js散度(Jensen-Shannon散度)
公式
- 针对KL散度的不对称性,詹森和香农提出一种新的相对熵的计算方法,定义于两个概率分布上,根据KL散度构造,同样描述了两个概率分布的差异,且具有对称性
DJS(p∣∣q)=12DKL(p∣∣m)+12DKL(q∣∣m) D_{JS}(p||q) = \frac{1}{2}D_{KL}(p||m)+\frac{1}{2}D_{KL}(q||m) DJS(p∣∣q)=21DKL(p∣∣m)+21DKL(q∣∣m)
其中概率分布m为p和q的均值
m(x)=p(x)+q(x)2 m(x) = \frac{p(x)+q(x)}{2} m(x)=2p(x)+q(x)
性质
-
JS三度非负,JS散度越大,两个概率分布之间的差异越大
-
JS散度具有对称性
-
当且仅当两个概率分布相同时,JS散度等于0
∵m(x)=p(x)+q(x)2=p(x)=q(x)∴DJS(p∣∣q)=12DKL(p∣∣m)+12DKL(q∣∣m)=12DKL(p∣∣p)+12DKL(p∣∣p)=DKL(p∣∣p)=0 \begin{aligned} \because m(x) &= \frac{p(x)+q(x)}{2}=p(x)=q(x)\\ \therefore D_{JS}(p||q) &= \frac{1}{2}D_{KL}(p||m)+\frac{1}{2}D_{KL}(q||m) \\ &= \frac{1}{2}D_{KL}(p||p)+\frac{1}{2}D_{KL}(p||p)\\ &=D_{KL}(p||p)\\ &=0 \end{aligned} ∵m(x)∴DJS(p∣∣q)=2p(x)+q(x)=p(x)=q(x)=21DKL(p∣∣m)+21DKL(q∣∣m)=21DKL(p∣∣p)+21DKL(p∣∣p)=DKL(p∣∣p)=0