Asymptotic Random Matrix Applications in Communicationweb.mit.edu/sea06/agenda/talks/Bliss.pdf ·...
Transcript of Asymptotic Random Matrix Applications in Communicationweb.mit.edu/sea06/agenda/talks/Bliss.pdf ·...
MIT Lincoln LaboratorymimoTechSem-1
bliss
Asymptotic Random Matrix Applications in Communication:
MIMO and More
Dr. Daniel W. Bliss
MIT-Lincoln Laboratory781-981-3300
This work was sponsored by the Air Force under Air Force Contract FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the United States Government.
MIT Lincoln LaboratorymimoTechSem-2
bliss
Topics
• Multiple-Input Multiple-Output (MIMO)– Introduction– Asymptotic results
• Survey of Communication Examples– Code-Division Multiple Access (CDMA)– Receiver performance in rich interference networks
• Wish List
MIT Lincoln LaboratorymimoTechSem-3
bliss
Complicated MultipathEnvironment
MIMO CommunicationMultiple-Input Multiple-Output
• Single transmitted data stream• Single received data stream• Employ array of antennas at both
transmitter and receiver• Employ multiple modes through
environment
Data…01110111001…
Data…01110111001…
TransmitAntenna
ArraySpace-Time
Coding
MultipleInput
ReceiveAntenna
Array
bounce offbuilding
Space-TimeReceiver
MultipleOutput
MIT Lincoln LaboratorymimoTechSem-4
bliss
The Channel Matrix
• Channel matrix, H, contains complex attenuation between each transmit and receive antenna
• Large channel matrix singular values are useful
Rel
ativ
e Po
wer
Channel Matrix Singular Values
HighCapacity
EnvironmentLowCapacity
Environment
Sorted by Mode Strength
Few Useful Modes
Many Useful Modes
Many Useful Modes
Scatterers
MIT Lincoln LaboratorymimoTechSem-5
bliss
Rich Scattering Channel
Rich ScatteringEnvironment
• Very complicated scattering environment
• Independent Rayleigh fading for each transmit-receive antenna pair
• Entries in matrix independently sampled from unit-norm complex Gaussian distribution
Scatterer
G
RMS antenna pairattenuation
H = a G
G =
g11 g12
g21 g22
gmn
⎛
⎝
⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟
MIT Lincoln LaboratorymimoTechSem-6
bliss
MIMO Taxonomy
• Stationarity required to pass channel estimate to transmitter
• Capacity formulation drives choice of transmitter rank
– Low Spec. Eff. ⇒ low rank TX– High Spec. Eff. ⇒ high rank TX
MIMO LinkTransmitter
Receiver MIMO System Selection
Spectral Efficiency
Stat
iona
rity
Stat
icD
ynam
ic
Low High
Low RankInformed Transmitter Informed Transmitter
Uninformed TransmitterUninformed Transmitter
Channel Matrix Estimate?
MIT Lincoln LaboratorymimoTechSem-7
bliss
Information Theoretic Capacity
Mutual Information
Receive-Signal Entropy Noise-Like Entropy
Signal Model
In Interference Environment
InterferenceCovariance
Matrix
Informed Transmitter Capacity
; Whitened Channel Matrix
Uninformed Transmitter Capacity
MIT Lincoln LaboratorymimoTechSem-8
bliss
Informed Transmitter MIMOWater Filling
• Model
• Modes in channel matrix live in singular values
• Number of employed modes depends on SNR– Higher SNR ⇒ Many modes– Lowest SNR ⇒ Single mode
ReceiveBeamformers
TransmitBeamformers
Mode
High SNR
Low SNR
Water-Filling Analogy
Single ModeInformed Transmitter
MIMO
MIT Lincoln LaboratorymimoTechSem-9
bliss
Topics
• Multiple-Input Multiple-Output (MIMO)– Introduction– Asymptotic results
• Survey of Communication Examples– Code-Division Multiple Access (CDMA)– Receiver performance in rich interference networks
• Wish List
MIT Lincoln LaboratorymimoTechSem-10
bliss
Convergence to the ContinuumHow Big is Big?
• Eigenvalue probability density of
• Ensemble of finite NxN eigenvalue spectra
• Converges quickly
Sort
edEi
genv
alue
[dB
]
Fraction of Eigenvalues
Sort
edEi
genv
alue
[dB
]
Fraction of Eigenvalues
Sort
edEi
genv
alue
[dB
]
Fraction of Eigenvalues
Eigenvalue Spectrum
Eigenvalue SpectrumEigenvalue Spectrum
4x4
16x168x8
HH† = a2GG†
MIT Lincoln LaboratorymimoTechSem-11
bliss
Continuous Uninformed Transmitter Capacity Approximation
• Replace discrete matrix with continuous approximation
• Shape of random finite matrixeigenvalue distributions converge quickly
• Use number of receivers normalized capacity
CUT = log2 I +Po
nTx
H H†
= log2 I + a2 PoGG†
nTx
= log2 I + a2Po μmm=1
nRx
∑
≈ nRx Φ a2 Po ;r( )
Φ x ; r( ) ≡ dμ fr (μ) log2 (1 + μ0
∞
∫ x)
Total TransmitPower
(noise-normalized)ChannelMatrix
EigenvalueDensity
continuous approximation
D. W. Bliss, K. W. Forsythe, and A. F. Yegulalp, “MIMO communication capacity using infinite dimension random matrix eigenvalue distributions,” Asilomar Conference, Nov., 2001
MIT Lincoln LaboratorymimoTechSem-12
bliss
Continuous Informed Transmitter Capacity Approximation
CIT = maxtr{P}= Po
log2 I + HP H†
= log2
Po + 1a2 M
tr Λ−1{ }L
a2 M Λ
≈ g M log2
a2 Po + dμ' f1(μ' ) 1μ'μcut
∞
∫g
⎡
⎣
⎢ ⎢ ⎢ ⎢ ⎢
⎤
⎦
⎥ ⎥ ⎥ ⎥ ⎥
+ M dμ f1(μ) log2 (μ)μcut
∞
∫
g = LM
≈ dμ f1(μ)μ cut
∞
∫ , μ cut =
dμ' f1(μ' )μ cut
∞
∫
a2Po + dμ f1(μ) 1μμ cut
∞
∫
EmployedEigenvalues
SmallestEigenvalueEmployed
Fraction of EigenvaluesEmployed
continuous approximation
• For this example assume that
• Employ top modes of contained in
M = nRx = nTx , r =1
L HH†
Λ
ImplicitlyDefined
MIT Lincoln LaboratorymimoTechSem-13
bliss
Large Random Matrix MIMO CapacityInformed vs. Uninformed Transmitter
Antenna Number Normalized MIMO Capacity
• Rich scattering environment
• M x M channel
• capacity using asymptotic eigenvalue distribution
• Low-SNR informed capacity higher
Informed
Uninformed
Mean Single Antenna Link SNR
x 4
MIT Lincoln LaboratorymimoTechSem-14
bliss
MIMO Capacity In InterferenceLarge Dimension Limit (UT)
• Capacity in presence of interference
• Equivalent MIMO Interferer
• Interference type– Cooperative– Uncooperative
• Cooperative interferencecan be demodulated
• Uninformed transmitter
• Square
Antenna Number Normalized MIMO Capacity
Mean Single Antenna Link SNR
No Interference
Uncooperative
H
CooperativeInterference
MIT Lincoln LaboratorymimoTechSem-15
bliss
Topics
• Multiple-Input Multiple-Output (MIMO)– Introduction– Asymptotic results
• Survey of Communication Examples– Code-Division Multiple Access (CDMA)– Receiver performance in rich interference networks
• Wish List
MIT Lincoln LaboratorymimoTechSem-16
bliss
Multiple Access
• Multiple users trying to talk to the base station at once• Separate users
– Frequency (frequency division multiple access)– Time (time division multiple access)– Code (code division multiple access)
FrequencyFDMA
ω
|F{s(t)}|2
TimeTDMA
�
|s(t)|2
CodeCDMA
1 → 00001111
1 → 11001100
1 → 10101010
MIT Lincoln LaboratorymimoTechSem-17
bliss
CDMA Performance Bounds
• Verdu’s multiuser detection bounds• Asymptotic in number of users and
spreading• Eigenvalue distribution in code
cross-correlation
Verdu and Shamai, “Spectral Efficiency of CDMA with Random Spreading,” IEEE Trans. On IT, March, 1999
Users
Spreading
Loading
SNR
Spectral Efficiency Bounds
Spec
tral
Effi
cien
cy (b
/chi
p)
MMSE
Optimal
MIT Lincoln LaboratorymimoTechSem-18
bliss
Multiple-Antenna Performance Bounds
• Random distribution of interferers on a plane• Spatial interference mitigation• Propagation
– Power law attenuation,– Rayleigh fading
• High interference, low noise regime
S. Govindasamy, et al, “The performance of linear multiple-antenna receivers with interferers distributed on a plain,” IEEE Workshop on SWAWC, June, 2005
LinkLength
InterferenceDensity
Number ofReceive
Antennas
Cap
acity
(b/s
/Hz/
user
)
Performance Bound
α = 3α = 4
α= 5
TX RX
InterferenceDensity
MIT Lincoln LaboratorymimoTechSem-19
bliss
Topics
• Multiple-Input Multiple-Output (MIMO)– Introduction– Asymptotic results
• Survey of Communication Examples– Code-Division Multiple Access (CDMA)– Ultra-Wideband (UWB)– Receiver performance in rich interference networks
• Wish List
MIT Lincoln LaboratorymimoTechSem-20
bliss
Wish List
What I would like for the holidays
• Densities for functions of random complex matrices – Finite or infinite
• Ratios, Products, Sums, Differences, Det’s, Traces– Signal in the mean– Colored matrices
• Make it easy for me…
MIT Lincoln LaboratorymimoTechSem-21
bliss
Backup Slides
MIT Lincoln LaboratorymimoTechSem-22
bliss
The Channel MatrixA Toy Model
• Toy MIMO channel model– 2x2 – line of sight
• Resolving individual antennas increases eigenvalue
• MIMO systems in real environments employ scatterers to increase effective aperture
• Toy MIMO channel model– 2x2 – line of sight
• Resolving individual antennas increases eigenvalue
• MIMO systems in real environments employ scatterers to increase effective aperture Channel Matrix,
Eigenvalues of Channel Matrix2x2 MIMO System
Vary
Ape
rtur
e
b =
2π
arccosh 1
†h 2
h 1
h 2
H = h 1 h 2( )= 2 a v 1
v 2( )Unit norm
steering vector
MIT Lincoln LaboratorymimoTechSem-23
bliss
Beamforming
• Phase multiple antenna elements to control direction of transmitted or received power
• Narrowband model - can’t resolve antennas,
wavefronts
k
x 1
x 2
x 3
x n
SteeringVector
c (τn − τ1)
MIT Lincoln LaboratorymimoTechSem-24
bliss
Multi-Antenna ProcessingSimple Example
How it works:• Adapt antenna array beam pattern• Distinct beam pattern for each user• Done simultaneously for all users• Spatial equivalent to adaptive filtering
in frequency domain
How it works:• Adapt antenna array beam pattern• Distinct beam pattern for each user• Done simultaneously for all users• Spatial equivalent to adaptive filtering
in frequency domain
Adaptive Spatial Beamforming User of
Interest
InterferingUser
Null inBeam Pattern
c2
c1
Σ
MIT Lincoln LaboratorymimoTechSem-25
bliss
Beamforming Pattern(Receive or Transmit)
• Phase multiple antenna elements to control direction of transmitted or received power
• Narrowband model• n = 10
v † k (θ = 0);
x 1 …
x n( ) v k (θ );
x 1 …
x n( )
2
SteeringVector
θ (degrees)
Rel
ativ
e Po
wer
MIT Lincoln LaboratorymimoTechSem-26
bliss
Minimize Mean Square Error Example
• 10 antenna array• 4 interfering users (10,-20,30,-40 degrees)• Signal of interest at 0 degrees
f †
v k (θ);
x 1…
x n( )
2
θ (degrees)
Rel
ativ
e Po
wer
InterferingUsers
SpatialFilter
MIT Lincoln LaboratorymimoTechSem-27
bliss
Minimize Mean Square Error ExampleMoving Interferer
• 10 antenna array• 1 interfering users • Signal of interest at 0 degrees
f †
v k (θ);
x 1…
x n( )
2
Rel
ativ
e Po
wer
θ (degrees)
InterferingUser
QuickTime™ and a Animation decompressor are needed to see this picture.
MIT Lincoln LaboratorymimoTechSem-28
bliss
Channel Complexity
Scatterers• Scatterers act as virtual elements in antenna array
• Channel matrix complexity increases with effective virtual aperture
• Virtual scattering array resolves antenna of other antenna array
VirtualArray
MIT Lincoln LaboratorymimoTechSem-29
bliss
The Rich Scattering ChannelLarge Dimension Limit
Eigenvalue Density,H H†
• Eigenvalue probability density of
• Entries in are drawn from unit norm complex Gaussian
• Eigenvalues probability density of
• PDF dependent upon,
fr=1(x) dx =x (1− x)
Rel
ativ
e Ei
genv
alue
(dB
)
Eigenvalue Spectrum
Fraction of Eigenvalues
π xdx
HH† = a2GG†
(1/ nTx) GG†
G
r = n
PeakNormalized
Rx / nTx
Prob
abili
ty D
ensi
ty
Eigenvalue [dB]
MIT Lincoln LaboratorymimoTechSem-30
bliss
MIMO Capacity Bound
• MIMO bound follows different theoretical limit• Divide total energy amongst transmitters
avoiding compressive regime of SISO Shannon limit
Spectral Efficiency(b/s/Hz)
E b/N
0(d
B)
Shannon Limit
SISO
4x4 MIMO
8x8 MIMO TransmitPowerMatrixChannel
Matrix
Determinant
ChannelLoss
BandwidthNormalized
Capacity Transmit Power(noise normalized)
Assuming FlatMIMO Channel
Assuming FlatMIMO Channel
MIT Lincoln LaboratorymimoTechSem-31
bliss
Uncooperative External Interference Uninformed Transmitter Redux
˜ C UT ,Un = log2 I + Po
nTx
˜ H ˜ H †
= log2 R +Po
nTx
H H† − log2 R
= log2 I +Pint
nint
JJ† +Po
nTx
HH† − log2 I +Pint
nint
JJ†
a2Po
nTx
=aint
2 Pint
nint
˜ C UT ,Un,Eq = log2 I +a2Po
nTx
AA† − log2 I +a2 Po
nTx
Γ Γ† ; A = G Γ( )
Assume same equal average receive power per antenna
Interference plus noise can written in the form
R = I +
Pint
nint
JJ† = I +aint
2 Pint
nint
Γ Γ†
Re-Express bound for uncooperative interference
MIT Lincoln LaboratorymimoTechSem-32
bliss
Uncooperative External Interference Continuous Uninformed Transmitter
˜ C UT ,Un,Eq = log2 I +a2Po
nTx
AA† − log2 I +a2 Po
nTx
Γ Γ†
≈ nRx Φ a2PonTx + nint
nTx
;nRx
nTx + nint
⎛ ⎝ ⎜
⎞ ⎠ ⎟ − Φ a2Po
nint
nTx
;nRx
nint
⎛ ⎝ ⎜
⎞ ⎠ ⎟
⎧ ⎨ ⎪
⎩ ⎪
⎫ ⎬ ⎪
⎭ ⎪
Continuous approximation
• Continuous form can be expressed as the difference of two UT bounds
• Assumptions– Interference has same channel statistics– Interference has same power per transmit antenna
MIT Lincoln LaboratorymimoTechSem-33
bliss
ReceiveArray
Cooperative InterferenceOther MIMO Systems
• Cooperative interference affects performance less• Receiver jointly demodulates signals from both interfering MIMO
system and system of interest– Discards information from interfering MIMO system
• Cooperative interference affects performance less• Receiver jointly demodulates signals from both interfering MIMO
system and system of interest– Discards information from interfering MIMO system
TransmitArray
OtherReceive
Array
InterferingTransmit
Array
Interfering MIMO System
MIT Lincoln LaboratorymimoTechSem-34
bliss
Cooperative External Interference Continuous Uninformed Transmitter
˜ C UT ,Co = γ log2 I +Po
nTx
H H† +Pint
nint
J J†
˜ C UT ,Co ,Eq =nTx
nTx + nint
log2 I +a2 Po
nTx
G Γ( ) G Γ( )† ;a2Po
nTx
=aint
2 Pint
nint
≈ nRx nTx
nTx + nint
Φ a2 PonTx + nint
nTx
; nRx
nTx + nint
⎛ ⎝ ⎜
⎞ ⎠ ⎟
Continuous approximation
• Interfering and intended MIMO systems allowed to coordinate
• Assumptions– Interference has same channel statistics– Interference has same power per transmit antenna
InterferingChannel Matrix
Fraction of data associatedwith system of interest