Covariance Matrix Estimation for Single User

Bellili Montassar
August 14, 2024

Abstract
In various signal processing and communication applications, the esti-
mation of the covariance matrix plays a critical role, particularly in scenar-
ios where direct channel estimation is challenging or suboptimal. Unlike
channel estimation, which focuses on determining the precise state of the
communication channel, covariance matrix estimation provides a statisti-
cal characterization of the channel, capturing the second-order statistics
of the signal. This approach is often preferred in environments with high
variability or insufficient data, as it offers a more robust and generalized
understanding of the signal characteristics. Covariance matrix estimation
is essential in scenarios with rapidly changing channel conditions, where
traditional channel estimation methods may fail to adapt quickly enough,
leading to performance degradation.

1 Introduction
Consider a multiuser MIMO channel formed by a base station (BS) with M
antennas and K single-antenna mobile users in a cellular network. The central
task at the BS side consists in estimating, for each user, a subspace containing a
significant amount of its received signal power. Since in a Hybrid digital analog
implementation we do not have direct access to all the M antennas, but only
to m ≪ M analog output observations, we need to estimate this subspace from
snapshots of a low-dim projection of the signal

1.1 Contribution
In this paper, we aim to design a subspace estimator for a base station (BS)
equipped with a large uniform linear array (ULA) where the number of antennas
M ≫ 1. The geometry of the array is illustrated in Fig. 1, with array elements
uniformly spaced by a distance d. We assume that the array serves the users
within the angular range [−θmax , θmax ] for some θmax ∈ (0, π/2), and we let
λ
d = 2 sin(θmax )
, where λ is the wavelength. In general, we assume that we can
observe only low-dimensional sketches of the received signal via m ≪ M linear
projections.

1
Figure 1: Array configuration in a multi-antenna receiver in the presence of a
scattering channel with discrete angle of arrivals.

Notation: Throughout the paper, the output of an optimization algorithm

arg minx f (x) is denoted by x∗ . We use T and T+ for the space of all M × M
Hermitian Toeplitz and Hermitian semi-definite Toeplitz matrices, respectively.
We always use I for the identity matrix, where the dimension may be explicitly
indicated for the sake of clarity (e.g., Ik denotes the k × k identity matrix).
We denote a k × k diagonal matrix with k diagonal elements α1 , . . . , αk with
diag(α1 , . . . , αk ). We define H(M, p) = {UM ×p ∈ CM ×p : UUH = Ip } as the
set of all unitary matrices of dimension M × p.
For matrices and vectors of appropriate dimensions, we define the inner
product by ⟨K, L⟩ = Tr(KLH ), pwhere Tr denotes the trace operator, and define
the induced norm by ∥K∥ = ⟨K, K⟩, also known as the Frobenius norm for
matrices.
For an integer k ∈ Z, we use the shorthand notation [k] for the set of non-
negative integers {0, 1, . . . , k − 1}, where the set is empty if k < 0.

2 Channel Model and Problem Statement

More general, the channel vector may be formed by the superposition of a
continuum of array responses. In order to include this case, we define the AoA
scattering function γ(u), which describes the received power density along the
sin(θ)
direction identified by u ∈ [−1, 1], where u = sin(θ max )
for θ ∈ [−θmax , θmax ].

2
 We denote the array vector in the u domain by a(u), where [a(u)]k = ejkπu .
Then, the channel model is given by
Z 1p
y(t) = γ(u)a(u)z(u, t)du + n(t),
−1

where z(u, t) is a white circularly symmetric Gaussian process with a covariance

function E[z(u, t)z ∗ (u′ , t′ )] = δ(u − u′ )δt,t′ . The covariance matrix of y(t) is also
given by Z 1
Cy = γ(u)a(u)a(u)H du + σ 2 IM = S + σ 2 IM .
−1
R1
where S = S(γ) := −1 γ(u)a(u)a(u)H du denotes the covariance matrix of the
signal part, and where σ 2 IM is the covariance matrix of the white additive noise.
We define the received SNR by
R1 R1
Tr(S(γ)) −1
γ(u)∥a(u)∥2 du γ(u)du
snr := 2
= 2
= −1 2 ,
Tr(σ IM ) Mσ σ
R1
where −1 γ(u)du is the total received signal power in a given array element.
For the ULA, S is a Toeplitz matrix.

We define the best p-dimensional beamforming matrix for the covariance

PM matrix
S as Vp = arg maxU∈H(M,p) Tr(S, UUH ). Letting S = VΛVH = i=1 λi vi viH
be the SVD of S, the matrix Vp is an M × p tall unitary matrix formed by the
first p columns ofPV. The signal power captured by this beamformer is given
p
by ⟨S, Vp VpH ⟩ = i=1 λi .
In this paper, we are concerned with the estimation of Vp for some appro-
priately chosen p, from the noisy snapshots of the projected channel (sketches)
X = BY obtained during a training period of length T .
In order to measure the “goodness” of estimators, we propose the following
performance metric, which is relevant to the underlying communication problem
of JSDM group separation beamforming. First, we define the efficiency of the
best p-dimensional beamformer by

⟨S, Vp VpH ⟩ Tr(VpH SVp )

ηp = = .
Tr(S) Tr(S)

If ηp ≈ 1 for some p ≪ M , then a significant amount of the signal’s power

is captured by a low-dimensional beamformer. Let now V e p = Vp (X) be an
estimator of Vp from the sketches X. We define the relative efficiency of Vp as

⟨S, V e H⟩
e pV
p
Γp = .
⟨S, Vp VpH ⟩

3
2.1 Algorithm 1: Approximate Maximum Likelihood (AML)
Estimator
We will use B as the projection matrix to produce the low-dimensional observa-
tions X = BY. In passing, this has the advantage that the projection reduces
to array subsampling, or ”antenna selection,” which is very easy to implement
in the analog RF domain by simple switches connecting the selected antennas
to the RF demodulation chains and A/D converters.

Let Cb x = 1 XXH be the sample covariance of the observations X. Then Cb x is

T
a sufficient statistic for estimating the signal covariance matrix S.
Proof: Recall that Cx = BCy BH = BSBH + σ 2 Im , where we have explic-
itly used the fact that BBH = Im . As the observations X are Gaussian, after
some simple algebra the likelihood function is given by
n  o
exp −T Tr C b x (BSBH + σ 2 Im )−1
p(X|S) = .
π T m det(BSBH + σ 2 Im )T

We always assume that the noise variance σ 2 can be estimated during the
system’s operation. In this section, for simplicity of the notation, we suppose
that the input signal is scaled by σ12 , and denote the resulting sample covariance
by Cb x̃ = C
σ 2 . Then, the Maximum-Likelihood (ML) estimator for the normal-
bx

ized subsampled data can be written as S b = arg minS∈T L(S), e where Se = S/σ 2 ,
+

and where L(S) e is the minus log-likelihood function given by

Note that: L(S) e is the sum of a concave function

e = log det(Im + BSBH )

Lcav (S)

and a convex function

 
e = Tr C
Lvex (S) b x̃ (Im + BSBH )−1 .

As L(S)
e is not convex, local optimization techniques such as gradient descent
are not guaranteed to converge to the globally optimal solution. Since S e scales
with SNR, it is possible to obtain a convex (indeed, linear) approximation of the
concave function Lcav (S),
e which is tight especially for low SNR. More precisely,
we have the following result.
e ≤ Tr(BSBH ) for all S
Lcav (S) e ∈ T+ . Moreover, for the low-SNR regime
e = Tr(BSBH ) + o(snr).
(snr ≪ 1), we have Lcav (S)
So far we stated that for low SNR, Tr(BSBH ) is the best linear approxima-
tion for Lcav (S),
e which implies that
 
e = Tr(BSBH ) + Tr C
Lapp (S) b x̃ (Im + BSBH )−1 .

Now we show that the AML estimation can be cast as an SDP.

4
  
e = Tr(BSBH ) + Tr C
Proposition : Let Lapp (S) b x̃ (Im + BSBH )−1 and
b x̃ = UΛUH be the SVD of C
let C b x̃ . Then the AML estimate can be obtained
from the following SDP:
e ∗ , W∗ ) = arg
(S min Tr(BMBH ) + Tr(W)
M∈T+ ,W

subject to " #
Im + BMBH ∆
b
bH ≥ 0,
∆ W

where ∆ b 1/2 = UΛ1/2 .

b =C
x̃

3 SIMULATION RESULTS
We consider an array of size M = 80 and θmax = 60 degrees (corresponding
to an angular sector of 120 degrees). We use a coprime sampling with q1 = 7,
q2 = 9, where we denote the set of indices of the sampled antennas with D, where
|D| = 19. Although there are still some array indices in [M ] = {0, . . . , 79} not
covered by ∆D, the simulations show that the estimators are quite insensitive
to the presence of a few unobserved elements. We also assume that only n = 20
RF chains are available at the BS, which would be enough to implement the
coprime sampling, and to serve up to 20 data streams in the UL or DL.
As an example, we consider a scattering channel with AoAs in the range
Θ = [θ1 , θ1′ ] ∪ [θ2 , θ2′ ], where θ1 = −50◦ , θ1′ = −40◦ , θ2 = 10◦ , and θ2′ = 20◦ . We
assume a uniform power distribution over Θ, thus, the total angular support is
20 degrees. The AoA scattering function γ(u) is given by
(
√ κ
1−u2
, u ∈ [u1 , u′1 ] ∪ [u2 , u′2 ]
γ(u) =
0, otherwise

where ui = sin(θi ) and u′i = sin(θi′ ) for i = 1, 2, and where κ > 0 is a

normalization
R1 constant. We calculate the vector of Fourier coefficients f =
−1
γ(u)a(u)du, from which we construct the Toeplitz signal covariance matrix
S. The i.i.d. channel vectors in each training period are generated according
to h = S1/2 n1 , and the corresponding observation at the array antennas is
y = h + σn2 , where n1 , n2 ∼ CN (0, IM ) are independent vectors, and where σ 2
denotes the noise variance.

5
 Performance comparison of AML estimator vs SNR for T = 100
1

0.9

0.8
Performance Metric Γp

0.7

0.6

0.5

0.4

0.3

0.2

0.1

−15 −10 −5 0 5 10 15
SNR [dB]
AML Estimator

Figure 2: Performance comparison of AML estimator vs SNR for T = 100.