Proof
lemma 1
含偶数个1的概率
假设：
f
(
x
)
=
∏
l
=
1
m
[
(
1
−
p
k
)
+
p
k
x
]
f\left( x\right) =\prod^{m}_{l=1} \left[ \left( 1-p_{k}\right) +p_{k}x\right]
f(x)=∏l=1m[(1−pk)+pkx]展开二项式，
x
2
k
x^{2k}
x2k项的系数就是出现2k个1的概率，展开式中偶数项和就是m长二元序列中含有偶数个1的概率。根据二项式的性质：偶数项的系数和
a
0
+
a
2
+
a
4
+
a
6
+
⋯
⋯
=
f
(
1
)
+
f
(
−
1
)
2
=
∏
k
=
1
m
[
(
1
−
p
k
)
+
p
k
]
+
∏
k
=
1
m
[
(
1
−
p
k
)
−
p
k
]
2
=
1
+
∏
k
=
1
m
[
1
−
2
p
k
]
2
\begin{matrix}a_{0}+a_{2}+a_{4}+a_{6}+\cdots \cdots &=\frac{f\left( 1\right) +f\left( -1\right) }{2} =\frac{\prod^{m}_{k=1} \left[ \left( 1-p_{k}\right) +p_{k}\right] +\prod^{m}_{k=1} \left[ \left( 1-p_{k}\right) -p_{k}\right] }{2} \\ &=\frac{1+\prod^{m}_{k=1} \left[ 1-2p_{k}\right] }{2} \\ \end{matrix}
a0+a2+a4+a6+⋯⋯=2f(1)+f(−1)=2∏k=1m[(1−pk)+pk]+∏k=1m[(1−pk)−pk]=21+∏k=1m[1−2pk]
含奇数个1的概率
1 − f ( 1 ) + f ( − 1 ) 2 = 1 − ∏ k = 1 m [ 1 − 2 p k ] 2 1-\frac{f\left( 1\right) +f\left( -1\right) }{2} =\frac{1-\prod^{m}_{k=1} \left[ 1-2p_{k}\right] }{2} 1−2f(1)+f(−1)=21−∏k=1m[1−2pk]
校验节点
校验节点将收集的变量节点信息（除
x
i
x_i
xi）校验结果概率发送给变量节点
x
i
x_i
xi
r
j
i
(
0
)
=
1
2
+
1
2
∏
i
′
∈
V
j
\
i
(
1
−
2
q
i
′
j
(
1
)
)
r_{ji}(0)=\frac{1}{2} +\frac{1}{2} \prod_{i^{\prime }\in V_{j}\backslash i} \left( 1-2q_{i^{\prime }j}(1)\right)
rji(0)=21+21i′∈Vj\i∏(1−2qi′j(1))
r
j
i
(
1
)
=
1
2
−
1
2
∏
i
′
∈
V
j
\
i
(
1
−
2
q
i
′
j
(
1
)
)
r_{ji}(1)=\frac{1}{2} -\frac{1}{2} \prod_{i^{\prime }\in V_{j}\backslash i} \left( 1-2q_{i^{\prime }j}(1)\right)
rji(1)=21−21i′∈Vj\i∏(1−2qi′j(1))
变量节点
变量节点接受来自校验节点的信息，发送给校验节点 f j f_j fj
q i j ( 0 ) = ( 1 − P i ) ∏ j ′ ∈ C i \ j r j ′ i ( 0 ) q_{ij}(0)=\left( 1-P_{i}\right) \prod_{j^{\prime }\in C_{i}\backslash j} r_{j^{\prime }i}(0) qij(0)=(1−Pi)j′∈Ci\j∏rj′i(0)
q i j ( 1 ) = P i ∏ j ′ ∈ C i \ j r j ′ i ( 1 ) q_{ij}(1)=P_{i}\prod_{j^{\prime }\in C_{i}\backslash j} r_{j^{\prime }i}(1) qij(1)=Pij′∈Ci\j∏rj′i(1)
lemma 2
p
(
x
i
=
x
|
y
)
=
p
(
x
i
=
x
,
y
)
p
(
y
)
=
p
(
y
|
x
i
=
x
)
p
(
x
i
=
x
)
p
(
x
i
=
1
)
p
(
y
|
x
i
=
1
)
+
p
(
x
i
=
−
1
)
p
(
y
|
x
i
=
−
1
)
p\left( x_{i}=x\middle| y\right) =\frac{p\left( x_{i}=x,y\right) }{p\left( y\right) } =\frac{p\left( y\middle| x_{i}=x\right) p\left( x_{i}=x\right) }{p\left( x_{i}=1\right) p\left( y\middle| x_{i}=1\right) +p\left( x_{i}=-1\right) p\left( y\middle| x_{i}=-1\right) }
p(xi=x∣y)=p(y)p(xi=x,y)=p(xi=1)p(y∣xi=1)+p(xi=−1)p(y∣xi=−1)p(y∣xi=x)p(xi=x)
代入已知条件可得：
p
(
y
|
x
i
=
x
)
p
(
y
|
x
i
=
1
)
+
p
(
y
|
x
i
=
−
1
)
=
1
2
π
σ
e
x
p
(
−
(
y
−
x
)
2
2
σ
2
)
1
2
π
σ
e
x
p
(
−
(
y
−
1
)
2
2
σ
2
)
+
1
2
π
σ
e
x
p
(
−
(
y
+
1
)
2
2
σ
2
)
=
e
x
p
(
x
y
σ
2
)
e
x
p
(
y
σ
2
)
+
e
x
p
(
−
y
σ
2
)
=
1
e
x
p
(
y
(
1
−
x
)
σ
2
)
+
e
x
p
(
−
y
(
1
+
x
)
σ
2
)
=
1
1
+
e
x
p
(
−
2
x
y
σ
2
)
\frac{p\left( y\middle| x_{i}=x\right) }{p\left( y\middle| x_{i}=1\right) +p\left( y\middle| x_{i}=-1\right) } =\frac{\frac{1}{\sqrt{2\pi } \sigma } exp\left( -\frac{\left( y-x\right)^{2} }{2\sigma^{2} } \right) }{\frac{1}{\sqrt{2\pi } \sigma } exp\left( -\frac{\left( y-1\right)^{2} }{2\sigma^{2} } \right) +\frac{1}{\sqrt{2\pi } \sigma } exp\left( -\frac{\left( y+1\right)^{2} }{2\sigma^{2} } \right) } \\ =\frac{exp\left( \frac{xy}{\sigma^{2} } \right) }{exp\left( \frac{y}{\sigma^{2} } \right) +exp\left( \frac{-y}{\sigma^{2} } \right) } =\frac{1}{exp\left( \frac{y\left( 1-x\right) }{\sigma^{2} } \right) +exp\left( -\frac{y\left( 1+x\right) }{\sigma^{2} } \right) } \\ =\frac{1}{1+exp\left( \frac{-2xy}{\sigma^{2} } \right) }
p(y∣xi=1)+p(y∣xi=−1)p(y∣xi=x)=2πσ1exp(−2σ2(y−1)2)+2πσ1exp(−2σ2(y+1)2)2πσ1exp(−2σ2(y−x)2)=exp(σ2y)+exp(σ2−y)exp(σ2xy)=exp(σ2y(1−x))+exp(−σ2y(1+x))1=1+exp(σ2−2xy)1
LLR formula proof
L
(
q
i
j
)
=
L
(
c
i
)
=
log
(
1
+
e
−
2
y
i
/
σ
2
)
−
1
(
1
+
e
2
y
i
/
σ
2
)
−
1
=
log
e
y
σ
2
(
e
y
σ
2
+
e
−
y
σ
2
)
e
−
y
σ
2
(
e
y
σ
2
+
e
−
y
σ
2
)
=
log
e
y
σ
2
e
−
y
σ
2
=
2
y
σ
2
\begin{array}{l}L\left( q_{ij}\right) =L\left( c_{i}\right) =\log \frac{\left( 1+e^{-2y_{i}/\sigma^{2} }\right)^{-1} }{\left( 1+e^{2y_{i}/\sigma^{2} }\right)^{-1} } \\ =\log \frac{e^{\frac{y}{\sigma^{2} } }\left( e^{\frac{y}{\sigma^{2} } }+e^{\frac{-y}{\sigma^{2} } }\right) }{e^{\frac{-y}{\sigma^{2} } }\left( e^{\frac{y}{\sigma^{2} } }+e^{\frac{-y}{\sigma^{2} } }\right) } =\log \frac{e^{\frac{y}{\sigma^{2} } }}{e^{\frac{-y}{\sigma^{2} } }} =\frac{2y}{\sigma^{2} } \end{array}
L(qij)=L(ci)=log(1+e2yi/σ2)−1(1+e−2yi/σ2)−1=logeσ2−y(eσ2y+eσ2−y)eσ2y(eσ2y+eσ2−y)=logeσ2−yeσ2y=σ22y
证明以下公式：
将校验节点中的公式改写为
2
r
j
i
(
0
)
=
1
+
∏
i
′
∈
V
j
\
i
(
1
−
2
q
i
′
j
(
1
)
)
⇒
1
−
2
r
j
i
(
1
)
=
∏
i
′
∈
V
j
\
i
(
1
−
2
q
i
′
j
(
1
)
)
\begin{aligned} &2 r_{j i}(0)=1+\prod_{i^{\prime} \in V_{j} \backslash i}\left(1-2 q_{i^{\prime} j}(1)\right) \\ &\Rightarrow 1-2 r_{j i}(1)=\prod_{i^{\prime} \in V_{j} \backslash i}\left(1-2 q_{i^{\prime} j}(1)\right) \end{aligned}
2rji(0)=1+i′∈Vj\i∏(1−2qi′j(1))⇒1−2rji(1)=i′∈Vj\i∏(1−2qi′j(1))
利用恒等公式
tanh
x
=
e
x
−
e
−
x
e
x
+
e
−
x
\tanh x=\frac{e^{x}-e^{-x}}{e^{x}+e^{-x}}
tanhx=ex+e−xex−e−x、
tanh
(
1
2
log
(
p
0
p
1
)
)
=
p
0
−
p
1
=
1
−
2
p
1
\tanh \left( \frac{1}{2} \log \left( \frac{p_{0}}{p_{1}} \right) \right) =p_{0}-p_{1}=1-2p_{1}
tanh(21log(p1p0))=p0−p1=1−2p1可得
tanh
(
1
2
L
(
r
j
i
)
)
=
tanh
(
1
2
log
(
r
j
i
(
0
)
r
j
i
(
1
)
)
)
=
1
−
2
r
j
i
(
1
)
=
∏
i
′
∈
V
j
\
i
(
1
−
2
q
i
′
j
(
1
)
)
=
∏
i
′
∈
V
j
\
i
tanh
(
1
2
L
(
q
i
′
j
)
)
\begin{array}{l}\tanh \left( \frac{1}{2} L\left( r_{ji}\right) \right) =\tanh \left( \frac{1}{2} \log \left( \frac{r_{ji}\left( 0\right) }{r_{ji}\left( 1\right) } \right) \right) \\ =1-2r_{ji}\left( 1\right) =\prod_{i^{\prime }\in V_{j}\backslash i} \left( 1-2q_{i^{\prime }j}(1)\right) \\ =\prod_{i^{\prime }\in V_{j}\backslash i} \tanh \left( \frac{1}{2} L\left( q_{i^{\prime }j}\right) \right) \end{array}
tanh(21L(rji))=tanh(21log(rji(1)rji(0)))=1−2rji(1)=∏i′∈Vj\i(1−2qi′j(1))=∏i′∈Vj\itanh(21L(qi′j))
得证
tanh ( 1 2 L ( r j i ) ) = ∏ i ′ ∈ V j \ i tanh ( 1 2 L ( q i ′ j ) ) \tanh \left(\frac{1}{2} L\left(r_{j i}\right)\right)=\prod_{i^{\prime} \in V_{j} \backslash i} \tanh \left(\frac{1}{2} L\left(q_{i^{\prime} j}\right)\right) tanh(21L(rji))=i′∈Vj\i∏tanh(21L(qi′j))
同时可以得到
L
(
r
j
i
)
=
2
tanh
−
1
(
∏
i
′
∈
V
j
\
i
tanh
(
1
2
L
(
q
i
′
j
)
)
)
L\left(r_{j i}\right)=2 \tanh ^{-1}\left(\prod_{i^{\prime} \in V_{j} \backslash i} \tanh \left(\frac{1}{2} L\left(q_{i^{\prime} j}\right)\right)\right)
L(rji)=2tanh−1⎝⎛i′∈Vj\i∏tanh(21L(qi′j))⎠⎞
ϕ \phi ϕ函数
我们可以
L
(
q
i
′
j
)
=
α
i
j
β
i
j
L\left( q_{i^{\prime }j}\right) =\alpha_{ij} \beta_{ij}
L(qi′j)=αijβij来表示，其中
a
i
j
≡
s
i
g
n
(
L
(
q
i
j
)
)
β
i
j
≡
∣
L
(
q
i
j
)
∣
\begin{array}{l}a_{ij}\equiv sign\left( L\left( q_{ij}\right) \right) \\ \beta_{ij} \equiv \left| L\left( q_{ij}\right) \right| \end{array}
aij≡sign(L(qij))βij≡∣L(qij)∣
上式等于
tanh
(
1
2
L
(
r
i
j
)
)
=
∏
i
′
α
i
j
⋅
∏
i
′
tanh
(
1
2
β
i
′
j
)
\tanh \left(\frac{1}{2} L\left(r_{i j}\right)\right)=\prod_{i^{\prime}} \alpha_{i j} \cdot \prod_{i^{\prime}} \tanh \left(\frac{1}{2} \beta_{i^{\prime} j}\right)
tanh(21L(rij))=i′∏αij⋅i′∏tanh(21βi′j)
我们可以得到：
L
(
r
j
i
)
=
(
∏
i
′
a
i
j
)
×
2
tanh
−
1
log
−
1
log
⏞
∏
i
′
tanh
(
1
2
β
i
′
j
)
=
(
∏
i
′
a
i
j
)
×
2
tanh
−
1
log
−
1
∑
i
′
log
tanh
(
1
2
β
i
′
j
)
=
(
∏
i
′
α
i
j
)
ϕ
(
∑
i
′
∈
R
j
/
i
ϕ
(
β
i
′
j
)
)
\begin{aligned}\begin{array}{ll}L\left( r_{ji}\right) &=\left( \prod_{i^{\prime }} a_{ij}\right) \times 2\tanh^{-1} \overbrace{\log^{-1} \log } \prod_{i^{\prime }} \tanh \left( \frac{1}{2} \beta_{i^{\prime }j} \right) \\ &=\left( \prod_{i^{\prime }} a_{ij}\right) \times 2\tanh^{-1} \log^{-1} \sum_{i^{\prime }} \log \tanh \left( \frac{1}{2} \beta_{i^{\prime }j} \right) \\ &=\left( \prod_{i^{\prime }} \alpha_{ij} \right) \phi \left( \sum_{i^{\prime }\in R_{j/i}} \phi \left( \beta_{i^{\prime }j} \right) \right) \end{array} \end{aligned}
L(rji)=(∏i′aij)×2tanh−1log−1log
∏i′tanh(21βi′j)=(∏i′aij)×2tanh−1log−1∑i′logtanh(21βi′j)=(∏i′αij)ϕ(∑i′∈Rj/iϕ(βi′j))
其中
ϕ
(
x
)
≡
−
log
tanh
(
1
2
x
)
=
log
e
x
+
1
e
x
−
1
\phi \left( x\right) \equiv -\log \tanh \left( \frac{1}{2} x\right) =\log \frac{e^{x}+1}{e^{x}-1}
ϕ(x)≡−logtanh(21x)=logex−1ex+1
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
