Abstract :
For a class of density functions q(x) on Rn we prove an inequality between relative entropy and the
weighted sum of conditional relative entropies of the following form:
D(p q) Const.
n
i=1
ρi
·D
pi (·|Y1, . . . , Yi−1,Yi+1, . . . , Yn)
Qi (·|Y1, . . . , Yi−1,Yi+1, . . . , Yn)
for any density function p(x) on Rn, where pi (·|y1, . . . , yi−1, yi+1, . . . , yn) and Qi (·|x1, . . . , xi−1,
xi+1, . . . , xn) denote the local specifications of p respectively q, and ρi is the logarithmic Sobolev constant
of Qi (·|x1, . . . , xi−1, xi+1, . . . , xn). Thereby we derive a logarithmic Sobolev inequality for a weighted
Gibbs sampler governed by the local specifications of q. Moreover, the above inequality implies a classical
logarithmic Sobolev inequality for q, as defined for Gaussian distribution by Gross. This strengthens a result
by Otto and Reznikoff. The proof is based on ideas developed by Otto and Villani in their paper on the
connection between Talagrand’s transportation-cost inequality and logarithmic Sobolev inequality.
© 2012 Elsevier Inc. All rights reserved.
Keywords :
Relative entropy , Wasserstein distance , Fokker–Planck equation , Gradient flow , Non-compact spin system , Gibbs sampler , Weakly dependent random variables , Logarithmic Sobolev inequality , Transportation-cost inequality