This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
1

Massive MIMO Radar for Target Detection

Stefano Fortunati, Member, IEEE, Luca Sanguinetti, Senior Member, IEEE,
Fulvio Gini, Fellow, IEEE, Maria S. Greco, Fellow, IEEE, Braham Himed, Fellow, IEEE

Abstract—Since the seminal paper by Marzetta from 2010, To solve (1), a decision statistic Λ(X) of the dataset X ,
the Massive MIMO paradigm in communication systems has [x1 , . . . , xK ] is needed and its value must be compared with
changed from being a theoretical scaled-up version of MIMO, a threshold:
with an infinite number of antennas, to a practical technology.
Its key concepts have been adopted in the 5G new radio H1
standard and base stations, where 64 fully-digital transceivers Λ(X) ≷ λ (2)
H0
have been commercially deployed. Motivated by these recent
developments, this paper considers a co-located MIMO radar to discriminate between the null hypothesis H0 and the al-
with MT transmitting and MR receiving antennas and explores ternative H1 . A common requirement in radar applications is
the potential benefits of having a large number of virtual spatial that the probability of false alarm has to be maintained below
antenna channels N = MT MR . Particularly, we focus on the
target detection problem and develop a robust Wald-type test a pre-assigned value, say PF A . Consequently, the threshold λ
that guarantees certain detection performance, regardless of the should be chosen to satisfy the following integral equation:
unknown statistical characterization of the disturbance. Closed- Z ∞
form expressions for the probabilities of false alarm and detection Pr {Λ(X) > λ|H0 } = pΛ|H0 (a|H0 )da = PF A , (3)
are derived for the asymptotic regime N → ∞. Numerical results λ
are used to validate the asymptotic analysis in the finite system where pΛ|H0 is the probability density function (pdf) of Λ(X)
regime with different disturbance models. Our results imply that
there always exists a sufficient number of antennas for which under the null hypothesis H0 .
the performance requirements are satisfied, without any a-priori
knowledge of the disturbance statistics. This is referred to as the A. Motivation
Massive MIMO regime of the radar system. Finding a solution to (3) is in general a challenge. The
Index Terms—Large-scale MIMO radar, robust detection, common way out relies upon some “ad-hoc” assumptions on
Wald test, misspecification theory, unknown disturbance distri- the statistical model of the dataset X. In order to clarify this
bution, dependent observations. point, let us have a closer look to the steps required to solve
(3). Firstly, a closed-form expression for pΛ|H0 is needed. By
definition, pΛ|H0 is a function of the chosen decision statistic
I. I NTRODUCTION
Λ(X) and of the joint pdf pX (X) of X. If all the ᾱk , ∀k
Consider a multiple antenna radar system characterized are modelled as deterministic unknown scalars, pX (X) is
by N spatial channels collecting K temporal snapshots fully determined by the joint pdf pC (C) of the disturbance
{xk }K N C = [c1 , . . . , cK ]. A first simplification comes from the
k=1 ∈ C from a specific resolution cell, defined in
an absolute reference frame. The primary goal of any radar assumption that the disturbance vectors {ck } are independent
system is to discriminate between two alternative hypotheses: and identically
QK distributed (i.i.d.) random vectors such that
the presence (H1 ) or absence (H0 ) of the target, in the pC (C) = k=1 pC (ck ) [2,3]. This assumption is, however,
resolution cell under test. Among others, a common model not always valid in practice. A second simplification that is
for the signal of interest is ᾱk vk , where vk ∈ CN is known commonly adopted in the radar literature (see e.g. [2]–[4]) is
at each time instant k ∈ {1, . . . , K} and ᾱk ∈ C is a to assume that the functional form of pC (ck ) ≡ pC (ck ; γ)
deterministic, but unknown, scalar that may vary over k. Any is perfectly known, up to a possible (finite-dimensional) de-
measurement process involves a certain amount of disturbance. terministic nuisance vector parameter γ; for example, the
In radar signal processing, the disturbance is produced by (vectorized) covariance matrix. In order to obtain a consistent
two components, the clutter and white Gaussian measurement estimate γ̂ of γ, a secondary dataset1 has to be exploited
noise, and it is modelled as an additive random vector, say (see e.g. [5]). Note that the required pΛ|H0 is a function
ck , whose statistics may vary over k. Formally, the detection of γ̂ as well. A third simplifying assumption is that the
problem can be recast as a composite binary hypothesis test signal parameters ᾱk remain constant over k, i.e. ᾱk ≡ ᾱ, ∀k
(HT) [1]: [2]–[4]. Under these three assumptions, a possible choice
for the decision statistic is the generalized likelihood ratio
H0 : xk = ck k = 1, . . . , K, (GLR) ΛGLR (X) (see e.g. [6, Ch. 11] and references therein).
(1) However, a closed-form solution to (3) can be found only
H1 : xk = ᾱk vk + ck k = 1, . . . , K.
for a very limited class of disturbance models for which the
Gaussianity assumption needs also to be imposed. An asymp-
S. Fortunati, L. Sanguinetti, F. Gini and M. S. Greco are with University of
Pisa, Dipartimento di Ingegneria dell’Informazione, Via Caruso, 56122, Pisa, totic approximation for the solution to (3) can be obtained
Italy.
Braham Himed is with Air Force Research Laboratory, Dayton OH, USA. 1 In radar terminology, a secondary dataset is a set of “signal-free” snapshots
This work has been partially supported by the Air Force Office of Scientific collected form resolution cells adjacent to the one under test and sharing the
Research under award number FA9550-17-1-0344. same statistical characterization.

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
2

by exploiting a well-known asymptotic property of the GLR. C. Relevant literature

Under the hypothesis H0 and for K → ∞, the pdf of ΛGLR (X) The MIMO paradigm has been the subject of intensive
converges to the one of a central χ-squared random variable research over the past 15 years in radar signal processing.
with 2 degrees of freedom, denoted as χ22 (0) [6, Ch. 11]. Initially introduced in wireless communications as a new en-
Hence, by using the properties of the χ-squared distribution, abling technology, the MIMO framework has been recognized
it is immediate to verify that (3) is asymptotically satisfied by to have a great potential in boosting the capabilities of classical
λ̄ = −2 ln PF A . This is a particularly simple result that has antenna array systems. Based on the array configurations used,
received a lot of attention in the literature. However, it relies MIMO radars can be classified into two main types. The
on the four simplifying assumptions previously introduced and first type uses widely separated antennas (so-called distributed
summarized as follows: MIMO) to capture the spatial diversity of the target’s radar
A1 The disturbance vectors {ck }K k=1 are i.i.d. over the ob- cross section (RCS) [20]. The second type employs arrays
servation interval. of closely spaced antennas (so-called co-located MIMO) to
A2 The pdf pC (ck ; γ) of the disturbance is perfectly known, coherently combine the probing signals in certain points of
up to a unknown nuisance parameter vector γ. the search area [21]. Hybrid configurations are also possible.
A3 The target complex amplitude ᾱk is maintained constant While the advantages in terms of spatial resolution, pa-
over the observation interval, i.e. ᾱk ≡ ᾱ, ∀k. rameter identifiability, direction-of-arrival estimation and in-
A4 The number of temporal snapshots K is assumed to be terference mitigation have been largely investigated in the
much larger that the spatial channels N . MIMO radar literature, the potential benefits that a large
Even if these assumptions make the (asymptotic) analysis of number of virtual spatial antenna channels can bring into
ΛGLR (X) analytically tractable, they are seldom satisfied in the target detection problem in terms of robustness with
practical applications. respect to the generally unknown disturbance model have not
been explored yet. Surprisingly, not only the highly desirable
robustness property has been somehow disregarded but, as
B. Contributions pointed out in [22], even the availability of reliable, non-
trivial, disturbance models is scarce. Remarkable exceptions
This paper considers a co-located MIMO radar with MT to the mainstream Gaussianity assumption have been recently
transmitting and MR receiving antennas and aims at deriving discussed in [23] and in [24]. Particularly, in [23] the perfor-
a detector that satisfies pre-assigned performance requirements mance of the Adaptive Normalized Matched Filter (ANMF),
without relying on the four assumptions above. Inspired by exploiting robust estimators for the disturbance covariance
the recent developments in Massive MIMO communications matrix, has been investigated with non-Gaussian disturbance.
[7]–[10], we aim at exploring the potential benefits of having Specifically, random matrix tools have been used to obtain
a very large number of antennas. Particularly, we assume asymptotic approximations of the probabilities of false alarm
that a single time snapshot, i.e. K = 1, is collected, and and misdetection of the ANMF for the regime in which both
operate in the asymptotic regime where the number of virtual N and K go to infinity with a non-trivial ratio N/K. Similar
spatial antenna channels N = MT MR grows unboundedly, random matrix tools have been adopted in [24] to derive
i.e., N → ∞. This makes the three assumptions A1, A3 and some asymptotic (in random matrix regime) results about
A4 no longer needed. Advances in robust and misspecified the direction-of-departure and direction-of-arrival estimation
statistics ([11]–[16] and [17,18]) are used to dispose of the in a non-Gaussian disturbance setting. Again, the random
cumbersome and unrealistic assumption A2. By adopting a matrix machinery has been exploited in [25] to investigate
very general disturbance model taking into account the spatial the asymptotic performance of a GLRT detector in cognitive
correlation structure of the observed samples, we propose a radio applications. Specifically, the HT problem tackled in [25]
robust Wald-type detector that is asymptotically distributed, is similar to the one in (1), but the steering vector vk and
when N → ∞, as a χ-squared random variable (under both ᾱk are assumed unknown. In the same spirit of [25], [26]
H0 and H1 ) irrespective of the actual and unknown distur- has recently investigated the possibility to derive eigenvalues-
bance pdf pC (c). This asymptotic result is achieved without based detectors for HT problem of the form in (1) for spectrum
the need of any secondary dataset. Although the theoretical sensing and sharing in cognitive radio. However, in both [25]
findings of this paper are valid for a very general disturbance and [26], the disturbance was assumed to be a simple white
model, numerical results are provided for two non-Gaussian, Gaussian process with distribution a priori known, up to its
stable auto-regressive disturbance models of order p = 3 and statistical power. The asymptotic analysis in [24]–[26] requires
6. It turns out that a pre-assigned value of PF A = 10−4 is that both N and K grow unboundedly. This is different from
achieved for N = MT MR ≥ 104 with both models. This this paper where the temporal dimension K is kept fixed;
number of virtual spatial antenna channels defines what we specifically, we assume to collect a single snapshot vector.
call the Massive MIMO regime of the radar system.
Compared to our previous paper [19], the main difference
lies in the absence of any a priori knowledge of the disturbance D. Outline and Notation
model. In fact, in [19] the analysis was developed for an au- The reminder of the paper is organized as follows. In
toregressive model of order 1, but with no a priori knowledge Section II, the system and signal models as well as the
of its statistics. resulting HT problem are introduced, by focusing our attention

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
3

Target, φ
and velocity. The complex, vector-valued, random process
n(t) ∈ CMR accounts for the disturbance. We assume that
s(t) is obtained as a linear transformation of a set of nearly
orthonormal signals so (t) ∈ CMT , i.e. s(t) = Wso (t), where
W = [w1 , . . . , wMT ]T ∈ CMT ×MT and wm ∈ CM is the
weighting vector of the transmit antenna element m with
power ||wm ||2 .
Let l = 1, . . . , L and k = 1, . . . , K be the indices
Receiver
characterizing each time l∆t and frequency k∆ω samples,
respectively. The output X(l, k) ∈ CMR ×MT of the linear
Transmitter
filter matched to so (t) can be expressed as [22,28]:

X(l, k) = ᾱaR (φ̄)aT (φ̄)T WS(l, k) + C(l, k), (5)

Fig. 1. Co-located MIMO radar.
where
on the disturbance model for co-located MIMO radars. Section Z T
III describes our main results. Specifically, the explicit form so (t − τ̄ )sH −j(k∆ω−ω̄)t
S(l, k) , o (t − l∆t)e dt (6)
of the proposed robust Wald-type test is provided along with 0
the theoretical derivation of its asymptotic distribution under takes into account potential “straddling losses”, that are losses
both H0 and H1 . Numerical validations and simulation results due to a not precise centering of the target in a range-Doppler
are provided in Section IV. Some concluding remarks and gate or to a not exact orthogonality between waveforms, and
discussions are drawn in Section V. Z T
Throughout this paper, italics indicates scalar quantities (a), −jk∆ωt
C(l, k) = n(t)sHo (t − l∆t)e dt. (7)
lower case and upper case boldfaces indicate column vectors 0
(a) and matrices (A), respectively. Each entry of a N × N After omitting the indexes (l, k) for ease of notation, we may
matrix A is indicated as ai,j , [A]i,j , while the i-th column rewrite (5) in vectorial form as:
vector of A is indicated as ai such that A = [a1 , . . . , aN ].
We use ∗ , T , and H to indicate complex conjugation, trans- CN 3 x = vec (X) = ᾱv(φ̄) + c, (8)
pose, and the Hermitian operators, respectively. For random
where N , MR MT and
variables or vectors, =d stands for “has the same distribution
a.s.
as”. Also, → indicates the almost sure (a.s.) convergence v(φ̄) = ST ⊗ IMR

 T 
N →∞
W aT (φ̄) ⊗ aR (φ̄) (9)
p
and → indicates the convergence in p-probability. We call and c , vec (C). While aT (φ) and aR (φ) depend on the
N →∞
Q1 (·, ·) the Marcum Q function of order 1. The symbol bac geometry of the transmitting and receiving arrays, respec-
defines the nearest integer less than or equal to a ∈ R. tively, the matrix W can be designed to shape arbitrarily the
transmitting beam; see [29, Ch. 4] and references therein. For
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION example, with W = IMT , the transmitted power is uniformly
Consider a co-located MIMO radar system equipped with distributed over all possible directions. On the other hand, with
MT transmitting antennas and MR receiving antennas [21]. W = aT (φ)∗ aT (φ)T , it is fully directed towards the direction
The transmitting array is characterized by the array manifold, φ. Intermediate cases can be obtained [28].
also called steering vector, aT (φ), where φ is the position We assume that n(t) is zero-mean and wide-sense station-
vector defined in an absolute reference frame [27]. Similarly, ary; that is, E{n(t)} = 0, ∀t and E{n(t)n(τ )H } = Σ(t − τ ).
the receiving array can be characterized by the steering vector Hence, from (7) it easily follows that E{c} = 0 and
aR (φ) since the positions of the antennas in the absolute Z TZ T
reference frame are known; see Fig. 1. Γ , E{cc } =H
[s∗o (t − l∆t) ⊗ IMR ] Σ(t − τ )×
0 0
H
× [s∗o (t − l∆t) ⊗ IMR ] e−jk∆ω(t−τ ) dtdτ
A. Signal model Z TZ T
 ∗
so (t − l∆t)sTo (t − l∆t) ⊗ Σ(t − τ ) ×

Given a target with position vector φ̄, the signal collected =
0 0
at the receiving array can be modelled as [22,28]:
× e−jk∆ω(t−τ ) dtdτ. (10)
x(t) = ᾱaR (φ̄)aTT (φ̄)s(t − τ̄ )ejω̄t + n(t), t ∈ [0, T ], (4)
As seen, Γ is a function of Σ(t − τ ) (i.e., the covariance
where x(t) ∈ CMR is the array output vector at time t, matrix of n(t)) and so (t). In the literature (e.g., [30],[31, Ch.
s(t) ∈ CMT is the vector of transmitted signals, ᾱ ∈ C 4]), a simple model for n(t) is to assume that its samples are
accounts for the radar cross section of the target and the uncorrelated in both spatial (along the receiving array) and
two-way path loss, which is the same for each transmitter temporal (along T ) domains. This implies that Σ(t − τ ) =
and receiver pair. This is generally verified in co-located σ 2 IMR δ(t−τ ). If so (t) is a vector of orthonormal waveforms,
MIMO radars [21]. The parameters τ̄ and ω̄ represent the it is thus immediate to verify that Γ in (10) reduces to Γ =
actual time delay and Doppler shift, due to the target position σ 2 IN . Under the assumption of uncorrelated samples in the

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
4

time domain only and perfect orthonormality of the transmitted Density (PSD) [40, Ch. 3]. Moreover, a non necessarily Gaus-
waveforms, we have that Γ = IMT ⊗ ΣR where ΣR denotes sian ARMA(p, q) is able to model the “spikiness” of heavy-
the receive spatial covariance matrix [32]. However, the above tailed data as well. Another disturbance model of practical
two conditions may not be satisfied in practice [22,33]. This interest satisfying Assumption 1 is the Compound-Gaussian
is why in this paper we do not make any a priori assumption (CG) model [41,42]. Indeed, recall that any CG-distributed
on the structure of Γ. We only assume that its (i, j)-th entry random vector c admits a representation:
goes to zero at least polynomially fast as |i−j| increases. This √
assumption will be discussed in the next section and formally c =d τ m (11)
introduced in Assumption 1. for some real-valued positive random variable τ , called texture,
independent of the zero-mean, N -dimensional, circular, com-
B. Disturbance model plex Gaussian random vector, called speckle, m ∼ CN (0, Γ),
As previously discussed, many simplified models have where Γ is its scatter matrix. Under a condition on the
been proposed in the literature to statistically characterize structure of Γ, it easily follows that the N entries of m
the disturbance vector c at the output of the matched filter can be interpreted as N random variables extracted from a
bank of a (co-located) MIMO radar system. We refer to circular, Gaussian, SOS ARMA(p, q) process {mn : ∀n}, with
[22,33] for a comparison among various statistical models and p, q < N .
for a discussion about the physical simplifying assumptions We conclude by noticing that Assumption 1 is more general
underlying them. Here, we simply note here that the two main than the one adopted in our previous work [19]. In fact, the
hypotheses usually made about the statistical characterization asymptotic results derived in [19] are obtained by assuming an
of the disturbance vector are: i) c is temporally and spatially AR disturbance model,3 driven by innovations with possibly
white, and ii) c is Gaussian-distributed. As we will show, unknown pdf. Unlike [19], this paper does not require any
advances in robust and mis-specified statistics allow us to a priori information on the specific disturbance model; as
drop these two strong assumptions in favor of much weaker stated in Assumption 1, only the polynomial decay of its auto-
conditions. correlation function is needed.
To formally characterize the class of random processes to
which the results of this paper apply, we need to introduce C. The hypothesis test problem
the concepts of uniform and strong mixing random sequences
Based on the assumptions discussed above, the HT problem
[34]–[37]. Roughly speaking, the uniform and strong mixing
for target detection in (1) can be expressed as:
properties characterize the dependence between two random
variables extracted from a discrete process separated by m H0 : x = c,
(12)
lags. Without any claim of completeness or measure-theoretic H1 : x = ᾱv + c,
rigor, the results of this paper apply to any random process
that satisfies a restriction on the speed of decay of its auto- where c ∈ CN is the disturbance vector whose entries are
correlation function. Specifically, we limit ourselves to the sampled from a complex random process {cn : ∀n} satisfying
following class of random processes: Assumption 1. Note that, in practical radar scenarios, (12)
needs to be solved for any radar resolution cell of interest.
Assumption 1. Let {cn : ∀n} be a stationary discrete and Specifically, let {φi ; i = 1, . . . , Q} be the set of position
circular complex-valued process [38] representing the true, vectors pointing at Q radar resolution cells, defined in an
and generally unknown, disturbance. Then, we assume that absolute reference frame. Then, the presence (or absence)
its autocorrelation function satisfies rC [m] , E{cn c∗n−m } = of a target has to be tested for all the Q resolution cells.
O(|m|−γ ), m ∈ Z, γ > %/(% − 1), % > 1.2 Consequently, (12) has to be solved for any different vector
Assumption 1 implies that the volume of the MIMO radar v(φi ), whose explicit form is defined in (9). Notice also that
must increase with the number of transmitting (MT ) and a single-snapshot, i.e., K = 1, is used in (12).
receiving (MR ) antennas. This means that the results of this To discriminate between H0 and H1 in the composite HT
paper do not apply to “space-constrained” array topologies. problem (12), a test statistic is needed. Classical model-based
That said, Assumption 1 is very general and allows to ac- test statistics such as the Generalized Likelihood Ratio (GLR)
count for most practical disturbance models. Indeed, any test and the Wald test (see e.g. [44]) cannot be used since the
(not necessary Gaussian) stable second-order stationary (SOS) functional form of the disturbance pdf pC and, consequently,
ARMA(p, q), and consequently any stable SOS AR(p) [39], of the data pdf pX are unknown. We propose the following
satisfies Assumption 1, since the auto-correlation function of approach. Since no a priori information on the functional form
any stable SOS ARMA decays exponentially. The generality of pX is available, let us choose the simplest estimator, say α̂,
of the ARMA model is because it can approximate, for p and for the signal parameter ᾱ. Then, building upon the asymptotic
q sufficiently large, the second-order statistics of any complex statistics of α̂ as N → ∞, we define a Wald-type test to
discrete random processes having a continuous Power Spectral solve the HT problem in (12). In fact, unlike the GLR statistic
that requires the explicit functional form of the data pdf, a
2 Given a real-valued function f (x) and a positive real-valued function g(x),
f (x) = O(g(x)) if and only if there exists a positive real number a and a 3 We considered only the AR(1) case, but the theory could be readily
real number x0 such that |f (x)| ≤ ag(x), ∀x ≥ x0 . extended to a general AR(p) as discussed in [43].

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
5

√
Wald-type statistic only needs an asymptotically normal, N - and asymptotically normal
consistent estimator of ᾱ and a consistent estimate of its error √ −1/2
covariance. As shown next, this key fact allows us to derive a N B̄N AN × (α̂ − ᾱ) ∼ CN (0, 1), (17)
N →∞
robust test statistic for Massive MIMO radar configurations.
where
III. M AIN R ESULTS AN , N −1 ||v||2 , (18)
The main result of this section is the derivation of a robust B̄N , N −1 H
v Γv, (19)
Wald-type test for (12) with the valuable property to have,
under H0 , an asymptotic distribution invariant w.r.t. pX . The and Γ , EpC {ccH }, with pC being the true (but generally
closed form expressions of its asymptotic distribution under unknown) disturbance pdf.
both H0 and H1 will be provided. Since (12) is a composite
HT problem (i.e., it depends on the unknown deterministic Proof: All the required regularity conditions and a
signal parameter ᾱ) a prerequisite for the implementation of measure-theoretic rigorous proof (for the real-valued case)
a decision statistic is the derivation of an estimator α̂ of ᾱ. can be found in [17] and [18], while in Appendix A of this
paper we provide the reader with an “easy-to-follow” but
still insightful sketch of the proof in the complex case. Here,
A. Estimation of ᾱ under dependent data only the principal facts underlying the proof of Theorem 1
By relying on the outcomes of [17,18], are discussed. To prove the consistency of α̂, we need an
√ in the following we
show that an asymptotically normal, N -consistent estimator extension of the Law of Large Numbers (LLN) to uniform
of ᾱ can be easily implemented under any general disturbance and strong mixing random sequences (see Assumption 1). This
models satisfying Assumption 1. result is stated in Theorem 2.3 of [17]. With this suitable
A standard procedure is to recast an estimation problem into generalization of the LLN, the (strong) consistency of the LS
a relevant (possibly constrained) optimization problem. In the estimator can be readily established as shown by Theorem 3.1
application at hand, an estimate of α can be obtained as: in [17]. Regarding the asymptotic normality, a generalization
to uniform and strong mixing random sequences of the Central
α̂ = argmin GN (x, α), x ∼ pX , (13) Limit Theorem (CLT) is required. This extension can be found
α∈C
in [17, Th. 2.4] for the scalar case, and in [45, Th. 2] for
where GN (·, ·) is a suitable objective function and x = the multivariate case. The asymptotic normality of α̂ can be
[x1 , . . . , xN ]T ∈ CN is the available dataset characterized by established by a direct application of this version of the CLT as
an unknown joint pdf pX . shown in [17, Th. 3.2]. Finally, the extension of these results
Given the measurement model (under H1 ) in (12), the to the complex field can be obtained simply by exploiting the
most natural choice for GN (·, ·) is a Least-Square (LS) based natural set isomorphism between C and R2 and by using the
objective function: fact that only circular random sequences are considered.
N
X Remark 1. Note that, since α̂ is obtained as a linear combi-
GN (x, α) , |xn − αvn |2 = (x − αv)H (x − αv) nation of circular observations x1 , . . . , xN , its second-order
n=1 (14) statistics are fully characterized by the variance EpX {|α̂ −
2
vH x |vH x|2 ᾱ|2 }, while its pseudo-variance EpX {(α̂ − ᾱ)2 } is nil [38].
= ||x||2 + ||v||2 α − −
||v||2 ||v||2
While AN in (18) is only a function of the known norm of v,
whose minimum is reached when the second addend vanishes. B̄N in (19) requires to compute the expectation w.r.t. the true,
This yields but unknown, disturbance pdf pC . It can be shown that, under
vH x a uniform (or strong) mixing condition for the disturbance
α̂ = , (15) process {cn : ∀n}, a consistent estimator of B̄N is [17]:
||v||2
which is a well-known result in the radar signal processing N
X
−1
literature addressing decision rules in Gaussian environment. B̂N ≡ B̂N (α̂) = N |vn |2 |ĉn |2
It can also be noted that, under a misspecified (white) Gaussian n=1
assumption on the disturbance vector c, the LS estimator in l
X N
X
+ 2N −1 ∗
ĉn ĉ∗n−m

(15) coincides with the Mismatched Maximum Likelihood Re vn vn−m (20)
estimator [13],[18]. m=1 n=m+1

By specializing the general findings about misspecified where l is the so-called truncation lag [17],
estimation under dependent data proposed in [17] and [18],
the asymptotic properties of the LS estimator in (15) can be ĉn = xn − α̂vn , ∀n (21)
obtained as shown in the following Theorem 1.
and α̂ is given in (15). The estimate B̂N in (20) can be
Theorem
√ 1. Under Assumption 1, the LS estimator α̂ in (15) rewritten in a more compact form as:
is N -consistent
a.s.
α̂ → ᾱ
N →∞
(16) B̂N = N −1 vH Γ
b l v, (22)

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
6

4
where where ς , 2|ᾱ|2 v||v||
H Γv .
 ĉi ĉ∗j

j−i≤l
[Γ
b l ]i,j , ĉ∗ ĉ i−j ≤l (23) Proof: The proof follows from Theorem 1 and a known
 i j result about circular Gaussian random variables [48]. In partic-
0 |i − j| > l
ular, if a ∼ CN (µa , σa2 ), then 2|a − µa |2 /σa2 ∼ χ22 (0). Details
for 1 ≤ i, j ≤ N . The consistency of B̂N is stated in the next are given in Appendix B.
theorem (see [17, Th. 3.5]). The above theorem shows that the pdfs of (25) under
Theorem 2. Under Assumption 1, if l → ∞ as N → ∞ such H0 and H1 converge to χ-squared pdfs with 2 degrees of
that l = o(N 1/3 ) then 4 : freedom when the number of virtual spatial antenna chan-
p nels N = MT MR goes to infinity. This means that (3) is
B̂N − B̄N → 0. (24) asymptotically satisfied by λ̄ = −2 ln PF A . This is valid
N →∞
for any pre-assigned PF A and, more importantly, for any
Proof: The proof is given in [46, Th. 6.20].
disturbance process {cn : ∀n} satisfying Assumption 1. In
Theorem 2 provides us with a useful criterion to choose the other words, we could say that ΛRW (x) achieves the CFAR
truncation lag l. In fact, to ensure the consistency of B̂N , l property w.r.t. all the disturbance distributions satisfying As-
has to grow with N , but more slowly than N 1/3 . sumption 1 when a sufficiently large number of transmitting
and receiving antennas is used. Numerical results will be used
B. A robust Wald-type test to show that such Massive MIMO regime is achieved for a
Given the asymptotic characterization of the LS estimator feasible large number of antennas. Some considerations on
presented in Theorem 1, the Wald-type statistic can be set up the asymptotic distribution of ΛRW (x) under H1 can also be
as: done. In particular, in order to make explicit the dependence of
2N |α̂|2 2|vH x|2 ς from aT (φ̄), aR (φ̄) and W, one can substitute the definition
ΛRW (x) = −2 = , (25) of v given in (9) into ς to obtain:
A B̂N vH Γ
blv
N
2|ᾱ|2 MR2 k(WS)T aT (φ̄)k4
where the entries of Γ b l are given in (23). The similarity ς= 
.
tr Γ (WS)T aT (φ̄)aH H

∗
between the proposed ΛRW statistic and the Adaptive Matched T (φ̄)(WS) ⊗ aR (φ̄)aR (φ̄)
Filter (AMF) proposed in [47] is evident. However, the fol- (28)
lowing comments are in order. Remark 2. Further manipulations to the expression of ς in
First, in [47] a set of homogeneous secondary snapshots, (28) are allowed if the model in (10) is adopted for the
i.e. a set of “signal-free” data collected from radar resolution covariance matrix Γ. Specifically, by substituting (10) in (28),
cells adjacent to the cell under test or at different time instants, we get:
is used to estimate the disturbance covariance matrix. This
2|ᾱ|2 MR k(WS)T aT (φ̄)k2
approach, however, requires that the disturbance statistics ς = RR T ,
remain constant over all the considered resolution cells or time 0
||so (t − ¯l∆t)||2 tr [Σ(t − τ )] e−j k̄∆ω(t−τ ) dtdτ
instants. Unlike the AMF, the decision statistic ΛRW in (25) (29)
does not need any secondary data since it is able to extract where ¯l and ω̄ defines the range-Doppler gate under test.
all the required information from the single snapshot collected Moreover, if so (t) is a vector of perfectly orthonormal wave-
in the cell under test. Secondly, the AMF in [47] is derived forms, i.e. S = IMT , and if Σ(t − τ ) = σ 2 IMR δ(t − τ ), (29)
under the Gaussian assumption for the disturbance vector can be further simplified as:
c. Conversely, ΛRW in (25) can handle all the disturbance 2|ᾱ|2 P (φ̄)2
distributions that satisfy Assumption 1, including the Gaussian ς= , (30)
σ2
one. Third, as shown in Theorem 3 below, ΛRW in (25) is ∗
where P (φ) , aH T
T (φ)W W aT (φ) is the beam pattern of
shown to have the Constant False Alarm Rate (CFAR) property
the transmitting array as a function of the position vector
as N → ∞ for all the disturbance distributions satisfying
φ. The same result can be found in [30], where a LR
Assumption 1 and without the need of any secondary data.
test, perfectly matched to white Gaussian disturbance, was
This is a great advantage w.r.t. the AMF that is CFAR
exploited as a detector.
only if the disturbance is Gaussian-distributed and a set of
homogeneous secondary data is available. The expression in (30) suggests us an interesting fact. Under
The asymptotic property of ΛRW (x) can be stated as fol- the specific and simplistic scenario of perfectly orthonormal
lows. waveforms and (spatial and temporal) white Gaussian distur-
Theorem 3. If Assumption 1 hold true, then bance, a perfectly matched LR-based detector has the same
(asymptotic) detection performance of the robust Wald-type
ΛRW (x|H0 ) ∼ χ22 (0), (26) test proposed here in (25). In other words, under the simplistic
N →∞
ΛRW (x|H1 ) ∼ χ22 (ς) , (27) scenario mentioned above, ΛRW in (25) does not present any
N →∞ loss in detection w.r.t. the “optimal” LR test even if ΛRW does
4 Given a real-valued function f (x) and a strictly positive real-valued
not require any a priori knowledge about the Gaussianity of the
function g(x), f (x) = o(g(x)) if for every positive real number a, there disturbance, while the LR test does. It is also worth mentioning
exists a real number x0 such that |f (x)| ≤ ag(x), ∀x ≥ x0 . that, in a more involved and realistic scenario, an LR-based test

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
7

0 0
ν1 = −0.2
ν2 = 0
10 log10 (PSD(ν)/ maxν PSD(ν))

10 log10 (PSD(ν)/ maxν PSD(ν))

−5 ν3 = 0.2 −5
PSD [dB]

−10 −10

−15 −15

ν1 = −0.2
−20 −20 ν2 = 0
ν3 = 0.2
PSD [dB]
−25 −25
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4
Spatial frequency: ν Spatial frequency: ν

Fig. 2. PSD of the AR(3) in Scenario 1. Fig. 3. PSD of the AR(6) in Scenario 2.

is practically unfeasible due to the lack of a priori information a transmitting array whose MT elements are spaced by MR d.
on the functional form of the disturbance pdf. Moreover, This implies that
even if such disturbance pdf were available, the derivation
aR (φ) = [1, ej2πν , . . . , ej2π(MR −1)ν ]T , (33)
of LR statistics would likely be met with the impossibility to
derive its closed form expression. On the contrary, the closed j2πMR ν j2π(MT −1)MR ν T
aT (φ) = [1, e ,...,e ] , (34)
form expression and the asymptotic detection performance of
ΛRW in (25) remain unchanged under any known or unknown where the spatial frequency ν is
disturbance process satisfying Assumption 1. f0 d
sin(g(φ)), ν, (35)
The following corollary is found. c
Corollary 1. If Assumption 1 holds true, then the detection where f0 is the carrier frequency of the transmitted signal, c is
probability of (25) is such that the speed of light and g(·) is a known function of the position
√ √  vector φ. By substituting (33) and (34) into (9), the vector
PD (λ) →N →∞ Q1 ς, λ , (31) v(φ) ∈ CN can be expressed as, for i = 1, . . . , N = MR MT
where Q1 (·, ·) is the Marcum Q function of order 1 [49] and [v(φ)]i = ej2π(i−1)ν (36)
4
ς is still given by ς = 2|ᾱ|2 v||v||
H Γv . which represents the steering vector of an equivalent phased-
Proof: By definition array with N elements [50].
Z ∞ Numerical results are obtained by averaging over 106 Monte
PD (λ) , Pr {Λ(X) ≥ λ|H1 } = pΛ|H1 (a|H1 )da. (32) Carlo simulations. Moreover, the truncation lag in Theorem 2
λ is chosen as l = bN 1/4 c.
Then, (31) follows directly from (27) and the properties of the
non-central χ2 distribution [49]. A. Disturbance models
Since Q1 (·, ·) is monotonic in its first argument, Corollary Two different models are considered for the disturbance. In
1 states that the PD of the RW test in (25) goes to 1 as Scenario 1, the disturbance vector c is generated according to
N → ∞. Moreover, it shows that PD depends on the true, an underlying circular, SOS AR(p)
but unknown, covariance matrix Γ of the disturbance vector Xp
c, the radar geometry, and the waveform matrix W (through cn = ρ̄i cn−i + wn , n ∈ (−∞, ∞), (37)
i=1
the vector v).
with p = 3, driven by i.i.d., t-distributed innovations wn whose
IV. N UMERICAL A NALYSIS pdf pw is [14]:
Numerical results are now used to validate the theoretical λ
 λ 
λ λ |wn |2
−(λ+1)
findings on the asymptotic properties of ΛRW as stated in The- pw (wn ) = 2 + 2 , (38)
σw π η η σw
orem 3. We consider a uniform linear array at the transmitter
2
and receiver, and a single target located in the far-field. We where λ ∈ (1, ∞) and η = λ/(σw (λ − 1)) are the shape
assume that W = IMT and the transmitted waveforms are and scale parameters. Specifically, λ controls the tails of
orthogonal, i.e., S = IMT . Following [50], we choose the pw . If λ is close to 1, then pw is heavy-tailed and highly
radar geometry that maximizes the parameter identifiability. non-Gaussian. On the other hand, if λ → ∞, then pw
This is achieved by using a receiving array characterized by collapses to the Gaussian distribution. In our simulations, we
2
MR antenna elements with an inter-element spacing of d and set λ = 2 and σw = 1. The AR(3) coefficient vector is

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
8

10−1 10−1
ν1 = −0.2 ν1 = −0.2
ν2 = 0 ν2 = 0
ν3 = 0.2 ν3 = 0.2
Nominal PF A Nominal PF A
10−2 10−2
PF A - AR(3)

PF A - AR(6)
10−3 10−3

10−4 10−4

102 103 104 105 102 103 104 105

Virtual spatial antenna channels: N = MR MT Virtual spatial antenna channels: N = MR MT

Fig. 4. Estimated PF A as a function of the virtual spatial Fig. 5. Estimated PF A as a function of the virtual spatial
antenna channels N in Scenario 1. antenna channels N in Scenario 2.
1
ρ̄ = [0.5ej2π0 , 0.3e−j2π0.1 , 0.4ej2π0.01 ]T . The normalized PSD
can be expressed as can be expressed as:
0.8
Xp −2
S(ν) , σw 2
1− ρ̄n e−j2πnν , p = 3, (39)
n=1 PD - AR(3)
0.6
and is shown in Fig. 2. As seen, Scenario 1 is characterized
by a disturbance whose power is mostly concentrated around
the spatial frequency ν = 0. 0.4
To prove the robustness of the proposed ΛRW w.r.t.
more general disturbance models, in Scenario 2 we in- SNR = -20 dB
crease the order of the AR process generating the distur- 0.2 SNR = -10 dB
SNR = -5 dB
bance vector c. Particularly, we consider a circular, SOS
Nominal PD
AR(6) driven (as before) by i.i.d., t-distributed innovations
0 2
wn and characterized by the following coefficient vector 10 103 104 105
ρ̄ = [0.5e−j2π0.4 , 0.6e−j2π0.2 , 0.7ej2π0 , 0.4ej2π0.1 , 0.5ej2π0.3 , Virtual spatial antenna channels: N = MR MT
0.6ej2π0.35 ]T . The normalized PSD is reported in Fig. 3 and
Fig. 6. Estimated and nominal PD as a function of the virtual
shows that, differently from Scenario 1, the disturbance power
spatial antenna channels N in Scenario 1 for different SNR
is spread over the whole range of ν. Moreover, it presents more
values, spatial frequency ν = 0, and nominal PF A = 10−4 .
than a single peak.
In both scenarios, the disturbance process is normalized
in order to have σ 2 = rC [0] = 1. Under hypothesis Fig. 6 considers Scenario 1 and illustrates the estimated
H1 , the signal-to-noise ratio is simply defined as SNR , and the closed-form expression of PD given in Corollary
10 log10 (|ᾱ|2 /σ 2 ). 1 for three distinct SNR values. Particularly, we assume
SNR = −20, −10 and −5 dB. As seen, the PD tends to 1 as
the number of virtual spatial antenna channels N increases.
B. Performance Analysis Specifically, for SNR ≥ −20 dB the PD approaches 1 for
N ≥ 104 . From Fig. 6, it is also immediate to verify that
Since the disturbance PSD in Figs. 2 and 3 is not constant
the PD estimated through Monte-Carlo runs is in perfect
w.r.t the spatial frequency ν, the performance of ΛRW (x) will
agreement with the theoretical one provided in Corollary 1.
be evaluated for three different values of ν corresponding to
We conclude by noticing that similar numerical results have
different disturbance power density levels (i.e. three different
been obtained for the CG disturbance model discussed in Sub-
target DOAs): ν1 = −0.2, ν1 = 0 and ν1 = 0.2. Figs. 4 and 5
section II-B. Since they are perfectly in line with the numerical
plot the PF A of the proposed robust Wald test in (25) for the
analysis reported above, we decided to not include them here
three considered values of ν. As we can see, in both scenarios
due to lack of space.
ΛRW in (25) achieves the nominal value of 10−4 for N ≥ 104 .
Moreover, in this Massive MIMO regime, i.e. for N ≥ 104 ,
the progress of the PF A curves for different scenarios and V. C ONCLUDING REMARKS AND DISCUSSIONS
for different values of spatial frequency are almost identical. The detection problem in co-located MIMO radar systems
This provides a numerical validation of the theoretical result was analysed in this paper. A robust Wald-type detector was
provided in (26) of Theorem 3. proposed and its asymptotic performance investigated when

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
9

the virtual spatial antenna channels N = MT MR goes to Theorem 4. If ḠN (α) has a unique minimum at α0,N ∈ C,
infinity. Specifically, the CFAR property of the proposed then from (41):
detector for the asymptotic regime N → ∞ and the wide 1 a.s.
family of disturbance processes satisfying Assumption 1 was GN (x, α̂) − ḠN (α0,N ) → 0 (42)
N N →∞
mathematically proved and validated through numerical simu-
and
lations. The purpose of analysing the asymptotic performance a.s.
α̂ − α0,N → 0 (43)
when N → ∞ is not that we advocate the deployment of N →∞
radars with a nearly infinite number of virtual antennas. The
where α0,N , argmin ḠN (α).
importance of asymptotics is instead what it tells us about α∈C
practical systems with a finite number of antennas. Indeed, Proof: See the proof of Theorem 2.4 in [52].
our main results imply that we can always satisfy performance Note that α0,N represents the counterpart of the pseudo-
requirements by deploying sufficiently many virtual antennas true parameter defined in [12] for the i.i.d. data case. The
N , without any a priori knowledge of the disturbance statistics. consistency of α̂ can finally be established by showing that
Our numerical results showed that a pre-assigned value of the pseudo-true parameter α0,N is equal to the true one ᾱ. To
PF A = 10−4 can be achieved with N = MT MR ≥ 104 this end, from (40) we obtain:
with non-Gaussian, stable autoregressive disturbance models
N
of order p = 3 and 6. This defines the so-called Massive X
EpX |xn − αvn |2

MIMO regime of the radar. ḠN (α) =
n=1
In Massive MIMO communications, linear combining, and
precoding schemes can entirely eliminate the interference = σc2 − |ᾱ − α|2 ||v||2 ×
N
 
as the number of antennas grows unboundedly even with P
2Re (ᾱ − α) EpX {xn − ᾱvn } vn . (44)
imperfect knowledge of propagation channels. We showed n=1
that a large-scale MIMO radar yields a target detector, which
By definition of data process in (12), EpX {xn − ᾱvn } =
is robust to the unknown disturbance statistics. We foresee
0, ∀n, then the minimum of Q̄N (α) is attained at α0,N = ᾱ.
that further breakthroughs can be obtained by extending the
This proves the consistency of α̂ under Assumption 1.
Massive MIMO concept to other radar problems [10]. Clearly,
by using very large arrays for waveform design, one can rad-
ically improve the spatial diversity gain and spatial resolution B. Asymptotic normality
for target detection, parameter estimation, and interference The proof for the asymptotic normality of the LS estimator
rejection. However, the lesson learned from the last decade α̂ mimics the one provided in standard statistical textbooks
of research in communications is that Massive MIMO is not (see e.g. [53, Ch. 9]) for the Maximum Likelihood estimator.
merely a system with many antennas but rather a paradigm Let us start by taking the Taylor’s expansion of the real-valued
shift with regards to the modelling, operation, theory, and objective function GN (x, α) around the complex parameter ᾱ
implementation of MIMO systems. Our vision is that this [54, Th. 3.3]:
paradigm can be applied also to radars, and open new research ∂GN (x, α)
directions. GN (x, α) w GN (x, ᾱ) + (α − ᾱ) +
∂α α=ᾱ
2
∂GN (x, α) 2 ∂ GN (x, α)
A PPENDIX A +(α − ᾱ)∗ + |α − ᾱ| ,
P ROOF OF T HEOREM 1 ∂α∗ α=ᾱ ∂α∂α∗ α=ᾱ
(45)
In this Appendix, the main ideas behind the proof √ of
Theorem 1 are discussed. Specifically, we show the N - where we used the fact that:
consistency and the asymptotic normality of the mismatched ∂ 2 GN (x, α) ∂ 2 GN (x, α)
LS estimator α̂ in (15) under Assumption 1. = = 0, ∀α. (46)
∂α2 ∂α∗2
By using the Mean Value Theorem [53, Ch. 9], for some α̃
A. Consistency such that |α̃ − ᾱ| < |α̂ − ᾱ| we have:
Let us recall here the explicitP
expression of the LS objective
N ∂GN (x, α) ∂ 2 GN (x, α)
function in (14): GN (x, α) , n=1 |xn − αvn |2 . Moreover, + (α̂ − ᾱ) = 0, (47)
let ḠN (α) be the following measurable and continuous func- ∂α∗ α=ᾱ ∂α∂α∗ α=α̃
tion: XN where α̂ is defined in (13). Consequently, we obtain
EpX |xn − αvn |2 .

ḠN (α) , (40)  2 −1
n=1 ∂ GN (x, α) ∂GN (x, α)
α̂ − ᾱ = − ∗
. (48)
Then, under Assumption 1, from [17, Th. 2.3] and [51, Th. ∂α∂α α=α̃ ∂α∗ α=ᾱ
1], we have that For simplicity, we define
1 a.s.
sup GN (x, α) − ḠN (α) → 0, (41) 1 ∂ 2 GN (x, α)
α∈Ω N N →∞ AN (x, α) , , (49)
N ∂α∂α∗
2
where Ω is a compact subset of C(≡ R ). For the LS estimator ∂GN (x, α)
α̂ in (15), the following convergence property holds. s(x, α) , , (50)
∂α∗

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
10

and rewrite (48) as From (58) and (59), by a direct application of the Slutsky’s
√

1
 Theorem and of the properties of the complex Gaussian
N (α̂ − ᾱ) = −AN (x, α̃)−1 √ s(x, ᾱ) . (51) distribution [48], we immediately obtain that:
N √ ∗
−1/2
Through direct calculation, it easily follows that ΛRW (x|H0 ) =2 N AN B̂N α̂ ×
| {z }
N
X ∼ CN (0,1)
AN (x, α) = N −1 |vn |2 = N −1 ||v||2 , AN , (52) √
N →∞

−1/2
n=1 N AN B̂N α̂ ∼ χ22 (0), (61)
N N N →∞
X X | {z }
s(x, ᾱ) = − vn∗ (xn − ᾱvn ) = − vn∗ cn , (53) ∼ CN (0,1)
N →∞
n=1 n=1
where χ22 (0) indicates a central χ-squared random variable
where cn = xn − ᾱvn ,∀n. By substituting (52) and (53) into with two degrees of freedom.
(51), we get:
N
!
√ N 1 X ∗ B. Asymptotic distribution of ΛRW under local alternatives
N (α̂ − ᾱ) = √ vn cn . (54)
||v||2 N n=1 Following [53, Ch. 9], we suppose that the alternative to
H0 is of the form:
The asymptotic normality of α̂ follows from the application √
of the Central Limit Theorem established for (strong) mixing H1 : α = d/ N , d ∈ C. (62)
processes in [17, Th. 2.4] for the real scalar case and in [45]
Note that, as stated in [53, Sec. 9.3], this choice is made
for the real multivariate case. In particular, let us define
  only for mathematical purposes, and has no direct physical
1 1 significance. Specifically, (62) allows us to approximate the
B̄N ≡ BN (ᾱ) , EpX √ s(x, ᾱ) √ s∗ (x, ᾱ)
N N power of the test (or, in radar terminology, the probability
N X N of detection) locally, i.e. in the neighbourhood
√ of the null
1 X (N )
= v ∗ vm EpX {cn c∗m }. (55) hypothesis in (56). By defining ᾱH1 , d/ N as the true
N n=1 m=1 n parameter under local alternatives, we clearly have
Under Assumption 1, by exploiting [17, Th.3.2] and recalling (N )
lim ᾱH1 = ᾱH0 . (63)
that {cn : ∀n} is a circular process, (17) follows. N →∞

If BN (ᾱ) in (55) is continuous in a neighbourhood of ᾱ, then

A PPENDIX B from (63) and (56) it follows that
A SYMPTOTIC DISTRIBUTION OF ΛRW (N )
lim BN (ᾱH1 ) = BN (0). (64)
N →∞
To establish the asymptotic distribution of ΛRW , we follow
the standard procedure discussed in [53, Ch. 9]. Under H1 in (62), from Theorem 1 we have
(N ) a.s.
α̂ − ᾱH1 → 0, (65)
N →∞
A. Asymptotic distribution of ΛRW under H0
√ h 
(N )
i−1/2 √
We start by defining N AN BN ᾱH1 (α̂ − d/ N ) CN (0, 1). ∼
N →∞
ᾱH0 , 0 (56) (66)
As before, by using the Continuous Mapping Theorem [55,
as the true parameter under the null hypothesis. Under H0 , Th. 2.7] and the limiting results in (64), we have
from Theorem 1, we have that: h i−1/2 h  i−1/2
(N )
a.s. B̂N (α̂) − BN ᾱH1
α̂ → ᾱH0 , (57)
N →∞ −1/2 −1/2 p
√ ≡ B̂N − [BN (0)] → 0. (67)
−1/2 N →∞
N AN [BN (ᾱH0 )] α̂ ∼ CN (0, 1). (58)
N →∞ By recasting the test in (25) as in (60), from (66) and (67), by a
The inverse operator and the principal square root are both direct applications of the Slutsky’s Theorem, we immediately
continuous operators on R+ , then their composition is contin- obtain that:
√ ∗
uous on R+ . Then, from (25) and by using the Continuous −1/2
ΛRW (x|H1 ) = 2 N AN B̂N α̂ ×
Mapping Theorem [55, Theo. 2.7], it follows that: | {z }
h i−1/2
−1/2 ∼ CN (AN [BN (0)]−1/2 d,1)
B̂N (α̂) − [BN (ᾱH0 )] N →∞
√ 
−1/2
−1/2 −1/2 p N AN B̂N α̂
≡ B̂N − [BN (0)] → 0. (59)
N →∞ | {z }
∼ CN (AN [BN (0)]−1/2 d,1)
Let us rewrite the test in (25) as: N →∞
 
√ ∗ √ −1
−1/2 −1/2
 ∼ χ22 2|d2 |A2N [BN (0)] . (68)
ΛRW (x) = 2 N AN B̂N α̂ N AN B̂N α̂ . (60) N →∞

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
11

Note that, from (55) we have BN (0) = N −1 vH Γv. This [19] S. Fortunati, L. Sanguinetti, M. S. Greco, and F. Gini, “Scaling up
result can be used to approximate the power of the test, i.e. MIMO radar for target detection,” in ICASSP 2019 - 2019 IEEE
√ International Conference on Acoustics, Speech and Signal Processing
the PD . By setting d ≡ N ᾱ, we have: (ICASSP), May 2019, pp. 4165–4169.
[20] A. M. Haimovich, R. S. Blum, and L. J. Cimini, “MIMO radar with
2|ᾱ|2 ||v||4
 
widely separated antennas,” IEEE Signal Processing Magazine, vol. 25,
ΛRW (x|H1 ) ∼ χ22 . (69)
N →∞ vH Γv no. 1, pp. 116–129, 2008.
[21] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal
By using the properties of the non-central χ-squared distribu- Processing Magazine, vol. 24, no. 5, pp. 106–114, Sept 2007.
[22] B. Friedlander, “On signal models for MIMO radar,” IEEE Transactions
tion with 2 degrees of freedom [49], a closed form expression on Aerospace and Electronic Systems, vol. 48, no. 4, pp. 3655–3660,
of the asymptotic probability of detection can be eventually October 2012.
expressed as: [23] A. Kammoun, R. Couillet, F. Pascal, and M. Alouini, “Optimal design of
the adaptive normalized matched filter detector using regularized Tyler
√ √ √  estimators,” IEEE Transactions on Aerospace and Electronic Systems,
PD (λ) = Q1 2|ᾱ|||v||2 / vH Γv, λ . (70) vol. 54, no. 2, pp. 755–769, April 2018.
[24] H. Jiang, Y. Lu, and S. Yao, “Random matrix based method for joint
DOD and DOA estimation for large scale MIMO radar in non-gaussian
R EFERENCES noise,” in 2016 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), March 2016, pp. 3031–3035.
[1] W. L. Melvin, “A STAP overview,” IEEE Aerospace and Electronic [25] P. Bianchi, M. Debbah, M. Maida, and J. Najim, “Performance of
Systems Magazine, vol. 19, no. 1, pp. 19–35, Jan 2004. statistical tests for single-source detection using random matrix theory,”
[2] E. Aboutanios and B. Mulgrew, “Hybrid detection approach for STAP in IEEE Transactions on Information Theory, vol. 57, no. 4, pp. 2400–
heterogeneous clutter,” IEEE Transactions on Aerospace and Electronic 2419, April 2011.
Systems, vol. 46, no. 3, pp. 1021–1033, July 2010. [26] H. Kobeissi, Y. Nasser, O. Bazzi, A. Nafkha, and Y. Louêt, “Elastic-
[3] A. L. Swindlehurst and P. Stoica, “Maximum likelihood methods in enabling massive-antenna for joint spectrum sensing and sharing: How
radar array signal processing,” Proceedings of the IEEE, vol. 86, no. 2, many antennas do we need?” IEEE Transactions on Cognitive Commu-
pp. 421–441, Feb 1998. nications and Networking, vol. 5, no. 2, pp. 267–280, June 2019.
[4] K. J. Sohn, H. Li, and B. Himed, “Parametric GLRT for multichannel [27] L. Xu, J. Li, and P. Stoica, “Target detection and parameter estimation for
adaptive signal detection,” IEEE Transactions on Signal Processing, MIMO radar systems,” IEEE Transactions on Aerospace and Electronic
vol. 55, no. 11, pp. 5351–5360, Nov 2007. Systems, vol. 44, no. 3, pp. 927–939, July 2008.
[5] J. Ward, “Space-time adaptive processing for airborne radar,” Mas- [28] B. Friedlander, “On transmit beamforming for MIMO radar,” IEEE
sachusetts Institute of Technology, Lincoln Labs, Cambridge, MA, Tech. Transactions on Aerospace and Electronic Systems, vol. 48, no. 4, pp.
Rep. TR-1015, Dec. 1994. 3376–3388, October 2012.
[6] S. M. Kay, Fundamentals of statistical signal processing, volume II: [29] F. Gini, A. De Maio, and L. Patton, Eds., Waveform Design and
detection theory. Prentice Hall, 1993. Diversity for Advanced Radar Systems, ser. Radar, Sonar & Navigation.
[7] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- Institution of Engineering and Technology, 2012.
bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, [30] I. Bekkerman and J. Tabrikian, “Target detection and localization using
no. 11, pp. 3590–3600, Nov. 2010. MIMO radars and sonars,” IEEE Transactions on Signal Processing,
[8] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO has vol. 54, no. 10, pp. 3873–3883, Oct 2006.
unlimited capacity,” IEEE Trans. Wireless Commun., vol. 17, no. 1, pp. [31] J. Li and P. Stoica, MIMO Radar Signal Processing. Hoboken, NJ:
574–590, Jan. 2018. Wiley, 2009.
[9] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO networks: [32] J. Li, L. Xu, P. Stoica, K. W. Forsythe, and D. W. Bliss, “Range
Spectral, energy, and hardware efficiency,” Foundations and Trends® compression and waveform optimization for MIMO radar: A cramé-rao
in Signal Processing, vol. 11, no. 3-4, pp. 154–655, 2017. [Online]. bound based study,” IEEE Transactions on Signal Processing, vol. 56,
Available: http://dx.doi.org/10.1561/2000000093 no. 1, pp. 218–232, Jan 2008.
[10] E. Björnson, L. Sanguinetti, H. Wymeersch, J. Hoydis, and T. L. [33] B. Friedlander, “Effects of model mismatch in MIMO radar,” IEEE
Marzetta, “Massive MIMO is a reality—What is next? Five promising Transactions on Signal Processing, vol. 60, no. 4, pp. 2071–2076, April
research directions for antenna arrays,” CoRR, vol. abs/1902.07678, 2012.
2019. [Online]. Available: http://arxiv.org/abs/1902.07678
[34] R. C. Bradley, “Basic properties of strong mixing conditions. A survey
[11] P. J. Huber, “The behavior of maximum likelihood estimates under non-
and some open questions,” Probab. Surveys, vol. 2, pp. 107–144, 2005.
standard conditions,” in Proceedings of the Fifth Berkeley Symposium on
Mathematical Statistics and Probability, Volume 1: Statistics. Berkeley, [35] T. W. Anderson, An Introduction to Multivariate Statistical Analysis.
Calif.: University of California Press, 1967, pp. 221–233. New York: John Wiley and Sons, Inc.,, 1958.
[12] H. White, “Maximum likelihood estimation of misspecified models,” [36] I. A. Ibragimov and Y. V. Linnik, Independent and Stationary Sequences
Econometrica: Journal of the Econometric Society, pp. 1 – 25, 1982. of Random Variables. The Netherlands: Wolters-Noordhoff, 1971.
[13] S. Fortunati, F. Gini, M. S. Greco, and C. D. Richmond, “Performance [37] M. Rosenblatt, “A central limit theorem and a strong mixing condition,”
bounds for parameter estimation under misspecified models: Funda- Proceedings of the National Academy of Sciences of the United States
mental findings and applications,” IEEE Signal Processing Magazine, of America, vol. 42, no. 1, pp. 43–47, 1956.
vol. 34, no. 6, pp. 142–157, Nov 2017. [38] B. Picinbono, “On circularity,” IEEE Transactions on Signal Processing,
[14] S. Fortunati, F. Gini, and M. Greco, “The Misspecified Cramér-Rao vol. 42, no. 12, pp. 3473–3482, 1994.
bound and its application to scatter matrix estimation in complex ellipti- [39] B. Picinbono and P. Bondon, “Second-order statistics of complex
cally symmetric distributions,” IEEE Transactions on Signal Processing, signals,” IEEE Transactions on Signal Processing, vol. 45, no. 2, pp.
vol. 64, no. 9, pp. 2387 – 2399, 2016. 411–420, Feb 1997.
[15] S. Fortunati, F. Gini, and M. S. Greco, “The constrained Misspecified [40] P. Stoica and R. L. Moses, Spectral Analysis of Signals, 2nd ed., 2005.
Cramér-Rao bound,” IEEE Signal Processing Letters, vol. 23, no. 5, pp. [41] M. Rangaswamy, D. D. Weiner, and A. Ozturk, “Non-gaussian random
718–721, 2016. vector identification using spherically invariant random processes,” IEEE
[16] A. Mennad, S. Fortunati, M. N. E. Korso, A. Younsi, A. M. Zoubir, and Transactions on Aerospace and Electronic Systems, vol. 29, no. 1, pp.
A. Renaux, “Slepian-bangs-type formulas and the related misspecified 111–124, Jan 1993.
Cramér-Rao bounds for complex elliptically symmetric distributions,” [42] K. J. Sangston, F. Gini, and M. S. Greco, “Coherent radar target
Signal Processing, vol. 142, pp. 320 – 329, 2018. detection in heavy-tailed compound-gaussian clutter,” IEEE Transactions
[17] H. White and I. Domowitz, “Nonlinear regression with dependent on Aerospace and Electronic Systems, vol. 48, no. 1, pp. 64–77, Jan
observations,” Econometrica, vol. 52, no. 1, pp. 143–161, 1984. 2012.
[18] I. Domowitz and H. White, “Misspecified models with dependent [43] T. Bollerslev and J. M. Wooldridge, “Quasi-maximum likelihood estima-
observations,” Journal of Econometrics, vol. 20, no. 1, pp. 35 – 58, tion and inference in dynamic models with time-varying covariances,”
1982. Econometric Reviews, vol. 11, no. 2, pp. 143–172, 1992.

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSP.2020.2967181, IEEE
Transactions on Signal Processing
12

[44] A. De Maio, S. M. Kay, and A. Farina, “On the invariance, coincidence,

and statistical equivalence of the GLRT, Rao test, and Wald test,” IEEE
Transactions on Signal Processing, vol. 58, no. 4, pp. 1967–1979, April
2010.
[45] A. Bulinski and N. Kryzhanovskaya, “Convergence rate in CLT for
vector-valued random fields with self-normalization,” Probab. Math.
Statist., vol. 26, no. 2, pp. 261–281, 2006.
[46] H. White, “Chapter VI - estimating asymptotic covariance matrices,”
in Asymptotic Theory for Econometricians, ser. Economic Theory,
Econometrics, and Mathematical Economics, H. White, Ed. San Diego:
Academic Press, 1984, pp. 132 – 161.
[47] F. C. Robey, D. R. Fuhrmann, E. J. Kelly, and R. Nitzberg, “A cfar
adaptive matched filter detector,” IEEE Transactions on Aerospace and
Electronic Systems, vol. 28, no. 1, pp. 208–216, Jan 1992.
[48] N. R. Goodman, “Statistical analysis based on a certain multivariate
complex gaussian distribution (an introduction),” Ann. Math. Statist.,
vol. 34, no. 1, pp. 152–177, 03 1963.
[49] A. Nuttall, “Some integrals involving the Qm function (corresp.),” IEEE
Transactions on Information Theory, vol. 21, no. 1, pp. 95–96, January
1975.
[50] J. Li, P. Stoica, L. Xu, and W. Roberts, “On parameter identifiability
of mimo radar,” IEEE Signal Processing Letters, vol. 14, no. 12, pp.
968–971, Dec 2007.
[51] B. M. Pötscher and I. R. Prucha, “A uniform law of large numbers for
dependent and heterogeneous data processes,” Econometrica, vol. 57,
no. 3, pp. 675–683, 1989.
[52] H. White, “Nonlinear regression on cross-section data,” Econometrica,
vol. 48, no. 3, pp. 721–746, 1980.
[53] D. R. Cox and D. V. Hinkley, Theoretical Statistics. Chapman and
Hall/CRC, 1979.
[54] J. Eriksson, E. Ollila, and V. Koivunen, “Essential statistics and tools for
complex random variables,” IEEE Transactions on Signal Processing,
vol. 58, no. 10, pp. 5400–5408, Oct 2010.
[55] P. Billingsley, Convergence of Probability Measures, 2nd ed. John
Wiley & Sons, 1999.

1053-587X (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.