Professional Documents
Culture Documents
net/publication/224246031
Conference Paper in Acoustics, Speech, and Signal Processing, 1988. ICASSP-88., 1988 International Conference on · June 2011
DOI: 10.1109/ICASSP.2011.5946710 · Source: IEEE Xplore
CITATIONS READS
6 379
4 authors, including:
Jan-Åke Larsson
Linköping University
103 PUBLICATIONS 2,394 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Jan-Åke Larsson on 18 August 2014.
fulfills the so called symmetry condition is common, see [2]. In the is perfectly known. In what follows, we assume NR ≥ NT since it
simple case with a binary additive white Gaussian noise (BIAWGN) simplifies the mathematics performed in this paper and is typical in
channel, the LLR distribution conditioned on the transmitted signal practice. Without loss of generality, assume BPSK (S = {−1, +1})
is well-known to be Gaussian. In a system with multiple inputs, such and consider detection of the first bit throughout this document. With
as MIMO, or with larger constellations, this is not the case though. uniform a priori probabilities, the LLR for the first bit is
However, it is common to use a Gaussian distribution in these cases ⎛ ⎞
1 2
as well, see [2] and [3]. The reason is its mathematical tractability i∈I1 exp − N0 y − Hsi
and in many cases applicability by the central limit theorem. Earlier Λ(y) log ⎝ ⎠, (2)
1 2
work regarding the LLR distribution consider only scalar channels, i∈I0 exp − N0 y − Hsi
see [3] and [4]. In [3], an expression of the asymptotic bit-error rate
is derived. In [4], an optimal linear approximation of the LLR values where the set Ix contains the indexes of symbol vectors si ∈ S NT
is derived in capacity sense for partial channel-state knowledge. for which the 1:st bit is equal to x ∈ {0, 1}. Before we continue, let
In MIMO detection, a full-blown LLR evaluation is computa- us look at the simple example mentioned in the introduction with a
tionally very expensive and the LLR values can not be computed ex- BIAWGN channel, i.e., NR = NT = 1,
actly. There are different detectors that approximate the LLR values
− N1 (y−h)2
in a computationally feasible fashion, such as the soft-output via par- p(y|s = +1) e 0 4yh
tial marginalization (PM) detector in [5], [6], and others in [7], [8]. Λ(y) = log = log = ,
p(y|s = −1) − N1 (y+h)2 N0
e 0
This work was supported in part by the Swedish Research Council (VR),
2 2
the Swedish Foundation of Strategic Research (SSF), and the Excellence which implies that, conditioned on s, Λ(y) ∼ N (± 4h , 8h ).
N0 N0
Center at Linköping-Lund in Information Technology (ELLIIT). E. G. Lars-
son is a Royal Swedish Academy of Sciences (KVA) Research Fellow sup- This is however not the case in MIMO nor in the scalar case with
ported by a grant from the Knut and Alice Wallenberg Foundation. The higher order constellations since (2) contains multiple summations.
Monte Carlo simulations in this paper were carried out at the National Su- In what follows, we present a slightly modified version of the
percomputer Centre in Linköping Sweden (www.nsc.liu.se). PM method in [5], see [6]. In order to explain PM, we use the fol-
Our aim is to obtain a Gaussian approximation of the distribution We can observe that:
of L(y) in (3a) conditioned on s, i.e., p(L(y)|s). To do this, we • The norm v 0 ≈ ||Π⊥ H̄(s̄ − s̄0 )|| is large when the
H̃
rewrite (3a) explicitly as a function of the noise e in a mathemati-
columns of H̄ are orthogonal to the column space of H̃. In
cally tractable fashion and linearize. Let us thus define
other words, the norm v 0 is likely to be large when H is
l(e) L(Hs + e), s̃zf (i, e) s̃ZF (i, Hs + e), (4a) well-conditioned.
• The spans of H̄ and of Π⊥ H̃
increase with r, which suggests
ˆs̃zf (i, e) ˆs̃ZF (i, Hs + e) = H̃ † (H̄s̄ + H̃ s̃ + e − H̄s̄i ) that v 0 and thus detection performance does the same.
† † ˆ
= s̃ + H̃ H̄(s̄ − s̄i ) + H̃ e, • The distribution
ofl(e) fulfills
thesymmetry condition in [9]
(4b) ˆ
since Vare l(e) = 2Ee l̂(e) , which is a fundamental
where s in (1) is known. Keep in mind that the functions with capital condition in density evolution in modern coding theory.
letters L and (·)ZF take the argument y whereas the same functions • The linear LLR in (6) can be associated with the binary de-
with small letters l and (·)zf take the argument e. Without loss of tection problem that distinguishes between the two symbol
generality, we have assumed that the true vector s̄ is found in the vectors s̄ and s̄0 . At high SNR, this is a good approximation
numerator of (3a) and that in the denominator, the vector s̄0 ∈ {s̄i : since the detector errors will mostly follow from confusing
i ∈ Ī0 } is closest to the true vector s̄ in terms of the metric v i , s̄ and s̄0 . This is equivalent to detection over a BIAWGN
channel, which in our case when ZF clipping is disregarded
v i H̄(s̄ − s̄i ) + H̃(s̃ − s̃zf (i, 0)) ∀i ∈ Ī0 ∪ Ī1 . (5)
consists of the two points Π⊥ H̃
H̄ s̄ and Π⊥
H̃
H̄s̄0 .
Recall that the intersection Ī0 ∩ Ī1 = ∅ is empty. The linearization • When we increase SNR, the constellation points in terms of
of l(e) in (4a) that we derive is LLR means are moving further away from each other faster
than what the LLR standard deviation is increasing, see (7).
v 0 2 2v T e
l̂(e) + 0 , (6) • According to Theorem 1, the approximation error is small
N0 N0 with a probability that goes exponentially fast to one, which
where v 0 argminvi ∀i∈Ī0 v i . We assume that v 0 is uniquely indicates analytically that the Gaussian distribution is a key
component in the LLR distribution.
defined1 with probability one. Note that the mean of ˆl(e) is positive
since s̄ is assumed to be found in the numerator of (3a). 1 Basically, the channel matrices that yield non-unique v lie in a space
0
with dimension strictly smaller than the whole space.
3233
CE = 27.3, r = 0 CE = 80.5, r = 0 CE = 27.3, r = 0 CE = 43.3, r = 0 ill-conditioned H (κ(H) ≈ 70)
rCE 27.3 43.3 61.4 80.5
p p p p
0 1.1327806 1.5195682 1.5345142 0.4181846
-60 -30 0 30 60 90 120 0 150 300 450 0 20 40 20 40 60 80 1 0.1184548 0.0383142 0.0176418 0.0020565
l̂(e) l̂(e) l̂(e) l̂(e)
2 0.1838109 0.0455121 0.0062638 0.0000569
CE = 27.3, r = 0 CE = 80.5, r = 0 CE = 27.3, r = 0 CE = 43.3, r = 0
3 0.1447524 0.0377483 0.0061711 0.0000563
p p p p 9 (ML) 0.1721505 0.0401805 0.0029932 0.0000073
-60 -30 0 30 60 90 120 0 150 300 450 0 20 40 20 40 60 80 well-conditioned H (κ(H) ≈ 4)
l(e) l(e) l(e) l(e) rCE 20.4 27.3 35 43.3
CE = 27.3, r = 1 CE = 80.5, r = 1 CE = 27.3, r = 1 CE = 43.3, r = 1
0 0.0187406 0.0029192 0.0000061 0.0000014
p p p p 1 0.0140127 0.0021493 0.0000002 0.0000021
2 0.0102622 0.0034864 0.0000469 0.0000036
-40 -20 0 20 40 60 80 100
l̂(e)
300 320
l̂(e)
340 360 0 20
l̂(e)
40 20 40
l̂(e)
60 80
3 0.0073968 0.0022437 0.0000260 0.0000007
CE = 27.3, r = 1 CE = 80.5, r = 1 CE = 27.3, r = 1 CE = 43.3, r = 1 9 (ML) 0.0054353 0.0001602 0.0000003 0.0000000
p p p p averaged over H
rCE 20.4 27.3 35 43.3
-40 -20 0 20 40 60 80 100
l(e)
300 320
l(e)
340 360 0
l(e)
20 40 20 40
l(e)
60 80
0 0.0614536 0.0597663 0.0563492 0.0520577
CE = 27.3, r = 9 CE = 80.5, r = 9 CE = 27.3, r = 9 CE = 43.3, r = 9 1 0.0360959 0.0325092 0.0312120 0.0305311
2 0.0254651 0.0207668 0.0191031 0.0186884
p p p p 3 0.0265090 0.0175610 0.0135940 0.0104095
0 20 40 180 200 220 0 20 40 20 40 60 80
4 0.0034151 0.0010697 0.0007791 0.0000195
l̂(e) l̂(e) l̂(e) l̂(e)
CE = 27.3, r = 9 CE = 80.5, r = 9 CE = 27.3, r = 9 CE = 43.3, r = 9
Table 1. Kullback Leibler Divergence for different SNR and r
p p p p values. Three different tables are presented for three different
0 20 40 180 200 220 0 20 40 20 40 60 80
settings: the ill-conditioned matrix H used in Fig. 1, the well-
l(e) l(e) l(e) l(e) conditioned matrix H used in Fig. 1, and an average over 3000
channel matrices
Hin Rayleighfading. The ergodic capacity
Fig. 1. Histogram comparison for l̂(e) and l(e). We have ran- CE EH log 2 det I + HN0H
T
is used as a proxy for SNR.
domly selected an ill-conditioned H with κ(H) ≈ 70 (the two
columns of histograms to the left with darker background) and a
well-conditioned H with κ(H) ≈ 4 (the two columns √ to the right).
The horizontal axis, l̂(e) and l(e), are scaled with N0 to keep the posed Gaussian distribution in Sec. 3. For small N0 , it converges in
variance of the histograms constant. The ergodic capacity defined in probability to this Gaussian distribution. The main idea behind the
Tab. 1 is used as a proxy for SNR. proposed linearization in (6) is essentially the log-max approxima-
tion. We can see that the PM distribution is multi-modal for small
r values when H is ill-conditioned, and that higher values of r or
One of the important practical benefits of (6) is that the minimization SNR are required for the LLR distribution to resemble a Gaussian
of interest in v 0 spans only over 2r possible candidate solutions and one than when H is well-conditioned. Additionally, on the average
can be efficiently solved. This simplifies numerical evaluations of over H, the KLD decays rapidly when increasing SNR or the pa-
the mutual information I(L(y); s), for instance. rameter r, which indicates that, in Rayleigh fading, the majority of
channel matrices are sufficiently well-conditioned for our approxi-
mation to be accurate. In Sec. 4, we give several interpretations on
5. NUMERICAL RESULTS the approximation we made in (6).
3234
vi
⎛ 2 ⎞
⎛ ⎞
− 1
H̄(s̄ − s̄i ) + H̃(s̃ − s̃zf (i, 0)) +e v i 2 +2v T
i e
⎜ i∈Ī1 exp ⎟ exp −
˜ = log ⎜ ⎟
N0 i∈Ī1 N0
l(e) ⎜ 2 ⎟ = log ⎝ ⎠
⎝ 1 ⎠ exp −
v i 2 +2v T
i
e
i∈Ī0 exp − H̄(s̄ − s̄i ) + H̃(s̃ − s̃zf (i, 0)) +e i∈Ī0 N0
N0
vi (8)
⎛ ⎞
v 2 +2v T e
v 0 2 2v T0 e 1 + i∈Ī1 \1 exp − i N0 i
= + + log ⎝ 2 2 T
⎠
N
0
N0
1 + i∈Ī0 \0 exp − vi −v0 N+2(v
0
i −v 0 ) e
Before we continue, let us define a more compact notation where: maximum does the same, d) P {A ∩ B} = P {B} − P{A ∩ B} for
events A and B, e) P {A ∪ B} ≤ P {A} + P {B} for events A and
ṽi v i 2 ∀i ∈ Ī1 , ẽi 2v Ti e ∀i ∈ Ī1 , B, f ) P {A ∩ B} ≤ P {A} for events A and B, g) the probabil-
2 2
ṽi v i − v 0 ∀i ∈ Ī0 , ẽi 2(v i − v 0 )T e ∀i ∈ Ī0 . ity P {e ∈ Z} goes exponentially fast to 1 with decreasing N0 and
Var {ẽi } is a multiple of N0 where is positive for small N0 . The
Now, a lower bound of P |l̂(e) − l(e)| < ∩ e ∈ Z is desired, rest follows from straight-forward calculations.
P |l̂(e) − l(e)| < ∩ e ∈ Z = P |l̂(e) − l̃(e)| < ∩ e ∈ Z B. REFERENCES
⎧ ⎛ ⎞ ⎫
⎨ 1 + i∈Ī1 \1 exp − ṽiN+ẽi
⎬ [1] M. Čirkić, D. Persson, and Erik G. Larsson, “Optimization of
= P log ⎝ ⎠ < ∩ e ∈ Z
0
computational resource allocation for soft MIMO detection,” in
⎩ 1+ exp − ṽi +ẽi ⎭ Proc. 43:rd Asilomar Conference on Signals, Systems and Com-
i∈Ī0 \0 N0
! puters, 2009, pp. 1488–1492.
(a)
≥ P log 1 + i∈Ī1 \1 exp − ṽiN+ẽ
0
i
[2] S.-Y. Chung, T.J. Richardson, and R.L. Urbanke, “Analysis of
" sum-product decoding of low-density parity-check codes using
a Gaussian approximation,” IEEE Transactions on Information
+ log 1 + i∈Ī0 \0 exp − ṽiN+ẽ i
<∩e∈Z
0
Theory, vol. 47, no. 2, pp. 657–670, Feb. 2001.
⎧ ⎫
(b) ⎨ # ⎬ [3] N. Sellami, A. Roumy, and I. Fijalkow, “A proof of convergence
≥P exp − ṽiN+ẽi
< ∩ e ∈ Z of the MAP turbo-detector to the AWGN case,” IEEE Transac-
⎩ 0 ⎭
i∈Ĩ(Ī1 ∪Ī0 )\{0,1} tions on Signal Processing, vol. 56, no. 4, pp. 1548–1561, Apr.
! " 2008.
≥ P max −(ṽi + ẽi ) < N0 log(/|Ĩ|) ∩ e ∈ Z [4] A. Alvarado, L. Szczecinski, R. Feick, and L. Ahumada, “Dis-
i∈Ĩ
! " tribution of L-values in Gray-mapped M 2 -QAM: Closed-form
≥ P max −ẽi < min ṽi + N0 log(/|Ĩ|) ∩e ∈ Z approximations and applications,” IEEE Transactions on Com-
i∈Ĩ i∈Ĩ munication, vol. 57, no. 7, pp. 2071–2079, Jul. 2009.
[5] E. G. Larsson and J. Jaldén, “Fixed-complexity soft MIMO de-
⎧ ⎫
⎨$ ⎬ tection via partial marginalization,” IEEE Transactions on Sig-
(c)
=P −ẽi < ∩ e ∈ Z nal Processing, vol. 56, no. 8, pp. 3397–3407, Aug. 2008.
⎩ ⎭
i∈Ĩ [6] D. Persson and E. G. Larsson, “Partial marginalization soft
⎧ ⎫
⎨% ⎬ MIMO detection with higher order constellations,” IEEE Trans-
(d)
= P {e ∈ Z} − P −ẽi > ∩ e ∈ Z actions on Signal Processing, vol. 59, no. 1, pp. 453–458, Jan.
⎩ ⎭ 2011.
i∈Ĩ
(e) # & ' [7] E. Viterbo and J. Boutros, “A universal lattice code decoder for
≥ P {e ∈ Z} − P −ẽi > ∩ e ∈ Z fading channels,” IEEE Transactions on Information Theory,
i∈Ĩ vol. 45, no. 5, pp. 1639–1642, Jul. 1999.
(f ) # & '
≥ P {e ∈ Z} − P −ẽi > [8] L.G. Barbero and J.S. Thompson, “Fixing the complexity of
the sphere decoder for MIMO detection,” IEEE Transactions
i∈Ĩ
on Wireless Communications, vol. 7, no. 6, pp. 2131–2142, Jun.
# (g) C
− N2 2008.
= P {e ∈ Z} − Q ( ≥ 1 − C1 e 0 ,
i∈Ĩ
Var {ẽi } [9] T.J. Richardson, M.A. Shokrollahi, and R.L. Urbanke, “De-
sign of capacity-approaching irregular low-density parity-check
where C1 and C2 are strictly positive constants. We have used the codes,” IEEE Transactions on Information Theory, vol. 47, no.
following mathematical relations: a) triangle inequality, b) log(1 + 2, pp. 619–637, Feb. 2001.
α) ≤ α for any α ≥ 0, c) all elements fulfill an < inequality iff their
3235