Linear MIMO Detection
It is considered that the MIMO system is a point to point MIMO without
CSI (or alternatively a Multiuser MIMO Uplink with single antennas at
the transmitters). With this, there is no precoding the transmitter(s). In
case of linear MIMO detection the MIMO channel is equalized with a linear
Filter G. It is assumed that the input symbols are taken from a discrete set
xk ∈ A. The output of the linear equalizer is described by
x̃ = G (Hx + w) , (1)
where x̃ ∈ CNt . Subsequently the symbol can be mapped to the closest
possible solution
x̃ → x̂, (2)
where x̂ ∈ ANt .
Moreover, it is considered that
E wxH = 0
(3)
H
Cx = E xx (4)
Cw = E wwH
(5)
Linear MMSE Filtering
The linear MMSE Filter corresponds to the following design criterion
n o
GMMSE = argminG E kx − x̃k22 , (6)
where
n o n o
E kx − x̃k22 = E k(INt − GH) x − Gwk22 (7)
n o n o
= E k(INt − GH) xk22 − E kGwk22 (8)
n o
= tr (INt − GH) Cx (INt − GH)H − tr GCw GH . (9)
The corresponding problem is convex, such that we can set the derivative
with respect to G equal to zero.
n o
dE kx − x̃k22
= − (INt − GH) Cx HH + GCw = 0 (10)
dG∗
Finally the linear MMSE filter is given by
−1
GMMSE = Cx HH HCx HH + Cw . (11)
By using the matrix inversion Lemma, it can be rewritten as
−1 H −1
GMMSE = C−1 H −1
x + H Cw H H Cw (12)
In the low SNR region, when C−1
w has very small entries the filter can be
approximated as follows
lim GMMSE → Cx HH C−1
w , (13)
SNR→0
which corresponds to the ‘matched filter’ solution (described later). In the
high SNR region, when C−1w has very large entries the filter can be approx-
imated as follows
−1 H −1
lim GMMSE → HH C−1 w H H Cw , (14)
SNR→∞
which corresponds to the ‘zero forcing filter’ solution (described later). The
2I Etx
common special case is given when for Cw = σw Nr and Cx = Nt INt . In
this case the MMSE filter is described by
2
−1
Nt σw H
GMMSE = INt + H H HH (15)
Etx
Linear Zero-Forcing Filtering
The zero-forcing approach implies that the equalizer is unbiased, which can
be expressed by
E {x̃|x} = x. (16)
The expected value of the estimator conditioned on the input symbol is
given by
E {x̃|x} = GHx = x, (17)
which implies that
GH = I. (18)
The constraint for the unbiasedness can be combined with the MMSE de-
sign criterion, such that the zero-forcing filter corresponds to the following
optimization problem
n o
GZF = argminG E kx − x̃k22 s.t. GH = I. (19)
Inserting the expression for the filter output yields
n o
GZF = argminG E kx − G (Hx + w)k22 s.t. GH = I. (20)
The expected value can be expressed with the trace operator as follows
n o n o
E kx − GHx − Gwk22 = tr (I − GH) Cx (I − GH)H + tr GCw GH .
(21)
Together with the zero-forcing constraint, GH = I, the expression reduces
to
n o
E kx − x̃k22 = tr GCw GH .
(22)
With this and the zero-forcing constraint the Lagrangian function can be
cast as
L (G, Λ) = tr GCw GH + 2Re {tr {Λ (I − GH)}}
= tr GCw GH + tr {Λ (I − GH)} + tr I − HH GH ΛH
= tr GCw GH + tr {Λ (I − GH)} + tr ΛH − HH GH ΛH
= tr GCw GH + tr {Λ (I − GH)} + tr ΛH − tr HH GH ΛH
= tr GCw GH + tr {Λ (I − GH)} + tr ΛH − tr ΛH HH GH ,
(23)
where the entries of Λ are the Lagrangian multipliers. The derivative of G∗
yields
dL(G, Λ)
= GCw − ΛH HH = 0. (24)
dG∗
Rearranging the (24) yields
G = ΛH HH C−1
w . (25)
Plugging (25) into the zero-forcing constraint (18) yields
−1
ΛH = HH C−1
w H . (26)
Inserting (26) into (25) results in the zero-forcing solution given by
−1 H −1
GZF = HH C−1
w H H Cw . (27)
2 I. In this case the zero-
A special case is given for the case when Cw = σw
foring receive filter is given by
−1 H
GZF = HH H H . (28)
Linear Matched Filtering
The matched filter approach has the objective to maximize the signal-to-
noise ratio (SNR) at the receiver. A measure for the SNR can be expressed
by
H 2
E x x̃
γ (G) = . (29)
Var {xH x̃}
2 n o n o
With E xH x̃ = |tr {GHCx }|2 and Var xH x̃ = E kxk22 E kGwk22
the SNR can be rewritten as
|tr {GHCx }|2
γ (G) = . (30)
tr {Cx } tr {GCx GH }
The filter G which maximizes γ corresponds to the matched filter
GMF = argmaxG γ (G) . (31)
The derivative with respect to G∗ yields
dγ(G) tr {GHCx } H |tr {GHCx }|2
= Cx H − GCw = 0
dG∗ tr {Cx } tr {GCw GH } tr {Cx } tr2 {GCx GH }
(32)
Finally the matched filter can be expressed as
GMF = αCx HH C−1
w , (33)
where α is a scalar which can be chosen arbitrarily. For the special case with
Cx = ENTx
t
I and Cw = σw 2 I the expression for the matched filter simplies to
GMF = α0 HH (34)
Linear MMSE Detection with 1-bit ADCs
In this special section it is assumed that the receiver is equipped with 1-bit
ADCs. This implies that the received signal is described by
y = Q {z}
= Q {Hx + w}
= sign (Re {Hx + w}) + j sign (Im {Hx + w}) (35)
where Q {·} describes the 1-bit quantization. The linear MMSE filter corre-
sponds to the following design objective
n o
GMMSE,1-bit = argminG E kx − x̃k22
n o
= argminG E kx − Gyk22 . (36)
The objective function can be expressed as follows
n o
E kx − Gyk22 = E xH − yH GH (x − Gy)
= E xH x − E xH Gy − E yH GH x + E yH GH Gy .
(37)
Only the last two terms on the right hand side of (37) depend on G∗ , which
can be rewitten as
−E yH GH x + E yH GH Gy = −tr E yH GH x + tr E yH GH Gy
= −tr E xyH GH + tr E GyyH GH
= −tr Cx,y GH + tr GCy GH
(38)
The derivation with respect to G∗ and equating it to zero yields
n o
dE kx − Gyk22
= GCy − Cx,y = 0. (39)
dG∗
With this the optimal filter is given by
GMMSE, 1-bit = Cx,y C−1
y . (40)
In the sequel it is described how to compute the matrices Cx,y and Cy .
According to the Bussgang decomposition the output of a nonlinear function,
in this case the 1-bit quantization can be described with a linear function and
an uncorrelated noise term, when the input signal is Gaussian distributed.
Assuming that the input of the quantizer is Gaussian distributed, the output
of the quantization can be described by
y = Az + wQ , (41)
where wQ represents the quantization noise which is uncorrelated with Az.
Inserting z = Hx + w in (41) yields
y = A (Hx + w) + wQ . (42)
With this, the crosscorrelation matrix can be represented by
Cx,y = E xyH = E x xH HH AH + wH AH + wQ H
= E xxH HH AH + E xwH AH + E xwQ
H
= Cx HH AH . (43)
The crosscorrelation matrix between 1-bit quantizer input and 1-bit quan-
tizer output is given by
r
2 1
Cy,z = diag {Cz }− 2 Cz , (44)
π
and
r
2 1
Cz,y = Cz diag {Cz }− 2 . (45)
π
According to the Bussgang decomposition the crosscorrelation matrix can
be also described by
n o
Cz,y = E zyH = E z (Az + wQ )H
= E zzH AH
= Cz AH . (46)
Equating (45) and (46) yields
r
2 1
Cz diag {Cz }− 2 = Cz AH , (47)
π
which implies that
r
2 1
A= diag {Cz }− 2 . (48)
π
Inserting (48) into (43) yields
r
2 1
Cx,y = Cx HH diag {Cz }− 2 . (49)
π
The 1-bit quantizer output covariance matrix can be obtained via the Price
theorem (arcsin law) as
2 1 1
Cy = arcsin diag {Cz }− 2 Re {Cz } diag {Cz }− 2
π
1 1
+ j arcsin diag {Cz }− 2 Im {Cz } diag {Cz }− 2 (50)
and the covariance matrix for the unquantized signal is given by
Cz = E zzH = HCx HH + Cw .
(51)
Inserting into
r
π − 1
Cx HH diag HCx HH + Cw 2
GMMSE, 1-bit =
2
− 1
arcsin diag HCx HH + Cw 2 Re HCx HH + Cw
H
− 12
diag HCx H + Cw
− 1
+ j arcsin diag HCx HH + Cw 2 Im HCx HH + Cw
−1
H
− 12
diag HCx H + Cw . (52)
