TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 2, Februa
ry 20
15, pp. 337 ~ 351
DOI: 10.115
9
1
/telkomni
ka.
v
13i2.700
3
337
Re
cei
v
ed Se
ptem
ber 5, 2014; Re
vi
sed
No
vem
ber 9,
2014; Accept
ed De
cem
b
e
r
6, 2014
A New Approa
ch to Linear Estimation Problem in Multi-
user Massive MIMO Systems
Muhammad
Ali Raza Anj
u
m
Arm
y
Pub
lic C
o
lle
ge of Ma
na
geme
n
t and Sc
ienc
es, Ra
w
a
lp
indi, Pak
i
stan
E-mail: ali.raz
a
.anjum@
g
ma
il.
com
A
b
st
r
a
ct
A novel
appr
o
a
ch for solvi
n
g
line
a
r esti
mati
on pro
b
l
e
m
in mu
lti-user mas
s
ive MIMO systems i
s
prop
osed. In th
is ap
proac
h, th
e difficu
lty of matrix inv
e
rsio
n i
s
attributed to t
he i
n
co
mp
lete defin
ition of
th
e
dot pro
duct. T
he g
ener
al d
e
finitio
n
of d
o
t product i
m
p
lies
that the col
u
mns of
chan
ne
l matrix
are a
l
w
a
y
s
orthog
on
al w
h
ereas, i
n
pr
actice, they
may
be n
o
t. If
the l
a
tter infor
m
ati
on ca
n b
e
i
n
c
o
rpor
ated
into
dot
prod
uct, then
the unk
now
ns
can be
direct
ly co
mpute
d
from
proj
ectio
n
s
w
i
thout invert
ing th
e cha
n
n
e
l
matrix. By
doi
ng so, th
e pr
o
pose
d
meth
od
is ab
le to
ach
i
eve
an
exact
soluti
on w
i
th a
25% r
e
d
u
ctio
n
i
n
computati
o
n
a
l
compl
e
xity as
compar
ed to t
he QR
meth
o
d
. Propos
ed
meth
od
is sta
b
le, offers a
n
extra
flexibi
lity of co
mp
utin
g any si
ngl
e unk
now
n, and ca
n be i
m
ple
m
ented i
n
ju
st tw
elve lin
es of code.
Ke
y
w
ords
:
est
i
matio
n
, large,
mass
ive, MIMO, multi-user
Copy
right
©
2015 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Multiple Input Multiple Output (MIMO)
sy
stem
s in
co
rpo
r
ate m
u
ltiple ante
nna
s at the
transmitter a
nd the
re
ceiv
er to
improve
the d
a
ta ra
te
of wi
rele
ss
communi
catio
n
sy
stem
s [1,
2].
Ho
wever, the
ever growi
n
g
desi
r
e for th
e increa
se
d d
a
ta rate
s can
hardly be
sat
i
ated and th
e
r
e
is a
dem
and
to increa
se
th
e data
rate
s
even furth
e
r.
As a
re
sult, t
he resea
r
che
r
s
are aimi
ng
to
achi
eve the
asymptotic li
mits in
cap
a
c
ity. Rec
ently, it has b
een
demo
n
strate
d that very h
i
gh
cap
a
citie
s
are a
c
hievabl
e
on fo
rward
and
reve
rs
e
links a
s
th
e
numbe
r
of transmit
anten
nas
approa
che
s
i
n
finity [3]. Such
system
s in whi
c
h a
relatively la
rge n
u
mbe
r
of base stati
o
n
antenn
as
serve a large nu
mber of u
s
ers are kn
own as
Mas
s
i
ve MI
MO s
y
s
t
ems
(M-MIMO) [4].
M-MIMO
systems provide
variou
s adv
antage
s
over the tradition
al MIMO sy
stem
s.
These sy
ste
m
s promi
s
e
highe
r data rates, high
er
antenn
a re
so
lution, lowe
r
error p
r
ob
abil
i
ties,
lowe
r therm
a
l
noise, an
d lo
wer tran
smit power
pe
r ant
enna. The
s
e
advantag
es
can be attribut
ed
to an ove
r
all
averagi
ng b
e
havior
of the
s
e
system
s.
But these
ad
vantage
s be
come eve
n
m
o
re
impre
s
sive i
n
the multi
-
use
r
scena
ri
o wh
ere
the
base
statio
n tran
smits
to seve
ral u
s
ers
simultan
eou
sl
y. Howeve
r,
multi-u
s
er
M-MIMO
systems in
cu
r costs
of their own: in
clud
ing
increa
se in
hard
w
a
r
e, in
cre
a
se in
po
wer con
s
um
ption, and i
n
cre
a
se in
ph
ysical
sp
aci
n
g
.
Eventually, th
e transceive
r
become
s
co
mplex as well
[3-6].
Tran
sceiver
compl
e
xity is currently an
issu
e of gre
a
t con
c
ern in
M-MIMO sy
stem
s. It
has be
en
sh
own
that fo
r
the poi
nt to
point
scen
ario, the d
e
cod
i
ng
compl
e
xity of the
re
cei
v
er
alone g
r
o
w
s
expone
ntially with increa
se in the
num
ber of tra
n
smit antenna
s [7]. In case
of a
multi-u
s
er
scenari
o
, the transmitte
r be
come
s
com
p
l
e
x as
well b
e
ca
use the t
r
an
smitter n
o
w
requi
re
s adv
anced codin
g
scheme
s
to manag
e si
multaneo
us t
r
an
smi
ssi
on
of informatio
n to
multiple u
s
e
r
s. Fo
rme
r
op
eration,
kn
o
w
n a
s
de
cod
i
ng, and
the
latter on
e,
calle
d p
r
e
c
o
d
ing,
gene
rally co
mpri
se the tra
n
sceiver o
p
e
r
ation [8].
For a larg
e n
u
mbe
r
of transmit anten
n
a
s,
linea
r de
cod
e
rs and p
r
ecode
rs have proven
almost
optim
al [9-1
1]. Th
ese
line
a
r
preco
ders/
de
co
ders
req
u
ire
the inversio
n
of a
potenti
a
lly
large
ch
ann
e
l
matrix. For example, a
pra
c
tica
l
preco
d
ing m
e
thod i
s
the Z
e
ro
-forcing
(ZF)
pre
c
odi
ng m
e
thod
whi
c
h
comp
utes the
pre
c
odi
ng m
a
trix by formi
ng the p
s
e
u
d
o
inverse of t
he
cha
nnel
matri
x
[9]. Similar to lin
ear p
r
e
c
odi
ng, a
si
mple lin
ea
r d
e
co
der i
s
the
MMSE d
e
co
der
whi
c
h
comp
u
t
es the
de
cod
i
ng matrix by
again fo
rm
in
g
the p
s
eud
oin
v
erse
of the
cha
nnel m
a
tri
x
[12]. If the di
mensi
o
n
s
of
a sy
stem a
r
e
very
la
rge, i
n
verting th
e
cha
nnel
matrix become
s
a
c
u
mbers
o
me tas
k
[4].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 337 – 351
338
Therefore, th
ere is a ne
e
d
to eliminate
the outrigh
t inversion.
Variou
s meth
ods for
approximate matrix
inversion
have
be
en propo
se
d
in literatu
r
e
:
includi
ng
Cayley-Hamil
ton
theore
m
met
hod, Neum
a
nn seri
es
ex
pan
sion me
t
hod,
Q
R
de
comp
ositio
n method, ran
dom
matrix metho
d
s, and met
hod
s ba
sed
on polynom
i
a
l and trun
cat
ed polynomi
a
l filters [13-2
1
].
These metho
d
s
still requi
re a lot of co
mputational
e
ffort and som
e
of them ha
ve proven to
be
even mo
re
co
mplex than
th
e direct i
n
version.
Ho
we
ve
r, it is impo
rta
n
t to note
tha
t
the challe
ng
e
is fun
dam
ent
ally different
i
n
at le
ast
thre
e po
ss
ible
wa
ys. To
begi
n
with, the
cha
nnel
matrix d
oes
not have
a
st
ructu
r
e. Se
co
ndly, inste
ad
of bein
g
dete
r
mini
stic, it i
s
ran
dom. Fi
n
a
lly, there i
s
no
guarantee
of spa
r
sity. So we can di
sp
ense with
th
e tradition
al method
s for
solving the li
nea
r
estimation p
r
oblem [22].
With this in view, we p
r
op
ose a n
o
vel method to so
lve such larg
e linear
syste
m
s. Th
e
prop
osed m
e
thod tre
a
ts th
e linea
r e
s
tim
a
tion p
r
obl
em
as
a
coo
r
din
a
te tran
sfo
r
m
a
tion p
r
obl
em.
By doing so, the method i
s
able to a
c
h
i
eve zero no
rms in erro
r a
nd re
sidu
e in
contra
st to the
state of
the
a
r
t iterative
me
thods for larg
e sy
stem
s –
LSQR,
LSMR, and
Kaczm
a
rz that
achie
v
e
much
highe
r
norm
s
- a
nd
a 25% red
u
cti
on in co
mput
ation co
mple
xity when co
mpared to th
e QR
method, the
leadin
g
prota
goni
st in exa
c
t met
hod
s [
23-2
5
]. Also
by employing
the pro
p
o
s
e
d
method, a p
a
r
ticula
r u
n
kno
w
n, which ma
y be the
only
informatio
n re
quire
d by a
si
ngle u
s
e
r
, ca
n
be
comp
uted
witho
u
t eval
uating th
e e
n
t
ire
solution
v
e
ctor
while no such flexibil
ity is availabl
e in
the traditio
nal
method
s. Fu
rtherm
o
re, th
e metho
d
is stable an
d lig
h
t
on comput
ation a
s
well a
s
it
only relie
s on
multiplicatio
n with sim
p
le
ran
k
-o
ne p
r
o
j
ection mat
r
ices for o
b
taini
ng the sol
u
tio
n
.
Finally, the method can be
prog
ram
m
ed
in just ten line
s
of cod
e
s.
Re
st of the
a
r
ticle i
s
org
a
n
ize
d
a
s
follo
ws
. Se
ction
2 begi
ns with
a sy
stem m
odel a
nd
define
s
the
u
nderlyin
g p
r
o
b
lem. Ba
si
c i
dea
behi
nd t
he p
r
o
posed
method
is ex
plore
d
in
Se
ction
3. The metho
d
is then explained in Sect
ion 4. Se
ction 5 contain
s
its compl
e
te p
r
oof. Section
6
provide
s
th
e
proof
of the
choi
ce
of
refl
ection
matri
c
es. A
co
mple
te algo
rithm f
o
r
step
by
step
solutio
n
of the line
a
r
est
i
mation p
r
obl
em a
c
cordi
n
g to the p
r
o
posed meth
o
d
is o
u
tlined
in
Section
7. Th
is
se
ction
also in
clud
es a
use
r-f
ri
endly
cod
e
fo
r the
algorith
m
. Co
mpari
s
o
n
of t
h
e
prop
osed me
thod with Q
R
deco
m
po
sition is carried
out in se
ction
8. The pro
c
e
ss of ge
ne
rati
on
of inverse ve
ctors i
s
al
so
discu
s
sed in
this
se
ction.
Comp
utation
a
l complexity of the p
r
o
p
o
s
ed
algorith
m
is
analyzed in
Section 9
which i
s
fo
llo
wed by its
sta
b
ility analysis in Section
10.
Section 11 co
nclu
de
s the article in whi
c
h
the co
mpa
r
ison of the pro
posed metho
d
is carrie
d out
with the state
of the art methods avail
abl
e in literature for so
lving la
rge syste
m
s.
2. Problem
A narrowban
d memo
ryless MIMO
cha
n
nel with
tran
smit and
re
ceive antenn
a
s
can
be model
ed a
s
a syste
m
of linear eq
uati
ons [13].
(1)
With,
⋮
⋮
⋮
(2)
is a large, random and no
n-spa
r
se matrix.
is the input vector.
is
the output vector. Zero
Forcing (ZF)
estimate
of the input vect
or is:
(3)
Re
cov
e
ry
of
requi
re
s the i
n
versi
on of a potentially large matrix
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New App
r
o
a
ch to Lin
ear
Estim
a
tion Problem
in Multi-user …
(Mu
ham
m
ad Ali
Ra
za Anjum
)
339
3. Basic Idea
Linea
r e
s
tim
a
tion proble
m
pre
s
e
n
ted
by Equation
(3)
ca
n be
viewed
as
a
coo
r
din
a
te
transfo
rmatio
n probl
em.
can be e
n
visag
ed as th
e vector that
undergoe
s the coo
r
di
nat
e
transfo
rmatio
n,
repre
s
ent
s its coo
r
din
a
t
es in the new coo
r
di
nate system, and
contai
ns
th
e
basi
s
vecto
r
s for the new coo
r
din
a
te sy
stem [
26]. The basi
s
vecto
r
s can be o
r
thogo
nal or n
on-
orthog
onal. For the first case, we are l
ead to t
he cla
ssi
c Le
ast Sq
uare
s
(LS) so
lution depi
cte
d
in
Figure 1.
Figure 1. Co
mputation of unkno
wn
s
in an ortho
gon
al
coordinate
system
In this
c
a
s
e
, the es
timates
,
,…,
are the proje
c
tion
s of
on the re
spe
c
tive
basi
s
vectors.
⋮
(4)
Whe
n
the basis vecto
r
s are not orthogo
nal,
is computed by forming the inverse of
matrix. This i
s
preci
s
ely what we de
si
re
to avoid. We would li
ke to comp
ute
from the direc
t
estimate
of th
e proje
c
tion
s. In o
r
de
r to
d
o
so, we fo
cu
s o
u
r
attentio
n on
Figu
re
2
for th
e
ca
se
of
non-orth
ogo
n
a
l basi
s
ve
ctors. A
s
ca
n
be ob
se
rved
from figure, e
m
ploying the
definition of
do
t
prod
uct in Eq
uation (4)
will
yield much l
onge
r
e
s
timates than th
e a
c
tual on
es.
Reason for th
e
s
e
inco
rrect e
s
ti
mates
ca
n b
e
attribute
d
to the p
r
e
s
en
ce of id
entity matrix in the
gene
ral d
e
fini
tion
of the dot pro
duct. We d
e
m
onst
r
ate thi
s
fact by
re-writing the first
term in Equat
ion (4
).
Figure 2. Co
mputation of first un
kno
w
n i
n
a non-orth
o
gonal
coo
r
din
a
te system
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 337 – 351
340
(5)
Identity matrix
depi
cts a set of orthogo
nal ba
sis ve
ctors. As they a
r
e not, we
ca
n
repla
c
e the
matrix in Equation (5
) by another m
a
trix, say
.
(6)
To s
o
lve for
,
we re
-a
rrang
e Equation (6
) a bit.
(7)
Or,
(8)
Equation (8) requi
re
s the
vectors
an
d
to have equal proje
c
tions o
n
a pla
ne
spa
nne
d by the colum
n
space of
. We observe fro
m
Figure 2 that
and
have equa
l
proje
c
tion
s o
n
a plane o
r
th
ogon
al to the se
con
d
ba
sis
vector
.
and
can b
e
proj
ected on
su
ch pla
ne b
y
a projectio
n
matrix of the
form:
(9)
We can subs
titute
in Equation (5
) to solv
e for
.
(10
)
Similarly
for
,
(11
)
will project
and
on a plane ortho
gona
l to
.
Using
Equation (1
0) and (11), we
can
dire
ct
ly
solv
e
f
o
r
without inverting the matrix
. Equation (10) and (11
)
also have a
con
notation t
hat
can be
solved indep
e
ndently of
.
4. Method
No
w we p
r
e
s
ent the method for solvin
g the gene
ral
proble
m
of the form
. We
begin by exp
andin
g
Equati
on (3
),
…
(12
)
Multiplying Equation (12
)
by a matrix
,
…
(13
)
In order for
to be ze
ro,
(14
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New App
r
o
a
ch to Lin
ear
Estim
a
tion Problem
in Multi-user …
(Mu
ham
m
ad Ali
Ra
za Anjum
)
341
Equation (13) becom
es,
…
(15
)
Multiplying Equation (15
)
by a matrix
,
…
(16
)
In order for
to be ze
ro,
(17
)
Equation (16) becom
es,
…
(18
)
Contin
uing in
the same fa
shion,
…
…
…
(91
)
or,
…
…
(20
)
Multiplying bo
th side
s by
,
…
…
(21
)
Or,
(22
)
With,
…
(23
)
For the
-th un
kno
w
n,
(24
)
With,
…
…
∃
(25
)
5. Proof
In this
se
ctio
n, we
provid
e the
pro
o
f o
f
the propo
se
d metho
d
. We be
gin the
proof
by
writing the b
a
s
ic le
ast squa
res p
r
o
b
lem.
(26
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 337 – 351
342
Expanding E
quation (26
)
,
⋮
…
⋮
(27
)
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(28
)
I
f
column
s of
are orth
ono
rmal, then Equation (2
8) b
e
com
e
s,
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(29
)
⋮
(30
)
I
f
column
s of
are orth
ogo
n
a
l, then Equa
tion (28
)
be
co
mes,
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(31
)
⋮
(32
)
I
f
column
s of
are n
e
ither
orthon
orm
a
l
nor o
r
tho
gon
al, then the o
ff-diagon
al te
rms i
n
Equation
(28
)
p
r
event the
dire
ct
soluti
on. If
we
ca
n someh
o
w
eliminate th
e
s
e te
rm
s, th
en a
dire
ct solutio
n
of the fo
rm
(32
)
is po
ssib
le.
In ord
e
r to
do that, we i
n
se
rt a bl
ock
matrix
matrix
in Equation (26) an
d re
-write it as:
(33
)
Expanding (3
3),
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New App
r
o
a
ch to Lin
ear
Estim
a
tion Problem
in Multi-user …
(Mu
ham
m
ad Ali
Ra
za Anjum
)
343
⋮
⋮
…
⋮
⋮
(34
)
⋮
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
⋮
(35
)
If only,
∃
(36
)
Then Equ
a
tio
n
(35
)
take
s the form,
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(37
)
And the soluti
on be
come
s,
⋮
(38
)
For Equatio
n (38
)
to hold, we can choo
se
to be:
…
…
∃
(39
)
is
a projec
tion matrix.
(40
)
proje
c
ts a ve
ctor on a
n
axis that
is orth
o
gonal to the p
l
ane defin
ed
by
. An important
prop
erty of
matrix is that
the pro
d
u
c
t
is zero. T
h
is i
s
be
ca
use th
e proj
ectio
n
o
f
any vector
on an axis o
r
thogo
nal to itself is alway
s
zero. Hen
c
e,
the prod
uct
i
s
always
zero.
(41
)
This p
r
op
erty of
matrix can
be can b
e
used to eliminat
e the off-diag
onal entri
es i
n
Equation (1).
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(42
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 337 – 351
344
Starting with
the last term
in first ro
w o
n
the left hand side of Eq. (42), we mul
t
iply th
e
top row
with
.
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(43
)
With,
(44
)
Equation (43) becom
es,
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(45
)
Similarly, to eliminate
we
multiply the top ro
w in Equ
a
tion (45
)
wit
h
,
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(46
)
Such that,
(47
)
We
contin
ue
this multipli
ca
tion in this m
anne
r until all
the off diago
nal entri
es i
n
the first
row a
r
e remo
ved.
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(48
)
With
,
…
…
(49
)
And,
…
…
…
…
(50
)
Followi
ng the
same proce
dure to rem
o
ve off diagonal entrie
s
for the remaini
n
g
rows of
Equation (48), it become
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New App
r
o
a
ch to Lin
ear
Estim
a
tion Problem
in Multi-user …
(Mu
ham
m
ad Ali
Ra
za Anjum
)
345
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
(51
)
With,
…
…
∃
(52
)
Multiplying bo
th side
s of Equation (52)
wi
th
to revert back
to our leas
t s
q
uares
s
o
lution as
in
Equation (35),
⋮
…
…
⋮⋮
⋱
⋮
⋮
…
…
⋮
⋮
(53
)
Solving for
in
Equation (53), we obtain Equation (38
)
.
⋮
(Re
peat
)
6. Proof of the Choic
e
of
Matrix
In this
section, we will
prov
ide a novel
factori
z
ation
of LS solution which confi
r
ms our
choi
ce of
ma
trix.
For the s
a
ke of brevit
y, we will provide the proof for
2
. The proof can be
simila
rly extended to any d
i
mensi
on. Su
bstituting
2
in
Equation (26) and expan
di
ng,
(54
)
(55
)
Proceedi
ng di
rectly for the
solutio
n
,
(56
)
This would yi
eld,
(57
)
(58
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 337 – 351
346
(59
)
The term
ca
n be in
written two
ways.
Firstly, multip
ly
and divide it by
,
(60
)
Secon
d
ly, multiply and divide it by
,
(61
)
Opting Equati
on (60
)
an
d Equation (61
)
for first an
d se
con
d
ro
ws of
Equation (59) resp
ectively,
(62
)
Term
s
and
are scal
ars. Can
c
ellin
g an
d simplifying
Equation (62),
(63
)
The factori
z
a
t
ion pro
c
ess
reveal
s the same
matrix
we employe
d
in Equation (38
)
as
the solution
s obtaine
d by
Equation
(3
8) and (63) a
r
e
exactly same.
This confirms our choi
ce of
matrix in Equation (38
)
.
7. Algorithm
In this se
ctio
n, we presen
t an algorith
m
for pro
p
o
s
ed method.
We will
start
with an
example, say
computation
of
, and onwa
r
ds b
u
ild a ge
neri
c
algo
rith
m. We begin by re-writing
Equation (15),
(64
)
is ge
ne
rated
from
by rem
o
ving its last
colum
n
. Thi
s
last colum
n
is used to
construct
whi
c
h is
then multiplied w
.
(65
)
and
are the
n
update
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.