Interference Management in Wireless Networks · 2015-08-12 · IS-95 standard) Interference from...
Transcript of Interference Management in Wireless Networks · 2015-08-12 · IS-95 standard) Interference from...
Interference Management in Wireless Networks
Venu Veeravalli
Coordinated Science LabDepartment of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
Aly El Gamal
Department of Electrical and Computer EngineeringPurdue University
2 / 177
Part 1: Introduction
ECE Illinois & Purdue
3 / 177
Interference in Wireless Networks
ECE Illinois & Purdue
4 / 177
Interference Management
Wireless World
Interference management is critical
ECE Illinois & Purdue
5 / 177
Interference Management in Cellular Networks:Historical Perspective
ECE Illinois & Purdue
6 / 177
Single Isolated Cell
• Manhattan mobile phone system (1946)• Interference managed by simply orthogonalizing users in
time-frequency
ECE Illinois & Purdue
7 / 177
Multiple Access Alternatives
ECE Illinois & Purdue
8 / 177
Cellular Concept (Bell Labs)
Reuse spectrum geographically through cell splitting
BASE STATION Mobile
Telephone Switching Office
ECE Illinois & Purdue
9 / 177
Narrowband Cellular (AMPS,GSM)
• Bandwidth divided into narrowbandchannels (200 Khz in GSM), andusers are assigned time slots (8 perchannel in GSM)
• No in-cell interference (orthogonalusers in cell)
• Interference across cells on samechannel is minimized by reusing samechannels only in cells far apart
ECE Illinois & Purdue
10 / 177
Frequency Reuse
• Spectral efficiency of narrowband cellular is reduced by reusefactor
• Interference localized to narrow band
ECE Illinois & Purdue
11 / 177
Wideband Cellular (CDMA)
• Universal frequency reuse – all users in allcells share same bandwidth (1.25 MHz inIS-95 standard)
• Interference from in-cell users can becontrolled by using orthogonal codes(downlink) or successive interferencecancellation (uplink)
• Interference from other cells is averaged andsimply raises the noise floor
• Advantages:• No reduction in capacity due to reuse (no frequency planning
needed)• Graceful degradation in performance with users (soft capacity)• Any technique that reduces power of interferers (soft handoff, voice
activity detection) increases capacity
ECE Illinois & Purdue
12 / 177
Disadvantages of CDMA
• In-cell interference reduces capacity (cannot be eliminatedcompletely)
• Tight power control is needed to manage interference, and maybe too expensive for data applications (with low duty cycle)
Can we simultaneously haveuniversal reuse and keep in-cellusers orthogonal?
ECE Illinois & Purdue
13 / 177
Wideband Cellular OFDM (LTE)
• Split up bandwidth into narrowband sub-channels, with every userhaving access to all sub-channels
• Basic unit of resource is virtual channel (hopping sequence) ?virtual channels are orthogonal in TF
From Tse & Viswanath, Fundamentals of Wireless Comm
ECE Illinois & Purdue
14 / 177
Wideband Cellular OFDM
• In-cell Orthogonalization• Hopping sequences of users within cell are
designed to be orthogonal• Users are assigned one or more virtual
channels
• Out-of-Cell Interference Averaging• Hopping patterns in adjacent cells are
chosen so that there is minimal overlapbetween any pairs
• Interference from out-of-cell is averagedover band rather than being localized
ECE Illinois & Purdue
15 / 177
Additional Resources for Interference Management
ECE Illinois & Purdue
16 / 177
Multiple Antennas (MIMO)
• Multiple antennas providediversity against fading andmultiplexing gain forpoint-to-point links
• Can also be used for spatialseparation of users –beamforming
• Additional degrees of freedomfor interference management!
ECE Illinois & Purdue
17 / 177
Basestation Cooperation
• Out-of-cell users become in-cell users
• Beamforming (downlink), joint decoding (uplink)
ECE Illinois & Purdue
18 / 177
Relaying and User Cooperation
• Potential interferers become helpers (relays of information)
• Particularly useful in distributed interference management (adhoc, mesh networks)
ECE Illinois & Purdue
19 / 177
Dynamic Spectrum Access (Cognitive Radio)
Primary Tx
range
guard band
secondary users
• Primary and secondary usersof spectrum
• Secondary users sensechannels to determinepresence/absence of primaryusers
• Probability of interfering withprimary is constrained
ECE Illinois & Purdue
20 / 177
Explosion in Wireless Data Traffic
How to accommodate exponential growth without new useful spectrum?
ECE Illinois & Purdue
21 / 177
Through Improved PHY?
• Point-to-Point wireless technology mature• Modulation/demodulation• Synchronization• Coding/decoding (near Shannon limits)• MIMO
• Centralized (in-cell) multiuser wireless technology also mature• Orthogonalize users when possible• Otherwise use successive interference cancellation
Spectral efficiency gains from further improvements in PHY are limited!
ECE Illinois & Purdue
22 / 177
By Adding More Basestations?
ECE Illinois & Purdue
23 / 177
Through Improved Interference Management
Several useful engineering solutions for managing interference
But...
What are fundamental limits?
ECE Illinois & Purdue
24 / 177
References
1 T.S. Rappaport. Wireless communications: Principles andpractice. New Jersey: Prentice Hall 1996.
2 A. J. Viterbi. CDMA: Principles of spread spectrumcommunication. Addison Wesley, 1995.
3 D. Tse and P. Viswanath. Fundamentals of wirelesscommunication. Cambridge University Press, 2005.
4 E. Biglieri et al. Principles of Cognitive Radio. CambridgeUniversity Press, 2012.
ECE Illinois & Purdue
25 / 177
Part 2: Information Theory for Interference Channels
ECE Illinois & Purdue
26 / 177
Claude Shannon’s Information Theory
Point-to-point Communication on AWGN channel
Capacity = log(1 + SNR) bits/sec/Hz
ECE Illinois & Purdue
27 / 177
Two User Gaussian Interference Channel
hX1
X2
Y1
Y2
Z2
Z1
• Gaussian noise with unit variance; transmit power constraint P
• Capacity region?
• Known when h ≥ 1
ECE Illinois & Purdue
28 / 177
Strong Interference Setting: h ≥ 1
hX1
X2
Y1
Y2
Z2
Z1
• Both receivers decode both messages (compound MAC)• Capacity region
R1 ≤1
2log(1 + P )
R2 ≤1
2log(1 + P )
R1 +R2 ≤1
2log(1 + P + h2P )
ECE Illinois & Purdue
29 / 177
Very Strong Interference Setting: h ≥√1 + P
hX1
X2
Y1
Y2
Z2
Z1
• Each user first decodes interference treating intended input asnoise; then subtracts interference to decode intended input
• Capacity region
R1 ≤1
2log(1 + P )
R2 ≤1
2log(1 + P )
same as when interference is absent
ECE Illinois & Purdue
30 / 177
Two User Gaussian Interference Channel
hX1
X2
Y1
Y2
Z2
Z1
• Gaussian noise with unit variance; transmit power constraint P
• Capacity region?
Open problem when h < 1
ECE Illinois & Purdue
31 / 177
What is Known?
• Simple schemes• Treat interference as noise• Orthogonalize users• Single user coding/decoding
• Sophisticated schemes• Exploit structure in interference• Joint coding/decoding
• Han-Kobayshi achievable scheme• Power splitting (common/private) and time-sharing• Best known inner bound to capacity region for two-user GIC
• Special case of H-K scheme achieves capacity to within 1 bit!• Etkin, Tse, Wang 2007
ECE Illinois & Purdue
32 / 177
Capacity in Low Interference Regime [Annapureddy&VVV ’09]
Treating interference as noise is optimal in low interference regime
0 5 10 15 20 25 30 35 40−10
−5
0
5
10
15
SNR
INR
Th
resh
old
ECE Illinois & Purdue
33 / 177
Main Result
Theorem
For the two user symmetric Gaussian interference channel satisfying
h(1 + h2P ) ≤ 0.5
treating interference as noise achieves the sum capacity, which is givenby
Csum = log
[1 +
P
1 + h2P
]
Requires new outer bound!
ECE Illinois & Purdue
34 / 177
Genie-aided Outer Bound
• Genie gives side-information to receivers
hX1
X2
Z1
Z2
Y1
Y2
S1
S2
• Sum capacity of genie-aided channel is obvious outer bound tooriginal channel
ECE Illinois & Purdue
35 / 177
Properties of Genie: Usefulness
• Useful: if sum capacity of genie-aided channel (easily) derivable
• Gaussian inputs are optimal=⇒ treating interference as noise is optimal
• Csum(genie-aided) = I(X1G;Y1G, S1G) + I(X2G;Y2G, S2G)
• Eg: S1 = X2 and S2 = X1
Y1
Z2
Z1
Y2
X1
X2
ECE Illinois & Purdue
36 / 177
Properties of Genie: Smartness
• Smart: if when Gaussian inputs are used, genie does not improvesum capacity
⇒ I(X1G;Y1G, S1G) = I(X1G;Y1G) andI(X2G;Y2G, S2G) = I(X2G;Y2G)
• Example: Genie that does not interact with receivers
not useful!
• We want a genie that is both useful and smart
⇒ treating interference as noise is optimal for original channel
Csum = I(X1G;Y1G) + I(X2G;Y2G)
ECE Illinois & Purdue
37 / 177
Quest for Divine Genie
• Restrict class of genies considered
S2 = X2 + ηW2
X1
X2
Z1
Z2
Y1
h
S1 = X1 + ηW1
Y2
• If Wi independent of Zi, i = 1, 2• genie is useful in proving one bit result of Etkin, Tse, and Wang
• Useful and asymptotically smart!
ECE Illinois & Purdue
38 / 177
Geometric Representation of Genie
Y
(√1 + h2P, 0)
(η, θ)
ηW
I + ZX
S
• cos θ: Correlation coefficient between I + Z and ηW
(equivalently between Z and W )
ECE Illinois & Purdue
39 / 177
Useful Genie
Y
(√1 + h2P, 0)
(η, θ)
ηW
I + ZX
S
(0, 1h )
Inside green curve: useful
ECE Illinois & Purdue
40 / 177
Condition: Useful Genie
• Mutual Information for the genie-aided channel
I(Xni ;Y n
i , Sni ) = I(Xn
i ;Sni ) + I(Xni ;Y n
i |Sni )
≤ h(Sni )− nh(SiG|XiG) + nh(YiG|SiG)− h(Y ni |Xn
i , Sni )
• Terms that are not maximized by Gaussian inputs:
h(Sn1 ) = h(hXn1 + hηWn
1 )
h(Y n2 |Xn
2 , Sn2 ) = h(hXn
1 + Zn2 |Wn2 )
and similar terms obtained by swapping “1” and “2”
• Difference between terms is maximized by X1G (worst case noiseresult) if:
|hη| ≤√
1− cos2 θ
ECE Illinois & Purdue
41 / 177
Smart Genie
YX
S
I + Z
ηW
Smart
On the blue line: smart
ECE Illinois & Purdue
42 / 177
Divine Genie
YX
S
I + Z
ηW
Smart
Useful
Green curve intersects with blue line:Treating interference as noise is optimal
ECE Illinois & Purdue
43 / 177
Divine Genie
Smart
Useful
√1 + h2P
0.5h(
√1+h2P)
Intersection if h(1 + h2P ) ≤ 0.5
• Also [Shang, Kramer, Chen 08] &[Motahari, Khandani 08]
ECE Illinois & Purdue
44 / 177
Capacity in Low Interference Regime [Annapureddy&VVV ’09]
INR = h2P and SNR = P
0 5 10 15 20 25 30 35 40−10
−5
0
5
10
15
SNR
INR
Th
resh
old
ECE Illinois & Purdue
45 / 177
Extensions – Outer Bound on Capacity Region
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80
0.5
1
1.5
2
R1
R2
HK Inner BoundETW Outer BoundBroadcast Outer BoundNew Outer Bound
ECE Illinois & Purdue
46 / 177
Interference Network
X3
YM
Y3
Y2
Y1
XM
X2
X1
• Is treating interference as noise still optimal in low interferenceregime?
• Yes, using same “scalar” genie we can find threshold below whichtreating interference as noise is optimal
ECE Illinois & Purdue
47 / 177
Many-One and One-Many Channels
X3
YM
Y3
Y2
Y1
XM
X2
X1
M∑i=2
h21i ≤ 1
X3
YM
Y3
Y2
Y1
XM
X2
X1
M∑i=2
h2i1P1 + h2
i1
h2i1P1 + 1
≤ 1
Treating interference as noise is optimal in low interference regime
ECE Illinois & Purdue
48 / 177
Symmetric M -User Interference Channel
X3
YM
Y3
Y2
Y1
XM
X2
X1
• Interference threshold is characterized by
h(1 + h2P ) ≤ 0.5, where h2 = (M − 1)h2
• Threshold on total interference at receiver below which it isoptimal to treat interference as noise remains constant as Mincreases!
ECE Illinois & Purdue
49 / 177
Case Study – Three User Symmetric Channel
X1
X3
X2
Y1
Y2
Y3
• Requires new genie construction
• No explicit equation for threshold on interference parameter h
• But can compute admissible values of h for given P numerically
ECE Illinois & Purdue
50 / 177
Interference Threshold
2 4 6 8 10 12 14 16 18 20−6
−5
−4
−3
−2
−1
0
1
2
3
4
SNR
INR
Th
resh
old
Two users
Three users
ECE Illinois & Purdue
51 / 177
Extension to MIMO Interference Channels
H21
X2
Z1
Z2
Y 1
Y 2
X1
H22
H11
H12
Input covariance constraints: Qi � 0,Tr (Qi) ≤ P
ECE Illinois & Purdue
52 / 177
Genie-Aided MIMO Interference Channel
Y 2
Hc
Hd
Hd
X2
Z1
Z2
X1
S1
Y 1
S2
Si = Hc Xi +W i
[ZiW i
]∼ N
(0,
[Σ A
A> Σ
])
Genie parameters: Ψ = {Σ, A}ECE Illinois & Purdue
53 / 177
Genie-aided outer bound
H21
X2
Z1
Z2
Y 1
Y 2
X1
H22
H11
H12
Y 2
Hc
Hd
Hd
X2
Z1
Z2
X1
S1
Y 1
S2
Csum ≤ CGA-ICsum (Ψ)
≥ Useful Genie=
maxQ
∑i
I(XiG;Y iG)Smart Genie
? maxQ
∑i
I(XiG;Y iG, SiG)
ECE Illinois & Purdue
54 / 177
Useful and Smart Genie Conditions
Useful Genie: If Σ � Σ−AΣ−1A> then:
CGA-ICsum (Ψ) = maxQ
RGA-ICTIN (Q,Ψ)
Smart Genie: If(A>(HcQH
>c + Σ)−1Hd −Hc
)Q = 0, then
RICTIN(Q) = RGA-IC
TIN (Q,Ψ)
i.e., the genie is smart w.r.t. input covariance Q.
ECE Illinois & Purdue
55 / 177
Genie-aided outer bound: Fix Q
H21
X2
Z1
Z2
Y 1
Y 2
X1
H22
H11
H12
Y 2
Hc
Hd
Hd
X2
Z1
Z2
X1
S1
Y 1
S2
Csum(Q) ≤ CGA-ICsum (Q,Ψ)
≥ Useful Genie=
RICTIN(Q)
Smart Genie w.r.t. Q= RGA-IC
TIN (Q,Ψ)
Sufficient Condition: Treating interference as noise is sum rate optimalif for every Q satisfying power constraint, there exists a genie that isuseful and smart w.r.t. Q [Shang, Chen, Kramer & Poor, Allerton 08]
ECE Illinois & Purdue
56 / 177
Low Interference Regime – Simpler Condition
Theorem
Let Q∗ be a global maximum of RICTIN(Q). If
• there exists a genie that is useful and smart w.r.t. Q∗
• Q∗ is full rank
then the channel is in low interference regime, i.e.,
Csum = RICTIN(Q∗).
ECE Illinois & Purdue
57 / 177
Proof Ilustration
Q∗
RGA−IC
RIC
Q
ECE Illinois & Purdue
58 / 177
MISO Channel
hc
Z1
Y1X1
Y2
Z2
X2 d
d
θ: angle between d and c.
ECE Illinois & Purdue
59 / 177
Application of General MIMO Result?
• Is Q∗ = arg maxRICTIN(Q) full rank?
No! Q∗ is unit rank
• Beamforming along direction b is optimal, i.e.,
Q∗ = Pbb>
where b is generalized eigenvector of matrix pair(Pdd>, I + h2Pcc>
)corresponding to largest generalized eigenvalue λmax
ECE Illinois & Purdue
60 / 177
Genie-aided outer bound
Csum ≤ CGA-ICsum (Ψ)
≥ Useful Genie=
maxQRICTIN(Q) maxQR
GA-ICTIN (Q,Ψ)
= ≤
maxQRICTIN(Q) maxQRrelaxed(Q,Ψ)
=Additional Constraints
=
RICTIN(Q∗)
Smart Genie w.r.t.Q∗= Rrelaxed(Q∗,Ψ)
ECE Illinois & Purdue
61 / 177
Outer Bound
0 20 40 60 80 100 120 140 160 1800
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Angle between Beamforming vector and cross channel vector
Su
m R
ate
Achievable Rate
Genie−aided outer bound
Relaxed outer bound
ECE Illinois & Purdue
62 / 177
Low Interference Regime
Theorem
The sum capacity of the MISO interference channel is achieved byusing Gaussian inputs, transmit beamforming, and treating interferenceas noise at the receivers, and is given by
CICsum =1
2log
(1 +
P cos2 θ
1 + h2P+ P sin2 θ
)if the channel gain parameter h satisfies:
|h| < h0(θ, P ) (1)
with h0(θ, P ) being the positive solution to the implicit equation
h2 − sin2 θ =
(cos θ
1 + h2P− h)2
+
. (2)
ECE Illinois & Purdue
63 / 177
Low Interference Regime
0 10 20 30 40 50 60 70 80 900
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
θ
Thr
esho
ld o
n h sin(θ)
P = 10
P = 1
For θ = 0, h(1 + h2P ) ≤ 0.5 (SISO result)
ECE Illinois & Purdue
64 / 177
Low Interference Regime - Simplified
Theorem
For any P , ifh ≤ sin(θ)
then sum capacity of MISO interference channel is achieved by usingGaussian inputs, transmit beamforming, and treating interference asnoise at receivers.
ECE Illinois & Purdue
65 / 177
SIMO Channel
Z2
d
d
hc
Z1
X1
X2 Y 2
Y 1
θ: angle between d and c.
No covariance optimization
ECE Illinois & Purdue
66 / 177
Improved usefulness condition
• h(hcX +W 1)− h(hcX +W 2) maximized by Gaussiandistribution?
• Σ1 � Σ2 is a sufficient condition, but not necessary.
• We can show that (c>Σ−11 c)−1 ≤ (c>Σ−1
2 c)−1 is sufficient.
ECE Illinois & Purdue
67 / 177
Genie Construction
• Symmetric genie signal
Si = hcXi +W i
[ZiW i
]∼ N
(0,
[I A
A> Σ
])• Useful Genie:(
c>Σ−1c)−1≤(c>(I −AΣ−1A>
)−1c
)−1
• Smart Genie:A>(h2Pcc> + I)−1d− hc = 0
• Goal: Find A and Σ that result in largest threshold on h for lowinterference regime
ECE Illinois & Purdue
68 / 177
Low Interference Regime
Theorem
The sum capacity of the SIMO interference channel is achieved byusing Gaussian inputs, receive beamforming, and treating interferenceas noise at the receivers, and is given by
CICsum =1
2log
(1 +
P cos2 θ
1 + h2P+ P sin2 θ
)if the channel gain parameter h satisfies:
|h| < h0(θ, P ) (3)
with h0(θ, P ) being the positive solution to the implicit equation
h2 − sin2 θ =
(cos θ
1 + h2P− h)2
+
. (4)
Same threshold as the MISO channel
ECE Illinois & Purdue
69 / 177
Low Interference Regime
0 10 20 30 40 50 60 70 80 900
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
θ
Thr
esho
ld o
n h sin(θ)
P = 10
P = 1
For θ = 0, h(1 + h2P ) ≤ 0.5 (SISO result)
ECE Illinois & Purdue
70 / 177
Low Interference Regime - Simplified
Theorem
For any P , ifh ≤ sin(θ)
then sum capacity of MISO interference channel is achieved by usingGaussian inputs, transmit beamforming, and treating interference asnoise at receivers.
ECE Illinois & Purdue
71 / 177
Interference Management – Lessons from IT
• Treating interference as noise is optimal as long as it is veryweak (low) – counterpart of very strong interference regime
• Low interference regime can be significant with multipleantennas
• What if interference is not low (but still weak)?
• SISO: Han-Kobayashi scheme with joint decoding in two-user case(one bit optimal [Etkin,Tse,Wang 07])
• Simple schemes that exploit structure in interferencewithout requiring joint decoding?
ECE Illinois & Purdue
72 / 177
References
1 T.S. Han and K. Kobayashi. “A new achievable rate region forthe interference channel.” IEEE Transactions on InformationTheory Jan 1981.
2 R.H. Etkin, D. NC Tse, and H. Wang. “Gaussian interferencechannel capacity to within one bit.” IEEE TransactionsInformation on Theory, Dec 2008.
3 X. Shang, G. Kramer, and B. Chen, “A new outer bound andnoisy-interference sum-rate capacity for the Gaussian interferencechannels, IEEE Trans. on Inform. Theory, Feb. 2009.
4 A. S. Motahari and A. K. Khandani, “Capacity bounds for theGaussian interference channel, IEEE Trans. on Inform. Theory,Feb. 2009.
5 V.S. Annapureddy and V.V. Veeravalli. “Gaussian interferencenetworks: Sum Capacity in the Low Interference Regime and NewOuter Bounds on the Capacity Region,” IEEE Trans. on Inform.Theory, July 2009.
6 V.S. Annapureddy and V.V. Veeravalli. “On the Sum Capacity ofMIMO Interference Channel in the Low Interference Regime.”IEEE Trans. on Inform. Theory, Mar. 2010.
ECE Illinois & Purdue
73 / 177
Part 3: Degrees of Freedom Characterization ofInterference Channels
ECE Illinois & Purdue
74 / 177
Information Theory for IC: State-of-the-art
• Exact characterization• Very hard problem, still open even after > 30 years
• Approximate characterization• Within constant number of bits/sec• Provides some architectural insights
• Degrees of freedom (or multiplexing gain)
DoF = limSNR→∞
sum capacity
log SNR
• Pre-log factor of sum-capacity in high SNR regime• Number of interference free sessions per channel use• Simplest of the three, but can provide useful insight
ECE Illinois & Purdue
75 / 177
K-user (SISO) Interference Channel
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
How many Degrees of Freedom (DoF)?
ECE Illinois & Purdue
76 / 177
Degrees of Freedom with Orthogonalization
• One active user per channel use• Every user gets an interference free channel once every K channel
uses• DoF per user is 1/K; total DoF equals 1
• Special Case: K = 2• Can easily show that outer bound on DOF equals 1
=⇒ TDMA optimal from DoF viewpoint for K = 2
ECE Illinois & Purdue
77 / 177
Degrees of Freedom for general K
• Outer Bound on DoF [Host-Madsen, Nosratinia ’05]
• There are K(K− 1)/2 pairs and each user appears in (K− 1) pairs• Thus DoF ≤ K/2 or per user DoF ≤ 1/2
• Amazingly, this outer bound is achievable via linear interferencesuppression!
Interference Alignment [Cadambe & Jafar ’08]
ECE Illinois & Purdue
78 / 177
Linear Transmit/Receive Strategies
Interference Channel with Tx/Rx Linear CodingU1 0 00 U2 00 0 U3
†︸ ︷︷ ︸
Receive Beams
H1,1 H1,2 H1,3
H2,1 H2,2 H2,3
H3,1 H3,2 H3,3
︸ ︷︷ ︸
Channel
V 1 0 00 V 2 00 0 V 3
︸ ︷︷ ︸
Transmit Beams
End-to-End matrix is Diagonal =⇒ No Interference!
# streams = Size of the Diagonal matrix
ECE Illinois & Purdue
79 / 177
DoF of Linear Strategies
U1 0 00 U2 00 0 U3
† H1,1 H1,2 H1,3
H2,1 H2,2 H2,3
H3,1 H3,2 H3,3
V 1 0 00 V 2 00 0 V 3
H i,j : NT ×NT block-diagonal matrix
• (Symmetric) MIMO:N = # antennas
• Symbol Extensions (Time or Frequency)T = # symbol extensions
DoF (T ) = (#streams)/T
ECE Illinois & Purdue
80 / 177
Interference Alignment with Symbol Extensions [Cadambe& Jafar ’08]
Tx1
Tx2
Tx3
Rx1
Rx2
Rx3
Hi, j =× 0 00 × 00 0 ×
⎛
⎝
⎜⎜
⎞
⎠
⎟⎟
3 Symbol Extensions
4 interference free streams =⇒ PUDoF = 4/9
ECE Illinois & Purdue
81 / 177
Interference Alignment with Symbol Extensions
Tx1
Tx2
Tx3
Rx1
Rx2
Rx3
Hi, j =× 0 00 × 00 0 ×
⎛
⎝
⎜⎜
⎞
⎠
⎟⎟
• Set v1a to some fixed direction
• Align v2 with v11 at Rx 3:
H3,2v2 = H3,1v1a
v2 = H−13,2H3,1v1a
• Align v3 with v2 at Rx 1:
H1,3v3 = H1,2v2
v3 = H−11,3H1,2v2
• Align v1b with v3 at Rx 2:
H2,1v1b = H2,3v3
v1b = H−12,1H2,3v3
ECE Illinois & Purdue
82 / 177
Interference Alignment with Symbol Extensions
• Signaling matrix for transmitter `: V ` = [v`1 v`2 · · ·v`m]
• Fix Ik as the interference space at receiver k
• Alignment conditions are:Receiver 1
H1,1V 1 ∩ I1 = ∅H1,2V 2 ⊆ I1
...
H1,KV K ⊆ I1
Receiver 2
H2,1V 1 ⊆ I2
H2,2V 2 ∩ I2 = ∅...
H2,KV K ⊆ I2
. . .
Receiver K
HK,1V 1 ⊆ IKHK,2V 2 ⊆ IK
...
HK,KV K ∩ IK = ∅• For alignment we require m/2 dimensions each for V k and Ik• May not be possible in general
ECE Illinois & Purdue
83 / 177
Asymptotic Interference Alignment
• Enumerate all S∆= K(K − 1) cross-channels with single index:
T = {T i} = {Hk,` : k 6= `}
• Use same signal space V(m) at all Tx’s, defined recursively as:
V(0) = {1}V(m) =
{v,T 1v, . . . ,T Sv : v ∈ V(m−1)
}={T α1
1 T α22 · · ·T αS
S 1 : α1 + α2 + · · ·+ αS ≤ m}
• Dimension of signal space (# symbol extensions):∣∣∣V(m)∣∣∣ =
(m+ S
m
)
ECE Illinois & Purdue
84 / 177
Asymptotic Interference Alignment
• Interference space at Rx k is Ik = ∪ 6=kHk,`V(m) ⊂ V(m+1)
• Desired signal space at Rx k is Hk,kV(m)
• Almost surely, no overlap between desired signal space andinterference space as long as channel coefficients are changingover symbol extensions and are generic
• Also ∣∣V(m)∣∣∣∣V(m+1)∣∣ =
(m+Sm
)(m+1+Sm+1
) =m+ 1
m+ 1 + S
m→∞−→ 1
• Desired signal space gets half the dimensions asymptotically:∣∣Hk,kV(m)∣∣∣∣Hk,kV(m)
∣∣+ |Ik|m→∞−→ 1
2
ECE Illinois & Purdue
85 / 177
Complexity of asymptotic Interference Alignment
# symbol extensions 0 20 40 60 80 100
0.44
0.45
0.46
0.47
0.48
0.49
0.5
PUDoF
PUDoF of 0.5 is achieved asymptotically
ECE Illinois & Purdue
86 / 177
Complexity of asymptotic Interference Alignment
# symbol extensions
PUDoF
0 200 400 600 800 1000 1200 14000.32
0.34
0.36
0.38
0.4
0.42
0.44
0.46
0.48
0.53 User
4 User
[Choi, Jafar, and Chung, ’09]
ECE Illinois & Purdue
87 / 177
Interference Alignment Summary
+ Achieves optimal PUDoF for fully connected channel
- Requires global channel state information (CSI)
- Requires large number of symbol extensions
ECE Illinois & Purdue
88 / 177
References
1 V.R. Cadambe and S.A. Jafar. “Interference Alignment andDegrees of Freedom of the K-User Interference Channel.” IEEETrans. Inform. Theory, August 2008.
2 S. A. Jafar. “Interference Alignment – A New Look at SignalDimensions in a Communication Network.” In Foundations andTrends in Communications and Information Theory, NOWPublications, 2010.
3 A. Host-Madsen and A. Nosratinia, “The multiplexing gain ofwireless networks,” in Proceedings of ISIT, 2005.
4 S.W. Choi, S.A. Jafar, and S.-Y. Chung. “On the beamformingdesign for efficient interference alignment.” IEEE CommunicationsLetters, 2009.
ECE Illinois & Purdue
89 / 177
Part 4: Finite Diversity and Iterative Algorithms
ECE Illinois & Purdue
90 / 177
Finite Diversity and Interference alignment
• Asymptotic interference alignment schemes provide degree offreedom gains by letting number of symbol extensions go toinfinity
• In practice, we only have finite number of symbol extensions fromdiversity in time, frequency, or multiple antennas
• How do we apply idea of interference alignment to finite diversitycase?
ECE Illinois & Purdue
91 / 177
Outline
• Look at two cases of IA with finite diversity
1 Spatial diversity through multiple antennas2 Time and frequency diversity
• Provide theoretical guarantees for DoF using IA schemes
• Iterative algorithms inspired by IA for MIMO IC
ECE Illinois & Purdue
92 / 177
Constant MIMO Channel
• Fully connected K-user MIMO interference channel
• N (k)t transmit antennas and N
(k)r receive antennas for user k
• Usual channel model for receiver k given by
Y k =
K∑j=1
HkjXj + Zk
• Y k - signal at receiver k (N(k)r × 1)
• Hkj - channel from transmitter j to receiver k (N(k)r ×N (j)
t )
• Xj - signal sent by transmitter j (N(j)t × 1)
• Zk - noise at receiver k (N(k)r × 1)
• Channels {Hkj} drawn once from continuous joint distributionhence name constant MIMO channel
ECE Illinois & Purdue
93 / 177
Constant MIMO Channel
Tx1 Rx1
Tx2 Rx2
Tx3 Rx3
ECE Illinois & Purdue
94 / 177
Interference Alignment for Constant MIMO Channel
• Consider linear transmit and receive strategies:
U†kY k = U†k
K∑j=1
Hkj Vjxj︸ ︷︷ ︸Xj
+U†kZk
• Transmitter constructs channel input Xj = Vjxj with Vj size
N(j)t × dj
• Receiver processes channel output U†kY k with Uk size N
(k)r × dk
• IA for Constant MIMO IC: Construct {Vk} and {Uk} such thatreceivers can zero force interference yielding interference freechannels between each Tx/Rx pair k
ECE Illinois & Purdue
95 / 177
Interference Alignment for Constant MIMO Channel
• Goal: to create dk interference free streams between Tx/Rx pair k
• To zero force interference, we need
U†kHkjVj = 0 ∀ k 6= j
• Zero forcing all interference at Rx k yields effective channel
U†kY k = U†kHkkVkxk + U†kZk
• This is MIMO channel with channel matrix U†kHkkVk
• To create dk interference free channels, we need
rank(U†kHkkVk
)= dk
ECE Illinois & Purdue
96 / 177
Interference Alignment for Constant MIMO Channel
• For IA in constant MIMO channel with dk interference freechannels between Tx/Rx pair k, we need
rank(U†kHkkVk
)= dk ∀k
U†kHkjVj = 0 ∀k 6= j
• Choose transmit and receive vectors {Vk} and {Uk} using onlyinterfering channels {Hkj}k 6=j
• Combined with our assumption that {Hkj} are drawn from jointdistribution, (almost surely) it holds that
rank(U†kHkkVk
)= dk
Rank condition is automatic
ECE Illinois & Purdue
97 / 177
Solving IA Equations
• We only need to solve IA equations
U†kHkjVj = 0 ∀k 6= j
When is this possible?
• Feasibility of solving IA equations will depend on K, {N (k)t },
{N (k)r }, and {dk}
• Easy to derive necessary conditions for solving IA equations usingvariable and equation counting [Yetis10,Razaviyayn12,Gonzalez12]
• For some specific system configurations, possible to derivesufficient conditions using algebraic geometry (branch of mathfocused on answering when polynomial equations have solutions)
ECE Illinois & Purdue
98 / 177
Symmetric Constant MIMO IC - Necessary
• To illustrate necessary condition, consider symmetric constantMIMO IC:
1 N(k)t = N
(k)r = N
2 dk = d
• In order to solve IA equations, number of variables should be noless than number of equations
• This intuitive idea can be proved rigorously using algebraicgeometry or algebraic field theory
• The no interference equation for Tx j at Rx k
U†kHkjVj = 0
consists of d2 equations
• There are total of K(K − 1)d2 equations to ensure no interference
ECE Illinois & Purdue
99 / 177
Symmetric Constant MIMO IC - Necessary
• Counting number of variables is little bit trickier
• Satisfying IA equations only depends on subspaces {span(Uk)}and {span(Vk)} and not particular choice of {Uk} and {Vk}
• To illustrate this point, suppose that U†kHkjVj = 0 ∀k 6= j
• For any other basis {Vj} and {Uj} of {span(Uk)} and{span(Vk)}, there exist d× d full rank matrices {Qk} and {Pk}such that
Vj = VjQj
Uk = UkPk
• Then we have U†kHkjVj = P†kU†kHkjVj︸ ︷︷ ︸
=0
Qj = 0
ECE Illinois & Purdue
100 / 177
Symmetric Constant MIMO IC - Necessary
• Since satisfying IA equations only depends on subspaces spannedby {Vk} and {Uk}, number of variables is less than 2K ×Nd
• By preceding argument, due to freedom in choosing Pk and Qk infact we have
dim (span (Vk)) = dim (span (Uk)) = Nd− d2 = (N − d)d
• This implies that number of variables is 2K × d(N − d)
• Necessary condition for IA in symmetric constant MIMO ICbecomes
(K + 1)d ≤ 2N
• This implies that
DoF ≤ K⌊
2N
K + 1
⌋≤ 2N
K
K + 1≤ 2N
ECE Illinois & Purdue
101 / 177
Solving IA Equations - Feasability
• This necessary condition is not always sufficient [Yetis10]
• N(k)t = N
(k)r = 3, dk = 2, and K = 2 system satisfies necessary
condition but is not achievable
• For symmetric constant MIMO channel with K ≥ 3, necessarycondition (K + 1)d ≤ 2N is sufficient [Bresler14]
• Necessary and sufficient conditions for feasibility of IA for otherconfigurations too using similar ideas[Razaviyayn12,Ruan13,Gonzalez12]
• Conditions for feasibility of solving IA equations for constantMIMO IC not known in general
ECE Illinois & Purdue
102 / 177
Finite Time and Frequency Diversity
• To model finite diversity, suppose that we have L symbolextensions, equivalently a diagonal channel matrix
Hkj =
Hkj(1) 0. . .
0 Hkj(L)
• For IA in K-user IC, L ≥ 2K
2sufficient [OzgurTse09]
• Three user case has been exactly characterized exactly usingvector IA achievable scheme [BreslerTse09]
DoF =3
2
(1− 1
4L− 2bL/2c − 1
)= O
(3
2
(1− 1
L
))• As L→∞, recover usual 3
2 DoF
ECE Illinois & Purdue
103 / 177
Finite Diversity and IA
• For K ≥ 4 case with finite diversity, there are sum DoF upperbounds [LiOzgur14]
DoF ≤ K
2
(1− 1
11√L
)• Achievable scheme yields
DoF ≥ K
2
(1− C1
C2√L/2
)
with C1 a constant and C2 = (K − 1)(K − 2)− 1
• Substantial gap between upper bound and achievable scheme
• As L→∞, recover usual K2 DoF
ECE Illinois & Purdue
104 / 177
Time/Frequency vs. Spatial Diversity
• For time and frequency diversity, when amount of diversity (L) islarge, DoF ≈ K
2
• DoF gains for time/frequency diversity scale with K but forK ≥ 4 may need large L to get close to K
2• Coherence time of channel can be issue for large L
• For spatial diversity in symmetric constant MIMO IC, DoF ≈ 2N
• DoF gains for spatial diversity do not scale with K but no need tocode over large number of time/frequency slots
• Finally, for time-varying symmetric MIMO channel with Nantennas at all Tx/Rx [CadambeJafar08]
DoF =KN
2
ECE Illinois & Purdue
105 / 177
IA Inspired Algorithms for Constant MIMO IC
• Examined when we can solve IA equations
rank(U†kHkkVk
)= dk ∀k
U†kHkjVj = 0 ∀k 6= j
• Analysis does not tell us how to actually solve IA equations
• Iterative algorithms based on idea of IA for constant MIMO IC
• Channel State Information: Assume each transmitter and receiverknows all connected channels
ECE Illinois & Purdue
106 / 177
IA Inspired Algorithms for Constant MIMO IC
• Look at two algorithms inspired by IA [Gomadam08]
• Designing transmit vectors to minimize interference is difficultbecause each transmit vector affects interference at all receivers
• Similarly, designing transmit vectors to maximize sum rate isnon-convex problem and difficult to solve
• In contrast, each receive vector is affected only by interferenceseen at receiver and is easy to design
• Idea of both algorithms is to only design receive vectors butalternate the direction of communication
• Role of transmit and receive vectors alternates• Appropriate for TDD systems with natural reversal of directions of
communication
ECE Illinois & Purdue
107 / 177
Constant MIMO IC Algorithm Overview
Tx1 Rx1
Tx2 Rx2
Tx3 Rx3
−→V1(n)−→V1(n)
−→U1(n)←−V1(n)
←−U1(n)
←−V1(n)
−→V1(n+ 1)
−→V2(n)−→V2(n)
−→U2(n)←−V2(n)
←−U2(n)
←−V2(n)
−→V2(n+ 1)
−→V3(n)−→V3(n)
−→U3(n)←−V3(n)
←−U3(n)
←−V3(n)
−→V3(n+ 1)
Forward Direction - Design Receive Vectors Reverse Direction ofCommunication: ←−Vk(n) =
−→Uk(n) Reverse Direction - Design Receive
Vectors Reverse Direction of Communication and Repeat:−→Vk(n+ 1) =
←−Uk(n)
−→←−ECE Illinois & Purdue
108 / 177
Min Leakage Algorithm
• For fixed transmit vectors {Vk}, interference leakage power at
receiver k given by Ik = trace(U†kQkUk) with
Qk , P
K∑j=1j 6=k
HkjVjV†jH†kj
• If Ik = 0, then we achieve IA
• To minimize interference leakage power, choose dk least dominanteigenvectors of Qk, i.e, those eigenvectors corresponding to dksmallest eigenvalues of Qk
• Alternate directions reversing role of transmit and receive vectors
• Sum interference leakage power converges
ECE Illinois & Purdue
109 / 177
Min Leakage Algorithm
1 Start with arbitrary {−→Vk(0)}2 Set
−→Uk(n) to be dk least dominant eigenvectors of
−→Qk(n)
3 Reverse direction of communication
←−Vk(n) =
−→Uk(n)
4 Set←−Uk(n) to be dk least dominant eigenvectors of
←−Qk(n)
5 Reverse direction of communication and repeat
−→Vk(n+ 1) =
←−Uk(n)
ECE Illinois & Purdue
110 / 177
Max SINR Algorithm
• At high SNR, Min Leakage provides good performance -interference limited regime
• At lower SNR, Min Leakage does not provide good performance -noise becomes more important
• Max SINR same form as Min Leakage algorithm but with MMSEreceive vectors
Uk =
I + P
K∑j=1j 6=k
HkjVjV†jH†kj
−1
HkkVk
ECE Illinois & Purdue
111 / 177
Max SINR Algorithm
• At high SNR, MMSE filter becomes zero forcing filter - focuses oninterference
• At low SNR, MMSE filter becomes matched filter - focuses onnoise
• At intermediate SNR, MMSE filter trades off between interferenceand noise
• Max SINR has better performance at low SNR than Min Leakageor other purely IA oriented algorithms
• No proof of convergence but appears to converge in practice
ECE Illinois & Purdue
112 / 177
Performance Comparison
SNR Per User [dB]0 5 10 15 20 25 30 35 40
Ave
rage
Sum
Rat
e [b
its/u
se]
0
10
20
30
40
50
60
70Max SINRMin Leakage
Similar performance at high SNR Max SINR much better at low SNR
ECE Illinois & Purdue
113 / 177
Other Algorithms
• Min Leakage and Max SINR make sense in TDD system due tonatural alternation of transmitting and receiving
• IA Algorithms for FDD channels have been developed[PetersHeath09]
• Similar in spirit to Min Leakage algorithm but allows for FDD
• Convergent version of Max SINR [WilsonVVV11]• Power control step to produce non-decreasing sum rate• Same performance as Max SINR in simulations
• Several other similar iterative algorithms inspired by IA inliterature
ECE Illinois & Purdue
114 / 177
References
1 C. Yetis, T. Gou, S. Jafar, and A. Kayran, “On feasibility ofinterference alignment in MIMO interference networks,” IEEETrans. Signal Process., 2010.
2 M. Razaviyayn, G. Lyubeznik, and Z.Q. Luo, “On the Degrees ofFreedom Achievable Through Interference Alignment in a MIMOInterference Channel,” IEEE Trans. Signal Process., 2012
3 O. Gonzalez, C. Beltran, and I. Santamaria, “A feasibility test forlinear interference alignment in MIMO channels with constantcoefficients,” IEEE Trans. on Info. Theory, 2014
4 G. Bresler, D. Cartwright, and D. Tse, “Feasibility of InterferenceAlignment for the MIMO Interference Channel,” in IEEE Trans.Info. Theory, 2014
5 L. Ruan, V.K.N. Lau, M.Z. Win, “The feasibility conditions forinterference alignment in MIMO networks,” IEEE Trans. SignalProcess., 2013
6 A. Ozgur and D. Tse, “Achieving Linear Scaling with InterferenceAlignment”, ISIT, 2009
7 G. Bresler and D. Tse, “Degrees-of-freedom for the 3-userGaussian interference channel as a function of channel diversity”,Allerton Conf., 2009
ECE Illinois & Purdue
115 / 177
References
8 V. Cadambe and S. Jafar, “Interference alignment and degrees offreedom of the-user interference channel,” IEEE Trans. Info.Theory, 2008
9 C.T. Li, and A. Ozgur, “Channel Diversity needed for VectorInterference Alignment”, on arXiV, 2014
10 K. Gomadam, V. R. Cadambe, and S. A. Jafar, “Approaching thecapacity of wireless networks through distributed interferencealignment,” in GLOBECOM, 2008
11 S. W. Peters and R. W. Heath, “Interference alignment viaalternating minimization, ICASSP, 2009
12 C. Wilson and V. Veeravalli, “A Convergent Version of the MaxSINR Algorithm for the MIMO Interference Channel,” IEEETrans. Wireless Comm., 2011
13 C. Wilson and V. Veeravalli, “Degrees of Freedom for theConstant MIMO Interference Channel with CoMP Transmission,”IEEE Trans. Comm., 2014
ECE Illinois & Purdue
116 / 177
Part 5: Coordinated Multi-Point Transmission
ECE Illinois & Purdue
117 / 177
K-User Interference Channel
Channel State Information known at all nodes.
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
ECE Illinois & Purdue
118 / 177
Coordinated Multi-Point (CoMP) Transmission
Messages are jointly transmitted using multiple transmitters.
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
ECE Illinois & Purdue
119 / 177
CoMP Transmission
• Each message is jointly transmitted using M transmitters
• Message i is transmitted jointly using the transmitters in Ti• For all i ∈ [K], |Ti| ≤M
• We consider all message assignments that satisfy the cooperationconstraint
ECE Illinois & Purdue
120 / 177
Degrees of Freedom (DoF)
DoF = limSNR→∞
sum capacity
log SNR
Objective: Determine the DoF as a function of K and M
PUDoF(M) = limK→∞
DoF(K,M)
K
Is PUDoF(M)>PUDoF(1) for M > 1?
ECE Illinois & Purdue
121 / 177
Example: Two-user Interference Channel
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
No Cooperation, DoF=1, Time Sharing
Full Cooperation, DoF=2, ZF Transmit Beamforming
ECE Illinois & Purdue
122 / 177
No Cooperation (M = 1)
• For M = 1, outer bound = K/2
• The outer bound can be achieved by jointly coding across multipleparallel channels [Cadambe & Jafar ’08]:
DoF(K,M = 1) = limL→∞
DoF(K,M = 1, L)
L= K/2
where L is the number of parallel channels
Corollary
Without cooperation, the Per User DoF number is given by
PUDoF(M = 1) =1
2
ECE Illinois & Purdue
123 / 177
Full Cooperation (M = K)
• In this case, the channel is a MISO Broadcast channel.
• Each message is available at K antennas, and hence, can becanceled at K − 1 receivers.
• Each user achieves 1 DoF,
DoF(K,M = K) = K.
What happens with partial cooperation (1 < M < K)?
ECE Illinois & Purdue
124 / 177
Clustering
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
W4 W4Tx4 Rx4
No Degrees of Freedom Gain
ECE Illinois & Purdue
125 / 177
Spiral Message Assignment
Ti = {i, i+ 1, . . . , i+M − 1}
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
ECE Illinois & Purdue
126 / 177
Spiral Assignment: Matrix Interpretation
ECE Illinois & Purdue
127 / 177
Spiral Assignment: Matrix Interpretation
M : # of non-zero blocks in the columns of V
ECE Illinois & Purdue
128 / 177
Example: K = 3,M = 2
PUDoF= 23
ECE Illinois & Purdue
129 / 177
Spiral Message Assignment: Results
Theorem
The DoF of interference channel with a spiral message assignmentsatisfies
K +M − 1
2≤ DoF(K,M) ≤
⌈K +M − 1
2
⌉
Proof of Achievability: First M − 1 users are interference-free, andinterference occupies half the signal space at each other receiver
Generalizes the Asymptotic Interference Alignment scheme
ECE Illinois & Purdue
130 / 177
Vector Space Interference Alignment [Cadambe-Jafar ’08]
• Asymptotic interference alignment works only if the (Signal +Interference) matrix is full rank
• The generic channel coefficients assumption (they have a jointpdf) is crucial
• With CoMP, the same point to point channel can carry bothdesired and interfering signals
• It is not clear whether the generic channel coefficients assumptionsuffices to prove full rankness of the (Signal + Interference) matrix
ECE Illinois & Purdue
131 / 177
Key Tool: Algebraic Independence
Consider the system of polynomial equations
s1 = f1(t1, t2, · · · , tn)
s2 = f2(t1, t2, · · · , tn)
...
sm = fm(t1, t2, · · · , tn).
Definition
The polynomials f1, f2, · · · , fm are said to be algebraically dependentif and only if there exists an annihilating polynomial F (s1, s2, · · · , sm)such that F (f1, f2, · · · , fm) = 0.
ECE Illinois & Purdue
132 / 177
Jacobian Criterion
The polynomials f1, f2, · · · , fm are algebraically independent if andonly if the Jacobian matrix(
∂fi∂tj
)1≤i≤m,1≤j≤n
has full structural row rank equal to m.
ECE Illinois & Purdue
133 / 177
Algebraic Independence: Applications
Consider the system of polynomial equations
s1 = f1(t1, t2, · · · , tn)
s2 = f2(t1, t2, · · · , tn)
...
sm = fm(t1, t2, · · · , tn).
Theorem
If the variables {t1, . . . , tn} have a continuous joint pdf and thepolynomials f1, f2, . . . , fm are algebraically independent, then thevariables {s1, . . . , sm} have a continuous joint pdf.
ECE Illinois & Purdue
134 / 177
Outline of the Achievable Scheme
OriginalChannel
ZFEncoder
IAEncoder
IADecoder
Derived Channel
Approach:
1 ZF Step: Exploit cooperation to transform the interferencechannel into a derived channel (with single-point transmission)
2 IA Step: Use the known IA techniques to design beams forderived channel
3 Prove that the asymptotic IA step works for generic channelcoefficients
ECE Illinois & Purdue
135 / 177
Example: M = 2
12 DoF
12 DoF
W(1)1
W(2)1
X1 Y1 1 DoF
12 DoF W
(1)2
X2 Y212 DoF
12 DoF W
(1)3
X3 Y312 DoF
h1,1
h2,2
h3,3
ECE Illinois & Purdue
136 / 177
Derived Channel Example: M = 2
X(1)1
X(2)1
Y1 1 DoF
X(1)2
Y212 DoF
X(1)3
Y312 DoF
g(1)1,1
g(2)1,1
g(1)2,2
g(1)3,3
Asymptotic Interference Alignment is used to pack the interference inhalf the signal space at Rx 2 and 3.
ECE Illinois & Purdue
137 / 177
Asymptotic IA for Derived Channel
• (Signal + Interference) matrix has full column rank if the
variables {g(m)i,j } have a continuous joint pdf
• {g(m)i,j } have a continuous joint pdf because {hi,j} have a
continuous joint pdf and the polynomials defining thetransformations from original to derived channel coefficients arealgebraically independent
Prove algebraic independence using Jacobian Criterion
ECE Illinois & Purdue
138 / 177
DoF Outer Bound
W1 W1Tx1 Rx1
W2 W2Tx2 Rx2
W3 W3Tx3 Rx3
W4 W4Tx4 Rx4
W5 W5Tx5 Rx5
ECE Illinois & Purdue
139 / 177
DoF Outer Bound
W1 W1
W2 W2
W3 W3
W4 W4
W5 W5
Tx1
Tx2
Tx3
Tx4
Tx5
Rx1
Rx2
Rx3
Generic channel coefficients → DoF ≤ 3
ECE Illinois & Purdue
140 / 177
DoF Outer Bound: Bipartite Graph Representation
W1
W2
W3
W4
W5
Tx1
Tx2
Tx3
Tx4
Tx5
|Neighborhood({Tx1,Tx2})| = 3 ≥ K − |{Tx1,Tx2}| =⇒ DoF ≤ 3
ECE Illinois & Purdue
141 / 177
DoF Outer Bound: Bipartite Graph Representation
Theorem
DoF ≤ minS⊆{1,...,K}
max (|Neighborhood(S)|,K − |S|)
ECE Illinois & Purdue
142 / 177
DoF Outer Bound: Bipartite Graph Representation
Definition
DoFout(K,M) = max{Ti}
minS⊆{1,...,K}
max(|Neighborhood(S)|,K − |S|)
PUDoFout(M) = limK→∞
DoFout(K,M)
K
DoF(K,M) ≤ DoFout(K,M)
PUDoF(M) ≤ PUDoFout(M)
ECE Illinois & Purdue
143 / 177
DoF Outer Bound: Results
Definition
We say that a message assignment satisfies a local cooperationconstraint if and only if ∃r(K) = o(K), and for all K−user channels,
Ti ⊆ {i− r(K), i− r(K) + 1, . . . , i+ r(K)},∀i ∈ [K]
Theorem
With the restriction to local cooperation,
PUDoFloc(M) =1
2
Local cooperation cannot achieve a scalable Dof gain
ECE Illinois & Purdue
144 / 177
DoF Outer Bound: Results
Theorem
For M ≥ 2,
PUDoF(M) ≤ PUDoFout(M) ≤ M − 1
M
Corollary
PUDoF(2) =1
2
Assigning each message to two transmitters cannot achieve a scalableDoF gain
ECE Illinois & Purdue
145 / 177
Fully Connected IC with CoMP: Summary
• Considered general cooperation constraint that limits the numberof transmitters at which each message can be available
• Asymptotic Interference Alignment can be used to achieve DoFgains
• Dual results can be obtained with receiver cooperation butrequires sharing of analog received signals
• Iterative Max SINR type algorithms algorithms can be designed inconstant MIMO channel setting [WilsonVVV-14]
• Scalable DoF gains cannot be achieved through:
• Local Cooperation
• Arbitrary assignment of each message to two transmitters
ECE Illinois & Purdue
146 / 177
References
1 P. Marsch and G. P. Fettweis “Coordinated Multi-Point in MobileCommunications: from theory to practice,” First Edition,Cambridge, 2011.
2 A. Host-Madsen and A. Nosratinia, “The multiplexing gain ofwireless networks,” in Proc. IEEE International Symp. Inf. Theory(ISIT), 2007.
3 V. Cadambe and S. A. Jafar, “Interference alignment and degreesof freedom of the K-user interference channel,” IEEE Trans. Inf.Theory, 2008.
4 V. S. Annapureddy, A. El Gamal, and V. V. Veeravalli, “Degreesof Freedom of Interference Channels with CoMP Transmission andReception,” IEEE Trans. Inf. Theory, 2012.
5 A. El Gamal, V. S. Annapureddy, and V. V. Veeravalli, “OnOptimal Message Assignments for Interference Channels withCoMP Transmission,” in Proc. CISS, 2012
6 C. Wilson and V. Veeravalli, “Degrees of Freedom for theConstant MIMO Interference Channel with CoMP Transmission,”IEEE Trans. Comm., 2014
ECE Illinois & Purdue
147 / 177
Part 6: Locally Connected Networks
ECE Illinois & Purdue
148 / 177
Locally Connected Model
Tx i is connected to receivers {i, i+ 1, . . . , i+ L}.
Wyner Model: L =1
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx5 Rx5
L = 2
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx5 Rx5
ECE Illinois & Purdue
149 / 177
Results for Wyner Model [Lapidoth-Shamai-Wigger ’07]
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx1 Rx1
Rx5 Tx5
Rx6 Tx6
W2
W3
W4
W1
W5
W6
Backhaul load factor =1
PUDoF (L=1,M=2) = 2/3 > 1/2
W1
W2
W4
W5
ECE Illinois & Purdue
150 / 177
Results for Wyner Model [Lapidoth-Shamai-Wigger ’07]
• Spiral transmit sets: Ti = {i, i+ 1, . . . , i+M − 1}
• PUDoF(L = 1,M) = MM+1
Backhaul load factor = M2
• Local cooperation can achieve PUDoF gains for locally connectedchannels
• Achievable scheme relies on only:• Zero-forcing transmit beamforming• Local CSI• Fractional reuse
Is spiral message assignment optimal?
ECE Illinois & Purdue
151 / 177
Example: No Cooperation
PUDoF(L =1,M =1) = 1
2 PUDoF(L =1,M =1) = 23
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
W1
W2
W3
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
W1
W2
W3
Interference-aware message assignment + Fractional reuse
ECE Illinois & Purdue
152 / 177
Locally Connected IC with CoMP: Main Result
Theorem
Under the general cooperation constraint |Ti| ≤M, ∀i ∈ {1, 2, . . . ,K},
2M
2M + L≤ PUDoF(L,M) ≤ 2M + L− 1
2M + L
and the optimal message assignment satisfies a local cooperationconstraint.
Corollary
PUDoF(L = 1,M) =2M
2M + 1
ECE Illinois & Purdue
153 / 177
DoF Achieving Scheme
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx1 Rx1
Rx5 X5
W2
W4
W1
W2
W4
W1
W5
W3
W5
Backhaul load factor =6/5 PUDoF (L=1,M=2) = 4/5 > 2/3
ECE Illinois & Purdue
154 / 177
DoF Outer Bound
Have to consider all possible message assignments satisfying|Ti| ≤M, ∀i ∈ [K]
1 First simplify the combinatorial aspect of the problem byidentifying useful message assignments
2 Then derive an equivalent model with fewer receivers and sameDoF
ECE Illinois & Purdue
155 / 177
DoF Outer Bound: Useful Message Assignments
An assignment of a message Wx to a transmitter Ty is useful only ifone of the following conditions holds:
1 Signal delivery: Ty is connected to the designated receiver Rx
2 Interference mitigation: Ty is interfering with anothertransmitter Tz, both carrying the message Wx
ECE Illinois & Purdue
156 / 177
DoF Outer Bound: Useful Message Assignments
Tx1 Rx1
Tx2 Rx2
Tx3 Rx3
Tx4 Rx4
W3 W3
Assigning W3 to Tx1 is not useful.
ECE Illinois & Purdue
157 / 177
DoF Outer Bound: Useful Message Assignments
Corollary
An assignment of a message Wx to a transmitter Ty is useful only ifthere exists a chain of interfering transmitters carrying Wx thatincludes Ty and another transmitter Tz that is connected to Rx
Proves optimality of local cooperation
ECE Illinois & Purdue
158 / 177
Extensions: Multiple-Antenna Transmitters
Theorem
For the locally connected interference channel with N−antennatransmitters, if MN ≥ L+M then PUDoF(L,M,N) = 1, otherwise,
2MN
M(N + 1) + L≤ PUDoF(L,M,N) ≤ M(N + 1) + L− 1
M(N + 1) + L
and the optimal message assignment satisfies a local cooperationconstraint.
ECE Illinois & Purdue
159 / 177
CoMP Transmission for IC: Summary
• Local Cooperation• no PUDoF gain for fully connected channel• is optimal for locally connected channel
• Interference aware message assignments allow for higherthroughput
• Fractional reuse and zero-forcing transmit beam-forming aresufficient to achieve PUDoF gains, without need for symbolextensions and interference alignment
ECE Illinois & Purdue
160 / 177
References
1 V. S. Annapureddy, A. El Gamal, and V. V. Veeravalli, “Degreesof Freedom of Interference Channels with CoMP Transmission andReception,” IEEE Trans. Inf. Theory, 2012.
2 A. Lapidoth, N. Levy, S. Shamai (Shitz) and M. A. Wigger“Cognitive Wyner networks with clustered decoding,” IEEE Trans.Inf. Theory 2014
3 A. Wyner, “Shannon-theoretic approach to a Gaussian cellularmultiple-access channel,” IEEE Trans. Inf. Theory, 1994.
4 S. Shamai and M. A. Wigger, “Rate-limited TransmitterCooperation in Wyner’s Asymmetric Interference Network,” inProc. IEEE Int. Symp. Inf. Theory, 2014
ECE Illinois & Purdue
161 / 177
Part 7: Cellular Networks and Backhaul Load Constraint
ECE Illinois & Purdue
162 / 177
Backhaul Load Constraint
More natural cooperation constraint that takes into account overallbackhaul load: ∑
i∈[K] |Ti|K
≤ B
Solution under transmit set size constraint |Ti| ≤M, ∀i ∈ [M ], can beused to provide solutions under backhaul load constraint
ECE Illinois & Purdue
163 / 177
Wyner’s Model with Backhaul Load Constraint
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
Theorem
Under cooperation constraint∑i∈[K] |Ti| ≤ BK,
PUDoF(B) =4B − 1
4B
Recall that |Ti| ≤M,∀i ∈ [K]⇒ PUDoF(M) = 2M2M+1
ECE Illinois & Purdue
164 / 177
Coding Scheme: B = 1
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
W1
W2
W3
B = 2
3 PUDoF = 2
3
B = 6
5 PUDoF = 4
5
3K
8users
5K
8users
PUDoF (B =1) = 3
4
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx1 Rx1
Rx5 X5
W2
W4
W1
W3
W5
ECE Illinois & Purdue
165 / 177
Application in Denser Networks
Tx i is connected to receivers {i, i+ 1, . . . , i+ L}.
Rx1 Tx1
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx5 Rx5
L = 2
Result: Using only zero-forcing transmitbeamforming and fractional reuse:
PUDoF(L,B = 1) ≥ 1
2,∀L ≤ 6.
without need for interference alignmentand symbol extensions
ECE Illinois & Purdue
166 / 177
Application in Denser Networks
PUDoF(M = 1) = 12 PUDoF(B = 1) ≥ 5
9
ECE Illinois & Purdue
167 / 177
Interference in Cellular Networks
Locally (partially) connected interference channel!
ECE Illinois & Purdue
168 / 177
Interference Graph for Single Tier
Tx,Rx pair
Each node represents a Tx-Rx pair
ECE Illinois & Purdue
169 / 177
No Intra-sector Interference
ECE Illinois & Purdue
170 / 177
Partition into Noninterfering Tx-Rx Pairs
ECE Illinois & Purdue
171 / 177
Full Cooperation (M = 6)
Rx2 Tx2
Rx3 Tx3
Rx4 Tx4
Tx1 Rx1
Rx5 Tx5
Rx6 Tx6
W2
W3
W4
W1
W5
W6
2
3 4 5
6
1
B = 6 × 6
9= 4; PUDoF = 6
9= 2
3
ECE Illinois & Purdue
172 / 177
Under-utlizing Backhaul Resources (M = 2)
Rx2 Tx2
Rx3 Tx3
Rx4
Tx1
Rx5
Tx6 Rx6
Tx5
W3
Tx4
W5
W2
W4
W6
Rx1 W1
1
2
3 4 5
6
B = 6
9= 2
3; PUDoF = 4
9
ECE Illinois & Purdue
173 / 177
No Extra Backhaul Load
Backhaul Load B = 1, PUDoF= 715
ECE Illinois & Purdue
174 / 177
Discussion: Cooperation through Backhaul
• Similar gains in DoF for other cellular interference models, withonly zero-forcing and fractional reuse
• Gains improve with asymmetric cooperation and interferenceaware message assignment
• Gains in DoF can also be obtained for uplink with decodedmessages being exchanged through backhaul [Ntranos et al ’14]
• Requires multiple antennas at both mobiles and basestations• For same backhaul load factor, gain is smaller than on downlink
with Tx cooperation
ECE Illinois & Purdue
175 / 177
Summary
• Infrastructure enhancements in backhaul can be exploited throughcooperative transmission to lead to significant rate gains
• Minimal or no increase in backhaul load• Fractional reuse and zero-forcing transmit beam-forming are
sufficient to achieve rate gains• No need for symbol extensions and interference alignment
• Open Questions:• Partial/unknown CSI• Network dynamics and robustness to link erasures• Joint design with message passing schemes for uplink
ECE Illinois & Purdue
176 / 177
References
1 A. El Gamal and V. V. Veeravalli, “Flexible Backhaul Design andDegrees of Freedom for Linear Interference Channels,” in Proc.IEEE Int. Symp. Inf. Theory, 2014.
2 M. Bande, A. El Gamal, and V. V. Veeravalli, “Flexible BackhaulDesign with Cooperative Transmission in Cellular InterferenceNetworks,” in Proc. IEEE Int. Symp. Inf. Theory, 2015
3 V. Ntranos, M. A. Maddah-Ali, and G. Caire, “CellularInterference Alignment,” arXiv, 2014.
4 V. Ntranos, M. A. Maddah-Ali, and G. Caire, “OnUplink-Downlink Duality for Cellular IA,” arXiv, 2014.
ECE Illinois & Purdue
177 / 177
We are writing a book!
A. El Gamal, V. V. Veeravalli “Cellular Interference Management:Flexible Backhaul Design and Cooperation”, Artech House, 2016
ECE Illinois & Purdue