ℹ

Differential Divergence

Definition (Differential divergence)

This is identical to divergence, where for RVs XX and YY with pdfs fX,fYf_{X}, f_{Y}, with supports SX⊂SY⊂RS_{X}\subset S_{Y}\subset \mathbb{R}. Then the divergence between XX and YY is D(X∥Y):=EfX[log⁡2fX(X)fY(Y)]=∫SXfX(t)log⁡2fX(t)fY(t)dt\begin{align*} D(X\|Y):&=E_{f_{X}}\left[\log_{2}\frac{f_{X}(X)}{f_{Y}(Y)}\right]\\\\ &=\int_{S_{X}}f_{X}(t)\log_{2}\frac{f_{X}(t)}{f_{Y}(t)}dt \end{align*} Multivariate version: D(Xn∥Yn):=∫SXnfXn(t1,⋯ ,tn)log⁡2fXn(t1,⋯ ,tn)fYn(t1,⋯ ,tn)dt1⋯dtnD(X^{n}\|Y^{n}):=\int_{S_{X^{n}}}f_{X^{n}}(t_{1},\cdots,t_{n})\log_{2}\frac{f_{X^{n}}(t_{1},\cdots,t_{n})}{f_{Y^{n}}(t_{1},\cdots,t_{n})}dt_{1}\cdots dt_{n}

Remark

For continuous RVs X∼fXX\sim f_{X} and Y∼fYY\sim f_{Y}, with SX⊂SY⊂RS_{X}\subset S_{Y}\subset\mathbb{R} for nn and mm sufficiently large lim⁡n→∞D([X]n∥[Y]n)=∫SXfX(t)log⁡2fX(t)fY(t)dt=D(X∥Y)\lim_{n\to\infty} D([X]_{n}\|[Y]_{n})=\int_{S_{X}}f_{X}(t)\log_{2}\frac{f_{X}(t)}{f_{Y}(t)}dt=D(X\|Y)

Conclusion

Divergence is a universal information measure in Information Theory.

Linked from