MIMO I: Spatial Diversity

COS 463: Wireless Networks

Lecture 16
Kyle Jamieson

[Parts adapted from D. Halperin et al., T. Rappaport]

 What is MIMO, and why?
• Multiple-Input, Multiple-Output (MIMO) communications
– Sends and receive more than one signal on different transmit
and receive antennas

• We’ve already seen frequency, time, spatial multiplexing in 463:

– MIMO is a more powerful way to multiplex wireless medium

in space

– Transforms multipath propagation from an impediment to

an advantage

2
 Many Uses of MIMO
• At least three different ways to leverage space:

1. Spatial diversity: Send or receive redundant streams of

information in parallel along multiple spatial paths
– Increases reliability and range (unlikely that all paths will be
degraded simultaneously)

2. Spatial multiplexing: Send independent streams of

information in parallel along multiple spatial paths
– Increases rate, if we can avoid interference

3. Interference alignment: “Align” two streams of interference

at a remote receiver, resulting in the impact of just one
interference stream
 MIMO-OFDM

(5) QPSK Modulated subcarriers

f
Symbols
(6) Mapped onto Subcarriers as OFDM Symbol

aphical view offading:

• Multipath the OFDM
differentencoding
effects onprocess
differentfor the 18 Mbps rate (QPSK, 3/4)
frequencies
) are encoded by a rate-1/2 convolutional code (1) and then optionally punctured b
– OFDM: Orthogonal Frequency Domain Multiplexing
higher coding rates (here, 3/4) that send fewer redundant bits (2). The remaining
spread–theDifferent subcarriers
redundancy acrossare independent
subcarriers of each other
and protect against frequency-selective fa
into symbols (4) based on the modulation (QPSK encodes 2 bits per symbol), mod
ped• onto
Channel model forsubcarriers
the di↵erent OFDM: y = to h∙xform
+ w an OFDM symbol (6).
– A single complex number h captures the effect of the channel on data
h11 iny1 a=particular
h11 x subcarrier h11 y = ( h11+ h21) x h11 y1 = h11x 1
x x 1
h12
h21
•h For MIMO:
Rx
Think about each subcarrier, independent
Tx
of other subcarriersRx
12 Tx h21 Rx
x x2
h22
 Plan
1. Today: Diversity in Space
– Receive Diversity
– Transmit Diversity

2. Next time: Multiplexing in Space

3. Next time: Interference Alignment

5
 Path Diversity: Motivation
1. Multi-Antenna Access Points (APs), especially 802.11n,ac:

2. Multiple APs cooperating with each other:

Wired backhaul

3. Distributed Antenna systems, separating antenna from AP:

Antenna 2
Antenna 1

AP Coaxial / Fiber
backhaul
6
 Review: Fast Fading
• Typical outdoor multipath propagation environment, channel h

• On one link each subcarrier’s power level experiences Rayleigh fading:

! "

7
 Uncorrelated Rayleigh Fading
• Suppose two antennas, separated by distance d12

• Channels from each to a distant third antenna (h13, h23) can be uncorrelated
– Fading happens at different times with no bias for a simultaneous fade

!" # , !# #

8
 When is Fading Uncorrelated, and Why?

≫, !"#

• Channels from each antenna (h13, h23) to a third antenna

– Channels are uncorrelated when !"# > ≈ &. ()
– Channels correlated, fade together when !"# < ≈ )

• This correlation distance depends on the radio environment

around the pair of antennas
– Increases, e.g., atop cellular phone tower
9
 Plan
1. Today: Diversity in Space
– Receive Diversity
• Selection Diversity
• Maximal Ratio Combining
– Transmit Diversity

2. Next time: Multiplexing in Space

3. Next time: Interference Alignment

10
 Channel Model for Receive Diversity
• One transmit antenna sends a symbol to two receive antennas
– Receive diversity, or Single-Input, Multi-Output (SIMO)
p
y1=3ei3π/4x 9/ 13
h1
Receive antenna 1
p
rotate, scale by 3/ 13

xx n1 p
h2 1
p
4/ 13
Receive antenna 2
p
rotate, scale by 2/ 13
n expected 1
n2
y2=2e-iπ/6x

Figure 3: MRC operation on a sample channel. The channel gains are ~h = h3ei3⇡/4 , 2e i⇡/6 i, wit
re 1: A graphical view of the OFDM encoding process for the 18 Mbps rate (Q
• Each receive antenna gets own copy of transmitted signal via
noise ~n = hn1 , n2 i of expected power 1. The antennas have respective SNRs of 9 and 4. To implem
data bits
the(0) are encoded
receiver multiplies the byreceived
a rate-1/2
signalconvolutional
y = ~hx + ~n by the code
~ (1) and
unit vector then
~h⇤ /||~h||, optionally
where ~h⇤ denotes tp
– Different path
ain bits conjugate
for higher coding
of ~h. rates (here,
This operation 3/4)antenna’s
scales each that send fewer
signal by itsredundant
magnitude, and bitsrotates
(2). The
the
– Potentially different channel
aved (3)the
tosame
spread
rather into
phasethe
than symbols
redundancy
reference
at 0). The resulting
across
before adding
sumonhasthe
subcarriers
them. (For graphical
magnitude
p and
13, and
protect
clarity, we depict
expectedencodes
against frequency
the common
noise power
phas
1 because th
are grouped (4) based modulation (QPSK 2 bits per sy
normalized. Thus, by coherently combining received signals from di↵erent antennas, the MRC
finally mapped
the expectedonto SNR the di↵erent
of 13. subcarriers
In systems with OFDM, to MRCform an OFDM
is performed symbol
separately (6).subcarr
for each
 Selection Diversity
p
y1=3ei3π/4x 9/ 13
h1
Rx 1 p
rotate, scale by 3/ 13

xx n1 Select stronger p
h2 13
p Radio
4/ 13
Rx 2 p
rotate, scale by 2/ 13
n expected 1
n2
y2=2e-iπ/6x

Figure 3: MRC operation on a sample channel. The channel gains are h = h3e ~ i3⇡/4
, 2e i⇡/6
i, with Gaussia
A graphical view of thepower
OFDM encoding process forSNRs
the of 189 Mbps rate (QPSK, 3
noise• Two
she(0)
~
n = hn1 , nreceiveby
are encoded
2 i of antennas
expected 1.
a rate-1/2 share~ one receiving
The antennas have respective radio and 4. To implement MRC
receiver multiplies the received signalconvolutional
y = hx + ~n by the code
~ (1)
unit vector and then
~h⇤ /||~h||, optionally
where punctur
~h⇤ denotes the comple
for higher ~coding rates (here, 3/4) that send fewer redundant bits
conjugate of h. This operation scales each antenna’s signal by its magnitude, and rotates the signals int (2). The remain
to•same
he spreadphasethe
Chooses redundancy
the
reference
antenna across
before adding
with subcarriers
them.
stronger
(For graphical
p and protect
signal,
clarity,
sends
we depictagainstthat frequency-selectiv
to
the common
the radio
phase vertically
rather
ped intothan symbols
at 0). The resulting
(4) based sumon hasthe
magnitude
modulation 13, and expectedencodes
(QPSK noise power 1 because
2 bits the scaling i
per symbol),
normalized.
mapped – onto
he expected
Helps
Thus, byreliability
SNR theof 13. In systems
(both
coherently combining
di↵erent subcarriersunlikely
received bad)
with OFDM, to MRC
signals from di↵erent antennas, the MRC output ha
form an OFDM
is performed symbol
separately (6).subcarrier.
for each
– Wastes received signal from other antenna(s)
h11 This is yknown h h11
econd node. 1 = h11 as
x a 1x2 system. Real systems11 y =0( h11+ h21) x y1 =
er (dB)

may have more than two receive antennas, but xtwo will suf-
x1
h12
fice for our explanation. With this setup, each receive an- -5 h21
enna receives a copy of the transmitted signal modified by
 Selection Diversity:
Performance Improvement
• In general, might have M receive Probability (%) that Selected
Antenna’s SNR Exceeds Threshold γ
antennas (average SNR Γ)
– !" : SNR of the i th receive antenna

• Probability selected SNR is less

than some threshold γ:
– Pr !%, ⋯ , !( ≤ γ = Pr !" ≤ γ (

• One more “9” of reliability per

additional selection branch

Higher probability
(better) ↓

ß lower threshold SNR 13

 Leveraging All Receive Antennas
p
y1=3ei3π/4x 9/ 13
h1 Rx 1 p
rotate, scale by 3/ 13

xx n1 p
h2 13
p
4/ 13
Rx 2 p
rotate, scale by 2/ 13
n expected 1
n2
y2=2e-iπ/6x

• Want
3: MRCtooperation
justofadd on the two channel.
received Thesignals
processtogether
~ i3⇡/4 i⇡/6
Figure a sample channel gains h = h3e i, with Gaussia
A graphical view the OFDM encoding forarethe 18 Mbps
noise ~n = hn1 , n2 i of expected power 1. The antennas have respective SNRs of 9 and 4. To implement MRC
, 2e
rate (QPSK, 3
she(0) – But
are
receiver encoded if wethe
multiplies bydid a the
received signals
rate-1/2 would
signalconvolutional
y = ~hx
~ often
+ ~n by the code cancel ~h⇤ /||out
(1) and
unit vector then
~h||, optionally
where punctur
~h⇤ denotes the comple
for higher ~coding rates (here, 3/4) that send fewer redundant bits
conjugate of h. This operation scales each antenna’s signal by its magnitude, and rotates the signals int (2). The remain
tosame
he spread
phasethe redundancy
reference across
before adding subcarriers
them. (For graphical
p and protect
clarity, we depictagainst frequency-selectiv
the common phase vertically
• into
rather
ped Solution:
than at 0). The
symbols Receive
resulting
(4) based M on
sum radios,
hasthe align13,
magnitude
modulation signal
and phases,
expected
(QPSK noise then
encodes power 1add
2 bitsbecause the scaling i
per symbol),
normalized. Thus, by coherently combining received signals from di↵erent antennas, the MRC output ha
mapped – onto
he expectedRequires
SNR theof 13. Msystems
di↵erent
In receive withradios,
subcarriers
OFDM, to inform
MRC generalan OFDM
is performed symbol
separately (6).subcarrier.
for each

h 11 This is yknown h h11

econd node. 1 = h11 as
x a 1x2 system. Real systems11 y =0( h11+ h21) x y1 =
er (dB)

may have more than two receive antennas, but xtwo will suf-
x1
h12
fice for our explanation. With this setup, each receive an- -5 h21
enna receives a copy of the transmitted signal modified by
 How to Choose Weights?
p
y1=3ei3π/4x 9/ 13
h1 Rx 1 p
rotate, scale by 3/ 13

xx n1 p
h2 y 13
p
4/ 13
Rx 2 p
rotate, scale by 2/ 13
n expected 1
n2
y2=2e-iπ/6x

• Suppose phase of incoming gains areish!= h3e

th
OFDM signal on the i branch
Figure 3: MRC operation on a sample channel. The channel ~ i3⇡/4
, 2e i⇡/6
i, with Gaussia
A graphical view of the encoding process for the "18 Mbps
noise ~n = hn1 , n2 i of expected power 1. The antennas have respective SNRs of 9 and 4. To implement MRC
rate (QPSK, 3
she(0) are encoded
receiver multiplies the byreceived
a rate-1/2
signalconvolutional
y = ~hx + ~n by the code
~ (1) and
unit vector then
~h⇤ /||~h||, optionally
where punctur
~h⇤ denotes the comple
( +,-
for higher
• To ofalign
conjugate ~coding
h. This rates
{ yioperation (here,
} in phase, 3/4)
scaleslet that
theantenna’s
each send
combiner fewer
output
signal = ∑"&' )"and
by its# magnitude,
redundant * rotates
bits .(2). The remain
the signals int
tosame
he spread
phasethe redundancy
reference across
before adding them.subcarriers
(For graphical and protect
clarity, we depictagainst frequency-selectiv
the common phase vertically
rather
ped into
– How
than symbols to choose
at 0). The resulting amplitudes
(4) based sumon hasthe ai?
magnitude
modulation
p
13, and expectedencodes
(QPSK noise power 1 because
2 bits the scaling i
per symbol),
normalized. Thus, by coherently combining received signals from di↵erent antennas, the MRC output ha
mapped onto
he expected SNR the di↵erent
of 13. subcarriers
In systems with OFDM, to MRCform an OFDM
is performed symbol
separately (6).subcarrier.
for each
• Idea: Put more weight into branches with high SNR: Let /0 = 10
h–11 This
This isis1 calledx aMaximal
1x2 system. Ratio Combining
y =(MRC)
econd node. yknown
= h11 as h11
Real systems 0( h11+ h21) x h11 y1 =
er (dB)

may have more than two receive antennas, but xtwo will suf-
x1
h12
fice for our explanation. With this setup, each receive an- -5 h21
enna receives a copy of the transmitted signal modified by
 MRC: Performance Improvement
Probability that MRC’s SNR is Under Threshold γ0
0
10

−1
10

M=1
−2
Pout

10

M=2

M=3
−3
10 M=4

Lower probability M = 10

(better) ↓ M = 20
−4
10
−10 −5 0 5 10 15 20 25 30 35 40

10log'( )/+(
10log (γ/γ )
10 0
Lower threshold SNR à
Figure 7.5: Pout for MRC with i.i.d. Rayleigh fading.

• Two “9”s of reliability improvement between one (i.e., no MRC) and two
Rayleigh fading, where pγ (γ) is given by (7.16), it can be shown that [4, Chapter 6.3]
MRC branches Σ

$M M −1
& '#
1−Γ 1+Γ m
# $
M −1+m
! ∞ " %
Pb = Q( 2γ)pγΣ (γ)dγ = , (7.18) 16
0 2 m=0
m 2
d power 1. The antennas have respective SNRs of 9 and 4. To implement MRC,
y = ~hx + ~n by the unit vector ~h⇤ /||~h||, where ~h⇤ denotes the complex
eceived signal ~
tion scales each antenna’s signal by its magnitude, and rotates the signals into
fore adding them. (For graphical
p clarity, we depict the common phase vertically,

Selection Diversity, in Frequency

lting sum has magnitude 13, and expected noise power 1 because the scaling is
rently combining received signals from di↵erent antennas, the MRC output has
systems with OFDM, MRC is performed separately for each subcarrier.

a 1x2 system. Real systems 0

Normalized power (dB)

antennas, but two will suf-
this setup, each receive an- -5
nsmitted signal modified by
mitter and itself. The chan-
ers that represent both the -10
channel as well as the path-
e 3 for a graphical example). -15
A
nel gains based on training
C
ote that the gains di↵er for B and SEL
-20
ective fading) as well as for AB (MRC)
w is how to combine the two ABC (MRC)
e of them. -25
niques to show the extremes. -20 -10 0 10 20
e antenna with the strongest Subcarrier index
o receive the packet and ig-
Figure 4: Frequency-selective fading over testbed
• Antennas A and links:
C experience
s method SEL, for selection
at is done by 802.11a/g APs
different
the figure shows, fades5.2on
for an example different
GHz link, subcarriers
ps with reliability, because the received power measured on each subcarrier for
bad, but it wastes perfectly individual antennas and under SEL and MRC diver-
• Selection Combining (“SEL”) improves but certain subcarriers still experience fading
nnas that are not chosen. sity, normalized to the strongest subcarrier power.
d the signals from the two
s cannot be done by simply tudes of 3 and 2. With expected noise power 1, these gains
• MRC increases power and flattens nulls, leading to fewer bit errors
we will have just recreated correspond to SNRs of 9 and 4, given that a signal’s power is
Rather, the signals should the square of its magnitude. The MRC receiver scales each
n the same phase; then, the antenna’s signal by its magnitude, normalized to the total;
ne coherently. To do this, delays the signals to a common phasepreference; and then
F chain for each antenna to adds them. The result has magnitude 13, and the normal-
es the hardware complexity ized weighted sum of noise still has expected power 1. The
 MRC’s Capacity Increase
• MRC with M branches increases SNR
– Increased Shannon capacity

• Sub-linear (logarithmic) capacity increase in M:

– !"#$ = &' ( log 1 + . ( /01 bits/second/Hz
 Plan
1. Today: Diversity in Space
– Receive Diversity
– Transmit Diversity
• Channel reciprocity
• Transmit beamforming
• Introduction to Space-Time Coding: Alamouti’s Scheme

2. Next time: Multiplexing in Space

3. Next time: Interference Alignment

19
 An Aside: Radio Channels are “Reciprocal”
a2,d2,τ2

Transmitter T Receiver R
a1,d1,τ1

• Forward channel (T to R) is ℎ"# = %& '()*+,/. + %)'()*+0/.

• Switch T and R roles, changing nothing else:

– Reverse channel (R to T) is ℎ#" = %& '()*+,/. + %)'()*+0/. = ℎ"#
– The reverse radio channel is “reciprocal”

• Practical radio receiver circuitry induces differences between ℎ"# , ℎ#"

20
 Transmit Diversity: Motivation
• More space, power, processing capability available at the
transmitter?
– Yes, likely! e.g. Cellular base station, Wi-Fi AP transmitting
downlink traffic to mobile

• But, a (possible) requirement: May need to know the radio

channel at the transmitter before the transmission commences
– cf. receive diversity: channel from preamble reception

• Then, a tension: Separate transmit antennas for path diversity

– Antenna 1, Antenna 2, transmit radio non co-located
• Then, harder to move transmit signals, radio channel measurements
i.e. channel state information (CSI) between the three locations

21
 Transmit Beamforming: Motivation

“receiver”

• Suppose the transmitter knows the CSI to receivers

• Transmitters align their signals so that constructive

interference occurs at the single receive antenna
– Align before transmission, not after reception (receive
beamforming)

22
 Transmit Beamforming
• Leverage channel reciprocity, receive beamforming “in reverse”

• Send one data symbol x from two antennas

,'() p
y1%# &i3π/4x
=3e ℎ# = 9/ 13
p
%# & '(
Tx 1
rotate, scale by 3/ 13 )
x n1 p
Receive 13
p & '(+
4/ %13
#
=
rotate, 2 by 2/pℎ13*
Tx scale n expected 1
n2
-iπ/6
y2=2e,'( x
%* & +
MRC operation on a sample channel. The channel gains are ~h = h3ei3⇡/4 , 2e i⇡/6 i, with Gaussian
1 , n2 i of expected power 1. The antennas have respective SNRs of 9 and 4. To implement MRC,
• Multiply (pre-code) x by the complex conjugate of each channel
y = ~hx + ~n by the unit vector ~h⇤ /||~h||, where ~h⇤ denotes the complex
multiplies the received signal ~
f ~h. This operation scales each antenna’s signal by its magnitude, and rotates the signals into
ase reference before adding them. (For graphical
p clarity, we depict the common phase vertically,
at 0). The resulting sum has magnitude 13, and expected noise power 1 because the scaling is 23
 Plan
1. Today: Diversity in Space
– Receive Diversity
– Transmit Diversity
• Channel reciprocity
• Transmit beamforming
• Introduction to Space-Time Coding: Alamouti’s Scheme

2. Next time: Multiplexing in Space

3. Next time: Interference Alignment

24
 Alamouti Scheme: Motivation
• Suppose transmitters don’t know CSI information to receiver:
what to do?

1. Naïve beamforming (just send same signals)

– Signals would often cancel out

2. Repetition
– Each antenna takes turns transmitting same symbol
• Receiver combines coherently

– Use M symbol times

• Increases diversity (“SNR” term in Shannon capacity)
• Cuts Shannon rate by 1/M factor
25
 Alamouti Scheme
• Scope: A two-antenna transmit diversity system (M = 2)

• Sends two symbols, s1 and s2, in two symbol time periods:

Symbol Time Period 1 2

Antenna 1: Send !" Send −!$∗
Antenna 2: Send !$ Send !"∗

• Then, by superposition the receiver hears:

Symbol Time Period 1 2

Receiver hears: ℎ"!" + ℎ$!$ −ℎ"!$∗ + ℎ$!"∗

26
 Alamouti Receiver Processing
Symbol Time Period 1 2
Receiver hears: y[1] = ℎ'(' + ℎ*(* +[2] = −ℎ'(*∗ + ℎ*('∗
+∗ 2 = ℎ*∗(' − ℎ'∗(*

+[1] ℎ' ℎ* ('

= ∗
+∗[2] ℎ* −ℎ'∗ (*

• Rewrite into two equations in two unknowns (s1 and s2):

– (Receiver has CSI information)
(' ℎ'∗ ℎ* +[1]
(* ∝ ℎ*∗ −ℎ' +∗[2]

• But, what’s happening in terms of the physical wireless channel?

27
 Intuition for Alamouti Receiver Processing
• Start with the inverted channel matrix:
!" &∗" &( ,[1]
#$ ∝ ℎ$∗ −ℎ+ ,∗[2]

• Consider the computation for s1:

– Rotate ,[1] by −1"
– Rotate ,∗[2] by 1(
– Sum the result

28
 Alamouti: Impact of Phase Rotations
• Consider the computation for s1:
– Rotate ![1] by −&'
– Rotate !∗[2] by &*
– Sum the result

Symbol Time Period 1 2

Receiver hears: y[1] = ℎ./. + ℎ1/1 !∗ 2 = ℎ1∗/. − ℎ.∗/1
↓ ↓ ↓ ↓
Phase after rotation: 2 &* − &' 2 &* − &'

29
 Alamouti: Receiver-Side Picture
Symbol Time Period 1 2
Receiver hears: y[1] = ℎ'(' + ℎ*(* +∗ 2 = ℎ*∗(' − ℎ'∗(*
↓ ↓ ↓ ↓
Phase after rotation: / 01 − 02 / 01 − 02

• Receiver then sums all terms above:

Received signal: Q

∡4* − 4'
s2
s2 s1 s1 I

30
 Alamouti: Interpretation
!" ℎ"∗ ℎ# ([1]
!# ∝ ℎ#∗ −ℎ" (∗[2]

• Two new signal dimensions:

1. Multiply two received symbols by the top column of H

– Name this dimension ℎ"∗ ℎ# -
– s1 arrives along this dimension (only!)

2. Multiply two received symbols by the lower column of H

– Name this dimension ℎ#∗ −ℎ" -
– s2 arrives along this dimension (only!)
31
 Alamouti: Performance
• Two dimensions: ℎ"∗ ℎ$ % , ℎ$∗ −ℎ" %
Received Q
signal:

s1 s1 I
• Send half power on each antenna
|,|.-/|,|..
– For both symbols, '() = $0.

• Rate gain from enhanced SNR, and maintains one symbol per
symbol time

32
 Multi-Antenna Diversity: Summary
• Leverage path diversity
– Decrease probability of “falling into” to deep Rayleigh fade on
a single link

• Defined new “dimensions” of independent communication

channels based on space

33
 Thursday Topic:
MIMO II: Spatial Multiplexing

Friday Precept:
Exploiting Doppler

34