Information-theoretic inequalities on unimodular Lie groups (Q613106)

A variety of unimodular Lie group analogues of Euclidean information theoretic inequalities are established. Unimodularity substitutes for translation invariance, while the group property allows for additivity estimates and coset decompositions. Analogues of DeBruijn's identity, inequalities for Fisher information (applying log-Sobolev inequalities due to \textit{L. Gross} [Ill. J. Math. 36, No.3, 447--490 (1992; Zbl 0761.46019)]) and properties of Kullback-Leibler divergence are studied. The analogue of the Fisher information matrix for the case when \(f(g, \theta)\) is a parametric distribution on a Lie group \(G\) with Haar measure \(dg\) is \[ F_{ij} (f,\theta) = \int_G \frac{1}{f}\, \frac{\partial f}{\partial \theta_i} \, \frac{\partial f}{\partial \theta_j}\, dg . \] When \(G\) is parametrized by \(g(\theta)= \exp(\sum_i \theta_i X_i)\) with \(\{X_i\}\) a basis for the Lie algebra, and \(f(g,\theta)=f(g\circ \exp(\sum_i \theta_i X_i))\), \[ \frac{\partial f}{\partial \theta_i}\bigg|_{\theta=0}=\widetilde{X}_i^r f \] with \(\widetilde{X}_i^r\) the right Lie derivative of \(X_i\) and \(F_{ij}^r(f,0)=F_{ij}^r(f)=\int_G \frac{1}{f} (\widetilde{X}_i^r f)(\widetilde{X}_j^r f)\, dg\). The convolution inequality \[ F_{ii}^r(f_1\ast f_2)\leq \min\{F_{ii}^r(f_1),F_{ii}^r( f_2)\} \] (and similarly for the left matrix) is established as a consequence of the Cauchy-Schwarz inequality and the bi-invariance of the integral under translations. The result is sharpened to \[ \operatorname{tr} [F_{ii}^r(\rho_1\ast \rho_2)\,\Lambda] \leq \operatorname{tr} [ F_{ii}^r( \rho_2)\,\Lambda ] \] (and similarly for \(F_{ii}^{l}\)) in which \(\Lambda\) is any positive definite diagonal matrix having the same size as \(F_{ii}^r\). In addition, in case \(\rho_1\ast\rho_2=\rho_2\ast\rho_1\) one obtains \[ \frac{1}{\operatorname{tr}F^r(\rho_1\ast \rho_2)\,\Lambda} \leq \frac{1}{\operatorname{tr} F^r( \rho_1)\,\Lambda} + \frac{1}{\operatorname{tr} F^r( \rho_2)\,\Lambda} \] and similarly for \(F^l\). The following analogue of the DeBruijn identity is proved, in which \(f_{D,h,t}=f(g,t;D,h)\) denotes a solution to the diffusion equation \[ \frac{\partial\rho}{\partial t} (g,t)=\sum_{i=1}^d h_i(t)\widetilde{X}_i^r \rho(g; t) +\frac{1}{2}\sum_{i,j=1}^d D_{ij} \widetilde{X}_i^r \widetilde{X}_j^r \rho(g; t) \] with constant \(h\) satisfying initial condition \(f(g, 0;D, h) =\delta(g)\): With entropy \(S(f)=-\int_G f\log{f}\, dg\), for any well-behaved pdf \(\alpha(g)\) one has the identity \[ \frac{d}{dt} S(\alpha\ast f_{D,h,t}) = \frac{1}{2}\operatorname{tr}[DF^r \big(\alpha \ast f(D,h,t)\big)]. \] A relationship between entropy powers and Fisher information is given by using logarithmic Sobolev inequalities. The particular inequality given here is \[ \frac{2\pi e}{C_G}\exp\bigg(-\frac{2}{n}S(f)\bigg)\leq \frac{1}{n}\operatorname{tr}F . \] A number of explicit motivating examples are discussed. Concrete physical applications are noted as motivation for the general approach here.