The overall architecture of the wireless semantic communications
system is shown in Fig. 1. At the transmitter, a
joint source-channel coding (JSCC) encodes the source image
s ∈ R3×Hs×Ws into a semantic feature vector x ∈ R2k, which
is then transformed into a complex-valued vextor xc ∈ Ck
with normalized power by combining the preamble segment
as the real part and the postamble segment as the imaginary
part. The signal xc is then transmitted through the wireless
channel. Here k represents the number of channel uses for
transmiting xc, and Hs and Ws denote the height and width
of the source image, respectively.
In this paper, we consider the additive white Gaussian
noise (AWGN) channel and the Rayleigh fading channel.
Specifically, there exists an unknown additive interference zc
in the wireless channel. Without loss of generality, we consider
that zc has the same dimension as xc and continuously affectsthe reception of xc. Therefore, for i = 1, 2, · · · , k, the i-th
element of the received signal yc is given by
yc,i =
p
Pxhx,i xc,i +
p
Pzhz,i zc,i + nc,i, (1)
where Px and Pz are the transmit power of the desired signal
and the interference, respectively. Under the Rayleigh fading
channel, the gains hx,i and hz,i are independent and identically
distributed (i.i.d) as CN(0, 1), while both gains are unity under
the AWGN channel. nc,i ∼ CN(0, σ2) are i.i.d AWGN sample.
The received signal yc is then equalized and represented
as a real vector y ∈ R2k. We consider that the receiver
knows the channel state hx = [hx,1, hx,2, · · ·, hx,k] but not
hz = [hz,1, hz,2, · · ·, hz,k]. Therefore, in this paper, we apply
the classic minimum mean square error (MMSE) equalizer
under the Rayleigh fading channel. After equalization, ICDM
is applied to y to cancel the interference, and the resulting
signal is fed into the JSCC decoder for reconstruction as ˆs.
B. Theorem for Interference Cancellation
We formulate the interference cancellation task as a MAP
estimation problem. Specifically, we seek estimates ˆx, ˆz that
maximize the posterior probability of x and z given the
observations and known channel state, i.e.,
(ˆx,ˆz) = arg max
x,z
log px,z|y,hx (x, z|y, hx), (2)
where z =
Re(hzzc)
Im(hzzc)
. To illustrate the effectiveness of the
formulation, we first clarify their real-valued relationship.Lemma 1. The real-valued vectors x, y and z satisfy
y =
p
PxWxx +
p
PzWzz +Wnn, (3)
where n ∼ N(0, σ2
2 I2k). For the Rayleigh fading channel with
MMSE equalizer, H = diag(
|hx|
|hx|
), Hr
x = diag(Re(hx)),
HI
x = diag(Im(hx)), Wz =
Hr
x,HI
x
−HI
x,Hr
x
(H2 + σ2I)−1,
Ws = H2(H2 + σ2I)−1, and Wn = H(H2 + σ2I)−1. For
the AWGN channel, Ws =Wz =Wn = I2k.
Proof. For the Rayleigh fading channel with MMSE equalization,
the i-th symbol of the equalized signal yeq,i is
yeq,i =
hH
x,i
|hx,i|2 + σ2 yc,i =
√
PxhH
x,ihx,i
|hx,i|2 + σ2 xc,i
+
√
PzhH
x,ihz,i
|hx,i|2 + σ2 zc,i +
hH
x,i
|hx,i|2 + σ2 nc,i. (4)
The coefficient hH
x,ihx,i
|hx,i|2+σ2 = |hx,i|2
|hx,i|2+σ2 is a real and hence
extends directly to the full vector x, yielding the matrix Ws.
For the second term, both hz,i and zc,i are unknown, so
we group them into the vector z and express the product
hH
x,ihz,izc,i in block-matrix form. This expression then extends
naturally to the full vector z, yielding the matrixWz. The third
term contains the Gaussian noise sample nc,i. By applying
the resampling trick, in which the hH
x,inc,i is equivalent to
|hx,i|nc,i, we obtain the matrix WnFor the AWGN channel, the equalized signal coincides
exactly with the received signal, thus we can easily derive
the expression.
Lemma 1 illustrates the structure of z in (2). By combining
the unknown channel gains hz into the interference signal
zc, the coefficient matrices Wx, Wz and Wn depend only
on the known gains hx, thereby eliminating all unknown
multiplication factors. Therefore, the problem reduces to one
of additive interference. We then introduce two assumptions
for theoretical analysis.
Assumption 1. The transmitted signal x and the interference
z are independent.
Assumption 2. The functions log px(x) and log pz(z) are
each locally strongly convex, with convexity parameters μ > 0
and ν > 0, respectively. Moreover, the ground truth x∗ and
z∗, which satisfy y =
√
PxWxx∗ +
√
PzWzz∗ +Wnn, are
local optimums of px(x), pz(z). 不太理解这些
最新发布