:=
J
offset
F
H P
J
offset
(12)Since
Λ
offset
is banded,
J
offset
becomes a spread version of
J
. Considering
J
offset
as an approximately sparse vector, wecan use the CSbased approach to estimate it as done inSection IVA for the case of mobile jammers. This techniqueis labelled as ”without estimating
α
” in Section VII. However,if
Λ
offset
is known at the receiver, then the CS approach canbe applied to recover
J
instead of
J
offset
. This is achieved byabsorbing
Λ
offset
into the measurement matrix. To show this,we write
Y
as
Y
=
W
Y
=
W
J
offset
+
W
Z
=
WΛ
offset
J
+
W
Z
.
(13)If
Λ
offset
is not known at the receiver, then we apply CS theoryto estimate
J
offset
. In this case, the measurement matrix is
W
.On the other hand, if
Λ
offset
is known at the receiver, then weapply CS theory to estimate
J
. In this case, the measurementmatrix is
WΛ
offset
. Since
J
is exactly sparse unlike
J
offset
,the solution of the
l
1
norm constrained minimization problemfor
J
will be more accurate than that for
J
offset
. This isbecause CSbased algorithms are designed for sparse vectors.Consequently, better NBI estimation and cancellation can beachieved if
Λ
offset
is known at the receiver. Note that the matrix
Λ
offset
is completely determined by
α
. Hence, we propose thefollowing technique to estimate
α
for singletone jammers andbelieve that the extension to multipletone jammers is possible.First, we apply CS theory to get an initial estimate of
J
offset
.Since this estimate is sparse, we extract its largest
s
nonzeroelements. These extracted elements can be expressed as
ˆ
J
offset
(
l
n
) =
d
k
1
−
e
i
2
πα
/P
1
−
e
i
2
πP
(
k
−
l
n
+
α
)
+
e
(
l
n
)
,
1
≤
n
≤
s
(14)where
e
(
l
n
)
is some error term at the
(
l
n
)
th
element,
d
k
and
k
denote the amplitude and the location of the singlejamming tone, respectively. The idea is to estimate
α
from
{
ˆ
J
offset
(
l
n
)
,n
= 1
,
2
,..,s
}
using the nonlinear least squaresestimation technique [12] because
ˆ
J
offset
(
l
n
)
is a nonlinearfunction of
α
. The estimate of
α
, denoted by
ˆ
α
, is chosen tomaximize the following cost function
ζ
(
α
) =
sn
=1
ˆ
J
∗
offset
(
l
n
)(
1
−
e
i
2
πα
)
1
−
e
i
2
πP
(
k
−
ln
+
α
)
2
(1
−
e
i
2
πα
)
1
−
e
i
2
πP
(
k
−
ln
+
α
)
2
(15)where

.

denotes the absolute value. This maximization canbe implemented using grid search over possible values of
α
as follows. We quantize the interval
−
12
,
12
into
N
q
levels,substitute each level for
α
in
ζ
(
α
)
, and set
ˆ
α
to the levelthat maximizes the cost function. Note that
ζ
(
α
)
depends on
k
which is the location of the jammer tone. As an estimateof
k
, we use
ˆ
k
=
arg
max
n
ˆ
J
offset
(
l
n
)
. Using
ˆ
α
, we forman estimate for
Λ
offset
, denoted by
ˆ
Λ
offset
. Then, we formthe measurement matrix as
WΛ
offset
and solve the
l
1
normconstrained minimization problem to get an estimate for thefrequency support of
J
. If this estimate differs from
ˆ
k
, we reestimate
α
using the new estimate of
k
. This reﬁned estimateof
α
is only used to recalculate
ˆ
Λ
offset
and the measurementmatrix. Finally, we perform the WLS step to estimate theamplitude of the jammer and output the estimate of
J
denotedby
ˆ
J
. The new estimate for
J
offset
is given by
˜
J
offset
=ˆ
Λ
offset
ˆ
J
.
(16)Next,
J
offset
is cancelled from the received signal prior to thedetection and decoding steps. This technique is labelled as”with estimating
α
” in Section VII. In [14], we use anothertechnique to deal with asynchronous jammers,
viz.
we usereceiver windowing to spectrally contain the jammer.V. A
LTERNATIVE
N
ULLING
M
ETHOD
In Section III, we set
W
to be the projection matrix onto theleft nullsubspace of
˜
Λ
to null out (cancel) the unknown dataterm. However, computing
W
requires the inversion of a
P
×
P
matrix which is computationally intense although the diagonalstructure of
Λ
can be utilized to reduce the complexity. In thissection, we propose a simpler method to null out the unknowndata term where we utilize the tall structure of
F
zp
. First, wemultiply
Y
by
G
1
F
H P
Λ
−
1
to get
Y
1
=
G
1
Y
=
F
zp
X
+
G
1
Z
+
G
1
J
(17)where
Λ
−
1
is easily computed since
Λ
is diagonal and themultiplication by
F
H P
is efﬁciently implemented using theinverse FFT operation. Recalling the structure of
F
zp
, wesimply multiply
Y
1
by
G
2
0
ν
×
N
I
ν
to null out theunknown data term as follows
Y
2
=
G
2
Y
1
=
G
2
G
1