文章目录
AWGN信道模型
考虑一个随机变量 x ∼ p X ( x ) x \sim p_X(x) x∼pX(x),信道模型为
q = x + v , v ∼ N ( 0 , μ ) q = x + v, \ \ \ v \sim \mathcal N(0, \mu) q=x+v, v∼N(0,μ)
已知观测值 q q q,将后验估计的均值表示为 F i n ( q , μ ) = E [ x ∣ q ] F_{in}(q,\mu)=\mathbb E[x|q] Fin(q,μ)=E[x∣q],方差表示为 E i n ( q , μ ) = Var [ x ∣ q ] \mathcal E_{in}(q,\mu)=\text{Var}[x|q] Ein(q,μ)=Var[x∣q]。
后验均值与方差的关系
后验均值 F i n ( q , μ ) F_{in}(q,\mu) Fin(q,μ)与方差 E i n ( q , μ ) \mathcal E_{in}(q,\mu) Ein(q,μ)满足如下关系式
∂ ∂ q F i n ( q , μ ) = 1 μ E i n ( q , μ ) \frac{\partial}{\partial q} F_{in}(q,\mu)=\frac{1}{\mu} \mathcal E_{in}(q,\mu) ∂q∂Fin(q,μ)=μ1Ein(q,μ)
证明:对 μ > 0 \mu > 0 μ>0,定义函数
A r ( q ) = ∫ x r p X ( x ) ϕ ( q − x ; μ ) d x r = 0 , 1 , 2 , … A_r(q) = \int x^r p_X(x) \phi(q-x;\mu) \mathrm{d} x \ \ \ r=0,1,2,\ldots Ar(q)=∫xrpX(x)ϕ(q−x;μ)dx r=0,1,2,…
其中 ϕ ( q − x ; μ ) \phi(q-x;\mu ) ϕ(q−x;μ)表示似然分布 p Q ∣ X p_{Q|X} pQ∣X,均值为 x x x方差为 μ \mu μ的高斯分布,即
ϕ ( q − x ; μ ) ≡ N ( x , μ ) \phi(q-x;\mu ) \equiv \mathcal {N}(x, \mu) ϕ(q−x;μ)≡N(x,μ)
特殊地,先考虑 A 0 ( q ) A_0(q) A0(q)
A 0 ( q ) = ∫ p X ( x ) ϕ ( q − x ; μ ) d x = ∫ p X ( x ) p Q ∣ X ( q ∣ x ) d x = p Q ( q ) \begin{aligned} A_0(q) &= \int p_X(x) \phi(q-x;\mu) \mathrm{d} x \\ &= \int p_X(x) p_{Q|X}(q|x) \mathrm{d} x \\ &= p_Q(q) \end{aligned} A0(q)=∫pX(x)ϕ(q−x;μ)dx=∫pX(x)pQ∣X(q∣x)dx=pQ(q)
根据期望的定义,可以写出
F i n ( q , μ ) = A 1 ( q ) A 0 ( q ) F_{in}(q,\mu)=\frac{A_1(q)}{A_0(q)} Fin(q,μ)=A0(q)A1(q)
根据 Var[w] = E [ w 2 ] − E [ w ] 2 \text{Var[w]} =\mathbb E[w^2] - \mathbb E[w]^2 Var[w]=E[w2]−E[w]2,可以写出
E i n ( q , μ ) = A 2 ( q ) A 0 ( q ) − A 1 2 ( q ) A 0 2 ( q ) \mathcal E_{in}(q,\mu) = \frac{A_2(q)}{A_0(q)} - \frac{A^2_1(q)}{A^2_0(q)} Ein(q,μ)=A0(q)A2(q)−A02(q)A12(q)
对高斯分布求导可得
∂ ∂ q ϕ ( q − x ; μ ) = x − q μ ϕ ( q − x ; μ ) \frac{\partial}{\partial q} \phi(q-x;\mu) = \frac{x-q}{\mu} \phi(q-x;\mu) ∂q∂ϕ(q−x;μ)=μx−qϕ(q−x;μ)
基于此,再对 A r ( q ) A_r(q) Ar(q)求导可得
∂ ∂ q A r ( q ) = 1 μ ( A r + 1 ( q ) − q A r ( q ) ) \frac{\partial}{\partial q} A_r(q) = \frac{1}{\mu}(A_{r+1}(q)-qA_r (q)) ∂q∂Ar(q)=μ1(Ar+1(q)−qAr(q))
因此,
∂ ∂ q F i n ( q , μ ) = ∂ ∂ q A 1 ( q ) A 0 ( q ) = ( A 2 ( q ) − q A 1 ( q ) ) A 0 ( q ) − ( A 1 ( q ) − q A 0 ( q ) ) A 1 ( q ) μ A 0 2 ( q ) = A 2 ( q ) A 0 ( q ) − A 1 2 ( q ) A 0 2 ( q ) = 1 μ E i n ( q , μ ) \begin{aligned} \frac{\partial}{\partial q} F_{in}(q,\mu) &=\frac{\partial}{\partial q} \frac{A_1(q)}{A_0(q)} \\ &= \frac{(A_2(q) - qA_1(q))A_0(q) - (A_1(q) - qA_0(q))A_1(q)}{\mu A^2_0(q)} \\ &= \frac{A_2(q)}{A_0(q)} - \frac{A^2_1(q)}{A^2_0(q)} \\ &= \frac{1}{\mu} \mathcal E_{in}(q,\mu) \end{aligned} ∂q∂Fin(q,μ)=∂q∂A0(q)A1(q)=μA02(q)(A2(q)−qA1(q))A0(q)−(A1(q)−qA0(q))A1(q)=A0(q)A2(q)−A02(q)A12(q)=μ1Ein(q,μ)
证毕。
总结
其实直接想的话也不难想出来,只需要记住对积分外求导也就是对积分内求导。更重要的应该是直接记住这个结论。