3.1 Specular- and dense multipath components
To describe the geometric properties of the electromagnetic waves of the MIMO propagation channel in terms of AoD, AoA, and ToA, multidimensional frequency domain channel sounding must be performed. This can be done with (virtual) MIMO array systems, consisting of
M
T
and
M
R
antennas at transmitter (Tx) and receiver (Rx), sampled at
M
f
frequency points. As such, the total amount of samples can be defined by
M as follows:
$$ M = M_{T} M_{R} M_{f}. $$
(1)
An observation of the frequency response of a MIMO radio channel
h can be modeled as the superposition of a deterministic part
x(
θsmc) (determined by the SMC parameter set
θsmc) and a stochastic part
d(
θdan) (diffuse scattering and noise; determined by the DMC and noise (DAN) parameter set
θdan). Both parameter sets will be described later in this section.
$$ \begin{aligned} \boldsymbol{h} &=\boldsymbol{x}(\boldsymbol{\theta}_{\text{smc}}) + \boldsymbol{d}(\boldsymbol{\theta}_{\text{dan}})\\ \boldsymbol{h} & \in \mathbb{C}^{M \times 1}. \end{aligned} $$
(2)
The deterministic part
x(
θsmc) of the data model acts as the first order statistics of the radio channel, so that it can be interpreted as the mean of
h, whilst the stochastic part
d(
θdan) describes the second-order statistics by means of its covariance matrix
\(\boldsymbol {R}(\boldsymbol {\theta }_{\text {dan}}) \in \mathbb {C}^{M \times M}\), which will be discussed in Section
5.1.2. A realization of the radio channel
h can be considered as a random variable distributed according to a complex multivariate Gaussian distribution
\(\boldsymbol {h} \sim \mathcal {N}_{c} (\boldsymbol {x}(\boldsymbol {\theta }_{\text {smc}}), \boldsymbol {R}(\boldsymbol {\theta }_{\text {dan}}))\) [
44] as follows:
$$ p(\boldsymbol{h}) = \frac{1}{\pi^{M} \det(\boldsymbol{R})} e^{-(\boldsymbol{h}-\boldsymbol{x})^{H} \boldsymbol{R}^{\text{-1}} (\boldsymbol{h}-\boldsymbol{x})}. $$
(3)
An estimate of the most likely SMC and DMC and noise parameters can be found by maximizing the likelihood function of Eq. (
3). Since this is not a trivial task, estimation frameworks such as the RiMAX algorithm will estimate
\(\hat {\boldsymbol {\theta }}_{\text {smc}}\) and
\(\hat {\boldsymbol {\theta }}_{\text {dan}}\) of the deterministic and stochastic arrays, such that they maximize the likelihood of observing the measured frequency response
h of the radio channel. The objective is thus to find the parameters
\(\hat {\boldsymbol {\theta }}_{\text {smc}}\) and
\(\hat {\boldsymbol {\theta }}_{\text {dan}}\) that maximize the correlation with the measurement data. A maximum likelihood (ML) estimator for the parameters
θsmc and
θdan has been proposed in [
7,
22], exploiting the fact that the parameters of the two components of the channel model are asymptotically independent. Therefore, one can decouple the estimation problem into two separate estimation problems. The resulting RiMAX algorithm is iterative and alternates between the maximization of the likelihood function with respect to the parameters
θsmc and
θdan. It has an approximately linear computational complexity in the number of propagation paths
P and in the number of data samples
M [
7].
Based on the capability of the RiMAX algorithm to extract both parameter sets from the (virtual) array measurement data, the following structures for the deterministic and stochastic parameter sets (
\(\boldsymbol {\theta }_{\text {smc}} \in \mathbb {C}^{P \times 4S}\) and
\(\boldsymbol {\theta }_{\text {dan}} \in \mathbb {R}^{4 \times S}\), respectively) can be adopted:
$$\begin{array}{*{20}l} \boldsymbol{\theta}_{\text{smc}} & = \left[ \begin{array}{c} \boldsymbol{\varphi_{D}}^{T} \\ \boldsymbol{\varphi_{A}}^{T} \\ \boldsymbol{\tau_{A}}^{T} \\ \boldsymbol{\gamma}^{T} \end{array}\right]^{T} \enskip \begin{array}{l} \leftarrow \text{SMC angle of departure [rad]} \\ \leftarrow \text{SMC angle of arrival [rad]} \\ \leftarrow \text{SMC time-delay of arrival [s]} \\ \leftarrow \text{SMC complex amplitude [/].} \end{array} \end{array} $$
(4)
In (
4),
φ
D
,
φ
A
,
τ
A
, and
γ are
P×
S matrices, where
P is the number of SMCs extracted from the measurement data, and
S is the number of sub-bands in which the total UWB bandwidth was partitioned (in analogy with the UWB-SAGE algorithm). As such,
θsmc is of size 4
S×
P. We will discuss the frequency-dependency of the geometrical parameters of the propagation paths in Section
4.2. Each row in the aforementioned matrices
φ
D
,
φ
A
, and
τ
A
contains the corresponding specular parameter for each of the
\(p \in \mathcal {P}\) propagation paths (
\(P = |\mathcal {P}|\) in total) and describes its frequency-dependency in each of the
\(s \in \mathcal {S}\) sub-bands (
\(S = |\mathcal {S}|\) in total). We note that the angular modeling was limited to that of the azimuthal plane, which is acceptable since most measurement campaigns are only performed with a planar array, with which the estimation of elevation parameters is not possible. The extension of the data model to the elevation domain is straightforward, where the AoD and AoA of the SMC will now have an extra elevation component together with an azimuthal component. We also consider only a single snapshot of the channel, such that the covariance matrix
R(
θdan) is only averaged over one observation of the channel. Whilst it would be more reliable to use several snapshots of the channel in a real measurement environment, we would then have to impose a parametric model to handle the time dependence of the SMC, or assume them to be independent and independent and identically distributed (i.i.d.) across snapshots in time. Since this is out of the scope for the purpose of this paper, we will leave this up for future work.
$$\begin{array}{*{20}l} \boldsymbol{\theta}_{\text{dan}} & = \left[ \begin{array}{c} \boldsymbol{\alpha_{1}} \\ \boldsymbol{\tau_{d}} \\ \boldsymbol{\tau_{r}} \\ \boldsymbol{\alpha_{0}} \end{array}\right] \enskip \begin{array}{l} \leftarrow \text{DMC peak power [W]} \\ \leftarrow \text{DMC onset time [s]} \\ \leftarrow \text{DMC reverberation time [s]} \\ \leftarrow \text{Noise power [W].} \\ \end{array} \end{array} $$
(5)
In (
5),
θdan contains the DMC parameters
\(\left (\left [\boldsymbol {\alpha _{1}} \in \mathbb {R}^{1 \times S}, \boldsymbol {\tau _{d}} \in \mathbb {R}^{1 \times S}, \boldsymbol {\tau _{r}} \in \mathbb {R}^{1 \times S}\right ]\right)\) and the noise parameters (
\(\boldsymbol {\alpha _{0}} \in \mathbb {R}^{1 \times S}\)) for each sub-band
\(s \in \mathcal {S}\). A discussion of the model for the DMC can be found in [
7]. The model is based on the observation that the PDP
ψ(
τ) [W] of the DMC and noise, which describes how the power of a signal is distributed over the time-delay domain, has a base time-delay
τ
d
related to the distance between the transmitter and receiver, together with an exponential decay over time-delay (see Eq. (
6)), corrupted by complex additive white Gaussian noise with power
α0:
$$ \psi (\tau) = \left\{ \begin{array}{ll} \alpha_{1} \hspace{1mm} e^{\left(- \frac{\tau-\tau_{d}}{\tau_{r}} \right)}+\alpha_{0}, & \text{if } \tau > \tau_{d}\\ \alpha_{0}, & \text{otherwise}. \end{array}\right. $$
(6)
In Eq. (
6),
α1,
τ
d
,
τ
r
, and
α0 are the four parameters which fully describe the DMC and noise characteristics of each sub-band and are gathered in the DMC and noise parameter vector
θdan. In this work, we will assume that the DMC is spatially white at the transmit and receive side of the measurement system, meaning that they have constant angular power densities. It should be noted that recent works will assume the DMC to be spatially correlated with the SMC. For example, [
25,
45] report a correlation between the location of SMC and DMC in the angular domain. In [
27], the DMC is modeled as local clusters around the SMC. In [
20,
46], it is proposed that the Power Angular Profile (PAP) is to be modeled by a uni-modal Von-Mises distribution.
The covariance matrix
R(
θdan) can be constructed by converting the model in Eq. (
6) from the time-delay domain to the frequency domain. To do so,
\(\boldsymbol {\kappa }(\boldsymbol {\theta }_{\text {dan}}) \in \mathbb {C}^{N_{F} \times 1}\) [W] is first introduced, which denotes a band-limited sampled version of the Fourier transform of Eq. (
6), and can be defined for a bandwidth
B and
M
f
frequency samples as follows:
$$ {\begin{aligned} {}\boldsymbol{\kappa}(\boldsymbol{\theta}_{\text{dan}}) =& \alpha_{0}~\boldsymbol{e_{0}} + \frac{\alpha_{1}}{M_{f}} \left[\frac{1}{\beta_{d}}, \frac{e^{-j2\pi \tau_{d}}}{\beta_{d} + j2\pi\frac{1}{M_{f}}} \cdots \right.\\& \qquad \qquad\qquad \qquad \qquad \quad \left. \frac{e^{-j2\pi \tau_{d} \left(M_{f}-1\right)}}{\beta_{d} + j2\pi\frac{\left(M_{f}-1\right)}{M_{f}}}\right], \end{aligned}} $$
(7)
in which
\(\boldsymbol {e_{0}} \in \mathbb {N}^{M_{f} \times 1}\) is a unit vector defined as follows:
$$ \boldsymbol{e_{0}} = \left[1,0, \cdots, 0\right]. $$
(8)
Furthermore, in Eq. (
7),
β
d
is the normalized coherence bandwidth (dimensionless), calculated as follows:
$$ \beta_{d} = \frac{B_{d}}{B} = \frac{1}{B \cdot \tau_{r}}, $$
(9)
in which
B
d
is the coherence bandwidth (Hz), normalized by the bandwidth per sub-band
B (Hz). Furthermore, the frequency sampling interval
Δ
f
(Hz) can be written as a function of
B and
M
f
as follows:
$$ \Delta_{f} = \frac{B}{M_{f}-1}. $$
(10)
After calculating
κ(
θdan), the covariance matrix
R(
θdan) of the DMC and noise can be calculated by applying the Toeplitz-operator [
7] as follows:
$$ \boldsymbol{R}(\boldsymbol{\theta}_{\text{dan}}) = \text{toep} \left(\boldsymbol{\kappa}(\boldsymbol{\theta}_{\text{dan}}), \boldsymbol{\kappa}(\boldsymbol{\theta}_{\text{dan}})^{H}\right). $$
(11)
The original data model of the RiMAX estimation algorithm follows the narrowband assumption, hence stating that the SMC and DMC are Kronecker-separable in the spatial and frequency domains in order to keep the algorithm computationally viable [
7]. In our measurements section, prior to the processing of the measurement data, we will check for the uncorrelated scattering assumption, which needs to hold in order to apply the Kronecker model. For a large MIMO configuration in a given communication system, the dimensions of the covariance matrix
R(
θdan) becomes too large to allow for a reliable estimation (remember that it has a size of [
M×
M]). This becomes a processing burden to accurately model the interaction between transmitter and receiver, so that both ends of a communication system need to be decoupled by applying the well-known Kronecker model. We refer to [
47] for a discussion of its applicability and limitations.
In the Kronecker model, the covariance matrices at transmitter and receiver are assumed independent and separable, allowing the full covariance matrix of the channel to be expressed as the Kronecker-product of several smaller matrices. Thus, the covariance matrix
R is assumed to have the following structure [
7]:
$$ \boldsymbol{R} = \boldsymbol{R}_{T} \otimes \boldsymbol{R}_{R} \otimes \boldsymbol{R}_{f} + \alpha_{0} \boldsymbol{I}_{M}, $$
(12)
in which \(\boldsymbol {R}_{T} \in \mathbb {C}^{M_{T} \times M_{T}}\) and \(\boldsymbol {R}_{R} \in \mathbb {C}^{M_{R} \times M_{R}}\) are the covariance matrices at the transmitter and receiver, respectively, describing the spatial distribution of the DMC at both ends, whereas \(\boldsymbol {R}_{f} \in \mathbb {C}^{M_{f} \times M_{f}}\) is the covariance matrix in the frequency domain. The term α0I denotes the amount of complex additive white Gaussian measurement noise, which contributes to the stochastic part of the data model. The matrix I
M
is an identity matrix of size [M×M].
3.2 Modeling propagation paths
In this extension of the RiMAX algorithm, we will assume that the transmitter and receiver are equipped with a uniform circular array (UCA), given that its angular resolution capability is uniform since the effective aperture does not change with azimuth angle. The extension to other array configurations is straightforward since only the steering matrices in the following subsection have to be adjusted. It should be noted that other configurations for the antenna array can be used in this algorithm as well, without much modification to the hereinafter explained matrices. The broadband MIMO radio channel model can be expressed in matrix notation to map the time-delay and angles of a propagation path to its complex notation in the frequency domain. We will therefore define the matrix
\(\boldsymbol {B}_{f}(\boldsymbol {\tau }_{A}) \in \mathbb {C}^{M_{f} \times P}\) (dimensionless), which maps the time-delays
τ
A
of each propagation path
\(p \in \mathcal {P}\) to its related complex exponential
\(\phantom {\dot {i}\!}e^{-j2\pi \cdot m \Delta _{f} \cdot \tau _{A,p}}\) notation as follows:
$$ {} \boldsymbol{B}_{f}(\boldsymbol{\tau}_{A})\! =\! \left[ \begin{array}{ccc} e^{-j2\pi\Delta_{f} \left(-\frac{M_{f}-1}{2}\right)\tau_{A,1}} & \cdots & e^{-j2\pi\Delta_{f}\left(-\frac{M_{f}-1}{2}\right)\tau_{A,P}} \\ \vdots & \ddots & \vdots \\ e^{-j2\pi\Delta_{f} \left(+\frac{M_{f}-1}{2}\right)\tau_{A,1}} & \cdots & e^{-j2\pi\Delta_{f}\left(+\frac{M_{f}-1}{2}\right)\tau_{A,P}} \end{array}\right], $$
(13)
Similarly, we can define the mapping of both the departing and the arriving angles
φ
D
and
φ
A
to the transmitting and receiving array responses
\(\boldsymbol {B}^{s}_{T}(\boldsymbol {\varphi }_{D}) \in \mathbb {C}^{M_{T} \times P}\) (dimensionless) and
\(\boldsymbol {B}^{s}_{R}(\boldsymbol {\varphi }_{A}) \in \mathbb {C}^{M_{R} \times P}\) (dimensionless) in each sub-band
\(s \in \mathcal {S}\) as follows:
$$ {} \boldsymbol{B}^{s}_{T}(\boldsymbol{\varphi}_{D})\! =\! \left[\! \begin{array}{ccc} e^{-j\frac{2\pi}{\lambda_{s}}r\cos(\varphi_{D,1}-\rho_{1})} & \cdots \!& \! e^{-j\frac{2\pi}{\lambda_{s}}r\cos(\varphi_{D,P}-\rho_{1})} \\ \vdots & \ddots & \vdots \\ e^{-j\frac{2\pi}{\lambda_{s}}r\cos\left(\varphi_{D,1}-\rho_{M_{T}}\right)} & \cdots & e^{-j\frac{2\pi}{\lambda_{s}}r\cos\left(\varphi_{D,P}-\rho_{M_{T}}\right)} \end{array}\right], $$
(14)
and
$$ {{} \begin{aligned} \boldsymbol B^{s}_{R}(\boldsymbol{\varphi}_{A}) \,=\, \left[\! \begin{array}{ccc} e^{-j\frac{2\pi}{\lambda_{s}}r\cos(\varphi_{A,1}-\rho_{1})} & \cdots \!& e^{-j\frac{2\pi}{\lambda_{s}}r\cos(\varphi_{A,P}-\rho_{1})} \\ \vdots & \ddots & \vdots \\ e^{-j\frac{2\pi}{\lambda_{s}}r\cos\left(\varphi_{A,1}-\rho_{M_{R}}\right)} & \cdots & e^{-j\frac{2\pi}{\lambda_{s}}r\cos\left(\varphi_{A,P}-\rho_{M_{R}}\right)} \end{array}\!\right], \end{aligned}} $$
(15)
with
r being the radius of the UCA,
λ
s
(m) the wavelength in each sub-band
\(s \in \mathcal {S}\), and the vector function
\(\boldsymbol {\rho }(M_{T/R}) \in \mathbb {R}^{M_{T/R} \times 1}\) (rad) mapping the angle between each antenna in the UCA and a chosen reference axis, defined as follows:
$$ \boldsymbol{\rho}(M_{T/R}) = (0:1:M_{T/R}-1)\frac{2\pi}{M_{T/R}}. $$
(16)
Both matrices B
T
(φ
D
) and B
R
(φ
A
) describe the complex frequency-dependent far-field beam pattern at each antenna array port at the transmitting and receiving array, respectively.
Now let us consider a measurement snapshot of the multidimensional UWB-MIMO channel
h. This snapshot can be defined as the instantaneous frequency domain transfer function of the channel between each MIMO antenna pair and is modeled as a superposition of
P discrete paths
x plus a contribution of DMC and noise, written as follows:
$$ \boldsymbol{h} = \sum\limits_{p=1}^{P} \boldsymbol{x}_{p}\left(\boldsymbol{\theta}_{\text{smc},p}\right) + \boldsymbol{d}(\boldsymbol{\theta}_{\text{dan}}). $$
(17)
The model for a single propagation path
p in sub-band
s is given by:
$$ \boldsymbol{x}^{s}_{p}\left(\boldsymbol{\theta}^{s}_{\text{smc},p}\right) = \texttt{vec} \left(\boldsymbol{B}_{T}^{s}(\varphi_{D,p}) \Diamond \boldsymbol{B}_{R}^{s}(\varphi_{A,p}) \Diamond \boldsymbol{B}_{f}(\tau_{A,p})\right) \gamma^{s}_{p}, $$
(18)
where the operator ♢ denotes Khatri-Rao (column-wise Kronecker) product and the variable
\(\gamma ^{s}_{p}\) denotes the complex amplitude of path
p in sub-band
s. The superposition of
P paths in sub-band
s can then be written as follows:
$$ \begin{aligned} \boldsymbol{x}^{s}\left(\boldsymbol{\theta}^{s}_{\text{smc}}\right) &= \sum\limits_{p=1}^{P} \boldsymbol{x}^{s}_{p}\left(\boldsymbol{\theta}^{s}_{\text{smc},p}\right) \\ &= \left(\boldsymbol{B}^{s}_{T}(\boldsymbol{\varphi}_{D}) \Diamond \boldsymbol{B}^{s}_{R}(\boldsymbol{\varphi}_{A}) \Diamond \boldsymbol{B}_{f}(\boldsymbol{\tau}_{A})\right) \boldsymbol{\gamma}^{s} \\ &= \boldsymbol{B}^{s}\left(\boldsymbol{\varphi}_{D},\boldsymbol{\varphi}_{A},\boldsymbol{\tau}_{A}\right) \boldsymbol{\gamma}^{s} \in \mathbb{C}^{M \times 1}, \end{aligned} $$
(19)
in which B
s
(φ
D
,φ
A
,τ
A
) represents the steering vector in space and frequency for the entire array, and for all propagation paths, in sub-band s.