Slides

MIMO Fundamentals and Signal Processing Course
Erik G. Larsson
Linköping University (LiU), Sweden
Dept. of Electrical Engineering (ISY)
Division of Communication Systems
www.commsys.isy.liu.se
slides version: September 25, 2009

Linköping University, ISY, Communication Systems, E. G. Larsson MIMO Fundamentals and Signal Processing
Objectives and Intended Audience
➮ This short course will give an introduction to the basic principles and
signal processing for MIMO wireless links.
➮ Intended audience are graduate students and industry researchers
➮ Prerequisites:
General mathematical maturity.
Solid knowledge of linear algebra and probability theory.
Good understanding of digital and wireless communications.
Some basic coding/information theory.
1
Organization
➮ 4 lectures:
Le 1: MIMO fundamentals
Le 2: Low-complexity MIMO
Le 3: MIMO receivers
Le 4: MIMO in 3G-LTE (by P. Frenger, Ericsson Research)
➮ Reading, in addition to course notes:

➠ Le 1: D. Tse & P. Viswanath, Fundamentals of Wireless
Communications, Cambridge Univ. Press 2005, Chs. 7–8
➠ Le 2: E. G. Larsson & P. Stoica, Space-time block coding for wireless
communications, Cambridge Univ. Press 2003, Chs. 2, 4–8
➠ Le 3: E. G. Larsson, ”MIMO detection methods: How they work”,
IEEE SP Magazine, pp. 91–95, May 2009
➮ Examination (for Ph.D. students): TBD

2
Le 1: MIMO fundamentals
3
Basic channel model
➮ Flat fading; linear, time-invariant channel

TX RX
1 1
nt nr
➮ Complex data {x1, . . . , xnt } are transmitted via the nt antennas

nt
X
➮ Received data: ym = hm,nxn + em
n=1 4
Basic MIMO Input-Output Relation
➮ Transmission model (single time interval):

       
y1 h1,1 · · · h1,nt x1 e1
 ..  =  .. ..   ..  +  .. 
ynr h · · · hnr ,nt xnt enr
| {z } | nr ,1 {z } | {z } | {z }
y (RX data) H (channel) x (TX data) e (noise)
➮ Transmission model (N time intervals):

y h x ··· x1,N
e
1,1 · · · y1,N 1,1 · · · h1,nt 1,1 1,1 ··· e1,N
.. .. .. .. .. .. .. ..
. . = . . . . + . .
ynr ,1 · · · ynr ,N h ··· h x ··· xn ,N enr ,1 · · · enr ,N
| {z } | nr ,1 {z nr ,nt } | nt,1 {z t } | {z }
Y H X∈X E
“code matrix”
➮ AWGN: ek,l ∼ N (0, N0), i.i.d.
5
MIMO Design Space
➮ Fast fading: codeword spans ∞ number of channel realizations

Channel can be time- or frequency-variant (e.g., MIMO-OFDM), or both
➮ Slow fading: codeword spans one channel realization
➮ For point-to-point MIMO, four basic cases (reality inbetween)

Fast fading Slow fading
Channel V-BLAST optimal V-BLAST suboptimal
unknown at TX no coding across antennas coding across antennas
D-BLAST optimal
Channel waterfilling waterfilling
known at TX over space& time over space
6
Slow fading, full CSI at TX
➮ Deterministic channel, known at TX and RX
y = Hx + e H ∈ Cnr ×nt
➮ Power constraint: E[||x||2] ≤ P
➮ e is white noise, N (0, N0I)
➮ SVD: H = U ΛV H , U H U = I, V HV = I
➮ Dimensions: U ∈ Cnr ×nr , Λ ∈ Rnr ×nt , V ∈ Cnt×nt
7
H
➮ Introduce transform: y = |U ΛV
{z } x + e
H
ỹ , U H y = Λ V H
x + H
| {z } | {z e}
U
,x̃ ẽ
➮ ẽ is white noise, N (0, N0I)
➮ x̃ is precoded TX data. Note: E[||x̃||2] = E[||x||2]
➮ Equivalent model with parallel channels: (n = rank(H))
ỹ1 = λ1x̃1 + ẽ1
...
ỹn = λnx̃n + ẽn
8
➮ No gain by coding across streams ➠ x̃k independent
➮ Operational meaning is multistream beamforming:
n
X
x = V x̃ = x̃k v k
k=1
➠ n indep. streams {x̃k } are transmitted over orthogonal beams {v k }

➠ We’ll assume that each stream x̃k has power Pk
➠ In the special case of n = 1 (rank one channel), then we have MRT:
H = λuv H ⇒ x = x̃v
9
Architecture
e1
x̃1 x1 y1 ỹ1
x̃2 x2 y2 ỹ2
H
V :,1:n H enr
U :,1:n
x̃n xnt ynr ỹn
➮ x̃k are independent streams with powers Pk and rates Rk
10
Optimal power allocation over n subchannels
➮ Each subchannel can offer the rate

Pk 2
Ck = log2 1 + λ (bits/cu)
N0 k
➮ The power constraint is

n
X
E[||x̃||2] = E[||x||2] = Pk ≤ P
k=1
➮ Optimal power allocation P1∗, ..., Pn∗
n n
X X Pk 2
Pmax Ck ⇔ Pmax log2 1+ λ
Pk , n
k=1 Pk ≤P Pk , n
k=1 Pk ≤P N0 k
k=1 k=1
11
Waterfilling solution
n
X Pk 2
➮ Problem: Pmax log2 1 + λk
Pk , n
k=1 Pk ≤P k=1 N0
➮ Solution: (need solve for µ)

n
X Pk∗λ2k
C= log2 1 +
N0
k=1
+ n
N0 X
Pk∗ = µ− 2 , Pk∗ = P
λk
k=1
➮ Special case: H = λuv H . Transmit one stream

P 2 P 2
C = log2 1 + λ = log2 1 + kHk 12
N0 N0
Waterfilling solution
P5∗ = 0
N0
λ2
k
P1∗ = 0
P3∗ = 0
P4∗
P2∗
... n k
1 2 3
13
Waterfilling at high SNR

P
➮ At high SNR, the water is deep, so Pk∗ ≈ n and
n n n
X X P λ2k P X λ2k
C= Ck ≈ log2 1+ ≈ n · log2 + log2
N0 n N0 n
k=1 k=1 k=1
➮ With SNR , P/N0, we have C ∼ n log2(SNR). We say that
The channel offers n = rank(H) degrees of freedom (DoF)
➮ Best capacity for well conditioned H (all λk ’s equal)
➮ Transmit n equipowered data streams spread on orthogonal beams
➮ Channel knowledge ∼unimportant

14
➮ No coding across streams
Waterfilling at high SNR

N0
λ2
k
000
0000
1111
111 0000
1111
000
111
0000
000
1111
111 0000
1111
000
111
0000
000
1111
111
000 0000
1111
000
111
0000
1111
111
000 0000
1111
000
111
0000
1111
111 0000
1111
000
111
1111111111111
0000000000000
0000
000
1111
111
0000
000
1111
111
0000000000000
1111111111111
0000
000
1111
111
000
0000
1111
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000000000000
1111111111111
0000
000
1111
111
0000
000
1111
111
0000000000000
1111111111111
0000
000
1111
111
000
0000
1111
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000000000000
1111111111111
0000
000
1111
111
0000
000
1111
111
0000
000
1111
111
000
0000
1111
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
0000
000
1111
111 0000
1111
000
111
0000
000
1111
111
000 0000
1111
000
111
0000
1111
111
000 0000
1111
000
111
0000
1111
111
0000
000
1111
111
0000
1111
000
111
0000
000
1111
111
0000
1111
000
111
000
0000
000
1111
111
0000
1111
111
000
000
111
0000
1111
111
0000
1111
000
111
0000
1111
000
111
0000
1111 0000
1111
000
111
000
111
0000
1111
000 0000
1111
000
111
0000
1111
111
000 0000
1111
000
111
111
1111
0000
000
111
0000
1111
000
111
0000
1111
000
111
0000
1111
000
111
000
1111
0000
000
111
0000
1111
111
000
0000
1111
000
111
0000
1111
111
0000
1111
000
111
0000
1111
000
111 0000
1111
000
111
0000
1111
000
111
0000
1111 0000
1111
000
111
000
111
1111
0000 0000
1111
000
111
000
111
0000
1111
000 0000
1111
000
111
111
0000
1111
000
111
0000
1111
000
111
000
111
0000
1111
000
111
000
000
111
0000
1111
111
0000
1111
000
111
000
111
000 0000
1111
000
111
111
000
111
0000
1111
000
111
000
111
0000
1111
000
111
000
111
0000
1111
000
111
0000
1111
000
111 P
000
111
000
111
0000
1111
000
111
0000
1111
000
111 Pk∗ ≈ n
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
... n k
1 2 3
15
Waterfilling at low SNR
➮ At low SNR, the water is shallow. Then

(
∗ P, k = argmax λ2k
Pk =
0, else

P 2 P 2
C = log2 1 + λmax ≈ λmax · log2(e)
N0 N0
➮ MIMO provides an array gain (power gain of λ2max) but no DoF gains.
➮ Channel rank does not matter, only power matters.
➮ Transmit one beam in the direction associated with largest λk
➮ Knowing H is very important! (to select what beam to use)

16
Waterfilling at low SNR
N0
λ2
k
P4∗ = P
µ
... n k
1 2 3
17
In practice
➮ Feedback of channel state information, requires quantization
➮ Potentially, by scheduling only “good” users, one may always operate at

high SNR
➮ Selection of modulation scheme

— e.g., M -QAM per subchannel, different M
— better channel, larger constellation
— should be done with outer code in mind
➮ Imperfect CSI ➠ cross-talk!
18
MIMO channel models
➮ MIMO channel modeling is a rich research field, with both empirical

(measurement) work and theoretical models.
➮ We will explore the main underlying physical phenomena of MIMO

propagation and how they connect to the DoF concept.
19
Line-of-sight SIMO channel
➮ Consider m-ULA at TX and RX, wavelength λ = c/fc, ant. spacing ∆
RX array
φ
TX
 
1
−j 2π
 e λ ∆ cos φ 
 
➮ Let u(φ) ,  .. 
 
2π
e−j(m−1) λ ∆ cos φ
➮ Signal from point source impinging on RX array (large TX-RX distance):

20
y = αu(φ) · s + e, (α ∈ C, dep. on distance)
Line-of-sight MIMO channel
φr
RX array
φt
TX array
H
➮ MIMO channel: y = α · u(φr ) · u(φt) ·x + e, n = rank(H) = 1
| {z } | {z }
| nr ×1 {z nt×1}
H
➮ The LoS-MIMO channel has rank one, so no DoF gain! 21

Lobes and resolvability
➮ Consider unit-power point sources at φ1, φ2 with sign. u(φ1), u(φ2).

How similar do these signatures look?
 
1 2  ∗ 1 H 
ks1u(φ1) − s2u(φ2)k = 2 − 2Res1 s2 · u (φ1)u(φ2)

m |m {z }

|·|=f (·)
1 H
where the lobe pattern f (cos(φ1) − cos(φ2)) , m |u (φ1)u(φ2)|.
➮ If f (·) < 1, then φ1, φ2 resolvable.

2π
➮ Resolvability criterion: | cos φ1 − cos φ2| ≥ A, A , (m − 1)∆
λ λ
➮ Grating lobes avoided if ∆ ≤ ⇒ A ≤ (m − 1) .
2 2 22
Two separated point sources and an m-receive-array
TX2
φ1
φ2
RX array
TX1
➮ Define H = [h1 h2],

hi = αiu(φi)
s
λmax(H) 1 + f (cos(φ1) − cos(φ2))
➮ Condition number κ(H) = =
λmin(H) 1 − f (cos(φ1) − cos(φ2))
2π
➮ κ(H) is small if f (· · · ) 6= 1 ⇔ φ1, φ2 resolvable ⇔ | cos φ1 −cos φ2| > A
23
MIMO with two plane scatters
RX array
TX array
➮ Here, H TX-RX = H AB-RX · H TX-AB
➮ We have rank(H TX-RX) = 2

only if rank(H AB-RX) = 2 and rank(H TX-AB) = 2
➮ For H to offer 2 DoF, A and B must be sufficiently separated in angle,

24
as seen both from TX and RX
Angular decomposition of MIMO channel
➮ For φ1, φ2, ..., φm define

1
U , √ [u(φ1) · · · u(φm)]
m
➮ Can show: With cos(φk ) = k/m, {u(φk )} forms ON-basis.

Then U H U = I.
➮ Let U r and U t be the U matrices associated with the TX and RX

arrays. Note that U H H
r U r = I and U t U t = I
➮ If ∆ = λ/2, then u(φi) correspond to simple, perfectly resolvable beams,

with a single mainlobe.
➮ We assume ∆ = λ/2 from now on.

The case of ∆ 6= λ/2 is more involved. 25
Angular decomposition, cont.
5 3
4
2
3 RX
1
1
26
TX
Angular decomposition, cont.
➮ Now define
Ha , U H
r HU t
⇒ Ha,(k,l) = uH (φrk )Hu(φtl)
" #
X
H r
u (φk ) αiu(φ′r H ′t t
i )u (φi ) u(φl )
i
| {z }
physical model
  
X  
 H r ′r H ′t t
= αi u (φk )u(φi )  ·
 u (φi )u(φl ) 

| {z } | {z }
i
=0 unless φ′r
i falls in lobe φr
k
′t t
=0 unless φi falls in lobe φl
➮ Elements of H a correspond to different propagation paths
➮ Ha,(k,l)=gain of ray going out in TX lobe l and arriving in RX lobe k

27
Angular decomposition, key points
➮ “Rich scattering” if all angular bins filled (H a has “no zeros”)
➮ “Diversity order”
= measure of error resilience
= number of propagation paths
= number of nonzero elements in H a
➮ Number of DoF
= rank(H)
= rank(H a)
➮ If H a,(k,l) are i.i.d. then Hk,l are i.i.d.

P
➮ With i.i.d. H a and many terms in , then we get i.i.d. Rayleigh fading.
28
Fast fading, no CSI at TX
➮ Each codeword spans ∞ number of H
➮ The V-BLAST architecture is optimal here

Note: Reminiscent of architecture for slow fading and full CSI@TX
➮ Transmit vectors
x = Qx̃
where x̃1, ..., x̃n are independent streams with powers Pk and rates Rk
29
V-BLAST architecture
e1
x̃1 x1 y1
x̃2 x2 y2
optimal
Q H enr receiver
x̃n xnt ynr
30
 
P1 0 ··· 0
 0 P2 ... .. 
➮ Transmit covariance: K x , cov(x) = Q QH
 .. . . . ... 0
0 ··· 0 Pn
➮ Achievable rate, for fixed H:

1
H
R = log2 I +
HK xH
N0
➮ Intuition: Volume of noise ball is |N0I|N .

Volume of signal ball is |HK xH H + NoI|N .
31
➮ Fast fading, coding over ∞ number of H matrices gives ergodic capacity

1
H
C = E log2 I +
HK xH
N0
➮ Choose Q and Pk to

1
H
max E log2 I +
HK xH
K x ,Tr(K x )≤P N 0
➮ Optimal K x depends on the statistics of H
32
➮ In i.i.d. Rayleigh fading, K ∗x = P

nt I (i.i.d. streams) and
n
P 1 X SNR 2
C = E log2 I + HH H = E log2 1 + λk
N 0 nt nt
k=1
where
n = rank(H) = min(nr , nt)
P
SNR ,
N0
{λk } are the singular values of H
➠ Antennas then transmit separate streams.

➠ Coding across antennas is unimportant.
33
Some special cases
➮ SISO: nt = nr = 1
2

C = E log2(1 + SNR|h| )
At high SNR, the loss is -0.83 bpcu relative to AWGN channel
➮ SIMO: nt = 1 (power gain relative to SISO)

" n
!#
X r
C = E log2 1 + SNR |hk |2
k=1
➮ MISO: nr = 1 (no power gain relative to SISO)

" n
!#
t
SNR X
C = E log2 1 + |hk |2
nt
k=1 34
Large arrays (infinite apertures)
➮ Large MISO (nt TX, 1 RX) becomes AWGN channel:

" n
!#
t
SNR X
C = E log2 1 + |hk |2 → log2(1 + SNR)
nt
k=1
➮ Large SIMO (1 TX, nr RX)

" nr
!#
X
C = E log2 1 + SNR |hk |2 ≈ log2(nr SNR) = log2(nr )+log2(SNR)
k=1
➮ Large square MIMO (nt TX, nr RX, nr = nt = n): Linear incr. with n:
Z 4 r ! !
1 1 1
C ≈n· log2(1 + t · SNR) − dt 35
π 0 t 4
Fast fading, no CSI at TX, high SNR
➮ Here
n
X SNR 2
C= E log2 1 + λ ≈ n log2(SNR) + const
nt k
k=1
➮ Both nr and nt must be large to provide DoF gain
➮ “Capacity grows as min(nr , nt)”
36
Fast fading, no CSI at TX, low SNR
➮ Here
n n
X SNR 2 SNR X
C= E log2 1 + λk ≈ log2(e) · · E[λ2k ]
nt nt
k=1 k=1
SNR
= log2(e) · · E[||H||2] = log2(e) · nr · SNR
nt | {z }
=nr nt
➮ Capacity independent of nt!
➮ No DoF gain. All what matters here is power
➮ Relative to SISO, a power gain of nr (array/beamforming gain)
➮ Multiple TX antennas do not help here

37
(but with CSI at TX, things are very different)
V-BLAST in practice
➮ Transmitter architecture “simple” but the receiver must separate the

streams ➠ major challenge
➮ Problems are conceptually similar to uplink MUD in CDMA and to

equalization for ISI channels
➮ Stream-by-stream receivers: Successive-interference-cancellation

➠ MMSE-SIC is theoretically optimal but suffers from error propagation
➠ Rate allocation necessary
➮ Iterative architectures
➠ Iteration between outer code and demodulator
➠ Demodulator design is major problem
38
➮ Receivers for MIMO to be discussed more in lecture 3
Fast fading, full CSI at TX
➮ The transmitter can do waterfilling over both space and time
➮ Parallel channels: ỹk [m] = λk [m]x̃k [m] + ẽk [m]

➠ Waterfilling over space (k) and time (m).
➠ Optimal powers Pk∗[m]
➠ Capacity
n ∗ 2

X P (λk )λk
C= E log2 1 +
N0
k=1
39
P
➮ High SNR: P ∗(λk ) ≈ n (equal power)
n
X SNR 2
C≈ E log2 1 + λ , n D.o.F.
n k
k=1
An SNR gain (compared to no CSI) of

nt nt nt
= = , if nt ≥ nr
n min(nt, nr ) nr
➮ Low SNR: Even larger gain, so here multiple antennas do help!
40
Slow fading, no CSI at TX

1 H
➮ Reliable communication for fixed H if log2 I + N0 HK xH > R
h i
H
= P log2 I + N10 HK xH < R

➮ Outage probability, for fixed R: Pout
➮ Optimal K x as function of H’s statistics:

1
K ∗x = argmin P log2 I + HK xH H < R
K x ,Tr K x ≤P N0
For H i.i.d. Rayleigh fading:

➠ K ∗x = P
nt I optimal at large SNR
➠ K ∗x = P
n′ diag{1, ..., 1, 0, ..., 0} at low SNR (n′ < nt)
41
➮ Notion of diversity: Pout behaves as SNR−d where d=diversity order
➮ Maximal diversity: d = nr nt
➮ To achieve diversity, we need coding across streams
➮ V-BLAST does not work here.

Each stream has diversity at most nr , while the channel offers nr nt
➮ Architectures for slow fading:

➠ Theoretically, D-BLAST is optimal
➠ Pragmatic approaches include STBC combined with FEC
42
Example: Outage probability at rate R = 2 bpcu
nt=1, nr=1 (SISO)

nt=2, nr=1
−1 nt=1, nr=2
10 nt=2, nr=2
−2
10
FER
−3
10
−4
10
−5 0 5 10 15 20 25 30 35 40
43
SNR
replacements
D-BLAST architecture
xB (1) xB (2)
xA(1) xA(2) xA(3)
➮ Decoding in steps:
1. Decode xA(1)
2. Decode xB (1), suppressing xA(2) via MMSE
3. Strip off xB (1), and decode xA(2)
➮ One codeword: x(i) = [xA(i) xB (i)]
➮ Requires appropriate rate allocation among xA(i), xB (i)
➮ In practice, error propagation and rate loss due to initialization

44
Le 2: Low-complexity MIMO
45
Antenna diversity basics
➮ Recall transmission model (single time interval):

      
y1 h1,1 · · · h1,nt x1 e1
 ..  =  .. ..   ..  +  .. 
ynr h · · · hnr ,nt xnt enr
| {z } | nr ,1 {z } | {z } | {z }
y (RX data) H (channel) x (TX data) e (noise)
➮ Transmission model (N time intervals):

y h x ··· x1,N
e
1,1 · · · y1,N 1,1 · · · h1,nt 1,1 1,1 ··· e1,N
.. .. .. .. .. .. .. ..
. . = . . . . + . .
y ··· y h ··· h x ··· xn ,N enr ,1 · · · enr ,N
| nr ,1 {z nr ,N } | nr ,1 {z nr ,nt } | nt,1 {z t } | {z }
Y H X∈X E
“code matrix”
46
Introduction and preliminaries
➮ Transmitting with low error probability at fixed rate requires N large.
➮ For practical systems, it is often of interest to design short space-time

blocks (small N ) with good error probability performance. Outer FEC
can then be used over these blocks.
➮ Throughout, we will assume Gaussian noise,
e ∼ N (0, N0I)
Usually, we assume i.i.d. Rayleigh fading,
Hi,j i.i.d. N (0, 1)
Sometimes, for SIMO/MISO, we take
h ∼ N (0, Q) 47
Receive diversity (nt = 1)
➮ Suppose s transmitted, and h known at RX.
➮ Receive: y = hs + e
➮ Detection of s via maximum-likelihood (in AWGN):
2

2
2 hH y
ky − hsk = ... = khk · s − ŝ + const., where ŝ ,
khk2
➮ MRC+scalar detection problem!
➮ Distribution of ŝ determines performance:

H
h y N0 khk2 2 2 P
ŝ , ∼ N s, , SNR|h = · E |s| = khk ·
khk2 khk2 N0 | {z } N0
|{z}
=P
SNR48
Diversity order (n × 1 fading vector h)

q
2
➮ P (e|h) = Q SNR · khk and h ∼ N (0, Q) (SNR up to a constant)
−1 n −1
SNR Y SNR
➮ Then P (e) = E[P (e|h)] ≤ I + Q = 1+ λk (Q)
2 2
k=1
− rank(Q)
SNR 1
➮ As SNR → ∞, P (e) ≤ · Qrank(Q)
2 λk (Q)
k=1
log P (e)
➮ Diversity order d , − = rank(Q)
log SNR
n
!1/n n
Y 1X 1 1 2
➮ Note that λk ≤ λk = Tr{Q} = E[khk ]
n n n
k=1 k=1
49
with eq. if λ1 = · · · = λn so Q ∝ I minimizes the bound on P (e)
Diversity order
P (e)
10−1 d=1
d=2
d=3
10−2
10−3
10−4
10−5
0 10 20 30 40 50
SNR 50
Transmit diversity, H known at transmitter
➮ Try transmit w · s where w is function of H! (as we did in Le 1)
➮ RX data is y = Hws + e and optimal decision minimizes the ML metric:
ky − Hwsk2 = ... = kHwk2 · |s − ŝ|2 + const.
wH H H y N0
where ŝ , ∼ N s,
kHwk2 kHwk2
➮ The SNR|H in ŝ is max for w=normalized dominant RSV of H

2 2 2
1 H 1 kHk
➮ Resulting SNR|H = N0 λmax(H H) ·E |s| ≥ N0 nt · E |s| .
| {z }
≥kHk2 /nt
➮ Diversity order: d = nr nt
h∗ khk2 2 51
➮ For nr = 1 take wopt = khk so SNR|h = N0 E |s| (same as for RX-d)
Transmit diversity, H unknown at transmitter
➮ From now on, TX does not know H!
➮ Consider Y = HX + E. Optimal receiver in AWGN (H known at RX):

2
max P (X|Y , H) ⇔ min kY − HXk
X X
q
kH(X 0 −X)k2
➮ Pairwise error probability P (X 0 → X|H) = Q 2N0
➮ Consider P (X 0 → X) = E[P (X 0 → X|H)]. For i.i.d. Rayleigh fading,

1 −nr
(X 0 − X)(X 0 − X)H

P (X 0 → X) ≤ I +
| 4N0 {z }
“ ”−d −d
∼ N1 ∼SNR
0
d=“diversity order”. Note: d ≤ nr nt and d = nr nt if X 0 − X full rank

52
Linear space-time block codes (STBC)
➮ STBC maps ns complex symbols onto nt × N matrix X:
{s1, . . . , sns } → X
➮ Linear STBC:
ns
X
X= (s̄nAn + is̃nB n)
n=1
where {An, B n} are fixed matrices
➮ Typically N small. Need N ≥ nt for max diversity (why?)
N
➮ Rate: R , bits/channel use
ns 53
STBC with a single symbol
➮ Transmit one symbol s during N time intervals, weighted by W :
X = W · s, Y = HX + E = HW s + E
➮ Average error probability in Rayleigh fading:

1 −nr nt
H −nr −2nr nt
P (s0 → s) ≤ |W W | |s − s0|
4N0
➮ What is the optimum W ? Try to maximize:
max |W W H |
W
H
2
s.t. Tr W W = kW k ≤ 1 (power constraint)
➮ Solution: W W H = 1
nt I, (antenna cycling). Diversity but rate 1/N ! 54
Alamouti scheme for nt = 2
∗
Time
√1 Time√2
s s2
➮ X = √12 1 . That is: Ant 1 s1/√ 2 s∗2 /√2
s2 −s∗1
Ant 2 s2 2 −s∗1 / 2

y1 h1s1 + h2s2 e
➮ RX data: = √1 + 1
y2 2 h1s∗2 − h2s∗1 e2

y1 √1
h1s1 + h2s2 e1 √1
h1 h2 s1 e1
➮ Consider ∗ = + ∗ = + ∗
y2 2 h∗1 s2 − h∗2 s1 e2 2 −h∗2 h∗1 s2 e2
➮ ML detector 2

1
y1 h1 h2 s1
min ∗ − √
−h∗2 h∗1 s2

s1 ,s2 y 2 2
| {z } | {z } |{z}
y ,G s
55
H
h1 −hT2 h1 h2 kh1 k2 +kh2 k2
➮ Observation: GH G = 12 H = I
h2 hT1 −h∗2 h∗1 2
2 2 GH y
➮ Hence min ky − Gsk ⇔ min kŝ − sk , ŝ = 2 2 2
kh1k + kh2k
➮ Distribution of ŝ:
!
H H
G y G (Gs + e) 2N0
ŝ = 2 2 2 =2 2 2 ∼ N s, 2I
kh1k + kh2k kh1k + kh2k kHk
2
kHk
➮ SNR|H = . For 2 × 1 system, 3 dB less than 1 × 2 with MRC
2N0
➮ Diversity order: 2nr
56
Overview of 2-antenna systems

Method SNR rate TX knows h1, h2
|h1|2 + |h2|2
1 TX, 2 RX, MRC 1 no
N0
|h1| + |h2|2
2
2 TX, 1 RX, BF 1 yes
N0
|h1| + |h2|2
2
2 TX, 1 RX, ant. cycl. 1/2 no
2N0
|h1|2 + |h2|2
2 TX, 1 RX, Alamouti 1 no
2N0
|h1|2 + |h2|2
2 TX, 1 RX, Ant. sel. ≥ 1 partly
2N0
➮ For antenna selection, note that

max |hn|2 1 |h1|2 + |h2|2
≥
N0 2 N0
57
2-antenna systems, cont

P (e)
1 RX, 1 TX
10−1 1 RX, 2 TX, Alamouti
1 RX, 2 TX, antenna selection
10−2 2 RX, 1 TX, MRC
or 1 TX, 2 RX, BF
10−3
10−4
10−5
0 10 20 30 40 50 58
SNR
Orthogonal STBC (OSTBC)
➮ Important special case of linear STBC:
ns
X ns
X 2
X= (s̄nAn+is̃nB n) for which XX H = |sn|2 · I = ksk · I
n=1 n=1
¯
Notation: (·)=real ˜
part, (·)=imaginary part
➮ This is equivalent to requiring for n = 1, . . . , ns, p = 1, . . . , ns
An AH
n = I, B B H
n n =I
An AH H
p = −Ap An , B nB H H
p = −B p B n , n 6= p
An B H
p = B p A H
n
59
Proof
➮ To prove ⇒), expand:

ns X
X ns
XX H = (s̄nAn + is̃nB n)(s̄pAp + is̃pB p)H
n=1 p=1
ns
X
= (s̄2nAnAH
n + s̃2
n n n)
B B H
n=1
ns
X ns
X
+ s̄ns̄p(AnAH
p + Ap A H
n) + s̃ns̃p(B nB H
p + B pB H
n)
n=1 p=1,p>n
ns X
X ns
+i s̃ns̄p(B nAH H
p − Ap B n )
n=1 p=1
➮ Proof of ⇐), see e.g., EL&PS book. 60

Some properties of OSTBC
➮ Manifests the intuition that unitary matrices are good

√
➮ Alamouti code is an OSTBC (up to 1/ 2 normalization)
➮ Pros
➠ Diversity of order nr nt
➠ Detection of {sn} is decoupled
➠ Converts space-time channel into ns AWGN channels
➠ Combination with outer coding is straightforward
➮ Cons
ns
➠ Rate loss for nt > 2, i.e., nt > 2 ⇒ R = <1
N
➠ Information loss except for when nt = 2, nr = 1
61
Diversity order of OSTBC
➮ Suppose {s0n}nn=1
s
are true symbols and {sn} are any other symbols.
Then ns
X
X − X0 = (s̄n − s̄0n)An + i(s̃n − s̃0n)B n
n=1
ns
X
⇒ (X − X 0)(X − X 0)H = |sn − s0n|2 · I
n=1
➮ Full rank ➠ full diversity for i.i.d. Gaussian channel
62
Derivation of decoupled detection
➮ Write the ML metric as

kY − HXk2
H
=kY k − 2ReTr Y HX + kHXk2
2
ns
X ns
X
H H
=kY k2 − 2 ReTr Y HAn s̄n + 2 Im Tr Y HB n s̃n
n=1 n=1
+ kHk2 · ksk2
ns
X
H H
= − 2ReTr Y HAn s̄n + 2 Im Tr Y HB n s̃n + |sn|2kHk2
n=1
+ const.
ns
H H 2
X ReTr Y HAn − i Im Tr Y HB n
=kHk2 ·

sn − + const.
n=1
kHk2 63
Decoupled detection, again
➮ Linearity: Y = HX + E ⇔ y = F s′ + e, s′ , [s̄T s̃T ]T
➮ Theorem: X is an OSTBC if and only if

H
Re F F = kHk2 · I ∀ H
➮ ML metric:
′ 2 2
H ′
′T H ′

ky − F s k = kyk − 2Re y F s + Re s F F s
H ′
= kyk − 2Re y F s + kHk2 · ks′k2
2
= kHk2 · ks′ − ŝ′k2 + const.

H !
′ ˆs̄ Re F y s̄ N0/2
where ŝ , ˆ = ∼N , I
s̃ kHk2 s̃ kHk2
64
➮ F is a “Spatial/temporal (code) matched filter”
Interpretation of decoupled detection
➮ Space-time channel decouples into ns AWGN channels

AWGN
signal 1
AWGN
signal 2
AWGN
signal ns
(a) nt × nr space-time channel (b) ns independent AWGN channels
2
N kHk P
➮ SNR per subchannel: SNR|H = · · 65
ns nt N0
Example: Alamouti’s code is an OSTBC
➮ Consider the Alamouti code (re-normalized):

∗

s s2
X= 1 ∗ , XX H = (|s1|2 + |s2|2)I
s2 −s1
➮ Identification of An and B n gives

1 0 0 1
A1 = A2 =
0 −1 1 0

1 0 0 −1
B1 = B2 =
0 1 1 0
66
Examples of OSTBC
➮ Best known OSTBC for nt = 3, N = 4, ns = 3:

 
s1 0 s2 −s3
X= 0 s1 s∗3 s∗2 
−s∗2 −s3 s∗1 0
Code rate: 3/4 bpcu
➮ For nt = 4, N = 4, ns = 3:
 
s1 0 s2 −s3
 0 s1 s∗3 s∗2 
X= ∗
 ∗

−s2 −s3 s1 0 
s∗3 −s2 0 s∗1
Rate is 3/4 bpcu.

67
Summary of OSTBC Relations
ns
X
XX H = |sn|2 · I = ksk2 · I
n=1
An AHn = I , B nB H
n = I
An AHp = −ApAH n , B nB H
p = −B pB H
n, n 6= p
An B H
p
H
= B p An
m
H

Re F F = kHk2 · I where F is such that
s̄
vec(Y ) = F · + vec(E)
s̃ 68
Mutual Information Properties of OSTBC
➮ Average transmitted energy per antenna and time interval = 1/nt
➮ Channel mutual information, with i.i.d. streams of power 1/nt:

1 HH H
CMIMO(H) = log I +
nt N 0
➮ Mutual information of OSTBC coded system:

2

ns N kHk
COSTBC(H) = log 1 +
N ns nt N 0
➮ Theorem:
CMIMO ≥ COSTBC, equality only for nt = 2, nr = 1

69
Capacity comparison (1% outage)
1
10
Capacity [bits/sec/Hz]
0
10
Average capacity, 1 TX, 1 RX

Outage capacity, 1 TX, 1 RX
Average capacity, 2 TX, 2 RX
Average capacity, 2 TX, 2 RX − OSTBC
Outage capacity, 2 TX, 2 RX
−1
Outage capacity, 2 TX, 2 RX − OSTBC
10
0 5 10 15 20 25 30 35 40
SNR 70
Non-orthogonal linear STBC
➮ Also called linear dispersion codes
➮ Different approaches:
➠ Optimization of mutual information between the TX & RX:
1 h 2 H i

max EH log2 I + Re F F
2 N0
(no explicit guarantee for full diversity here)
➠ Quasi-orthogonal codes
➠ Codes based on linear constellation (complex-field) precoding
s′ = Φs
71
Example: A non-OSTBC
➮ Consider the following diagonal code, where |sn| = 1:

s 0
X= 1
0 s2
➮ Then

1 0 0 0 1 0 0 0
A1 = , A2 = , B1 = B2 =
0 0 0 1 0 0 0 1
➮ ML metric for symbol detection:

H H
kY − HXk =kY k − 2ReTr X H Y + kHk2
2 2
H H
= − 2Re [H Y ]1,1 · s1 − 2Re [H Y ]2,2 · s2 + const.
72
➮ Decoupled detection, but not OSTBC, and not full diversity
More examples of linear but not orthogonal STBC
➮ Alamouti code with forgotten conjugates

s s2
X= 1
s2 −s1
➮ “Spatial multiplexing” (R = nt, N = 1, ns = nt, d = nr ).

For nt = 2:
1 0
A1 = , A2 =
0 1

1 0
B1 = , B2 =
0 1
73
Linearly precoded STBC
➮ Transmit W X where W ∈ {W 1, . . . , W K }. Data model:
Y = HW X + E
➮ Fact: If rank{W k } = nt then same diversity order as but without W
➮ Consider correlated fading: R = E[hhH ] = RTt ⊗ Rr , h = vec(H)
➮ Error probability:
˛ 1 H H
˛−nr
EH [P (X 0 → X)] ≤ const. · ˛I + (X 0 − X)(X 0 − X) · W RtW ˛
˛ ˛
N0
ns H
➮ For OSTBC, (X 0 −X)(X 0 −X)H

∝ I. Hence, min I + W RtW
W N0
74
Ex. OSTBC with One-Bit Feedback for nt = 2
➮ One bit used to choose between

p
|a| p 0 1− |a|2 0
W1 = , W2 =
0 1 − |a|2 0 |a|
| {z } | {z }
if kh1 k>kh2 k if kh2 k>kh1 k
➮ Let Pc be the probability that the feedback bit is correct
➮ For Pc = 1 (reliable feedback), a = 1 is optimal ➠ antenna selection

W may be multiplied with fixed unitary matrix ➠ grid of beams
➮ For Pc < 1 (erroneous feedback),

2 P
c 1 − Pc
EH [P (X 0 → X)] ≤ 2 · |a|2 + 1 − |a|2 (for nr = 1)
SNR
75
−1
10
Unweighted OSTBC.
Optimal weighting (No feedback error).
Optimal weighting with feedback error.
Error tolerant weighting (No feedback error),
Error tolerant weighting with feedback error
−2
10
BER
−3
10
−4
10
−5
10
4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
SNR
76
MIMO with feedback - optimized transmission

n
x s y
Encoder W (I) s̃ U H Decoder
I I(H̃) H̃
➮ Here
➠ U depends on long-term feedback
➠ I depends on short-term (few bits) feedback
➮ State-of-the art designs rely on vector quantization techniques

77
Frequency-selective channels
➮ Maximum diversity order (with ML detection) will be nr ntL where

L=length of CIR
➮ Variety of techniques to achieve maximum diversity
➮ Most widely used transmission technique is MIMO-OFDM

➠ coding across multiple OFDM symbols
➠ coding across subcarriers within one OFDM symbol
➮ Basic model per subcarrier is
Y n = H nX n + E n
78
Le 3: MIMO receivers
79
Summary of MIMO receivers
➮ Optimal architectures (from Le 1):

➠ CSI@TX (any fading): linear processing, ỹ = U H y, separates streams
➠ no CSI@TX, fast fading: V-BLAST, optimal receiver is more involved
- linear receiver (channel inversion) is grossly suboptimal
- successive interference cancellation (SIC)
- using soft MIMO demodulator + decoder, possibly iterative
➠ no CSI@TX, slow fading: D-BLAST
➮ Architectures with STBC+outer FEC (from Le 2)

➠ With OSTBC, decoupled detection and thing are simple:
2 2

min kY − HXk ∼ min |s1 − ŝ1| + |s2 − ŝ2|
➠ With non-OSTBC, min kY − HXk does not decouple

- problem similar to for V-BLAST 80
Theoretically optimal V-BLAST receiver based on SIC

decoded stream 1
MMSE 1 decoder 1
y = Hs + e
decoded stream 2
MMSE 2 decoder 2
decoded stream 3
MMSE 3 decoder 3
decoded stream nt
(MMSE nt) decoder nt
➮ Optimality only for fast fading. Requires rate allocation on streams.
➮ Major drawback: Requires long codewords. Prone to error propagation.81

Theoretically optimal D-BLAST receiver based on SIC
xB (1) xB (2)
xA(1) xA(2) xA(3)

➮ One codeword split as x(i) = [xA(i) xB (i)], with rate allocation
➮ Decoding in steps:
1. Decode xA(1)
3. Strip off xB (1), and decode xA(2)
➮ Drawbacks:
➠ error propagation
➠ rate loss due to initialization
➠ requires long codewords
82
Receivers for linear STBC architectures
Training Data 1 Data 2

➮ Received block: Y = HX + E.
➮ X linear in {s1, ..., sns } so with appropriate F , the ML metric is

2
2 vec(Ȳ ) s̄
kY − HXk =
−F
vec(Ỹ ) s̃
2
➮ For OSTBC, F T F = kHk I so detection decouples
➮ Spatial multiplexing (V-BLAST) can be seen a degenerated special case

of this architecture, with  
s1
X =  .. 
snt 83
Demodulator+decoder architectures
y = Hs + e soft output P (bi|y)

MIMO demodulator channel decoder
a priori information P (bi)
➮ Demodulator computes P (bi|y) given a priori information P (bi)
➮ Decoder adds knowledge of what codewords are valid
➮ Added knowledge in decoder is fed back to demodulator as a priori
➮ Iteration until convergence (a few iterations, normally)

84
MIMO demodulation (hard)
➮ General transmission model, with Gi,j ∈ R
|{z} G · |{z}
y = |{z} e ,
s + |{z} sk ∈ S
m×1 m×n n×1 m×1
➮ Models V-BLAST architectures, and (non-O)STBC architectures
➮ Other applications: multiuser detection, ISI, crosstalk in cables, ...
➮ Typically, m ≥ n and G is full rank and has no structure.
85
The problem
➮ If e ∼ N (0, σI) then the problem is to detect s from y
2
minn ky − Gsk , y ∈ Rm , G ∈ Rm×n
s∈S
(
Q ∈ Rm×n is orthonormal (QT Q = I)
➮ Let G = QL where
L ∈ Rn×n is lower triangular
2 2
2
Then ky − Gsk = QQT (y − Gs) + (I − QQT )(y − Gs)

2 2
T T
= Q y − Ls + (I − QQ )y
2
so minn ky − Gsk ⇔ minn kỹ − Lsk where ỹ , QT y
s∈S s∈S 86
Some remarks
➮ Integer-constrained least-squares problem, known to be NP hard
➮ Brute force complexity O(|S n|)
➮ Typical dimension of problem: n ∼ 8−16, so |S| ∼ 2–8, |S n| ∼ 256–1014
➮ Needs be solved
➠ in real time
➠ once per received vector y
➠ in power-efficient hardware (beware of heavy matrix algebra)
➠ possibly fixed-point arithmetics
➠ preferably, in a parallel architecture
➮ In communications, we can accept a suboptimal algorithm that finds the

correct solution quickly, with high probability
87
Some remarks, cont.
➮ For G ∝ orthogonal (OSTBC), the problem is trivial.
➮ Our focus is on unstructured G
➮ If G has structure (e.g., Toeplitz) then use algorithm that exploits this
➮ Generally, slow fading (no time diversity) is the hard case
88
Zero-Forcing
➮ Let
s̃ , arg minn ky − Gsk = arg minn kỹ − Lsk = L−1ỹ
s∈R s∈R
E.g., Gaussian elimination: s̃1 = ỹ1/L1,1
s̃2 = (ỹ2 − s̃1L2,1)/L2,2

..
➮ Then project onto S: ŝk = [s̃k ] , arg min |sk − s̃k |

sk ∈S
➮ This works very poorly. Why? Note that
s̃ = s + L−1QT e = s + ẽ, where cov(ẽ)=σ · (LT L)−1
ZF neglects the correlation between the elements of ẽ 89

Decision tree view

min {f1(s1) + f2(s1, s2) + · · · + fn(s1, . . . , sn)}
{s1 ,...,sn }
sk ∈S !2
k
X
where fk (s1, ..., sk ) , ỹk − Lk,lsl
l=1
root node
s1 = −1 s1 = +1
f1(−1) = 1 f1(1) = 5
1 5
s2 = −1 s2 = +1 s2 = −1 s2 = +1
f2(−1, −1) = 2 f2(−1, 1) = 1 f2(1, −1) = 2 f2(1, 1) = 3
3 2 7 8
s3 = −1 s3 = +1 s3 = −1 s3 = +1 s3 = −1 s3 = +1 s3 = −1 s3 = +1
f3(· · · ) = 4 f3(· · · ) = 1 3 4 3 1 1 9
7 4 5 6 10 8 9 17 leaves
{−1, −1, −1} {−1, −1, 1} {−1, 1, −1} {−1, 1, 1} {1, −1, −1} {1, −1, 1} {1, 1, −1} {1, 1, 1} 90
Zero-Forcing with Decision Feedback (ZF-DF)
➮ Consider the following improvement

ỹ1
i) Detect s1 via: ŝ1 = = arg min f1(s1)
L1,1 s1∈S
ỹ2 − ŝ1L2,1
ii) Consider s1 known and set ŝ2 = = arg min f2(ŝ1, s2)
L2,2 s2 ∈S
iii) Continue for k = 3, ..., n:
" Pk−1 #
ỹk − l=1 Lk,lŝl
ŝk = = arg min fk (ŝ1, ..., ŝk−1, sk )
Lk,k sk ∈S
➮ This also works poorly. Why? Error propagation.

Incorrect decision on si ➠ most of the following sk wrong as well.
➮ Optimized detection order (start with the best) does not help much.
91
Zero-Forcing with Decision Feedback (ZF-DF)
r1
1 5
1 5
2 1 2 3
3 2 7 8
4 1 3 4 3 1 1 9
7 4 5 6 10 8 9 17
92
Sphere decoding (SD)
➮ Select a sphere radius, R. Then traverse the tree, but once encountering
a node with cumulative metric > R, do not follow it down
2
➮ Enumerates all leaf nodes which lie inside the sphere kỹ − Lsk ≤ R
➮ Improvements:
➠ Pruning: At each leaf, update R according to R := min(R, M )
➠ Improvements: optimal ordering of sk
➠ Branch enumeration
(e.g., sk = {−5, −3, −1, −1, 3, 5} vs. sk = {−1, 1, −3, 3, −5, 5})
➮ Known facts:
➠ The algorithm solves the problem, if allowed to finish
➠ Runtime is random and algorithm cannot be parallelized
➠ Under relevant circumstances, average runtime is O(2αn) for α > 0
93
SD, without pruning, R = 6
r2
r2a
1 5
1 5
2 1 2 3
3 2 7 8
4 1 3 4 3 1 1 9
7 4 5 6 10 8 9 17
94
SD, with pruning, R = ∞
r3
r3a
1 5
1 5
2 1 2 3
3 2 7 8
4 1 3 4 3 1 1 9
7 4 5 6 10 8 9 17
95
“Fixed complexity” sphere decoding (FCSD)
➮ Select a user parameter r, 0 ≤ r ≤ n
➮ For each node on layer r, consider {s1, ..., sr } fixed and solve
(∗) min {fr+1(s1, ..., sr+1) + · · · + fn(s1, ..., sn)}

{sr+1 ,...,sn }
sk ∈S
➮ Subproblem (*) solved using |S|r times
➮ Low-complexity approximation (e.g. ZF-DF) can be used.

Why? (*) is overdetermined (equivalent G is tall)
➮ Order can be optimized: start with the “worst”
➮ Fixed runtime, fully parallel structure

96
FCSD, r = 1
r4
r4a
1 5
1 5
2 1 2 3
3 2 7 8
4 1 3 4 3 1 1 9
7 4 5 6 10 8 9 17
97
Semidefinite relaxation (for sk ∈ {±1})

T T

s T s T L L −L ỹ
➮ Let s̄ , , S , s̄s̄ = s 1 , Ψ,
1 1 −ỹ T L 0
Then
2 2 2
kỹ − Lsk = s̄T Ψs̄ + kỹk = Trace{ΨS} + kỹk
so the problem is to
min Trace{ΨS}
diag{S}={1,...,1}
rank{S}=1
s̄n+1 =1
➮ SDR proceeds by relaxing rank{S} = 1 to S positive semidefinite
➮ Interior point methods used to find S
➮ s recovered, e.g., by taking dominant eigenvector and project onto S n 98

Lattice reduction
➮ Extend S n to lattice. For example, if S = {−3, −1, 1, 3},

then S̄ n = {. . . , −3, −1, 1, 3, . . .} × · · · × {. . . , −3, −1, 1, 3, . . .}.
➮ Decide on orthogonal integer matrix T ∈ Rn×n that maps S̄ n onto itself:
Tk,l ∈ Z, |T | = 1, and T s ∈ S̄ n ∀s ∈ S̄ n
➮ Find one such T for which LT ∝ I
➮ Then solve ŝ′ , arg min kỹ − (LT )s′k2, and set ŝ = T −1ŝ′
s′ ∈S̄ n
➮ Critical steps:
➠ Find suitable T (computationally costly, but amortize over many y)
➠ ŝ ∈ S̄ n, but ŝ ∈
/ S n in general, so clipping is necessary
99
MIMO demodulation (soft)
➮ Data model
G · |{z}
y = |{z}
|{z} e ,
s + |{z} sk = S(b1, ..., bp) ∈ S, |S| = 2p
m×1 m×n n×1 m×1
➮ Bits bi a priori indep. with

P (bi = 1)
L(bi) = log , i = 1, ..., np
P (bi = 0)
➮ Objective: Determine

P (bi = 1|y)
L(bi|y) = log
P (bi = 0|y)
100
Posterior bit probabilities
P ! P !
P (bi = 1|y) (a)
„ «
s:b (s)=1 P (s|y) (b) s:b (s)=1 p(y|s)P (s)
L(bi |y) = log = log P i = log P i
P (bi = 0|y) s:bi (s)=0 P (s|y) s:bi (s)=0 p(y|s)P (s)
“Q ”1
np
0P
(c) s:bi (s)=1 p(y|s) i′ =1
P (bi′ = bi′ (s))
= log P@ “Q ”A
np
s:bi (s)=0 p(y|s) i′ =1
P (bi′ = bi′ (s))
0P “Q ” 1
np
p(y|s) P (bi′ = bi′ (s)) · P (bi = 1)
B s:bi (s)=1 i′ =1,i′ 6=i
= log @ P
C
“Q ”
np
A
s:b (s)=0 p(y|s) ′ ′ P (b i′ = b i′ (s)) · P (b i = 0)
i i =1,i 6=i
0P “Q ”1
np
p(y|s) P (bi′ = bi′ (s))
B s:bi (s)=1 i′ =1,i′ 6=i
= log @ P ” A + L(bi )
C
“Q
np
s:bi (s)=0 p(y|s) i′ =1,i′ 6=i
P (bi′ = bi′ (s))
11

In Gaussian noise p(y|s) = exp − 2σ
(2πσ)m/2
ky − Gsk2 so
0P “ ”“Q ”1
1 2 np
exp − 2σ ky − Gsk P (bi′ = bi′ (s))
B s:bi (s)=1 i′ =1,i′ 6=i
L(bi |y) = log @ P ” A + L(bi )
C
“ ”“Q
1 2 np
s:b (s)=0 exp − 2σ ky − Gsk
i ′ ′ P (bi′ = bi′ (s))
i =1,i 6=i 101
➮ With a priori equiprobable bits

P 1 2
!
s:bi (s)=1 exp − 2σ ky − Gsk
L(bi|y) = log P 1

s:bi (s)=0 exp − 2σ ky − Gsk2
P
➮ can be relatively well approximated by its largest term
That gives problems of the type
min
n
ky − Gsk2
s∈S ,bi (s)=β
This is called “max-log” approximation

P
➮ If many candidates s are examined, then use these terms in
➠ List-decoding algorithm
102
Incorporating a priori probabilities (BPSK/dim case)
➮ Consider sk ∈ {±1}, and let
sk = 2bk − 1
1 1
γk , log (P (sk = −1)P (sk = 1)) = log (P (bk = 0)P (bk = 1))
2 2
P (sk = 1) P (bk = 1)
λk , log = log = L(bk )
P (sk = −1) P (bk = 0)
➮ The prior is linear in sk :

1
log(P (sk = s)) = [(1 + s) log(P (sk = 1)) + (1 − s) log(P (sk = −1))]
2
1 1
= γk + λk sk
2 2
103
➮ Write
 
P 1 2 1
Pn
 s:sk =1 exp − σ ky − Gsk + 2 i=1,i6=k (γi + λisi) 
L(sk |y) = log  P P +λk
1 2 1 n
s:sk =0 exp − σ ky − Gsk + 2 i=1,i6=k (γi + λi si )

T G
T
➮ Define ỹ , [y 1 · · · 1] and G̃ , where
Λk
σ σ σ σ

Λk , diag 4 λ1, · · · , 4 λk−1, 4 λk+1, . . . , 4 λn
➮ Then
 
P 1 2 γi Pn σλ2i
 s:sk =1 exp − σ kỹ − G̃sk + i=1,i6=k 16 + 2 
L(sk |y) = log  P 2 +λk
1 2
Pn σλ γ
s:sk =0 exp − σ kỹ − G̃sk + 16 + 2
i i
i=1,i6=k
➮ A priori information on sk ➠ “virtual antennas”

104
Example w. iter. decod. 4×4, r = 1/2-LDPC, 1000 bits
Practical method, r = 3, no iteration

Practical method, r = 3, 1 iteration
Frame-error-rate (FER)
Practical method, r = 3, 2 iterations
0.1
Brute-force, no iteration
Brute-force, 1 iteration
Brute-force, 2 iterations
0.01
−4 −3.5 −3 −2.5 −2
105
Normalized signal-to-noise-ratio (SNR) [dB]
Channel (H) estimation and associated receivers
➮ Very often pilots are used to form a channel estimate.

Consider
Y t = HX t + E t
➮ Estimate H via training:

➠ Maximum likelihood (in Gaussian noise):
Ĥ = argmin kY t − HX tk2 = Y tX H H −1
t (X t X t )
H
➠ Can show, estimate is the same also in colored noise (e.g. co-channel
interference in multiuser system)
➠ Can be somewhat improved by using MMSE estimation
106
➮ Estimate noise (co)variance:
1 ⊥ H
Λ̂ = Nt Y t Π XH
Y t , for colored noise
n t o
1 ⊥ H
N̂0 = Nt nr Tr Y t Π X H Y t , for white noise
t
➠ This estimate can be used to prewhiten the received signal to suppress

co-channel interference. E.g.
Ỹ = Λ−1/2Y
➠ At most nr − 1 rank-1 interferers can be suppressed.

The receive array has nr degrees of freedom.
At high SNR, Λ−1 becomes a projection matrix that projects out
interference.
107
More on training
➮ Training-based detector uses Ĥ, N̂0 and Λ̂ in the coherent detector
➮ Can be improved, e.g., cyclic detection
➮ How should training be designed?

Optimum pilots (in many respects) satisfy X tX H
t ∝I
➮ Inserting pilot-based estimate in LF is not optimal

Cf. the use of
p(X|Y , H) (coherent)
p(X|Y , H)|H:=Ĥ (training-based)
p(X|Y , Ĥ) (best possible given Ĥ)
p(X|Y , Y t) (best possible given training data)
Later, we will give p(X|Y , Ĥ) on explicit form 108

Example, joint detection and estimation schemes

Re−estimate
Initialize Detect Symbols Channel Convergence?
Yes
No
➮ Re-estimation step may make use of soft decoder output
109
Optimal training
➮ Recall: ML channel estimate Ĥ = Y tX H H −1

t (X t X t )
➮ Let h = vec(H), ĥ = vec(Ĥ). Can show:

E[ĥ] = h
“ ”
H H −T
Σ , E[(ĥ − h)(ĥ − h) ] = . . . = N0 (X tX t ) ⊗ I
˘ H −1 ¯
Tr {Σ} = nr Tr (X tX t ) N0
H

➮ Lemma: Suppose Tr XX ≤ nt. Then
H −1

Tr (XX ) ≥ nt with equality if and only if XX H = I
➮ Application of lemma ➠ optimal training block is (semi-)unitary:
X tX H
t ∝I 110
Channel estimation for frequency-selective channels
➮ MIMO channel as matrix-valued FIR filter:

L
X
H(z −1) = H lz −l
l=0
➮ L is the length of the channel, L = 0 for ISI-free channel
➮ Transfer function:
L
X
H(ω) = H le−iωl
l=0
➮ Transmission methods:
➠ Single-carrier (block-based)
➠ Multicarrier (e.g. OFDM)
111
Training for frequency selective channels
➮ Two basic approaches:

➠ Frequency-domain estimation (estimate H(ω))
➠ Time-domain estimation (estimate H l, then compute H(ω) via FT)
➮ Frequency-domain estimation of H(ω) is straightforward:

➠ Estimate in the frequency domain (via ML):
Ĥ(ω) = Y t(ω)X H H
t (ω)(X t (ω)X t (ω))
−1
➠ Problem: suboptimal because parameterizations is not parsimonious.

It does not exploit structure that
L
X
H(ω) = H le−iωl
l=0 112
Time-domain estimation of H(ω)
➮ Let x(n) be time-domain training, y(n) received (t-d) training and define
 T T

xt (0) ··· ··· ··· xt (−L)
 xTt (1) ... xTt (1 − L) 
Xt =  ... ... ... 
xTt (N − 1) ··· ··· xTt (0) ··· xTt (N − 1 − L)
 T  
H0 y Tt (0)
H =  .. , Yt= .. 
H TL y Tt (N − 1)
➮ Then the ML estimate of H(ω) is:

L
X
Ĥ = (X H
t X t )−1
X H
t Y t, Ĥ(ω) = Ĥ le−iωl
l=0
➮ Exploits structure (L unknowns but N equations) 113
Metrics with imperfect CSI (complex y, s, G, e)
➮ G not known perfectly ➠ replacing G with Ĝ in p(y|s, G) not optimal!
➮ Instead, need to work with p(y|s, Ĝ). Write
y = Gs + e y = (sT ⊗ I)h + e, h = vec(G), e = vec(E)

⇔
(
h ∼ N (0, ρI), e ∼ N (0, σI)
Suppose ksk2 = n and
ĥ = h + δ, δ ∼ N (0, ǫI)
T

y (nρ + σ)I ρ(s ⊗ I)
➮ Then ∼ N 0, so
ĥ ρ(s∗ ⊗ I) (ρ + ǫ)I
2!
1 1 1 ρ
p(y|ĥ, s) = n nǫ exp − nǫ y− Ĝs
π 1+ǫ/ρ + σ 1+ǫ/ρ + σ ρ + ǫ
114
Example: 4 × 4 slow Rayl. fading MIMO, QPSK, est. G
Practical detector, mismatched metric

Practical detector
0.1 Imperfect CSI
Practical detector, optimal metric

Brute-force
Brute-force, optimal metric
Perfect CSI Brute-force, mism.

0.01
−5 −4 −3 −2 −1 0 1 2 3
115
Example: 4 × 4 slow Rayl. fad., QPSK, outdat. G
Imperfect CSI
Practical detector, mismatched metric

0.1 Practical detector, optimal metric
Brute-force, optimal metric

Perfect CSI
Brute-force, mism.
0.01 Practical detector

Brute-force
−5 −4 −3 −2 −1 0 1 2 3
116
Last slide
117

Slides

Uploaded by

Copyright:

Available Formats

Slides

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Slides

Uploaded by

Copyright:

Available Formats

MIMO Fundamentals and Signal Processing Course

slides version: September 25, 2009

Objectives and Intended Audience

➮ Intended audience are graduate students and industry researchers

➮ Reading, in addition to course notes:

➮ Examination (for Ph.D. students): TBD

Basic channel model

➮ Flat fading; linear, time-invariant channel

➮ Complex data {x1, . . . , xnt } are transmitted via the nt antennas

Basic MIMO Input-Output Relation

➮ Transmission model (single time interval):

➮ Transmission model (N time intervals):

➮ AWGN: ek,l ∼ N (0, N0), i.i.d.

MIMO Design Space

➮ Fast fading: codeword spans ∞ number of channel realizations

➮ Slow fading: codeword spans one channel realization

➮ For point-to-point MIMO, four basic cases (reality inbetween)

Slow fading, full CSI at TX

➮ Deterministic channel, known at TX and RX

➮ Power constraint: E[||x||2] ≤ P

➮ e is white noise, N (0, N0I)

➮ Dimensions: U ∈ Cnr ×nr , Λ ∈ Rnr ×nt , V ∈ Cnt×nt

➮ ẽ is white noise, N (0, N0I)

➮ x̃ is precoded TX data. Note: E[||x̃||2] = E[||x||2]

➮ Equivalent model with parallel channels: (n = rank(H))

ỹ1 = λ1x̃1 + ẽ1

ỹn = λnx̃n + ẽn

➮ No gain by coding across streams ➠ x̃k independent

➮ Operational meaning is multistream beamforming:

➠ n indep. streams {x̃k } are transmitted over orthogonal beams {v k }

➮ x̃k are independent streams with powers Pk and rates Rk

Optimal power allocation over n subchannels

➮ Each subchannel can offer the rate

➮ The power constraint is

➮ Optimal power allocation P1∗, ..., Pn∗

➮ Solution: (need solve for µ)

➮ Special case: H = λuv H . Transmit one stream

Waterfilling at high SNR

➮ With SNR , P/N0, we have C ∼ n log2(SNR). We say that

The channel offers n = rank(H) degrees of freedom (DoF)

➮ Best capacity for well conditioned H (all λk ’s equal)

➮ Transmit n equipowered data streams spread on orthogonal beams

➮ Channel knowledge ∼unimportant

Waterfilling at high SNR

Waterfilling at low SNR

➮ At low SNR, the water is shallow. Then

➮ Channel rank does not matter, only power matters.

➮ Transmit one beam in the direction associated with largest λk

➮ Knowing H is very important! (to select what beam to use)

Waterfilling at low SNR

➮ Feedback of channel state information, requires quantization

➮ Potentially, by scheduling only “good” users, one may always operate at

➮ Selection of modulation scheme

➮ Imperfect CSI ➠ cross-talk!

MIMO channel models

➮ MIMO channel modeling is a rich research field, with both empirical

➮ We will explore the main underlying physical phenomena of MIMO

Line-of-sight SIMO channel

➮ Consider m-ULA at TX and RX, wavelength λ = c/fc, ant. spacing ∆

➮ Signal from point source impinging on RX array (large TX-RX distance):