Critical Summary "Generalized Tangential Sphere Bound On The ML Decoding Error Probability of Linear Binary Block Codes in AWGN Interference"
Critical Summary "Generalized Tangential Sphere Bound On The ML Decoding Error Probability of Linear Binary Block Codes in AWGN Interference"
Critical Summary "Generalized Tangential Sphere Bound On The ML Decoding Error Probability of Linear Binary Block Codes in AWGN Interference"
On
"Generalized Tangential Sphere Bound on the ML Decoding Error Probability
of Linear Binary Block Codes in AWGN Interference"
By Shahram Yousefi and Amir K. Khandani.
Submitted by Einat Tevel (037339710)
Preface
As was observed during the past years, reaching a closed and convenient form for the
error performance of coded communications schemes rarely succeeds. The
alternative is to find tight analytical upper (and lower) bounds and to use them as
tools to estimate codes performances. Bounds such as the Fano and Gallager bounds
were introduced during the 1960's for evaluating the error exponents of the ensambles
of random codes. A most commonly used upper bound is the Union Bound (UB)
which is a general upper bound for error probability in any digital comm. System:
⎛ n ⎞ n
Pr ⎜ ∪ Ai ⎟ ≤ ∑ Pr ( Ai ) (1) for n events { Ai }i =1
n
⎝ i =1 ⎠ i =1
The UB was found to be quite accurate for high signal to noise ratios (SNRs) while
for long block codes with many codewords it becomes useless at low SNR values. It
also becomes useless at rates above the cutoff rate of the channel. The area between
the cutoff rate and the channel capacity (near Shannon's limit) is the interesting one,
as we want to estimate the performance of capacity approaching codes (such as turbo
codes, LDPC codes etc.) This motivates the search for other upper bounds for non-
random codes. The discussed paper continues previous works on establishing tight
upper bounds on the error probability of Maximum Likelihood (ML) decoding of
linear block codes in additive white Gaussian noise (AWGN) interference. The basic
concept in these works derives from what is referred to as "Gallager's first bounding
technique" (GFBT). Given a received signal r , the word error probability can be
written as Pw ( E ) = P { E , r ∈ ℜ} + P { E , r ∉ ℜ} ≤ P { E , r ∈ ℜ} + P {r ∉ ℜ} (2) where ℜ is
a "Gallager Region", namely an appropriate region chosen around the transmitted
signal point. The idea is to use the UB to upper bound the first term (the joint event
where there is a decoding error and the received signal vector falls in region ℜ ), while
the second term (which is dominant in very low SNRs) is calculated only once (not
using the UB). That leaves us with the choice of the region ℜ which is the significant
component in this bounding technique.
Overview
Following is a short review of bounds using the GFBT:
Tangential Bound (TB) of Berlekamp (1980): Choosing ℜ to be a half space
and separating the radial and tangential components of the Gaussian noise.
Sphere Bound (SB) of Herzberg (1994): Choosing ℜ to be a hyper-sphere
with radius r and then finding the optimization with regards to r .
Tangential Sphere Bound (TSB) of Poltyrev (1994): Choosing ℜ to be a
conical region.
It can be proven that the TB is not tighter than the TSB (since the TB is a private case
of the TSB when r → ∞ , r being the conical radius) and that the UB is not tighter
than the both of them. It was later shown that the TSB is, de facto, one of the tightest
bounds to date.
We also note that all the above regions have azimuthal symmetry along the radial
axis.
In this work Yousefi & Khandani (Y&K) extend the TSB to a General TSB (GTSB)
by generalizing the Gallager region to a generic one, which encompass all the above
cases. They apply a hyper-surface of revolution to the above scheme and show that
the optimal Gallager region is a right-circular hyper-cone, which coincides with the
TSB of Poltyrev. Consequently, most of their analytical calculations are the same as
the ones established in the development of the TSB, whereas at the end they apply
their optimization to reach their conclusions. It is therefore necessary to present a
brief overview on the TSB derivation.
signal set is S = {s0 , s1 … s2k −1} with Es being the symbol energy ( Es = R ⋅ Eb = ⋅ Eb ).
k
n
All the transmitted signals are interpreted as points on an n-dimensional sphere with
center at the origin and radius rc = n ⋅ Es . In this scheme the Euclidean distance is
proportional to the Hamming distance so that δ ij 2 = δ 2 ( si , s j ) = 4 ⋅ Es ⋅ d (ci , c j ) for two
signals si and s j . The output of the channel is r = si + n where n is an n-dimensional
vector whose elements are independent zero mean Gaussian random variables with
variance σ 2 (i.e. n = (n0 , n1 … nn −1 ), ni ∼ N (0, σ 2 ) ). The probability of a word error
2k −1
is Pw ( E ) = ∑ p( E | si ) ⋅ p( si ) , and for geometrically uniform (GU) signal sets (which
i =0
where Ek is the error event that the received vector r is closer to sk than to s0
n
(namely, pair-wise error event), y = ∑ zi2 is a random variable with Chi-square
i =2
distribution with (n − 1) degrees of freedom, and β k ( z ), r ( z1 ) can be calculated from
the figure (taken from the paper discussed):
Next, the tangential component z2 is extracted from the noise vector and we can write
in (4) : P ⎡⎣ Ek | z1 , y ≤ r 2 ( z1 ) ⎤⎦ = p ⎡⎣ β k ( z1 ) < z2 <| r ( z1 ) |, y1 ≤ (r 2 ( z1 ) − z22 ) ⎤⎦ (5)
, where
n
y1 = ∑ zi2 is a random variable with Chi-square distribution with ( n − 2) degrees of
i =3
freedom, whose pdf is known. Combining (3) , (4) and (5) we receive the TSB result: (6)
∞ ⎡ | r ( z1 )| r 2 ( z1 ) − z22 ∞ ⎤
Pw ( E ) ≤ ∫ ⎢ ∑ ( Ak ⋅ ∫ f z2 ( z2 ) ⋅ ∫ f y1 ( y1 )dy1 ⋅ dz2 ) + ∫ f y ( y )dy ⎥ f z1 ( z1 )dz1
−∞ ⎢
⎣ k :βk ( Z1 )<|r ( z1 )| β k ( Z1 ) 0 r 2 ( z1 ) ⎥⎦
The optimization
Defining the right hand side term in (6)
to be F [ r ( z1 ) ] we are obviously interested in
the minimization of it over r ( z1 ) , to achieve the tightest upper bound possible. K&Y
obtain this by using the fact that the functional of F [ r ( z1 ) ] will yield a stationary
point if ∂F [ r ( z1 ) + ε h( z1 ) ] / ∂ε |ε = 0 is zero for all choices of h( z1 ) . They apply this (see
∞
first remark) and end up with an expression simplified to the form: ∫ g ( z ) ⋅ h( z )dz .
−∞
1 1 1
Since they need it to equal zero for all h ( ⋅) they conclude that g ( z1 ) must equal zero
for all z1 . In the terms of the real equation optimized this concludes to (see second
( )
remark): r ( z1 ) = r0 ⋅ z1 − n where r0 is a constant, yielding a conical shaped region.
The optimizations equation obtained is the same as for the TSB of Poltyrev.
Results
Yousefi & Khandani showed that for a linear block code with a BPSK modulation of
equi-energy signals with soft decision ML decoding in AWGN channel, the best
decision region is a polyhedral cone with a single vertex at the origin of the n-space
and unboundedly extending in one (radial) direction. The application of their result is
not limited to either linear codes or binary alphabets or even GU constellations. The
only property necessary is for the signal set to be equi-energy (i.e. sphere code) or
else β k ( z1 ) will not receive its form. In case of a non-linear code, the signal spectrum
Ak depends on the transmitted codeword, and thus the application of the results is for
a specific code with a specific spectrum (i.e. finding the optimal bound for a specific
code). This is not mentioned in the paper.
This might be the place to mention that an important result of Sason & Shamai, which
didn't appear in the original paper of Poltyrev, is that the optimization does not
depend on the signal to noise ratio (SNR) but depends solely on the signal spectrum,
thus allowing us to perform the optimization only once, and then apply it on different
SNRs.
Another important result of Sason & Shamai, which is referred to in this paper, is that
in fact there exists another component in the bound for the case that z1 > n ⋅ Es . In
this case β k ( z1 ) < 0 and we are on the lower nappe of the cone. The component that
⎛ 2nEs ⎞
should be added to the original TSB (and is not) is Q ⎜⎜ ⎟⎟ . They also show that
⎝ N 0 ⎠
this component has a negligible effect on the bound and therefore is ignored.
Remarks
On a more critical note, I have two critiques regarding the optimization performed in
this paper:
The first one is regarding the use of equation [16] in the paper in the derivation
v( x)
d
of F [ r ( z1 ) ] . In [16] Y&K present a technique for performing
dx u (∫x )
f (t , x)dt . They
apply it by changing the order of the summation and the derivative in [17] .When
simplifying the formula it can be viewed as performing:
| x| | x|
d d
∑
dx k :βk ( z1 )<| x| β k ∫( z1 )
f (t , x)dt = ∑ ∫ f (t , x)dt where x = r ( z1 ) . This is a rather
k :β k ( z1 ) <| x| dx β k ( z1 )
References
In this work I used the following material:
Papers mentioned in the References of the discussed paper: [16], [17], [19].
"Performance Analysis of Linear Codes under Maximum-Likelihood
Decoding: A Tutorial" by Igal Sason and Shlomo Shamai. CCIT Report #569
December 2005
Material from the course "048934 - Coded communications" given by Prof.
Neri Merhav.