TELKOM
NIKA
, Vol.14, No
.2, June 20
16
, pp. 538~5
4
7
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i1.3147
538
Re
cei
v
ed
No
vem
ber 2
5
, 2015; Re
vi
sed
March 21, 20
16; Accepted
April 15, 201
6
Low Complexit
y
Sparse Channel Estimation Based on
Compressed Sensing
Fei Zhou, Yantao Su*, Xin
y
ue Fan
Cho
ngq
in
g Ke
y L
abor
ator
y
of
Optical Comm
unic
a
tion
and
Net
w
orks, Ch
o
ngq
ing U
n
iv
ers
i
t
y
of Posts an
d
T
e
lecommunic
a
tions, Ch
on
g
w
e
n
R
oad, 4
0
0
065, Ch
on
gqi
n
g
, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: s
y
tgps
200
0
@
16
3.com
A
b
st
r
a
ct
In w
i
reless co
mmu
n
icati
on,
chan
nel
esti
mation is
a key
techno
lo
gy to receiv
e sig
nal
precis
ely.
Rece
ntly, a
ne
w
meth
od
na
med c
o
mpress
e
d
se
nsin
g (C
S) has
be
en
pr
op
osed
to
esti
mat
e
sp
arse c
h
a
n
nel,
w
h
ich greatly
improves th
e spectru
m
effici
ency. How
e
ve
r, it is difficult to reali
z
e
it due to its hi
g
h
computati
o
n
a
l
compl
e
xity. Althou
gh the pr
o
pose
d
Orthog
ona
l Matchin
g
Pursuit (OMP) can reduc
e the
compl
e
xity of CS, the efficie
n
cy of OMP is still low
beca
u
se on
ly on
e i
ndex is i
d
e
n
tifi
ed per it
eratio
n.
T
herefore, to
solve this
prob
le
m, more effi
cient sch
emes
are pro
pos
ed
. At first, w
e
a
pply Ge
nera
l
i
z
ed
Orthogon
al M
a
tching
Pursu
i
t
(GOMP) to chann
el
esti
mati
on, w
h
ich
low
e
r co
mp
utatio
n
a
l co
mplex
i
ty
by
selecti
ng
multi
p
le i
ndic
e
s in
each
iter
atio
n. T
hen a more
effective sch
e
m
e th
at select
s indic
e
s by le
ast
squar
es (LS)
meth
od
is pro
p
o
sed to s
i
g
n
ific
antly re
duc
e t
h
e co
mp
utatio
n
a
l co
mplex
i
ty, w
h
ich is a
mod
i
fied
meth
od
of GOMP. Simul
a
ti
on resu
lts an
d theor
etical
ana
lysis sh
ow
the effectivity of the prop
o
s
ed
alg
o
rith
ms.
Ke
y
w
ords
: ch
ann
el esti
mati
o
n
, compresse
d
sensin
g, comp
utation
a
l co
mpl
e
xity, index, at
om
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The pe
rform
ance of wireless comm
unication co
nsid
era
b
ly depen
ds o
n
wirel
e
ss
cha
nnel.
Du
e
to the
comp
lexity and va
riability of
ge
ogra
phi
cal
e
n
vironm
ent, the p
r
op
agati
on
signal is li
kely to undergo multipath
pr
opagation and Doppl
er
shi
ft, generating frequency
sele
ctive fadi
ng and time selective fadin
g
whi
c
h di
sto
r
t the signal
severely [1]. In
orde
r to obtain
accurate
re
ceipt si
gnal, it
is
ne
ce
ssa
ry to ac
quire
the exa
c
t ch
annel
state
informatio
n (CSI).
Therefore, we need to kn
ow the preci
s
e chan
nel im
pulse re
spo
n
s
e (CIR) firstl
y. So
far, cha
nnel
estimation u
s
i
ng refe
ren
c
e
sign
als o
r
pilo
ts is
the mo
st commo
n method to obtain
CIR [2], which
transmits a g
r
oup of kno
w
n signal, that
is, referen
c
e signal
s o
r
pilots, and th
en ca
rrie
s
o
u
t
cha
nnel
esti
mation ba
se
d on the va
ri
ation bet
wee
n
re
ceiving
signal a
nd tra
n
smitting
sig
nal,
lastly obtainin
g
the CIR.
Cla
ssi
cal
pilot
s
b
a
sed
ch
an
nel e
s
timatio
n
metho
d
s in
clud
e le
ast
squares (LS),
minimum
mean
squ
a
red
error (M
MSE), and
discrete Fou
r
ie
r Trans
f
orm
(DFT). Noteworthily,
traditional
method
s do
n’
t con
s
ide
r
the
spa
r
sity of wi
rele
ss
ch
ann
el, which re
sult in huge
waste of spe
c
trum
resou
r
ce. Wit
h
the
develo
p
ment of
re
search, a
lot of
literatu
r
e
s
have sho
w
n
that
the wirel
e
ss
cha
nnel
s are gene
rally sp
a
r
se in p
r
a
c
tice. The CS
ba
sed
chan
nel
estimation util
ize
s
the cha
n
nel
spa
r
sity, re
d
u
ce
s the n
u
m
ber
of pilots an
d impr
o
v
es the
spe
c
trum
efficie
n
cy finally. Many
resea
r
chers
have pro
p
o
s
ed a lot of a
l
gorithm
s su
ch as
1
l
-no
r
m
minimization
based Ba
si
s
Pursuit (BP) algorithm [
3
], greedy it
eration
ba
se
d Matchi
ng
Pursuit (MP) algorithm
a
nd
Orthog
onal
Matchin
g
Pursuit (OMP)
algorith
m
[4
, 5]. BP algorithm has hi
gh re
con
s
tru
c
tion
accuracy, bu
t its comput
ation co
st is huge.
Com
pare
d
with
BP algorithm
, MP algorithm
improve
s
the
comp
utationa
l compl
e
xity,
but we
ak
en
s
its
r
o
bu
s
t
ne
ss
. O
M
P is
a
mo
d
i
fie
d
me
th
od
of MP algori
t
hm, which execute
s
an
orthog
onali
z
ation of the
sele
cted at
oms. Th
e O
M
P
algorith
m
ca
n brin
g a fa
st co
nverg
e
n
c
e a
nd go
o
d
rob
u
stn
e
ss as
well. In
addition, O
M
P
algorith
m
ha
s advantage
s
of high re
con
s
tru
c
tion a
c
cura
cy and co
mputation sp
eed.
Re
cent
ly
,
mo
st
re
se
ar
che
r
s
have
paid
their m
a
in
attenti
on on
the improve
m
ent of
estimation
accuracy. Aimin
g
at Ultra
Wi
deba
nd
(UWB) ch
ann
el, A.H. Muqaib
e
l
prop
osed m
o
re
pra
c
tical
di
ctionari
e
s to
en
han
ce th
e
sp
arsity
of UWB ch
annel,
a
nd the
n
im
proved the
a
c
cura
cy
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Low
Com
p
lexity Sparse Ch
annel Estim
a
tion Base
d on
Com
p
re
ssed
Sensin
g (Fei
Zhou
)
539
of received
signal [6]. S. Pramon
o ha
s imp
r
oved
t
he e
s
timatio
n
accu
ra
cy throu
gh a
pply
i
ng
multiple input multiple
output (MIM
O) to
ch
annel es
timation [7]. In literatures
[8]
,
a
reconfigu
r
a
b
le and
spa
r
se
array a
n
tenn
a wa
s propo
sed, whi
c
h
ca
n also im
prove the estimati
on
accuracy. Ad
ditionally, ma
ny other
re
searche
r
s hav
e pu
rsued
hi
gh e
s
timatio
n
accu
ra
cy
by
optimiz
ing pilots
’ s
e
ttings
[9, 10].
In fact, the chann
el estim
a
tion accu
ra
cy has
re
ache
d a relatively
high level by
utilizing
the existed
a
l
gorithm
s. So
the com
put
ational
compl
e
xity of algorithms al
rea
d
y
becom
es t
h
e
main con
c
ern in pra
c
tice. However,
up to now, only a few literatu
r
e
s
about re
du
cin
g
comp
utationa
l complexity have bee
n re
leased.
A modified OMP (MOMP) was
prop
osed in [
11],
it not only re
duces th
e co
mputational
complexity
effectively, but also h
a
s
a g
ood e
s
timatio
n
accuracy. Bu
t it assume
s
all cha
nnel ta
ps di
stri
b
u
tin
g
adja
c
ently, whi
c
h limits it
s appli
c
atio
n in
pra
c
tice. In [12], an expand
er graph ba
sed CS wa
s
propo
sed to sp
arse ch
ann
el, which is ben
efit
to red
u
ce the
com
putat
ional complexity to
((
)
)
OM
N
N
, where
M
and
N
are th
e
length o
f
pilots an
d ch
annel, re
sp
ectively. But the
method in [12] is sen
s
itive
to noise.
Therefore,
a
further resea
r
ch
i
s
n
eede
d to
solve th
e p
r
oble
m
of
high
comput
ationa
l
compl
e
xity.
As mentio
ne
d previo
usly,
OMP algo
rit
h
m ha
s adva
n
tage
s of hig
h
re
con
s
truct
i
on
accuracy
and
com
putation
sp
eed, b
u
t for p
r
a
c
tical
a
pplication, its com
putation
a
l complexity
i
s
still too hi
gh.
Hence, in thi
s
pa
per, we
do a furt
her research
on t
he basi
s of
OMP algorithm to
redu
ce th
e computation
a
l
compl
e
xity and en
su
re t
he estim
a
tion
pre
c
isi
on at
the sam
e
time.
Firstly, we
a
pply Gen
e
rali
zed
Orth
ogo
nal Mat
c
hi
ng
Pursuit (G
O
M
P) alg
o
rith
m [13] to sp
arse
cha
nnel e
s
ti
mation, whi
c
h is a modifi
ed metho
d
o
f
OMP. Owin
g to the sele
ction of multi
p
le
atoms p
e
r iteration, the
calculating
spee
d is
i
m
prove
d
an
d the estim
a
tion accu
ra
cy is
guarantee
d a
t
the same ti
me. The
n
ba
sed
on th
e id
ea of
GOMP
algo
rithm, a
better al
go
rith
m
named M
-
G
O
MP is p
r
o
posed. M-G
O
MP sele
ct
s
atoms with
a new p
e
rspe
ctive, avoidin
g
repe
ating ite
r
ation
s
of th
e greedy al
gorithm,
whi
c
h
can
grea
tly redu
ce th
e co
mputatio
nal
compl
e
xity. A numbe
r of compute
r
sim
u
lation
s
are
con
d
u
c
ted, showi
ng a bet
ter com
putati
on
spe
ed an
d a good recon
s
truction a
c
cura
cy of the both algorithm
s.
The
re
st of t
h
is
pap
er is
orga
nized
as fo
llows. Se
ct
ion 2
b
r
iefly
depi
cts th
e
CS theory
and the p
r
o
b
l
e
m is
stated i
n
Section
3. In Secti
on
4, we p
r
op
ose the efficie
n
t schem
es
of in
dex
sele
ction, na
mely, the GOMP and
M-GO
MP al
gorithm
s. Se
ction 5 p
r
e
s
ent
s comp
uter
simulatio
n
s a
nd com
p
lexity analysi
s
. The con
c
lu
sio
n
is dra
w
n in Se
ction 6 finally.
In
this
p
ape
r,
bold
u
ppe
r case and bold lowe
r ca
se
le
tters
d
enote matrices and vectors,
r
e
spec
tively.
T
A
denote
s
th
e
subm
atrix of
A
re
stricte
d
to
colum
n
s inde
xed by the
se
t
T
.
H
A
,
T
A
, and
†1
()
H
H
A
AA
A
denote
the he
rmitian
tran
spo
s
ition
,
transpo
sitio
n
, and
pse
u
d
o
-inve
r
se
of matrix
A
, res
p
ec
tively.
2
t
d
enote
s
2
l
-no
r
m
of
t
.
,
Ah
den
otes the in
ne
r p
r
o
duct
of mat
r
ix
A
and vector
h
.
ˆ
h
denote
s
the e
s
timated valu
e of
h
.
2. Compres
s
e
d Sensing
Bas
e
d Re
co
nstru
c
tion
The p
r
op
ositi
on of CS i
s
a
revolution
ary
brea
kth
r
ou
gh
in sign
al p
r
o
c
e
ssi
ng. It breaks the
rest
rictio
n of the Nyqui
st sampling th
eo
rem an
d im
p
r
oves th
e sp
ectru
m
efficie
n
cy greatly [14,
15]. The
pre
m
ise
of CS
a
pplication i
s
t
hat the
ob
served si
gnal
is
spa
r
se o
r
co
mpre
ssible,
whi
c
h
mean
s there are only a fe
w non
ze
ro va
lues o
r
si
gni
ficant value
s
i
n
a gro
up of observed
sig
nal.
Fortun
ately, a lot of
stati
s
tics a
nd
ob
servatio
n d
a
ta indi
cate
th
at the
wirele
ss chan
nel i
s
gene
rally sp
arse, whi
c
h l
a
ys found
ation of
the CS applicatio
n
.
Standard
CS mea
s
ure
m
ent
model is give
n as follo
ws:
rt
(1)
Whe
r
e
r
is a
1
M
observed
vector,
is
a
M
N
measureme
n
t mat
r
ix,
t
is a
1
N
K
-s
par
s
e
sign
al with
K
non
zero valu
es,
is a
1
M
noise vecto
r
wh
ose ele
m
ent
s are additive white
Gau
ssi
an va
riable
s
. CS i
s
mainly abo
ut
how to
re
co
n
s
tru
c
t ori
g
inal
sign
al
t
from
the kn
own
r
and
. Since
M
N
for mo
st of the com
p
re
ssive sen
s
in
g scenari
o
s, reco
nstru
c
ting
t
directly
from (1
) is
an
unde
rdete
r
m
i
ned p
r
obl
em. Fortun
ately,
that sign
al is
spa
r
se ma
ke
s it possibl
e to
solve thi
s
p
r
oblem. T
he l
i
terature [16]
pointed
out
that if
satisfies Restri
ct
ed Isometry
Prope
rty (RIP
), then
t
can b
e
recon
s
tru
c
t
ed accu
rately
. We write
RIP as lemma 1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 538 – 54
7
540
L
e
mma
1
: Fo
r any
K
-sparse sign
al
t
, if there exists a con
s
tant
(0
,
1
)
K
s
u
c
h
that:
22
2
22
2
(1
)
(
1
)
KK
tt
t
(2)
Then
sat
i
sf
ie
s RI
P
.
In the algo
rit
h
ms
pro
p
o
s
e
d
in this
pap
er, the n
u
mb
er of me
asurements
M
is a very
important fac
t
or. If
M
sat
i
sf
ie
s:
2
(1
)
(
1
)
l
o
g
1
N
MK
K
K
(3)
Whe
r
e
(0
,1
)
is
a
c
o
ns
tant, then
obey
s the
RIP con
d
ition
with overwh
elming p
r
ob
a
b
ility
[17]. This i
s
t
he lo
we
r bo
u
nd of
M
on th
e
pre
m
ise of
accurate
re
constructio
n
. Ho
wever,
the
magnitud
e
is not given in [17], so the exact lower bo
u
nd of
M
is not determin
ed eit
her.
With num
ero
u
s exp
e
rim
e
n
t
analyse
s
, Tropp a
nd
Gil
b
ert pointe
d
o
u
t that we
ca
n re
cove
r
t
from (1)
with large p
r
o
babil
i
ty if
M
is expre
s
sed a
s
follo
ws [18]:
lo
g
M
KN
(4)
3. Problem Statement
Con
s
id
er a
M
pilots
,
N
length
and
K
-spa
rse
cha
nnel
sy
stem. The
po
si
tions
of pilo
t
s
are de
noted
by
01
1
,,
,
M
kk
k
. CS based
sparse
cha
n
nel estimatio
n
can b
e
expressed a
s
:
MN
yX
F
h
(5)
Whe
r
e,
01
1
[(
)
,
(
)
,
,
(
)
]
M
d
i
a
g
xk
xk
xk
X
is the diago
nal matrix
of transm
i
tted pilots.
01
1
[(
)
,
(
)
,
,
(
)
]
T
M
yk
yk
y
k
y
and
[(
0
)
,
(
1
)
,
,
(
1
)
]
T
hh
h
N
h
are the
1
M
received pi
lot vector
and
1
N
spa
r
se chann
el vecto
r
, resp
ectively
. There
are
K
n
onzero val
u
e
s
in
h
.
is a
1
M
compl
e
x addi
tive white Ga
ussian n
o
ise vector.
M
N
F
is a
M
N
partial Fo
uri
e
r matrix, whi
c
h is
obtaine
d by
sele
cting
the
ro
ws of
Fo
urie
r m
a
trix
with in
dices
01
1
,,
,
M
kk
k
.
M
N
F
can
be
expre
s
sed a
s
:
00
0
11
1
11
1
01
(
1
)
01
(
1
)
01
(
1
)
MM
M
kk
k
N
VV
V
kk
k
N
VV
V
MN
kk
k
N
VV
V
ff
f
ff
f
ff
f
F
(6)
Whe
r
e
e
xp(
2
/
)
p
kq
Vp
f
jk
q
V
,
V
is a positive integer,
01
pM
,
01
qN
. Let
M
N
AX
F
, (5) ca
n be written as:
yA
h
(7)
Whe
r
e,
12
[,
,
,
]
N
A
aa
a
is a
M
N
matrix.
i
a
is the colum
n
of matrix
A
. Comparin
g (7
) with
(1), we
can see
that
y
,
A
,
h
are the
ob
serv
ed ve
ctor, m
easure
m
ent
matrix an
d
sp
arse
sign
al
in CS,
re
spe
c
tively. After obtaining
the
o
b
se
rved ve
ct
or
y
, by ad
opti
ng a
certain
algorith
m
, we
can recon
s
tru
c
t the cha
nne
l
h
from
y
.
The OMP al
g
o
rithm i
s
a commen
dable
method fo
r chann
el re
co
n
s
tru
c
tion. It is one of
the g
r
eedy
a
l
gorithm
s,
wh
ich
ca
n effici
ently and
sta
b
ly
re
cove
r sparse sig
nal from
o
b
serve
d
value. The followin
g
is the
detailed
step
s of the OMP algorith
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Low
Com
p
lexity Sparse Ch
annel Estim
a
tion Base
d on
Com
p
re
ssed
Sensin
g (Fei
Zhou
)
541
Algorithm 1: OMP Algorith
m
Input
: receiv
ed pilots
y
, measu
r
em
ent matrix
A
, s
pars
i
ty
K
, noise va
riance
2
.
Initializ
e
: resi
dual
0
ry
, estimated su
ppo
rt se
t
0
, iteration co
unt
1
t
.
Repe
at
: Ass
u
me in the
th
t
iteration
1:
1
kk
.
2: Identification:
1
arg
m
ax
,
,
tt
j
j
j
ra
a
A
.
3: Support M
e
rge
r
:
t1
{}
tt
.
4: Estimation: utilize the LS method:
1
ˆ
()
tt
t
HH
t
hA
A
A
y
.
5: Resi
dual
Update:
ˆ
t
tt
-
r
y
Ah
.
Until
:
2
2
t
r
or
tK
.
Outpu
t
: outp
u
t the estimat
ed value
ˆ
h
.
Carefully stu
d
ying the
O
M
P algo
rithm
,
we
can
fin
d
that the mo
st important
p
a
rt of the
whol
e al
gorit
hm i
s
to fin
d
the
K
atoms correspon
ding
to n
onzero v
a
lue
s
in
chan
nel
h
, which
can b
e
call
ed
as matching
atoms. As lo
n
g
as the
K
matchin
g
atom
s are fou
nd, we
can a
c
hiev
e
th
e
ac
cu
r
a
te
r
e
co
ns
tr
uc
tion
o
f
h
thro
ugh
LS metho
d
gi
ven in
step
4.
At the
sam
e
time, the mo
st
time of the O
M
P algorithm
are spe
n
t on see
k
in
g
matching atom
s, namely,
the identification st
ep,
whose effici
ency is quite low. We illuminate it by (8):
12
,,
,
a
r
g
,
1
,
2
,
Ni
t
i
Ci
N
ra
(8)
Whe
r
e
C
is an colle
ction
of
N
indice
s a
nd one in
dex
in
C
corre
s
po
n
d
to one ato
m
in
A
.
Gene
rally, th
e value
N
is la
rge,
so
obtai
ning
C
by
step 2
will cost
pl
enty of time. However,
even so, OM
P algorithm d
oesn’t make f
u
ll use of
C
. In each iteratio
n, only one of the biggest
indices is se
lected
with
the oth
e
r ind
i
ce
s a
ban
do
ned
co
mplet
e
ly, whi
c
h
n
o
t only
wa
stes
resou
r
ce, but
also m
a
kes
the algo
rithm
conve
r
ge
sl
owly. With th
e increa
se
of iteration
s
, the
disa
dvantag
e
s
of the OMP algorith
m
are
more hi
ghligh
t
ed.
4.
Efficient S
c
hemes o
f
Indices Selecti
on
Before giving
the prop
osed
scheme
s
, we
fi
rstly pre
s
ent
two theore
m
s as follo
ws:
Theo
rem
1
: In the
algo
rith
m of this pa
p
e
r, the
sel
e
ct
ed in
dice
s
,,
ii
t
iN
will not
be sel
e
cte
d
in the later iteration.
Proof
: The th
eore
m
1 sh
o
w
s that when
the iter
ation run
s
, we do
n'
t have to worry about
the sele
cted
atoms to be
selecte
d
agai
n
,
includin
g
the non-matchi
ng atoms.
For the
conv
enien
ce
of de
scription
of th
eore
m
2,
we
give an
expre
ssi
on of th
e l
o
catio
n
of matching a
t
oms at first.
{0
,
}
i
Ti
h
i
N
(9)
Whe
r
e
i
h
is a tap of cha
nnel
h
.
Theo
rem
2
: (Assu
ming the
r
e ha
s b
een
execute
d
K
iteration
s
in tot
a
l): The
cha
n
nel
h
can
be
pe
rfe
c
tly re
con
s
tru
c
ted
wh
en th
e re
sultin
g
su
pport
set
K
an
d the m
a
tchi
ng atom
s
set
T
s
a
tis
f
y relation as
follows
:
K
T
(10)
Specifically, (10) ha
s two
meanin
g
s: 1
)
The su
ppo
rt set
K
must cont
ain all eleme
n
ts of
the mat
c
hing
atom
s
set
T
. 2)
Th
e sup
p
o
rt
set
K
may co
ntain
othe
r ele
m
ent
s b
e
sid
e
s th
e
matc
hing atoms
set
T
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 538 – 54
7
542
4.1. GOMP Algorithm
The propo
siti
on of GOMP
algorithm
efficiently
improves the d
e
ficien
cy of the OMP
algorith
m
. Different from t
he OMP alg
o
r
ithm, the G
O
MP algo
rith
m sele
cts
n
(
1
n
) atoms p
e
r
iteration, that
is, sele
cting
n
indi
ce
s f
r
o
m
C
. Then
G
O
MP pe
rforms
spa
r
se e
s
timation
an
d
resi
dual
up
da
te. In this wa
y, we
ca
n n
o
t only m
a
ke fu
ll use
of the
i
ndices colle
ction
C
, but al
so
accele
rate th
e co
nverg
e
n
c
e of algo
rith
m, and imp
r
o
v
e the algo
rithm efficien
cy
as a
wh
ole.
We
have lemma
2 with re
spe
c
t to the value
n
.
L
e
mma
2
: (GOMP algo
rith
m): Supp
ose
sele
cting
n
ato
m
s p
e
r ite
r
ati
on, then the
GOMP
algorith
m
. Ca
n reali
z
e reco
vering the o
r
i
g
inal sparse
cha
nnel u
nde
r con
d
ition:
mi
n
{
,
}
M
nK
K
(11)
Whe
r
e
K
and
M
are the
spa
r
si
ty of channel
and the num
b
e
r of pilots, re
spe
c
tively [13].
The wh
ole p
r
oce
dure of the GOMP algo
rithm is sket
ched a
s
follows:
Algorithm 2: GOMP Algori
t
hm
Input
: receiv
ed pilots
y
, measure
m
ent
matrix
A
, s
pars
i
ty
K
, noise varian
ce
2
, the
numbe
r of ind
i
ce
s sel
e
ctio
n
per iteratio
n
n
.
Initializ
e
: resi
dual
0
ry
, estimated su
ppo
rt se
t
0
, iteration co
unt
1
t
.
Repe
at
: Ass
u
me in the
th
t
iteration
1:
1
kk
.
2: Identification:
S
indice
s o
f
the
n
highe
st amplitude
co
mpone
nts of
1
T
t
A
r
.
3: Support M
e
rge
r
:
t1
t
S
.
4: Estimation: utilize the LS method :
1
ˆ
()
tt
t
HH
t
hA
A
A
y
5: Resi
dual
Update:
ˆ
t
tt
-
r
y
Ah
.
Until
:
2
2
t
r
or
tK
.
Outpu
t
: outp
u
t the estimat
ed value
ˆ
h
.
T
o
mak
e
it ea
s
i
er
to
u
n
d
e
r
s
ta
nd
, w
e
g
i
ve the sche
matic representation of th
e GOMP
algorith
m
in Figure 1.
C
1
()
tt
t
HH
A
AA
y
1
T
t
A
r
ˆ
t
t
A
h
y
1
t
r
t
ˆ
t
h
ˆ
h
ˆ
t
t
A
h
C
o
rre
l
at
i
o
n O
p
erat
i
o
n
Ch
oo
se I
n
dices
o
f
t
h
e
L
a
rge
s
t
Value
s
o
f
n C
E
s
tim
a
t
ion
R
e
m
e
as
u
r
e
m
ent
S
1
t
t1
t
S
S
u
pp
ort Merg
er
Figure 1. Sch
e
matic represent
ation of the GOMP algo
rithm
Comp
ari
ng th
e GOMP algo
rithm with the
OMP algorit
hm, we find that the most
different
part i
s
th
e
step 2,
whi
c
h
i
s
the
core
co
mpone
nt
of t
he G
O
MP
algorithm.
Note
worthily,
becau
se
more th
an o
n
e
atom i
s
sel
e
cted i
n
ea
ch
iterati
on,
we
inevitably put
the non
-mat
chin
g atom
s i
n
.
Even so, con
s
ide
r
ing th
eorem 2, we
still can
re
co
ver
the sp
arse
ch
annel a
c
cu
rat
e
ly. In addition,
becau
se the GOMP algo
rithm sele
cts
n
atoms pe
r iteration, gene
ral
l
y, the sparse
chan
nel ca
n
be re
con
s
tructed with
/
Kn
loop
s, whi
c
h grea
tly saves the runni
ng time
of algorithm.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Low
Com
p
lexity Sparse Ch
annel Estim
a
tion Base
d on
Com
p
re
ssed
Sensin
g (Fei
Zhou
)
543
4.2. M-GOM
P Algorithm
Based
o
n
the
idea
of
GOM
P algo
rithm,
we
propo
se
a
more
efficien
t schem
e, tha
t
is, M
-
GOMP algo
rithm. Compa
r
i
ng the GOM
P algorithm wi
th the OMP
algorith
m
, we
can se
e that the
key of th
e two alg
o
rithm
s
i
s
to
se
arch t
he mat
c
hin
g
atoms. A
s
lo
ng a
s
all the
matchin
g
ato
m
s
are a
c
qui
red,
we ca
n achi
eve accurate recon
s
tru
c
tio
n
by the step 4 of the OMP algorithm o
r
the
GOMP algo
ri
thm. Hence, the most critical thing
now i
s
ho
w to get all the matchi
ng atoms
with
no pl
entiful ti
me
co
sting
at the
sam
e
tim
e
. Co
nsi
deri
n
g the
sim
p
licit
y of LS al
gorit
hm, we u
s
e
the
LS algo
rithm
to se
arch
ato
m
s, a
nd th
en
estima
te
the
sp
arse
ch
an
nel by th
e
step 4
of G
O
M
P
,
whi
c
h g
ene
rates the
M-GOMP alg
o
rithm. T
he d
e
tailed impl
e
m
entation of
the M-GO
MP
algorith
m
is shown as follo
ws:
Algorithm 3: M-GO
MP Algorithm
Input
: receiv
ed pilots
y
, measu
r
em
ent matrix
A
, s
pars
i
ty
K
.
Initializ
e
: esti
mated su
ppo
rt set
.
Begin
:
1: Identification: solving a
simple LS e
s
t
i
mation
1
0
ˆ
hA
y
, let
indices of the
K
highe
st ampli
t
ude com
pon
ents of
0
ˆ
h
.
2: Estimation: utilize the LS method to re
con
s
tru
c
t:
1
ˆ
()
HH
hA
A
A
y
End
Outpu
t
: outp
u
t the estimat
ed value
ˆ
h
.
Comp
ari
ng M
-
GOMP al
gorithm with the
OMP and G
O
MP algo
rith
ms, we
can
see that
the M-G
O
MP
algorithm i
s
quite differe
n
t
fr
om the O
M
P and GO
MP algorithm
s. The M
-
GO
MP
algorith
m
o
n
l
y
has two
st
eps with
ide
n
t
ificati
on a
n
d
estim
a
tion,
and
wh
at’s
more, th
ere i
s
n
o
loop
s in M-G
O
MP. For the convenie
n
ce
to understan
d, we give the schemati
c
repre
s
e
n
tation
of
the M-GO
MP algorithm in
Figure 2.
1
Ay
0
ˆ
h
0
Choo
se
I
ndic
e
s
of
the
ˆ
La
r
g
e
s
t
V
a
l
u
e
s
o
f
K
h
1
()
HH
AA
A
y
ˆ
h
y
LS
E
s
t
i
ma
t
i
o
n
Ch
a
nne
l
Re
c
onstru
c
t
i
on
Figure 2. Sch
e
matic representat
ion of the M-GO
MP algorithm
From Fig
u
re 2 we can see
the M-GOM
P algorit
hm p
e
rform
s
o
n
ly one iteratio
n from left
to right. Hen
c
e it can g
r
e
a
tly improve the efficien
cy of reco
nstru
c
tion. In term
s of estimati
on
pre
c
isi
on, be
cau
s
e
we
ca
n acquire the
requi
re
d
K
atoms from e
s
t
i
mated
0
ˆ
h
, we are a
b
le t
o
reali
z
e a
c
curate recon
s
tru
c
tion by the step 2 in M-G
O
MP.
In this p
ape
r,
we
also a
d
o
p
t the comm
only used
LS algo
rithm a
n
d
the hi
gh
accuracy
estimated M
M
SE algorith
m
as the refe
ren
c
e
s
of per
forman
ce an
alysis. He
re,
first of all, we put
down their ex
pre
ssi
on
s as
follows:
1
ˆ
()
HH
LS
hA
A
A
y
(12)
21
1
1
ˆ
((
)
)
H
MM
S
E
HH
H
H
hR
R
A
A
A
y
(13)
W
h
er
e
H
H
R
den
otes th
e
aut
ocova
r
ian
c
e
matrix of
ch
annel
vecto
r
H
,
H
is th
e fre
quen
cy
domain rep
r
e
s
entatio
n
of
h
and
denotes
the stand
ard
deviation of n
o
ise.
5. Simulations and An
aly
ses
In this sectio
n, the Mean
Square Error (MSE)
simul
a
tion, run
n
in
g time sim
u
la
tion an
d
compl
e
xity analysi
s
are d
e
scrib
ed to verify
the goo
d perfo
rma
n
ce of the pro
p
o
se
d GOMP
and
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 538 – 54
7
544
M-GO
MP alg
o
rithm
s
. The
classi
cal m
e
thod
s su
ch
as LS algo
rithm, MMSE
algorith
m
, OMP
algorith
m
are
also a
pplied f
o
r co
mpa
r
iso
n
at the same
time.
5.1. MSE Si
mulation
In this
simul
a
tion, we
con
s
ider
a
spa
r
se
ch
ann
el
with
length
496
N
, s
pars
i
ty
12
K
,
and utilize G
aussia
n
pilots, who
s
e nu
mber is d
e
te
rmine
d
by (4). In order to
guara
n
tee the
accurate re
constructio
n
, we choo
se
256
M
. The noi
se i
s
compl
e
x ad
ditive white
Gau
ssi
an
noise. Additionally, consi
d
ering the lem
m
a 2, we
ch
oose the nu
mber of indi
ces sele
ction with
2,
5
,
9
n
in GO
MP alg
o
rithm. In o
r
d
e
r to
avoid th
e interfe
r
en
ce
ca
used by
ra
ndomi
c
ity of signal
,
we ado
pt the idea of taking
average with
multip
le loop
s. In addition, normali
zed
MSE is adopted
in the simulati
on. The sim
u
l
a
tion re
su
lts
are sho
w
n in
Figure 3 and
Figure 4.
Figure 3 i
s
th
e MSE simul
a
tion un
der
di
fferent SNR. As can b
e
se
en from
Figu
re 3, the
MSE curve
s
of different
algorith
m
s
decrea
s
e
wh
en the SNR increa
se
s, whi
c
h me
an
s a
n
improvem
ent
of cha
nnel
recon
s
tru
c
tio
n
accu
ra
cy. Additionally, comp
ared
wi
th the cla
s
si
cal
algorith
m
s, th
e recon
s
tru
c
tion a
c
cu
ra
cie
s
of
the prop
ose
d
G
O
MP and M-G
O
M
P
algo
rithm
s
are
not only fa
r b
e
tter, but
also very
clo
s
e t
o
the
OMP al
gorithm. Eve
n
when
2
n
, the MSE c
u
rv
e
of GOMP alg
o
rithm
compl
e
tely catch
e
s up with that
of OMP algo
rithm. All of th
ese d
e
mo
nstrate
the effectivity
of the propo
sed algo
rithm
s
.
Figure 3. The
compa
r
i
s
on
on MSE with
varying SNR
Figure 4. The
compa
r
i
s
on
on MSE with
varying sp
arsity
Figure 4 i
s
th
e MSE perfo
rmance si
mul
a
tion with
va
rying sp
arsity. From Fi
gu
re
4, we
can
see tha
t
the chann
el re
con
s
tru
c
tion a
c
cura
cy decre
ase
s
with
K
increasi
ng. T
h
i
s
phen
omen
on
can b
e
expl
ained by the
increa
se
of chann
el den
sit
y
. When the
cha
nnel d
e
n
s
ity
increa
se
s, th
e num
ber
of
signifi
cant tap
s
be
co
me
s
la
rge
r
. Ho
weve
r, the nu
mbe
r
of pilots
doe
sn’t
increa
se at t
he sa
me tim
e
, which ma
kes u
s
not a
b
l
e
to get eno
ugh chan
nel
informatio
n a
n
d
lead
s to the
increa
se of e
s
timated e
r
ro
rs finally
. It’s worth
noting
that with the increa
sin
g
of
spa
r
sit
y
K
, although th
e pe
rforma
nce of the propo
se
d
algorith
m
s i
s
decli
ned, the
MSE curv
e
s
of the propo
sed al
gorith
m
s have foll
owe
d
the O
M
P well all the time. It i
m
plies
both of the
prop
osed alg
o
rithm
s
have
robu
stne
ss a
gain
s
t differe
nt chan
nel
sp
arsity. In addi
tion, from Fig
u
re
3 an
d Fig
u
re
4, we
can
see that th
e e
s
timati
on
accura
cy of the
MMSE algo
rithm is the
be
st.
Ho
wever, d
u
e
to its high
comp
utationa
l compl
e
xity,
we generally
can’t utili
ze it in practice. But
we can take the MSE curv
e of MMSE algorithm a
s
th
e lowe
r bou
n
d
of estimatio
n
accuracy.
Combi
n
ing F
i
gure 3
with
Figure 4, we can
con
c
l
ude that the
MSE of the
GOMP
a
l
g
o
r
ithm be
co
me
s
w
o
rs
e
w
i
th
th
e inc
r
ea
s
e
o
f
n
. This
is m
a
inly because
when
n
is larger, the
GOMP algori
thm will involve more nonmatching at
oms. Theoreti
cally, we
can utilize (16) to
demon
strate. When the value
n
is larger, the value
2
in (16) ca
nnot be igno
red, wh
o
s
e
0
5
10
15
20
25
30
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
S
NR(
d
B
)
MS
E
LS
OM
P
GOM
P
n
=
2
GOM
P
n
=
5
GOM
P
n
=
9
M
-
GOM
P
MM
SE
4
8
12
16
20
24
28
10
-8
10
-7
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
S
par
s
i
t
y
K
MS
E
LS
OM
P
GO
M
P
n
=
2
GO
M
P
n
=
5
GO
M
P
n
=
9
M-
G
O
MP
MMS
E
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Low
Com
p
lexity Sparse Ch
annel Estim
a
tion Base
d on
Com
p
re
ssed
Sensin
g (Fei
Zhou
)
545
existen
c
e lea
d
s to the de
cline of e
s
tim
a
ted accu
ra
c
y
.
H
o
w
e
v
e
r
,
i
t
i
s
a
l
s
o
l
i
k
e
l
y
t
o
s
h
o
w
u
p
a
spe
c
ial
ca
se.
When th
e value
n
is sma
ller, the co
rresp
ondi
ng M
SE is highe
r, as sho
w
n i
n
Figure 3. It is be
ca
use th
e GOMP
alg
o
rithm
with
smaller
n
executes m
o
re
times
of iterati
on,
and then p
u
ts more no
nma
t
ching atom
s
in.
5.2. Running
Time Simulation
In this
se
ctio
n, we
mainly
co
nsid
er th
e
run
n
ing tim
e
sim
u
lation
with varyin
g
cha
nnel
length. T
he
simulatio
n
p
a
r
amete
r
s a
r
e
set
a
s
SNR
5
dB
,
51
2
M
,
12
K
, the
s
i
mulation
results
are
sh
own i
n
Fig
u
re
5. Du
e to the
high
co
mple
xity of MMSE algo
rithm, we don’t
co
nsi
d
e
r
the MMSE algorithm in the
simulation.
As is sh
own in Figure 5, comp
ared wit
h
the OMP algorithm, the
prop
osed alg
o
rithm
s
have go
od p
e
rform
a
n
c
e.
Among them,
the GOMP
algorith
m
wit
h
5,
9
n
co
st only
about
1/
2
runni
ng time
as mu
ch
as t
hat of the O
M
P algorith
m
, namely, the
y
can
save th
e run
n
ing tim
e
by
more tha
n
50
%. What’s m
o
re, the M-G
O
MP algorit
h
m
can save the run
n
ing ti
me by more than
85% un
der di
fferent chan
n
e
l length
s
,
which
proves
the M
-
GOMP
algorith
m
to b
e
mo
re
sup
e
rior.
These sig
n
ificant improvem
ents mainly o
w
e
s
to
the efficient sch
e
me
s of indices
selectio
n.
Beside
s, the averag
e iterat
ions
simulatio
n
with
varying
sparsity is al
so given in Fi
gure 6.
From Fig
u
re
6 we ca
n se
e
,
the average
iterati
ons of the pro
p
o
s
ed
algorith
m
s a
r
e less than that
of the OMP
algorith
m
un
der diffe
rent
spa
r
sity.
The
com
putation
a
l com
p
lexity of algo
rithm
is
clo
s
ely
relate
d to th
e n
u
m
ber of ite
r
atio
ns,
so
t
he
si
mulation
re
su
lts of
Figu
re
6 not
only ve
rify
the su
peri
o
ri
ty of the GOMP and M
-
GOMP
algo
rithms
agai
n, but also d
e
mon
s
trate t
h
e
robu
stne
ss of both algorith
m
s at the sa
me time.
Figure 5. The
compa
r
i
s
on
on run
n
ing ti
me with
varying ch
an
nel length
Figure 6. The
compa
r
i
s
on
on avera
ge
iteration
s
with
varying spa
r
sity
5.3. Comple
xit
y
Analy
s
is
In this sectio
n, we
an
alyze the
comput
ational
com
p
l
e
xity of the
GOMP
and
M-GO
MP
algorith
m
s, t
o
p
r
ove th
e
sup
e
rio
r
ity of
the t
w
o
alg
o
rithm
s
q
uan
titatively. The
literature
[
13]
pointed
out that
the co
mputational
compl
e
xi
ty
of
GOMP alg
o
rithm ca
n be
a
pproxim
ately
expre
s
sed as
22
2(
2
)
GO
MP
Cs
M
N
n
n
s
M
, wh
ere
s
is t
he n
u
mbe
r
of iteratio
ns. B
e
ca
use th
e
OMP alg
o
rith
m is a
sp
eci
a
l ca
se
of th
e
GOMP al
go
rithm
with
1
n
, the c
o
mplexity of the OMP
algorith
m
can
be exp
r
e
s
se
d as
2
23
OMP
CK
M
N
K
M
. In [13], a multiplication an
d an
a
ddition
are
re
garded
as a
ba
sic o
peratio
n, respectively.
Ba
sed
on
the
same p
r
in
ciple
,
we
discu
s
s
the
comp
utationa
l complexity of the M-GOM
P algorithm.
Identification
:
This ste
p
ha
s high
com
p
l
e
xity due to the involveme
n
t of matrix inversi
on,
so
we figu
re
out the
cha
n
n
e
l freq
uen
cy resp
on
se
H
first
l
y, and then
convert
H
to time dom
ain
by IFFT. By this way, the total numbe
r o
f
operation
s
i
s
2(
2
1
)
M
MN
. In addition
,
0
ˆ
h
need
s to
be so
rted to find
K
best indi
ces, whi
c
h req
u
ire
s
((
1
)
)
/
2
NK
K
K
ope
rati
ons.
1
2
4
6
8
10
12
14
16
0
0.
01
0.
02
0.
03
0.
04
0.
05
0.
06
0.
07
0.
08
C
han
nel
L
e
ngt
h
L
(
x
62)
R
u
nn
i
n
g Ti
m
e
(
s
)
OM
P
M-
G
O
MP
GOM
P
n
=
2
GOM
P
n
=
5
GOM
P
n
=
9
4
8
12
16
20
24
28
0
5
10
15
20
25
30
S
pars
i
t
y
K
A
v
erage
I
t
e
r
at
i
o
n
s
OM
P
GO
M
P
n
=
2
GO
M
P
n
=
5
GO
M
P
n
=
9
M-
G
O
MP
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 538 – 54
7
546
Es
timation
: This
step i
s
al
so involved
wit
h
matrix
inversion, therefore, we
utilize t
he QR
factori
z
ation
and mo
dified
Gram
-Schmi
dt (MGS)
alg
o
rithm to get
a fast solutio
n
, whi
c
h lea
d
s
to
a c
o
mplexity
with
23
2
25
4
K
MK
M
K
K
K
.
The whole
M-GO
MP algorithm o
n
ly has o
ne
iteration,
so i
t
s total com
putationa
l
compl
e
xity can be app
roxi
mately expre
s
sed a
s
2
22
MG
O
M
P
CM
N
K
M
.
No
w let’s discu
ss the com
putational co
mplexity
of th
e three algo
ri
thms. Noting
that the
value
s
and
n
of GOMP algorithm a
r
e
small con
s
ta
nts, so the compl
e
xity of the GOMP
algorith
m
ca
n be exp
r
e
ssed a
s
()
Os
M
N
, and
the co
mplexit
y
of the OM
P algorith
m
can
be
expre
s
sed as
()
OK
M
N
. As can b
e
see
n
fro
m
the Fig
u
re
6,
the value
s
usually tend
s t
o
be
smaller than
K
, that is to
say, the G
O
MP algo
rit
h
m ha
s g
r
e
a
t advantag
e
s
in te
rm
s
of
comp
utationa
l complexity. Additionally, becau
se
K<M, M<N
, the
comp
utationa
l complexity o
f
M-
GOMP alg
o
rit
h
m can be
e
x
presse
d a
s
()
OM
N
. Obviou
sly, the M-GOMP
algorith
m
that
we p
u
t
forwa
r
d h
a
s a
better perfo
rmance.
6. Conclusio
n
In this pa
pe
r, in view o
f
the high com
putation
a
l
compl
e
xity of existing
chann
el
estimation m
e
thod
s, we a
pply GOMP algorith
m
to spa
r
se ch
an
nel estimatio
n
. Meanwhile
, a
more
effectiv
e sch
e
me
na
med M
-
G
O
M
P
algo
rithm i
s
p
r
o
p
o
s
ed
b
a
se
d o
n
the
idea
of GO
M
P
algorithm, which selects m
u
ltiple indi
ces in eac
h ite
r
a
t
ion. Compa
r
ed with the
OMP algo
rithm,
comp
uter sim
u
lation
s an
d
theoreti
c
al
an
alysis sh
o
w
t
hat the
pro
p
o
s
ed
two
algo
rithms not
on
ly
have goo
d e
s
timation a
ccura
cy, but also can g
r
eat
ly redu
ce th
e run
n
ing ti
me req
u
ired
for
cha
nnel e
s
ti
mation. Esp
e
c
ially the M-GOMP alg
o
rit
h
m,
as a
re
su
lt of efficient scheme
of in
dice
s
sele
ction, it
can
save mo
re tha
n
8
5
%
of t
he
run
n
i
ng time.
Con
s
eq
uently, G
O
MP alg
o
rith
m,
esp
e
ci
ally the M-GO
MP algorith
m
is
expec
te
d to be a co
mpet
itive candid
a
t
e sch
eme f
o
r
cha
nnel e
s
ti
mation in wi
reless commu
nicatio
n
.
Ackn
o
w
l
e
dg
ements
This
work was su
ppo
rte
d
by
th
e National Natu
ral
Sci
e
n
c
e Found
ation
of
Chi
na
(614
710
77,
6130
1126
), t
he Fu
ndam
e
n
tal and
Frontier
Re
se
arch Proje
c
t of Ch
ong
qi
ng
(cstc2013j
cyj
A
40034,
cstc2013j
cyjA400
41), th
e S
c
ie
nce
an
d T
e
chnolo
g
y Proj
ect of
Chong
qin
g
Munici
pal Ed
ucatio
n Com
m
issi
on (K
J1
4004
13).
Referen
ces
[1]
Hla
w
a
tsc
h
F
,
Matz G. W
i
rele
ss commu
nicat
i
ons
over
ra
pid
l
y time-var
yi
n
g
cha
nne
ls. US
A: Academ
ic
Press. 2011: 1
99-2
36.
[2]
Baoka
i
Z
u
, Xin
y
u
an
Xia, Ke
w
en
Xi
a, Chu
a
n
j
i
an B
a
i. Ch
an
n
e
l estimati
on
o
n
60GHz
w
i
rel
e
ss s
y
stem
base
d
on s
ubs
pace
p
u
rsuit.
T
E
LKOMNIKA Indo
nesi
a
n
Jo
u
r
nal
of El
ectric
al E
ngi
ne
erin
g
. 20
14;
12(9):
685
2-68
59.
[3]
Jianz
hon
g Hu
a
ng, Berger C
R
,
Shengl
i Z
hou
, Jie Huan
g.
Comparis
on of b
a
sis purs
u
it al
gorith
m
s for
sparse ch
an
nel
estimati
on i
n
u
nderw
a
ter aco
u
stic OF
DM
. OCEANS 2010 IEEE. S
y
dney
.
2010: 1-6.
[4]
T
eng Sun, Z
h
i
qun S
o
n
g
, Yo
ngji
e
Z
h
a
ng.
Matchin
g
purs
u
it bas
ed s
par
se cha
n
n
e
l est
i
matio
n
usi
n
g
pseu
dora
n
d
o
m
seque
nces
. Gl
oba
l S
y
m
posi
u
m on Millim
eter
Waves (GSMM). Harbin. 20
12: 33-3
7
.
[5]
Aboutor
ab N,
Hardj
a
w
a
na W
,
Vucetic B.
Ap
plicati
on
of co
mpr
e
ssive s
e
n
s
ing to c
han
ne
l
estimatio
n
of
high mobility OFDM system
s
. IEEE
International Confer
ence on
Com
m
unic
a
tions (I
CC). Budapest
.
201
3: 494
6-49
50.
[6]
Muqa
ibe
l
A
H
, Alkh
od
ar
y M
T
. Practical
a
pplic
atio
n
of
compressiv
e
s
ensi
ng t
o
u
l
tra-
w
i
d
e
b
a
n
d
chan
nels.
IET
on Co
mmunic
a
tions
. 201
2; 6(
16): 253
4-2
542
.
[7]
Pramon
o S, T
r
i
y
on
o E. Perfor
mance
of
channel
estimation in MIMO-OFD
M s
y
stems.
TEL
K
OMNIKA
T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol
. 20
13; 11(2): 3
55-
362.
[8]
Morabit
o
AF
, Iserni
a T
,
Donato LD. Optimal s
y
nthes
is of phase-
onl
y re
conf
ig
urab
le l
i
near sp
ars
e
arra
ys
h
a
vin
g
uniform-
ampl
itude e
x
citati
ons
.
Progress i
n
Electro
m
a
gneti
cs Rese
arch
. 201
2;
12
4(1):
405-
423.
[9]
Xu
e
y
u
n
H
e
, R
ongfa
n
g
So
ng,
W
e
ipi
n
g
Z
hu.
Pilot
all
o
cati
on
for sp
arse
ch
ann
el
estimati
on
in MIMO-
OFDM sy
stem
s.
IEEE
Transa
c
tions on C
i
rcu
i
ts and Syste
m
s II: Express Briefs
. 2013; 6
0
(
9
): 612-6
16.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Low
Com
p
lexity Sparse Ch
annel Estim
a
tion Base
d on
Com
p
re
ssed
Sensin
g (Fei
Zhou
)
547
[10]
Xi
an
g R
en, W
en
Che
n
, Me
i
x
ia T
ao,
Xi
aofei
Sha
o
.
C
o
mpr
e
ssed
ch
ann
el
esti
mati
on
w
i
th j
o
int
pi
lot
sym
bol and pl
acem
ent desi
g
n for high
m
o
bility OFDM system
s
. Intern
ation
a
l W
o
rks
hop
on Hi
gh
Mobil
i
t
y
W
i
re
le
ss Communic
a
tions (HMW
C). Beiji
ng. 20
14: 38-4
2
.
[11]
Ziji
Ma, Hon
g
li
Liu,
H
i
g
a
shi
n
o
T, Okada M, Furud
a
te H.
Lo
w
-
compl
e
xity c
han
nel
esti
mation for IS
DB-
T
over dou
bly-
selectiv
e fadi
n
g
chan
ne
ls
. Internati
ona
l S
y
mposi
u
m on Intelli
ge
nt Sign
al Process
i
ng
and C
o
mmun
i
c
a
tions S
y
stems
(ISPACS). Naha. 201
3: 114-
118.
[12]
Junji
e
P
an, F
e
ifei Ga
o.
Efficient ch
an
nel
estimatio
n
us
i
ng ex
pa
nd
er
grap
h b
a
sed
compressiv
e
sensi
n
g
. IEEE International Conferen
ce on Communications (ICC)
. S
y
dney
. 2014: 4542-4547.
[13]
Jian W
a
n
g
, K
w
on S, Shim B.
G
enera
lize
d
or
thogo
na
l matchin
g
purs
u
it.
IEEE Transactions on Signal
Processi
ng
. 20
12; 60(1
2
): 620
2-62
16.
[14]
Can
des EJ, R
o
mber
g JK, T
ao T
.
Robust uncerta
int
y
pri
n
cipl
es: e
x
act
sign
al rec
onstr
uction from
h
i
gh
ly
i
n
co
mp
le
te
frequency
information.
IEEE Transactio
n
s on Infor
m
ati
on Theory
. 2
0
0
6
; 52(2): 48
9-
509.
[15]
Can
des EJ, W
a
kin
MB. An
int
r
oducti
on to
co
mpressive
sam
p
lin
g.
IEEE Signal Process
i
ng Maga
z
i
ne
.
200
8; 25(2): 21
-30.
[16]
Can
des EJ. T
h
e restricted is
o
m
etr
y
pr
oper
t
y
and its imp
lic
ations for com
p
resses se
nsi
n
g.
Co
mptes
R
e
nd
u
s
Ma
them
a
t
iq
ue
. 200
8; 346(9): 58
9-5
92.
[17]
Baran
i
uk
R, D
a
ven
port M, D
e
Vore
R, W
a
ki
n M.
A sim
p
l
e
proof
of the
re
stricted is
omet
r
y
pro
pert
y
for
rand
om matric
es.
Constructiv
e
Approx
i
m
atio
n
. 2008; 2
8
(3): 253-
263.
[18]
T
r
opp JA, Gilb
ert AC. Sig
n
a
l
recover
y
fr
om
rand
om me
asu
r
ements vi
a
orthog
on
al m
a
tch
i
ng
purs
u
it.
IEEE Transactions on Infor
m
a
t
ion Theory
. 2
0
07; 53(1
2
): 465
5-46
66.
Evaluation Warning : The document was created with Spire.PDF for Python.