf with C2
f and C3f fixed is good for coding
34 / 38
Application 2: Lattice Index Coding
Capacity-achieving lattice index codes
• Let wk ∈ Fpk, k ∈ {1, . . . ,K}
• Encode the messages by C = Λf ∩ Vol(VΛc)
Λf , γq−1M(C1f , . . . , C
kf , . . . , C
Kf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
kc , . . . , C
Kc ) + γZN ,
• Here, M(v1, . . . , vL) , q1v1 + . . .+ qLv
L mod qZ
• Receiver l sees a codebook with messages in Sl fixed
Example 6 (3-User Case. S1 = {1}, S2 = {2, 3}, S3 = {1, 3})
Λf = γq−1M(C1f , C
2f , C
3f ) + γZN
Λc = γq−1M(C1c , C
2c , C
3c ) + γZN
R2 ≤ 12 log(1 + P/ρ23) if Λf with C1
f and C3f fixed is good for coding
34 / 38
Application 2: Lattice Index Coding
Capacity-achieving lattice index codes
• Let wk ∈ Fpk, k ∈ {1, . . . ,K}
• Encode the messages by C = Λf ∩ Vol(VΛc)
Λf , γq−1M(C1f , . . . , C
kf , . . . , C
Kf ) + γZN ,
Λc , γq−1M(C1c , . . . , C
kc , . . . , C
Kc ) + γZN ,
• Here, M(v1, . . . , vL) , q1v1 + . . .+ qLv
L mod qZ
• Receiver l sees a codebook with messages in Sl fixed
Example 6 (3-User Case. S1 = {1}, S2 = {2, 3}, S3 = {1, 3})
Λf = γq−1M(C1f , C
2f , C
3f ) + γZN
Λc = γq−1M(C1c , C
2c , C
3c ) + γZN
In general, need good Construction πA lattices with arbitrary levels fixed
34 / 38
Application 2: Lattice Index Coding
Sketch of the proof
Consider 2 levels: Λ = p2C1 + p1C
2 + p1p2ZN . Note that
Λ = p1C2 + p2
(
C1 + p1ZN)
= p1C2 + p2Λ1
= p2C1 + p1
(
C2 + p2ZN)
= p2C1 + p1Λ2
• Randomly picking C1 results in good Λ1 w.h.p.
• Randomly picking C2 results in good Λ2 w.h.p.
35 / 38
Application 2: Lattice Index Coding
Sketch of the proof
Consider 2 levels: Λ = p2C1 + p1C
2 + p1p2ZN . Note that
Λ = p1C2 + p2
(
C1 + p1ZN)
= p1C2 + p2Λ1
= p2C1 + p1
(
C2 + p2ZN)
= p2C1 + p1Λ2
• Randomly picking C1 results in good Λ1 w.h.p.
• Randomly picking C2 results in good Λ2 w.h.p.
• Note that p2Λ1 ⊂ Λ ⊂ Λ1
• Λ can be viewed as a Construction A lattice over base lattice Λ1
• Tailor (Loeliger’s version) Minkowski-Hlawka theorem specifically for thisconstruction
So picking C2 randomly results in good Λ w.h.p.
35 / 38
Future Directions
Outline
1 Construction A Lattices Review
2 Construction πA Multilevel Lattices
3 Application 1: Multistage Compute-and-Forward
4 Application 2: Lattice Index Coding
5 Future Directions
36 / 38
Future Directions
Future Directions
Construction πA lattices:
Seeking for interesting problems where Construction πA can be useful
37 / 38
Future Directions
Future Directions
Construction πA lattices:
Seeking for interesting problems where Construction πA can be useful
Number field lattices:
Construction A lattice codes over a general OK
• Kositwattanarerk-Ong-Oggier: Block fading wiretap channel
• Campello-Ling-Belfiore: Compound block fading channel
• Huang-Narayanan-Wang: Adaptive compute-and-forward
37 / 38
Future Directions
CF over block fading channel
RX: [y(1)m , . . . ,y
(B)m ] where y
(b)m =
∑Kk=1 h
(b)mkx
(b)k + z
(b)m
• Construct lattice code from OK where K has degree B
• M : Fpf → OK/p ring isomorphism
• Λ = M(C) + pn and Λ = Ψ(Λ) where Ψ is canonical embedding
• λ = Ψ(λ) = [σ1(λ), . . . , σB(λ)]
• Enable computing
[
K∑
k=1
a(1)mkx
(1)k , . . . ,
K∑
k=1
a(B)mkx
(B)k
]
where Ψ−1([a(1)mk, . . . , a
(B)mk ]) ∈ OK
• Seem to better match channel’s structure than Z-lattices
38 / 38
Future Directions
CF over block fading channel
RX: [y(1)m , . . . ,y
(B)m ] where y
(b)m =
∑Kk=1 h
(b)mkx
(b)k + z
(b)m
• Construct lattice code from OK where K has degree B
• M : Fpf → OK/p ring isomorphism
• Λ = M(C) + pn and Λ = Ψ(Λ) where Ψ is canonical embedding
• λ = Ψ(λ) = [σ1(λ), . . . , σB(λ)]
• Enable computing
[
K∑
k=1
a(1)mkx
(1)k , . . . ,
K∑
k=1
a(B)mkx
(B)k
]
where Ψ−1([a(1)mk, . . . , a
(B)mk ]) ∈ OK
• Seem to better match channel’s structure than Z-lattices
Good for coding has been proved by Campello-Ling-Belfiore. Need to show goodfor MSE quantization
38 / 38
Future Directions
CF over block fading channel
RX: [y(1)m , . . . ,y
(B)m ] where y
(b)m =
∑Kk=1 h
(b)mkx
(b)k + z
(b)m
• Construct lattice code from OK where K has degree B
• M : Fpf → OK/p ring isomorphism
• Λ = M(C) + pn and Λ = Ψ(Λ) where Ψ is canonical embedding
• λ = Ψ(λ) = [σ1(λ), . . . , σB(λ)]
• Enable computing
[
K∑
k=1
a(1)mkx
(1)k , . . . ,
K∑
k=1
a(B)mkx
(B)k
]
where Ψ−1([a(1)mk, . . . , a
(B)mk ]) ∈ OK
• Seem to better match channel’s structure than Z-lattices
Good for coding has been proved by Campello-Ling-Belfiore. Need to show goodfor MSE quantization
38 / 38
Chinese Remainder Theorem
Let p1, p2, . . . , pL be L distinct co-prime integers and q = p1.p2. . . . pL. Then,given
x ≡ a1(modp1)
x ≡ a2(modp2)
...
x ≡ aL(modpL)
there exists exactly one x ∈ Zq satisfying this system.
39 / 38
Chinese Remainder Theorem
Let p1, p2, . . . , pL be L distinct co-prime integers and q = p1.p2. . . . pL. Then,given
x ≡ a1(modp1)
x ≡ a2(modp2)
...
x ≡ aL(modpL)
there exists exactly one x ∈ Zq satisfying this system.
EX: q = 6, p1 = 2, p2 = 3
x ≡ 0(mod2)
x ≡ 1(mod3)
x = 439 / 38
Modified M.-H. Theorem
Consider Λ = p2C1 + p1C
2 + p1p2Zn = p1C
2 + p2Λ1
• Let M2(v2) , p1v
2 mod p2Λ1 where v2 ∈ Fp2
• Let σ , M−1 ◦ mod p2Λ1
• Let C2 be the ensemble of (N, k2) linear codes over Fp2
1
|C2|
∑
C2∈C2
∑
v∈γΛ\0
f(v)
=1
|C2|
∑
C2∈C2
∑
v∈Λ1\0:σ(v)=0
f(γv) +∑
v∈Λ1\0:σ(v)∈C2\0
f(γv)
=∑
v∈Λ1\0:σ(v)=0
f(γv) +1
|C2|
∑
C2∈C2
∑
c∈C2\0
∑
v∈Λ1:σ(v)=c
f(γv)
40 / 38
=∑
v∈Λ1\0:σ(v)=0
f(γv) +pk2
2 − 1
pN2 − 1
∑
c∈FNp2
∑
v∈Λ1:σ(v)=c
f(γv)
=∑
v∈Λ1\0:σ(v)=0
f(γv) +pk2
2 − 1
pN2 − 1
∑
v∈Λ1:σ(v) 6=0
f(γv)
(a)≈ pk2−N
2 γ−NVol(VΛ1)−1
∑
v∈γΛ1:σ(v) 6=0
f(v)γNVol(VΛ1)
(b)≈ Vol(VγΛ)
−1
∫
RN
f(v)dv,
• (a) requires p2 large and f(.) bounded
• (b) requires γNVol(VΛ1) small s.t. Riemann sum → Riemann integration
• Vol(VγΛ) = γNpk2−N2 Vol(VΛ1
)−1 = pk2−N2 pk1−N
1
41 / 38