Mutual Information for Electromagnetic Information Theory Based on Random Fields

Equation (?) fully depicts the autocorrelation of the noise field of three polarization orientations under the assumption of isotropic incidence. This form of autocorrelation is also employed to describe the channel autocorrelation [shafi2006polarized, eqn (19)], which is reasonable because the noise discussed here can be considered as unwanted signals received through a random channel.

This noise model is based on the isotropic scattering assumption, which can be easily extended to the non-isotropic case. If we assume that the radiation interference is not uniform in spatial angle but concentrated near a certain angle, we can use von Mises-Fisher (vMF) distribution [abdi2002space, gatto2007generalized] to depict such angular concentration. A similar idea was proposed in [pizzo2022spatial] to analyze the angular selectivity of the random channel modeling but the analytic solution was not given. For the 3-dimensional von Mises-Fisher distribution, which is a close approximation to the spherical analogue of the Gaussian distribution, the probability density function is given by

where C(\|\boldsymbol{\mu}\|) = \frac{\|\boldsymbol{\mu}\|}{2\pi (e^{\|\boldsymbol{\mu}\|}-e^{-\|\boldsymbol{\mu}\|})} is the normalization constant. Instead of (?) we have

Then, by replacing {\bf r}\in\mathbb{R}^3 in (?) by {\bf r} = {\bf r}_R - {\rm j}{\boldsymbol \mu}/\kappa \in\mathbb{C}^3. The imaginary part {\rm j}{\boldsymbol \mu}/k can be combined into the spatial harmonic factor e^{{\rm j}\boldsymbol{\kappa} \cdot {\bf r}} of the planar wave, in order to describe a vMF distribution of the incident angles \theta and \varphi. Equation (?) can then be extended to the non-isotropic case accordingly by analytic continuation [ahlfors1953complex] techniques, leading to

where r_1 = \sqrt{(\boldsymbol{r}-{\rm j}\boldsymbol{\mu}/\kappa)^{\rm T}(\boldsymbol{r}-{\rm j}\boldsymbol{\mu}/\kappa)} and {\bf{\hat{r}_1}} = \frac{\boldsymbol{r}-{\rm j}\boldsymbol{\mu}/\kappa}{r_1}.

For simplicity, we focus on the linear transceivers in the paper, so as to use the tools and conclusions in the classical theory of stochastic process. However, the analyzing schemes here are not limited to the scenario with linear transceivers, but can be extended to other scenarios like two-dimensional surfaces. The conclusions in stochastic processes should also be extended to those in random fields. For example, if the source is a rectangular surface with current density J({\bf s}) = J(s_x,s_y) and the destination receives electric field E({\bf r}) = E(r_x,r_y) in a rectangular surface. We can also derive the autocorrelation function R_E({\bf r},{\bf r}') of the electric field. Since the rectangular surface is a compact separable metric space if we define the area as the measure on the space, we can perform Mercer expansion on random field E(r_x,r_y) according to [mercer1909xvi]. The expansion is similar to (?) as R_E({\bf r},{\bf r}') = \sum_k \lambda_k \phi_k({\bf r})\phi_k({\bf r}'). Therefore, mutual information expression and the corresponding derivation can also be done under the scenario with two-dimensional transceivers. For three-dimensional transceivers, the extension is similar to the two-dimensional cases. Furthermore, if we want to consider the vector-valued fields as {\bf J}(\bf s) and {\bf E}(\bf r), we can decompose {\bf E}(\bf r) according to the extension of the Mercer theorem [de2013extension].

Our analysis here from the mutual information perspective is compatible with existing works about the capacity. The existing works decompose the current density and electric field on a set of bases, which is similar to the SVD decomposition on {\bf H}{\bf H}^\dagger, where {\bf H} is the channel matrix in discrete MIMO systems [miller2019waves]. Our work decomposes the current density and electric field according to their autocorrelation function, which finds the bases for the field with given autocorrelation function. This scheme is similar to the decomposition of the covariance matrix to obtain the mutual information between random vectors, since I(X;X+N) = {\rm logdet} \frac{K_X+K_N}{K_N} = \sum_{i=1}^m (1+\frac{\lambda_m}{n_0/2}), where \lambda_k is the eigenvalue of K_X and K_N = (n_0/2){\bf I}.

It is obvious that when fixing the field distribution which maximizes the mutual information, the capacity between continuous transceivers can be obtained. This conclusion is an extension from the discrete matrices to continuous operators. For discrete matrices we have C = \underset{K_J}{\rm max}{\rm logdet}\frac{HK_JH^{\rm H}+K_N}{K_N} with respect to {\rm Tr}(K_J) \leqslant P. The optimal solution is obtained by using singular value decomposition and Karush–Kuhn–Tucker (KKT) theorem. For operators, Mercer expansion is utilized instead of the singular value decomposition. Through this approach, an operator can be viewed as an infinitely large matrix [gruber2008new], and the optimization problem can be solved by extended KKT theorem [bachir2021finitely] for countably infinite variables. The result is in the form of water-filling method.

The Fredholm determinant \det(\cdot) is a homomorphism of G into the multiplicative group of the complex numbers \mathbb{C}, i.e., \det(AB)=\det(A)\det(B), \;\forall A,B\in G, which is similar to the determinant of a square matrix. Furthermore, suppose the trace-class operator T has eigenvalues \lambda_k, then \det({\bf 1}+T)=\prod_k (1+\lambda_k).

The discretization algorithm of the mutual information not only has the mathematical meaning but also has physical correspondence. Since -{\rm logdet}({\bf I}-{\bf K}_{\rm sep}) in Algorithm ? equals {\rm log}\,{{\rm det}({\bf K}_{E}+{\bf K}_{N})/{\rm det}({\bf K}_{N})}, which is the form of the mutual information of multiple Gaussian channel model. If we express {\bf K}_{E,i,j} and {\bf K}_{N,i,j} by \mathbb{E}\left[ \int_{a_{i-1}}^{a_i}E(x){\rm d}x \int_{a_{j-1}}^{a_j}E^{*}(y){\rm d}y \right] and \mathbb{E}\left[ \int_{a_{i-1}}^{a_i}N(x){\rm d}x \int_{a_{j-1}}^{a_j}N^{*}(y){\rm d}y \right] respectively, it is easy to find that the mutual information after discretization can be viewed as using several patch antennas in the destination region which receives the integral of the electric field. Therefore, the convergence of the numerical approximation of the mutual information shows how the discrete destinations approach the continuous one. Here we use Gauss-Legendre quadrature rule because it converges faster than the equally-spaced trapezoid or rectangular quadrature rules.