Introduction
Incorporating a large number of antennas in an array is required in beyond-5G devices owing to the demand for increased spectral efficiency. As a result, antennas are placed very closely in a compact space with element spacing smaller than half-wavelength, bringing about non-negligible mutual coupling among antennas [1 ] . This trend goes even further with continuous-like antennas [2 , 3 ] .
Mutual coupling (MC) has been considered a negative factor that worsens the performance of communications as well as radar sensing. Therefore, methods were designed to alleviate its impact, e.g. [4 , 5 ] . However, seminal studies including [6 , 7 ] demonstrated that MC can be beneficial. Such merits of MC have not been fully investigated yet. The MC of tightly-coupled antennas (TCAs) that are densely packed contributes to superdirectivity or wide operational bandwidth, but both cannot be achieved at the same time as there is a trade-off between directivity and operational bandwidth [8 , 9 ] .
Erfan Khordad, Sachitha C. Bandara and Rajitha Senanayake are with the Department of Electrical and Electronic Engineering, the University of Melbourne, Melbourne, VIC 3052, Australia (e-mail: khordade@yahoo.com; hbandara@student.unimelb.edu.au; rajitha.senanayake@unimelb.edu.au).
Peter J. Smith is with the School of Mathematics and Statistics, Victoria University of Wellington, Wellington 6140, New Zealand (e-mail: smithpe@ecs.vuw.ac.nz). (Corresponding author: Erfan Khordad )
We restrict our study to the impact of the wide operational bandwidth offered by TCAs on detection. The large operational bandwidth is shown in practical experiments [10 , 11 ] and by analysis using physically-consistent modeling that appropriately accounts for the antenna size and MC [1 , 12 , 13 , 14 ] . Physically-consistent models are produced from modeling the components of a communications system based on circuit theory, physical laws, electromagnetic theory and communication channel properties [15 , 16 ] .
Target detection can substantially benefit from the large operational bandwidth of TCAs, which we demonstrate in this paper. Another important topical application that can benefit from TCAs is integrated sensing and communications (ISAC). The wide bandwidth of TCAs is desirable for such paradigms as both communications and radar sensing gain from large bandwidths. For ISAC systems, target detection is of particular interest, which has been explored without considering physically-consistent models for the antenna elements. For instance, target detection was investigated in [17 , 18 ] for a multi-target scenario, in [19 ] for wideband signals and in [20 ] considering the clutter. In these works, various antenna topologies were investigated however without leveraging the merits of MC.
Outside the context of ISAC, target detection as a standalone problem is well known under a range of assumptions and system characteristics. Detection with unknown MC among the antenna elements was studied in [21 ] . However, MC was considered as a cause of performance degradation and a partially known signal model was used.
Energy detectors or multi-channel energy detectors with detectors in multiple sub-bands [22 ] can be used to detect unknown wideband signals. In energy detection, no assumptions are made about the dynamics of the signal. Also energy detection is vulnerable to uncertainty in the noise and is not optimal when the signal is correlated [23 ] . Wideband detection was studied in [24 ] assuming an extended target constituting several reflecting points. In [23 , 25 , 26 ] , detection was studied by factoring in the dynamics of the signal as a correlation model. In [23 ] for a multiple-antenna receiver, the noise was assumed to be independent and identically distributed (i.i.d.). Also in [25 ] , a single antenna receiver was considered and asymptotic results were provided using the central limit theorem. Moreover, approaches in [24 ] –[26 ] considered only white noise. All the above approaches considered detection in either frequency or time domains, making them attractive only for stationary signals, as for non-stationary signals joint frequency and time approaches are required. For non-stationary signals, joint frequency and time target detection was studied in [27 ] using an improved version of the Wigner-Ville transform. In [28 ] , energy detection was considered using the spectrogram and Wigner-Ville transform. In [29 ] , a detection method using a filter bank was developed considering colored noise, and the distribution of the detector was approximated using simulated data for a scalar signal. Yet [27 ] –[29 ] lack analysis on the distribution of their detectors.
Nevertheless, the benefits of TCAs for target detection, considering all possibilities about temporal and spectral dynamics of the signal, are still unexplored. As such, in [30 ] we studied the target detection problem using TCAs by categorizing the spectral dynamics of the signal into three regimes. We extend the analysis in [30 ] to a more complete signal model by factoring in both spectral and temporal variations of the signal, which results in nine cases as explained below.
In this paper, we propose a novel target detection approach. We also show the opportunity that TCAs offer in target detection due to the resultant wide operational bandwidth. Echo signals reflected by the target are assumed to be unknown wideband signals. Thus, any wideband waveform can be used for detection. Since our approach can be used for sensing the target with any wideband waveform, it could be beneficial for both communications-centric and radar-centric ISAC waveform designs. Also, MC introduces frequency-dependent noise correlation into the detection problem, which we take into account. We consider the temporal and spectral dynamics of the signal, classifying the dynamics as either constant, slowly or rapidly varying in frequency and time. This gives nine possible cases in total, given in Table I . Specifically, our novel contributions are summarized as follows:
Table I. Nine Cases for Target Detection
We consider an unknown wideband signal model in the presence of frequency-dependent correlated noise and take into the account temporal and spectral dynamics of the signal. This results in the nine possible cases discussed above. For each of the cases, we derive a detector and provide a thorough analysis of the distribution of the detector output under null and alternative hypotheses. We show the perfect agreement of our analytical results and simulations.
We propose a novel detection approach for the cases with a signal changing slowly in at least one domain. Our approach is based on estimation of the signal by averaging of the samples using a one-dimensional window either in frequency or time, as well as a two-dimensional window in both domains.
We derive a detector that gives an upper bound on the probability of detection, for comparison. In addition, we examine the performance of the nine detectors under various time/frequency variation rates. We derive an extended energy detector considering correlated noise as well as both frequency and time samples. We show that our proposed detection approach outperforms other detectors including the extended energy detector. We also highlight the benefit of TCAs over conventional weakly-coupled antennas.
Notation: χ A 2 ( B ) \chi_{A}^{2}(B) χ A 2 ( B ) denotes a non-central chi-square variable with A A A degrees of freedom (DoFs) and non-centrality parameter, B B B . χ A 2 \chi_{A}^{2} χ A 2 denotes a chi-square variable with A A A DoFs. Non-central chi-square and chi-square variables in the Theorems are assumed to be independent and denoted respectively by χ A ( z ) , z 2 ( B ( z ) ) \chi_{A(z),z}^{2}(B(z)) χ A ( z ) , z 2 ( B ( z )) and χ A ( z ) , z 2 \chi_{A(z),z}^{2} χ A ( z ) , z 2 for a general index z = 1 , … , Z z=1,\ldots,Z z = 1 , … , Z , with A ( z ) A(z) A ( z ) and B ( z ) B(z) B ( z ) being the DoFs and non-centrality parameter. Bold uppercase A A A and bold lowercase a a a denote a matrix and a column vector respectively. Non-bold typeface A A A denotes a scalar. The conjugate transpose and transpose are denoted by A H A^{H} A H and A T A^{T} A T . The operator blockdiag { D 1 , D 2 , … , D N } \operatorname{blockdiag}\{D_{1},D_{2},\ldots,D_{N}\} blockdiag { D 1 , D 2 , … , D N } indicates a block diagonal matrix with sub-matrices D 1 D_{1} D 1 to D N D_{N} D N . Summation ∑ c = c 1 c N \sum_{c=c_{1}}^{c_{N}} ∑ c = c 1 c N indicates ∑ c = c 1 c N \sum_{c=c_{1}}^{c_{N}} ∑ c = c 1 c N , for a general index c = c 1 , c 1 + 1 , … , c N c=c_{1},c_{1}+1,\ldots,c_{N} c = c 1 , c 1 + 1 , … , c N . Operators Re \operatorname{Re} Re and Im \operatorname{Im} Im denote the real and imaginary parts of a complex number respectively. Also, 0 0 0 denotes an all-zero vector and I N \mathbf{I}_{N} I N denotes an N N N by N N N identity matrix.
Our proposed detection approach and analysis in this paper are general in the sense that they can apply to any similar joint time and frequency detection problem with frequency varying correlated noise. However, we aim to unveil the merits of tight coupling in target detection. Thus we present a system model using circuit-theoretic parameters and pose the target detection problem for TCAs.
We assume the use of canonical minimum scattering (CMS) antennas that provide a broad bandwidth as they have only the first radiation mode with a low Q-factor. As mentioned in § 1 , TCA arrays can also achieve superdirectivity, but at the cost of larger Q-factor and reduced bandwidth [8 ] , [9 ] .
Our model is based on Chu’s theory, and the antennas are referred to as Chu’s CMS antennas [1 ] , [31 ] . Based on Chu’s theory, an antenna that is restricted in a sphere with a specific radius is modeled with an equivalent resistor-inductor-capacitor (RLC) circuit. Impedance values of the components of the RLC circuit are dependent on the radius. Using Chu’s model allows us to account for the impact of MC on the detection performance.
The mutual impedance between two Chu’s CMS antennas is given in [1 ] , Eq. (5) which was derived by establishing an equivalence between Hertz dipoles and Chu’s CMS antennas based on their power radiation. The expression takes into account the antenna spacing, which we denote by δ \delta δ , and the orientation of the dipoles. We adopt a colinear configuration for the dipoles which is a commonly-used configuration where the dipoles are put along the axis connecting the dipoles.
We consider a transmitter sending wideband signals which are reflected from a target and the echoes arrive at a sensing receiver equipped with a uniform linear array of N R N_{R} N R Chu’s CMS antennas. We denote by a R a_{R} a R the radius of the sphere within which a receiver antenna is enclosed. The antenna spacing for the array at the sensing receiver is assumed to be small so that there is tight coupling among the antennas, which results in a TCA array.
Wideband signals are used to leverage the wideband TCA array for target detection. Echoes from the target are assumed to be unknown wideband signals. Thus, any wideband waveform can be used for detection and we do not impose any other assumptions on the waveform.
Considering the equivalent circuit for a Chu’s CMS antenna, the sensing receiver’s antenna array is modeled as a network of RLC circuits, resulting in a circuit-theoretic model [1 ] . As such, the received signal is a vector of voltages v k , t v_{k,t} v k , t where k = 1 , … , K k = 1,\ldots,K k = 1 , … , K and t = 1 , … , T t = 1,\ldots,T t = 1 , … , T respectively denote k k k th frequency sample and t t t th time sample, with K K K and T T T denoting the total number of frequency and time samples, respectively, that the receiver collects.
Our objective is to detect the target using joint time and frequency samples at the sensing receiver. Hence, the detection problem is posed as
H 0 : v k , t = n k , t H 1 : v k , t = α k , t + n k , t , (1) \begin{aligned}
\mathcal{H}_0 &: \boldsymbol{v}_{k,t} = \boldsymbol{n}_{k,t}\\
\mathcal{H}_1 &: \boldsymbol{v}_{k,t} = \boldsymbol{\alpha}_{k,t} + \boldsymbol{n}_{k,t}, \tag*{(1)}
\end{aligned} H 0 H 1 : v k , t = n k , t : v k , t = α k , t + n k , t , (1)
with H 0 \mathcal{H}_0 H 0 and H 1 \mathcal{H}_1 H 1 being the null and alternative hypotheses respectively, where vectors α k , t \boldsymbol{\alpha}_{k,t} α k , t and n k , t \boldsymbol{n}_{k,t} n k , t denote the signal and noise components. The signal component, that is assumed to be unknown, models the echo signal from the target that has been impacted by the channel between the target and the sensing receiver. We employ a physically-consistent model for the channel in Section VI where we present numerical results. The noise component has a complex correlated Gaussian distribution that models the background noise as well as noise due to low-noise amplifiers and the receive circuit, with frequency-dependent complex covariance matrix R k \mathbf{R}_k R k given in [1 Eq. (16b)] . We assume noise is uncorrelated in time. The hypothesis testing problem in (1) is a physically-consistent target detection problem because the channel model and matrix R k \mathbf{R}_k R k both capture the impact of δ \delta δ and a R a_R a R as well as MC.
The signal α k , t \boldsymbol{\alpha}_{k,t} α k , t in frequency and time domains can be either constant, slowly or rapidly varying, which we denote by numeric representations 1, 2 and 3 respectively. This gives 9 possible cases in total. We denote such cases by a numeric pair ( a , b ) (a,b) ( a , b ) where a a a and b b b denote the numeric representations in frequency and time respectively. Table I summarizes the cases with their corresponding sections in this paper.
In the following, we first review how probabilities of false alarm and detection are calculated for a general detector. We then provide a detection framework and explain how to estimate the signal. In the sections that follow, we derive the 9 particular detectors and provide analysis.
Preliminaries: Probabilities of False Alarm and Detection
Calculating probabilities of false alarm and detection is reviewed here for a general detector, given that the distributions of the detector test statistic are known under the null and alternative hypotheses. In Section VI for the numerical results, the method explained here is used to compute probabilities of false alarm and detection for all the detectors.
Denote by Λ \Lambda Λ a general detector and denote by γ Λ \gamma_{\Lambda} γ Λ a general test threshold for Λ \Lambda Λ . Also, denote the probability density function (PDF) and cumulative distribution function (CDF) of the detector under the null hypothesis respectively by f Λ ∣ H 0 ( x ) f_{\Lambda|\mathcal{H}_0}(x) f Λ∣ H 0 ( x ) and F Λ ∣ H 0 ( x ) F_{\Lambda|\mathcal{H}_0}(x) F Λ∣ H 0 ( x ) . Similarly, denote the PDF and CDF of the detector under the alternative hypothesis respectively by f Λ ∣ H 1 ( x ) f_{\Lambda|\mathcal{H}_1}(x) f Λ∣ H 1 ( x ) and F Λ ∣ H 1 ( x ) F_{\Lambda|\mathcal{H}_1}(x) F Λ∣ H 1 ( x ) . The probability of false alarm is calculated as P F A = ∫ γ Λ ∞ f Λ ∣ H 0 ( x ) d x P_{\mathrm{FA}} = \int_{\gamma_{\Lambda}}^{\infty} f_{\Lambda|\mathcal{H}_0}(x)\mathrm{d}x P FA = ∫ γ Λ ∞ f Λ∣ H 0 ( x ) d x . Thus, for a fixed P F A P_{\mathrm{FA}} P FA , the test threshold γ Λ \gamma_{\Lambda} γ Λ is calculated as γ Λ = F Λ ∣ H 0 − 1 ( 1 − P F A ) \gamma_{\Lambda} = F_{\Lambda|\mathcal{H}_0}^{-1}(1-P_{\mathrm{FA}}) γ Λ = F Λ∣ H 0 − 1 ( 1 − P FA ) . Also, the probability of detection is computed by P D = ∫ γ Λ ∞ f Λ ∣ H 1 ( x ) d x P_{\mathrm{D}} = \int_{\gamma_{\Lambda}}^{\infty} f_{\Lambda|\mathcal{H}_1}(x)\mathrm{d}x P D = ∫ γ Λ ∞ f Λ∣ H 1 ( x ) d x and hence for a fixed P F A P_{\mathrm{FA}} P FA , it is given as P D = 1 − F Λ ∣ H 1 ( γ Λ ) P_{\mathrm{D}} = 1-F_{\Lambda|\mathcal{H}_1}(\gamma_{\Lambda}) P D = 1 − F Λ∣ H 1 ( γ Λ ) .
LLR for Joint Frequency and Time Samples
We derive a general detection framework in this subsection with no prior assumption about the variation of the signal in frequency or time, based on the log-likelihood ratio (LLR). Denote by V \mathbf{V} V the array of all received samples v k , t \boldsymbol{v}_{k,t} v k , t in time and frequency. We write the likelihood ratio test (LRT) with threshold γ \gamma γ for the detection problem in (1) next. Using the PDF of a correlated Gaussian distribution, the LRT is written as
Λ L R T ( V ) = exp ( ∑ k ∑ t ( v k , t − α k , t ) H R k − 1 ( v k , t − α k , t ) ) exp ( ∑ k ∑ t v k , t H R k − 1 v k , t ) ≷ H 0 H 1 γ . (2) \Lambda_{\mathrm{LRT}}(\mathbf{V}) =
\frac{\exp\left(\sum_k\sum_t(\boldsymbol{v}_{k,t}-\boldsymbol{\alpha}_{k,t})^{H}\mathbf{R}_k^{-1}(\boldsymbol{v}_{k,t}-\boldsymbol{\alpha}_{k,t})\right)}
{\exp\left(\sum_k\sum_t\boldsymbol{v}_{k,t}^{H}\mathbf{R}_k^{-1}\boldsymbol{v}_{k,t}\right)}
\mathop{\gtrless}_{\mathcal{H}_0}^{\mathcal{H}_1}\gamma. \tag*{(2)} Λ LRT ( V ) = exp ( ∑ k ∑ t v k , t H R k − 1 v k , t ) exp ( ∑ k ∑ t ( v k , t − α k , t ) H R k − 1 ( v k , t − α k , t ) ) ≷ H 0 H 1 γ . (2)
Since the signal is unknown, (2) cannot be used directly. Hence, we use the generalized likelihood ratio test (GLRT) method, where the unknown signal is replaced by its maximum likelihood estimate (MLE). The GLRT method is general and we will use it in Section IV for signals without slow variations in either of the domains, i.e. Cases (1,1), (1,3), (3,1) and (3,3). Denote the MLE estimate of α k , t \boldsymbol{\alpha}_{k,t} α k , t by α k , t M L E \boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}} α k , t MLE . By taking the log of (2) and replacing α k , t \boldsymbol{\alpha}_{k,t} α k , t with α k , t M L E \boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}} α k , t MLE , the test statistic for these cases is given as an LLR expression, with threshold γ 1 \gamma_1 γ 1 , as
Λ G L R T ( V ) = ∑ k ∑ t ( v k , t H R k − 1 α k , t M L E + ( α k , t M L E ) H R k − 1 v k , t − ( α k , t M L E ) H R k − 1 α k , t M L E ) ≷ H 0 H 1 γ 1 . (3) \Lambda_{\mathrm{GLRT}}(\mathbf{V}) =
\sum_k\sum_t\left(\boldsymbol{v}_{k,t}^{H}\mathbf{R}_k^{-1}\boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}}+
(\boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}})^{H}\mathbf{R}_k^{-1}\boldsymbol{v}_{k,t}-
(\boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}})^{H}\mathbf{R}_k^{-1}\boldsymbol{\alpha}_{k,t}^{\mathrm{MLE}}\right)
\mathop{\gtrless}_{\mathcal{H}_0}^{\mathcal{H}_1}\gamma_1. \tag*{(3)} Λ GLRT ( V ) = k ∑ t ∑ ( v k , t H R k − 1 α k , t MLE + ( α k , t MLE ) H R k − 1 v k , t − ( α k , t MLE ) H R k − 1 α k , t MLE ) ≷ H 0 H 1 γ 1 . (3)
For the other cases, where the signal is slowly changing in at least one domain, finding α k , t \boldsymbol{\alpha}_{k,t} α k , t depends on the precise model for the way in which α k , t \boldsymbol{\alpha}_{k,t} α k , t changes in time or frequency. Such precise knowledge is very hard to obtain and restricts the detection to certain scenarios. Ideally, we would have an estimate of α k , t \boldsymbol{\alpha}_{k,t} α k , t which is useful over a wide range of unknown signal types. A non-parametric estimator is the obvious choice and in particular smoothing estimates are simple, general-purpose and effective. Define α ^ k , t \widehat{\boldsymbol{\alpha}}_{k,t} α k , t as an estimate of α k , t \boldsymbol{\alpha}_{k,t} α k , t based on local averaging (LA) for which we give expressions in § 3 . Hence, using the LLR of (2), we create a new test statistic as follows
Λ L A ( V ) = ∑ k ∑ t ( v k , t H R k − 1 α ^ k , t + α ^ k , t H R k − 1 v k , t − α ^ k , t H R k − 1 α ^ k , t ) ≷ H 0 H 1 γ 1 . (4) \begin{aligned}
\Lambda_{\mathrm{LA}}(\mathbf{V})
&= \sum_k\sum_t\left(\boldsymbol{v}_{k,t}^{H}\mathbf{R}_k^{-1}\widehat{\boldsymbol{\alpha}}_{k,t}
+\widehat{\boldsymbol{\alpha}}_{k,t}^{H}\mathbf{R}_k^{-1}\boldsymbol{v}_{k,t}
-\widehat{\boldsymbol{\alpha}}_{k,t}^{H}\mathbf{R}_k^{-1}\widehat{\boldsymbol{\alpha}}_{k,t}\right)
\mathop{\gtrless}_{\mathcal{H}_0}^{\mathcal{H}_1}\gamma_1. \tag*{(4)}
\end{aligned} Λ LA ( V ) = k ∑ t ∑ ( v k , t H R k − 1 α k , t + α k , t H R k − 1 v k , t − α k , t H R k − 1 α k , t ) ≷ H 0 H 1 γ 1 . (4)
In the following section, we present our proposed detection approach using (4) .
Proposed Approach for Cases with a Slowly-Varying Signal in at Least One Domain
We propose a novel detection approach based on the test statistic in (4) , for all the cases in which the signal is varying slowly in at least one of the time or frequency domains and present analysis for them. These cases are Case (1,2), Case (2,1), Case (2,2), Case (2,3) and Case (3,2).
We take into consideration the limited knowledge about the slow variation property of the signal in the development of an enhanced detection approach. Our proposed approach is based on LA of the received voltage vectors in the domain where the signal is slowly changing. We adopt a window that slides through time and/or frequency and uniformly averages the samples within it, to estimate the signal using surrounding time and/or frequency samples. We use uniform averaging but the averaging could be non-uniform in general. We show by numerical results in Section VI that our proposed detectors that factor in signal variation outperform the MLE-based detectors, derived in Section IV, that lack the ability to use the information about signal variation.
We show that our proposed detectors under the null and alternative hypotheses are linear combinations of non-central chi-square and chi-square variables that are special cases of a generalized chi-square. The cumulative distribution function (CDF) of the generalized chi-square is efficiently generated with software using the Imhof method [32 , 33 ] .
Note that [25 ] and [34 ] also used a window sliding through samples in the frequency and time domains respectively. However, the detection in [34 ] was for a known signal with uncorrelated noise and [25 ] studied a special case as mentioned in § 1 .
Proposed Detector for Case (1,2): Constant in Frequency and Slowly-varying in Time
We make use of the slow variation in time and use a window in the time domain. As the signal is constant in frequency, for the ease of design and analysis we drop the index k k k from α k , t \alpha_{k,t} α k , t . The window slides only in the time domain and all the frequency bins are averaged uniformly. We estimate α t \alpha_t α t with α t ( 1 , 2 ) \alpha_t^{(1,2)} α t ( 1 , 2 ) , which is a weighted average of received voltage vectors in the time domain as
α t ( 1 , 2 ) = ∑ k ∑ ℓ 1 = − L 1 L 1 v k , t − ℓ 1 w k , t , ℓ 1 , (5) \alpha_t^{(1,2)}=\sum_k\sum_{\ell_1=-L_1}^{L_1}\mathbf{v}_{k,t-\ell_1}w_{k,t,\ell_1}, \tag*{(5)} α t ( 1 , 2 ) = k ∑ ℓ 1 = − L 1 ∑ L 1 v k , t − ℓ 1 w k , t , ℓ 1 , (5)
where w k , t , ℓ 1 w_{k,t,\ell_1} w k , t , ℓ 1 are the weights of the window and L 1 L_1 L 1 determines the length of the window, i.e., 2 L 1 + 1 2L_1+1 2 L 1 + 1 . For the averaging, the center of the window with ℓ 1 = 0 \ell_1=0 ℓ 1 = 0 lies on the t t t -th time sample. We calculate the weights based on uniform averaging of the voltage vectors, but (5) is general and any weighting approach can be utilized. To avoid edge effects, w k , t , ℓ 1 w_{k,t,\ell_1} w k , t , ℓ 1 is set to zero when t − ℓ 1 < 1 t-\ell_1<1 t − ℓ 1 < 1 or t − ℓ 1 > T t-\ell_1>T t − ℓ 1 > T . In these edge cases, the averaging is restricted to time samples within the range 1 ≤ t ≤ T 1\le t\le T 1 ≤ t ≤ T . Also for ℓ 1 < − L 1 \ell_1<-L_1 ℓ 1 < − L 1 or ℓ 1 > L 1 \ell_1>L_1 ℓ 1 > L 1 , w k , t , ℓ 1 w_{k,t,\ell_1} w k , t , ℓ 1 is set to zero. Hence, we obtain the weights w k , t , ℓ 1 w_{k,t,\ell_1} w k , t , ℓ 1 as
w k , t , ℓ 1 = { 1 / ( K ( 2 L 1 + 1 ) ) t ≥ L 1 + 1 , t ≤ T − L 1 1 / ( K ( t + L 1 ) ) t ≤ L 1 1 / ( K ( T − t + 1 + L 1 ) ) t ≥ T − L 1 + 1 , (6) w_{k,t,\ell_1}=
\begin{cases}
1/(K(2L_1+1)) & t\ge L_1+1,t\le T-L_1\\
1/(K(t+L_1)) & t\le L_1\\
1/(K(T-t+1+L_1)) & t\ge T-L_1+1,
\end{cases} \tag*{(6)} w k , t , ℓ 1 = ⎩ ⎨ ⎧ 1/ ( K ( 2 L 1 + 1 )) 1/ ( K ( t + L 1 )) 1/ ( K ( T − t + 1 + L 1 )) t ≥ L 1 + 1 , t ≤ T − L 1 t ≤ L 1 t ≥ T − L 1 + 1 , (6)
where we calculate the weights based on uniform averaging across both domains, with the constraint ∑ k ∑ ℓ 1 = − L 1 L 1 w k , t , ℓ 1 = 1 \sum_k\sum_{\ell_1=-L_1}^{L_1}w_{k,t,\ell_1}=1 ∑ k ∑ ℓ 1 = − L 1 L 1 w k , t , ℓ 1 = 1 . For the frequency domain, we take the average of all the frequency bins uniformly. Hence, for any k k k we have ∑ ℓ 1 = − L 1 L 1 w k , t , ℓ 1 = 1 / K \sum_{\ell_1=-L_1}^{L_1}w_{k,t,\ell_1}=1/K ∑ ℓ 1 = − L 1 L 1 w k , t , ℓ 1 = 1/ K . Note that calculation of weights for other cases in this section, i.e. Case (2,1), Case (2,2), Case (2,3) and Case (3,2), is also performed piecewise, considering the edge effects.
Our proposed detector for Case (1,2) is written by replacing (5) in (4) as
Λ ( 1 , 2 ) , p ( V ) = ∑ t [ ( ∑ k v k , t H R k − 1 ) ( ∑ k ′ ∑ ℓ 1 = − L 1 L 1 v k ′ , t − ℓ 1 ′ w k ′ , t , ℓ 1 ) ] + ∑ t [ ( ∑ k ′ ∑ ℓ 1 = − L 1 L 1 v k ′ , t − ℓ 1 H w k ′ , t , ℓ 1 ) ( ∑ k R k − 1 v k , t ) ] − ∑ t [ ( ∑ k ′ ∑ ℓ 1 = − L 1 L 1 v k ′ , t − ℓ 1 H w k ′ , t , ℓ 1 ) ( ∑ k R k − 1 ) ( ∑ k ~ ∑ ℓ ~ 1 = − L 1 L 1 v k ~ , t − ℓ ~ 1 w k ~ , t , ℓ ~ 1 ) ] , (7) \begin{aligned}
\Lambda_{(1,2),\mathrm{p}}(\mathbf{V})={}&
\sum_t\left[\left(\sum_k\mathbf{v}_{k,t}^{H}\mathbf{R}_k^{-1}\right)
\left(\sum_{k'}\sum_{\ell_1=-L_1}^{L_1}\mathbf{v}_{k',t-\ell_1}'w_{k',t,\ell_1}\right)\right]\\
&+\sum_t\left[\left(\sum_{k'}\sum_{\ell_1=-L_1}^{L_1}\mathbf{v}_{k',t-\ell_1}^{H}w_{k',t,\ell_1}\right)
\left(\sum_k\mathbf{R}_k^{-1}\mathbf{v}_{k,t}\right)\right]\\
&-\sum_t\left[\left(\sum_{k'}\sum_{\ell_1=-L_1}^{L_1}\mathbf{v}_{k',t-\ell_1}^{H}w_{k',t,\ell_1}\right)
\left(\sum_k\mathbf{R}_k^{-1}\right)
\left(\sum_{\tilde{k}}\sum_{\tilde{\ell}_1=-L_1}^{L_1}\mathbf{v}_{\tilde{k},t-\tilde{\ell}_1}w_{\tilde{k},t,\tilde{\ell}_1}\right)\right], \tag*{(7)}
\end{aligned} Λ ( 1 , 2 ) , p ( V ) = t ∑ [ ( k ∑ v k , t H R k − 1 ) ( k ′ ∑ ℓ 1 = − L 1 ∑ L 1 v k ′ , t − ℓ 1 ′ w k ′ , t , ℓ 1 ) ] + t ∑ [ ( k ′ ∑ ℓ 1 = − L 1 ∑ L 1 v k ′ , t − ℓ 1 H w k ′ , t , ℓ 1 ) ( k ∑ R k − 1 v k , t ) ] − t ∑ ( k ′ ∑ ℓ 1 = − L 1 ∑ L 1 v k ′ , t − ℓ 1 H w k ′ , t , ℓ 1 ) ( k ∑ R k − 1 ) k ~ ∑ ℓ ~ 1 = − L 1 ∑ L 1 v k ~ , t − ℓ ~ 1 w k ~ , t , ℓ ~ 1 , (7)
where subscript p indicates our proposed detector. After some algebra, we write (7) in a quadratic form as
Λ ( 1 , 2 ) , p ( V ) = ∑ k 1 = 1 K ∑ t 1 = 1 T ∑ k 2 = 1 K ∑ t 2 = 1 T v k 1 , t 1 H A k 1 t 1 k 2 t 2 v k 2 , t 2 . (8) \Lambda_{(1,2),\mathrm{p}}(\mathbf{V})=\sum_{k_1=1}^{K}\sum_{t_1=1}^{T}\sum_{k_2=1}^{K}\sum_{t_2=1}^{T}\mathbf{v}_{k_1,t_1}^{H}\mathbf{A}_{k_1t_1k_2t_2}\mathbf{v}_{k_2,t_2}. \tag*{(8)} Λ ( 1 , 2 ) , p ( V ) = k 1 = 1 ∑ K t 1 = 1 ∑ T k 2 = 1 ∑ K t 2 = 1 ∑ T v k 1 , t 1 H A k 1 t 1 k 2 t 2 v k 2 , t 2 . (8)
where we calculate matrix A k 1 t 1 k 2 t 2 \mathbf{A}_{k_1t_1k_2t_2} A k 1 t 1 k 2 t 2 as
A k 1 t 1 k 2 t 2 = R k 1 − 1 w k 2 , t 1 , t 1 − t 2 + w k 1 , t 2 , t 2 − t 1 R k 2 − 1 − ∑ t w k 1 , t , t − t 1 ( ∑ k R k − 1 ) w k 2 , t , t − t 2 . (9) \mathbf{A}_{k_1t_1k_2t_2}=\mathbf{R}_{k_1}^{-1}w_{k_2,t_1,t_1-t_2}+w_{k_1,t_2,t_2-t_1}\mathbf{R}_{k_2}^{-1}
-\sum_t w_{k_1,t,t-t_1}\left(\sum_k\mathbf{R}_k^{-1}\right)w_{k_2,t,t-t_2}. \tag*{(9)} A k 1 t 1 k 2 t 2 = R k 1 − 1 w k 2 , t 1 , t 1 − t 2 + w k 1 , t 2 , t 2 − t 1 R k 2 − 1 − t ∑ w k 1 , t , t − t 1 ( k ∑ R k − 1 ) w k 2 , t , t − t 2 . (9)
We further simplify our proposed detector in (8) and present it in a matrix form. Define a K T N R KTN_R K T N R by K T N R KTN_R K T N R block matrix A ‾ \overline{\mathbf{A}} A . Block ( i , j ) (i,j) ( i , j ) in A ‾ \overline{\mathbf{A}} A is A k 1 t 1 k 2 t 2 \mathbf{A}_{k_1t_1k_2t_2} A k 1 t 1 k 2 t 2 where the indices are
k 1 = ⌊ ( i − 1 ) / T ⌋ + 1 , t 1 = i − ( k 1 − 1 ) T , k_1=\lfloor(i-1)/T\rfloor+1,\quad t_1=i-(k_1-1)T, k 1 = ⌊( i − 1 ) / T ⌋ + 1 , t 1 = i − ( k 1 − 1 ) T ,
k 2 = ⌊ ( j − 1 ) / T ⌋ + 1 , t 2 = j − ( k 2 − 1 ) T , (10) k_2=\lfloor(j-1)/T\rfloor+1,\quad t_2=j-(k_2-1)T, \tag*{(10)} k 2 = ⌊( j − 1 ) / T ⌋ + 1 , t 2 = j − ( k 2 − 1 ) T , (10)
with i = 1 , 2 , … , K T i=1,2,\ldots,KT i = 1 , 2 , … , K T and j = 1 , 2 , … , K T j=1,2,\ldots,KT j = 1 , 2 , … , K T . Define a K T N R KTN_R K T N R by 1 vector v ‾ \overline{\mathbf{v}} v as
v ‾ = [ v 1 , 1 T , … , v 1 , T T , v 2 , 1 T , … , v 2 , T T , … , v K , 1 T , … , v K , T T ] T . (11) \overline{\mathbf{v}}=\left[\mathbf{v}_{1,1}^{T},\ldots,\mathbf{v}_{1,T}^{T},\mathbf{v}_{2,1}^{T},\ldots,\mathbf{v}_{2,T}^{T},\ldots,\mathbf{v}_{K,1}^{T},\ldots,\mathbf{v}_{K,T}^{T}\right]^{T}. \tag*{(11)} v = [ v 1 , 1 T , … , v 1 , T T , v 2 , 1 T , … , v 2 , T T , … , v K , 1 T , … , v K , T T ] T . (11)
Therefore, our proposed detector in (8) is written as
Λ ( 1 , 2 ) , p ( V ) = v ‾ H A ‾ v ‾ . (12) \Lambda_{(1,2),\mathrm{p}}(\mathbf{V})=\overline{\mathbf{v}}^{H}\overline{\mathbf{A}}\overline{\mathbf{v}}. \tag*{(12)} Λ ( 1 , 2 ) , p ( V ) = v H A v . (12)
Now we provide an analysis of the proposed detector for Case (1,2) in the following theorem.
Theorem 1. The detector Λ ( 1 , 2 ) , p ( V ) \Lambda_{(1,2),p}(\mathbf{V}) Λ ( 1 , 2 ) , p ( V ) in (12) under H 1 \mathcal{H}_1 H 1 and H 0 \mathcal{H}_0 H 0 is respectively a linear combination of non-central chi-squares and a linear combination of chi-squares, as
Λ ( 1 , 2 ) , p ∣ H 1 = ∑ j ′ = 1 K T N R λ ‾ j ′ χ 2 , j ′ 2 ( ∣ μ ‾ j ′ ∣ 2 + Im { μ ‾ j ′ } 2 ) , (13) \left.\Lambda_{(1,2),p}\right|\mathcal{H}_1=\sum_{j'=1}^{KTN_R}\overline{\lambda}_{j'}\chi^2_{2,j'}\left(\left\lvert\overline{\mu}_{j'}\right\rvert^2+\operatorname{Im}\{\overline{\mu}_{j'}\}^2\right), \tag*{(13)} Λ ( 1 , 2 ) , p H 1 = j ′ = 1 ∑ K T N R λ j ′ χ 2 , j ′ 2 ( μ j ′ 2 + Im { μ j ′ } 2 ) , (13) Λ ( 1 , 2 ) , p ∣ H 0 = ∑ j ′ = 1 K T N R λ ‾ j ′ χ 2 , j ′ 2 , (14) \left.\Lambda_{(1,2),p}\right|\mathcal{H}_0=\sum_{j'=1}^{KTN_R}\overline{\lambda}_{j'}\chi^2_{2,j'}, \tag*{(14)} Λ ( 1 , 2 ) , p H 0 = j ′ = 1 ∑ K T N R λ j ′ χ 2 , j ′ 2 , (14) where λ ‾ j ′ \overline{\lambda}_{j'} λ j ′ and μ ‾ j ′ \overline{\mu}_{j'} μ j ′ are defined above (18) and below (19) respectively.
Proof. We designed the detector for Case (1,2) when the signal is constant over frequency, but in practice detectors are applied where the assumptions are reasonable but not exact. Hence, we consider a general scenario where the signal is changing in frequency and time. Consequently, to keep our analysis general for Case (1,2) we include the frequency index k k k and we use α k , t \alpha_{k,t} α k , t in our analysis. Similarly for other detectors we conduct analysis for a general scenario with a signal changing in both domains.
Under H 1 \mathcal{H}_1 H 1 , v k , t \mathbf{v}_{k,t} v k , t in (1) has a Gaussian distribution with mean α k , t \boldsymbol{\alpha}_{k,t} α k , t and covariance matrix R k \mathbf{R}_k R k . Hence, we can write v k , t = α k , t + R k 1 2 v ~ k , t \mathbf{v}_{k,t}=\boldsymbol{\alpha}_{k,t}+\mathbf{R}_k^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} v k , t = α k , t + R k 2 1 v k , t where v ~ k , t ∼ C N ( 0 , I ) \widetilde{\mathbf{v}}_{k,t}\sim\mathcal{CN}(\mathbf{0},\mathbf{I}) v k , t ∼ C N ( 0 , I ) . Hence, for the alternative hypothesis we substitute α k , t + R k 1 2 v ~ k , t \boldsymbol{\alpha}_{k,t}+\mathbf{R}_k^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v k , t for v k , t \mathbf{v}_{k,t} v k , t in (12) to calculate Λ ( 1 , 2 ) , p ∣ H 1 \left.\Lambda_{(1,2),p}\right|\mathcal{H}_1 Λ ( 1 , 2 ) , p H 1 . Define
b ‾ = [ b ~ 1 , 1 T , … , b ~ 1 , T T , b ~ 2 , 1 T , … , b ~ 2 , T T , … , b ~ K , 1 T , … , b ~ K , T T ] T , (15) \overline{\mathbf{b}}=\left[\widetilde{\mathbf{b}}_{1,1}^{T},\ldots,\widetilde{\mathbf{b}}_{1,T}^{T},\widetilde{\mathbf{b}}_{2,1}^{T},\ldots,\widetilde{\mathbf{b}}_{2,T}^{T},\ldots,\widetilde{\mathbf{b}}_{K,1}^{T},\ldots,\widetilde{\mathbf{b}}_{K,T}^{T}\right]^{T}, \tag*{(15)} b = [ b 1 , 1 T , … , b 1 , T T , b 2 , 1 T , … , b 2 , T T , … , b K , 1 T , … , b K , T T ] T , (15)
with b ~ k , t = R k − 1 2 α k , t + v ~ k , t \widetilde{\mathbf{b}}_{k,t}=\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,t}+\widetilde{\mathbf{v}}_{k,t} b k , t = R k − 2 1 α k , t + v k , t . Also define R ‾ \overline{\mathbf{R}} R as a K T N R KTN_R K T N R by K T N R KTN_R K T N R block diagonal matrix, with blocks R k 1 2 \mathbf{R}_k^{\frac{1}{2}} R k 2 1 for k = 1 , … , K k=1,\ldots,K k = 1 , … , K and each block repeats T T T times on the diagonal, i.e.,
R ‾ = blockdiag { R 1 1 2 , … , R 1 1 2 , R 2 1 2 , … , R 2 1 2 , … , R K 1 2 , … , R K 1 2 } . (16) \overline{\mathbf{R}}=\operatorname{blockdiag}\{\mathbf{R}_1^{\frac{1}{2}},\ldots,\mathbf{R}_1^{\frac{1}{2}},\mathbf{R}_2^{\frac{1}{2}},\ldots,\mathbf{R}_2^{\frac{1}{2}},\ldots,\mathbf{R}_K^{\frac{1}{2}},\ldots,\mathbf{R}_K^{\frac{1}{2}}\}. \tag*{(16)} R = blockdiag { R 1 2 1 , … , R 1 2 1 , R 2 2 1 , … , R 2 2 1 , … , R K 2 1 , … , R K 2 1 } . (16)
Therefore, we give (12) as
Λ ( 1 , 2 ) , p ( V ) = b ‾ H R ‾ A R ‾ b ‾ . (17) \Lambda_{(1,2),p}(\mathbf{V})=\overline{\mathbf{b}}^{H}\overline{\mathbf{R}}\mathbf{A}\overline{\mathbf{R}}\overline{\mathbf{b}}. \tag*{(17)} Λ ( 1 , 2 ) , p ( V ) = b H R A R b . (17)
We replace matrix R ‾ A R ‾ \overline{\mathbf{R}}\mathbf{A}\overline{\mathbf{R}} R A R in (17) with its eigenvalue decomposition as Φ ‾ Δ ‾ Φ ‾ H \overline{\mathbf{\Phi}}\overline{\mathbf{\Delta}}\overline{\mathbf{\Phi}}^{H} Φ Δ Φ H with Φ ‾ \overline{\mathbf{\Phi}} Φ being a unitary matrix and Δ ‾ \overline{\mathbf{\Delta}} Δ being a diagonal matrix containing the eigenvalues, λ ‾ j ′ \overline{\lambda}_{j'} λ j ′ . Define
c ‾ = Φ ‾ H [ v ~ 1 , 1 T , … , v ~ 1 , T T , v ~ 2 , 1 T , … , v ~ 2 , T T , … , v ~ K , 1 T , … , v ~ K , T T ] T , (18) \overline{\mathbf{c}}=\overline{\mathbf{\Phi}}^{H}\left[\widetilde{\mathbf{v}}_{1,1}^{T},\ldots,\widetilde{\mathbf{v}}_{1,T}^{T},\widetilde{\mathbf{v}}_{2,1}^{T},\ldots,\widetilde{\mathbf{v}}_{2,T}^{T},\ldots,\widetilde{\mathbf{v}}_{K,1}^{T},\ldots,\widetilde{\mathbf{v}}_{K,T}^{T}\right]^{T}, \tag*{(18)} c = Φ H [ v 1 , 1 T , … , v 1 , T T , v 2 , 1 T , … , v 2 , T T , … , v K , 1 T , … , v K , T T ] T , (18)
and define vector μ ‾ \overline{\boldsymbol{\mu}} μ with K T N R KTN_R K T N R elements as
μ ‾ = Φ ‾ H [ p 1 , 1 T , … , p 1 , T T , p 2 , 1 T , … , p 2 , T T , … , p K , 1 T , … , p K , T T ] T , (19) \overline{\boldsymbol{\mu}}=\overline{\mathbf{\Phi}}^{H}\left[\mathbf{p}_{1,1}^{T},\ldots,\mathbf{p}_{1,T}^{T},\mathbf{p}_{2,1}^{T},\ldots,\mathbf{p}_{2,T}^{T},\ldots,\mathbf{p}_{K,1}^{T},\ldots,\mathbf{p}_{K,T}^{T}\right]^{T}, \tag*{(19)} μ = Φ H [ p 1 , 1 T , … , p 1 , T T , p 2 , 1 T , … , p 2 , T T , … , p K , 1 T , … , p K , T T ] T , (19)
with p k , t = R k − 1 2 α k , t \mathbf{p}_{k,t}=\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,t} p k , t = R k − 2 1 α k , t . In addition, denote the j ′ j' j ′ -th element of vectors μ ‾ \overline{\boldsymbol{\mu}} μ and c ‾ \overline{\mathbf{c}} c respectively by μ ‾ j ′ \overline{\mu}_{j'} μ j ′ and c ‾ j ′ \overline{c}_{j'} c j ′ . Substituting the eigenvalue decomposition into (17) gives
Λ ( 1 , 2 ) , p ( V ) = ( μ ‾ H + c ‾ H ) Δ ‾ ( μ ‾ + c ‾ ) = ∑ j ′ = 1 K T N R λ ‾ j ′ ∣ μ ‾ j ′ + c ‾ j ′ ∣ 2 . (20) \Lambda_{(1,2),p}(\mathbf{V})=(\overline{\boldsymbol{\mu}}^{H}+\overline{\mathbf{c}}^{H})\overline{\mathbf{\Delta}}(\overline{\boldsymbol{\mu}}+\overline{\mathbf{c}})=\sum_{j'=1}^{KTN_R}\overline{\lambda}_{j'}\left\lvert\overline{\mu}_{j'}+\overline{c}_{j'}\right\rvert^2. \tag*{(20)} Λ ( 1 , 2 ) , p ( V ) = ( μ H + c H ) Δ ( μ + c ) = j ′ = 1 ∑ K T N R λ j ′ μ j ′ + c j ′ 2 . (20)
The summation in (20) is a linear combination of K T N R KTN_R K T N R non-central chi-squares with 2 DoFs and non-centrality parameter Re { μ ‾ j ′ } 2 + Im { μ ‾ j ′ } 2 \operatorname{Re}\{\overline{\mu}_{j'}\}^2+\operatorname{Im}\{\overline{\mu}_{j'}\}^2 Re { μ j ′ } 2 + Im { μ j ′ } 2 . This proves (13) . To prove (14) , a similar approach can be taken by replacing v k , t \mathbf{v}_{k,t} v k , t in (12) with R k 1 2 v ~ k , t \mathbf{R}_k^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} R k 2 1 v k , t .
Proposed Detector for Case (2,1): Slowly-varying in Frequency and Constant in Time
We benefit from the slow variation in frequency and use a window in the frequency domain to develop our proposed detector. The signal is constant in time so that α k , t = α k \alpha_{k,t}=\alpha_k α k , t = α k . As such, the window moves only in the frequency domain and all samples are averaged uniformly in time. We use an estimate of α k \alpha_k α k consisting of the weighted average of the received voltage vectors as
α k ( 2 , 1 ) = ∑ t ∑ ℓ 2 = − L 2 L 2 v k − ℓ 2 , t w k , t , ℓ 2 . (21) \alpha_k^{(2,1)}=\sum_t\sum_{\ell_2=-L_2}^{L_2}\mathbf{v}_{k-\ell_2,t}w_{k,t,\ell_2}. \tag*{(21)} α k ( 2 , 1 ) = t ∑ ℓ 2 = − L 2 ∑ L 2 v k − ℓ 2 , t w k , t , ℓ 2 . (21)
In (21) , w k , t , ℓ 2 w_{k,t,\ell_2} w k , t , ℓ 2 are the weights and the window length is ( 2 L 2 + 1 ) (2L_2+1) ( 2 L 2 + 1 ) , with the center of the window at ℓ 2 = 0 \ell_2=0 ℓ 2 = 0 .
For ℓ 2 < − L 2 \ell_2<-L_2 ℓ 2 < − L 2 or ℓ 2 > L 2 \ell_2>L_2 ℓ 2 > L 2 , w k , t , ℓ 2 w_{k,t,\ell_2} w k , t , ℓ 2 is set to zero. Also since frequency indices are from 1 to K K K , w k , t , ℓ 2 w_{k,t,\ell_2} w k , t , ℓ 2 is set to zero for k − ℓ 2 < 1 k-\ell_2<1 k − ℓ 2 < 1 or k − ℓ 2 > K k-\ell_2>K k − ℓ 2 > K to avoid edge effects. Thus, there are three conditions for the weights, as in Section III-A, and we obtain the weights w k , t , ℓ 2 w_{k,t,\ell_2} w k , t , ℓ 2 as follows
w k , t , ℓ 1 = { 1 / ( T ( 2 L 2 + 1 ) ) k ≥ L 2 + 1 , k ≤ K − L 2 1 / ( T ( k + L 2 ) ) k ≤ L 2 1 / ( T ( K − k + 1 + L 2 ) ) k ≥ K − L 2 + 1. (22) w_{k,t,\ell_1}=
\begin{cases}
1/(T(2L_2+1)) & k\geq L_2+1,\ k\leq K-L_2\\
1/(T(k+L_2)) & k\leq L_2\\
1/(T(K-k+1+L_2)) & k\geq K-L_2+1.
\end{cases} \tag*{(22)} w k , t , ℓ 1 = ⎩ ⎨ ⎧ 1/ ( T ( 2 L 2 + 1 )) 1/ ( T ( k + L 2 )) 1/ ( T ( K − k + 1 + L 2 )) k ≥ L 2 + 1 , k ≤ K − L 2 k ≤ L 2 k ≥ K − L 2 + 1. (22)
We calculated the weights in (22) by uniform averaging across both time and frequency domains. Hence, ∑ t ∑ ℓ 2 = − L 2 L 2 w k , t , ℓ 2 = 1 \sum_t\sum_{\ell_2=-L_2}^{L_2}w_{k,t,\ell_2}=1 ∑ t ∑ ℓ 2 = − L 2 L 2 w k , t , ℓ 2 = 1 and owing to uniform averaging in the time domain for any t t t we also have ∑ ℓ 2 = − L 2 L 2 w k , t , ℓ 2 = 1 / T \sum_{\ell_2=-L_2}^{L_2}w_{k,t,\ell_2}=1/T ∑ ℓ 2 = − L 2 L 2 w k , t , ℓ 2 = 1/ T .
Now, using the weighted average in (21) and referring to (4) our proposed detector can be expressed as
Λ ( 2 , 1 ) , p ( V ) = ∑ k ( ∑ t v k , t H ) R k − 1 ( ∑ t ′ ∑ ℓ 2 = − L 2 L 2 v k − ℓ 2 , t ′ w k , t ′ , ℓ 2 ) + ∑ k ( ∑ t ′ ∑ ℓ 2 = − L 2 L 2 v k − ℓ 2 , t ′ H w k , t ′ , ℓ 2 ) R k − 1 ( ∑ t v k , t ) − T ∑ k ( ∑ t ′ ∑ ℓ 2 = − L 2 L 2 v k − ℓ 2 , t ′ H w k , t ′ , ℓ 2 ) R k − 1 ( ∑ t ∑ ℓ 2 = − L 2 L 2 v k − ℓ 2 , t w k , t , ℓ 2 ) . (23) \begin{aligned}
\Lambda_{(2,1),p}(\mathbf{V})={}&\sum_k\left(\sum_t\mathbf{v}_{k,t}^{H}\right)\mathbf{R}_k^{-1}\left(\sum_{t'}\sum_{\ell_2=-L_2}^{L_2}\mathbf{v}_{k-\ell_2,t'}w_{k,t',\ell_2}\right)\\
&+\sum_k\left(\sum_{t'}\sum_{\ell_2=-L_2}^{L_2}\mathbf{v}_{k-\ell_2,t'}^{H}w_{k,t',\ell_2}\right)\mathbf{R}_k^{-1}\left(\sum_t\mathbf{v}_{k,t}\right)\\
&-T\sum_k\left(\sum_{t'}\sum_{\ell_2=-L_2}^{L_2}\mathbf{v}_{k-\ell_2,t'}^{H}w_{k,t',\ell_2}\right)\mathbf{R}_k^{-1}\left(\sum_t\sum_{\ell_2=-L_2}^{L_2}\mathbf{v}_{k-\ell_2,t}w_{k,t,\ell_2}\right). \tag*{(23)}
\end{aligned} Λ ( 2 , 1 ) , p ( V ) = k ∑ ( t ∑ v k , t H ) R k − 1 ( t ′ ∑ ℓ 2 = − L 2 ∑ L 2 v k − ℓ 2 , t ′ w k , t ′ , ℓ 2 ) + k ∑ ( t ′ ∑ ℓ 2 = − L 2 ∑ L 2 v k − ℓ 2 , t ′ H w k , t ′ , ℓ 2 ) R k − 1 ( t ∑ v k , t ) − T k ∑ ( t ′ ∑ ℓ 2 = − L 2 ∑ L 2 v k − ℓ 2 , t ′ H w k , t ′ , ℓ 2 ) R k − 1 ( t ∑ ℓ 2 = − L 2 ∑ L 2 v k − ℓ 2 , t w k , t , ℓ 2 ) . (23)
Again, some algebra allows us to write (23) in a quadratic form as follows
Λ ( 2 , 1 ) , p ( V ) = ∑ k 1 = 1 K ∑ t 1 = 1 T ∑ k 2 = 1 K ∑ t 2 = 1 T v k 1 , t 1 H B k 1 t 1 k 2 t 2 v k 2 , t 2 , (24) \Lambda_{(2,1),p}(\mathbf{V})=\sum_{k_1=1}^{K}\sum_{t_1=1}^{T}\sum_{k_2=1}^{K}\sum_{t_2=1}^{T}\mathbf{v}_{k_1,t_1}^{H}\mathbf{B}_{k_1t_1k_2t_2}\mathbf{v}_{k_2,t_2}, \tag*{(24)} Λ ( 2 , 1 ) , p ( V ) = k 1 = 1 ∑ K t 1 = 1 ∑ T k 2 = 1 ∑ K t 2 = 1 ∑ T v k 1 , t 1 H B k 1 t 1 k 2 t 2 v k 2 , t 2 , (24)
where B k 1 t 1 k 2 t 2 \mathbf{B}_{k_1t_1k_2t_2} B k 1 t 1 k 2 t 2 is given by
B k 1 t 1 k 2 t 2 = R k 1 − 1 w k 1 , t 2 , k 1 − k 2 + w k 2 , t 1 , k 2 − k 1 R k 2 − 1 − T ∑ k w k , t 1 , k − k 1 ( ∑ k R k − 1 ) w k , t 2 , k − k 2 . (25) \mathbf{B}_{k_1t_1k_2t_2}=\mathbf{R}_{k_1}^{-1}w_{k_1,t_2,k_1-k_2}+w_{k_2,t_1,k_2-k_1}\mathbf{R}_{k_2}^{-1}-T\sum_k w_{k,t_1,k-k_1}\left(\sum_k\mathbf{R}_k^{-1}\right)w_{k,t_2,k-k_2}. \tag*{(25)} B k 1 t 1 k 2 t 2 = R k 1 − 1 w k 1 , t 2 , k 1 − k 2 + w k 2 , t 1 , k 2 − k 1 R k 2 − 1 − T k ∑ w k , t 1 , k − k 1 ( k ∑ R k − 1 ) w k , t 2 , k − k 2 . (25)
We can now write (24) in a matrix form by defining a K T N R KTN_R K T N R by K T N R KTN_R K T N R block matrix B ‾ \overline{\mathbf{B}} B . Block ( i , j ) (i,j) ( i , j ) in B ‾ \overline{\mathbf{B}} B is B k 1 t 1 k 2 t 2 \mathbf{B}_{k_{1}t_{1}k_{2}t_{2}} B k 1 t 1 k 2 t 2 by indices k 1 k_{1} k 1 , t 1 t_{1} t 1 , k 2 k_{2} k 2 and t 2 t_{2} t 2 which are computed referring to (10) , and i = 1 , 2 , … , K T i=1,2,\ldots,KT i = 1 , 2 , … , K T and j = 1 , 2 , … , K T j=1,2,\ldots,KT j = 1 , 2 , … , K T . Therefore, referring to (24) , we give our proposed detector as
Λ ( 2 , 1 ) , p ( V ) = v ‾ H B ‾ v ‾ , (26) \Lambda_{(2,1),\mathrm{p}}(\mathbf{V})=\overline{\mathbf{v}}^{H}\overline{\mathbf{B}}\overline{\mathbf{v}}, \tag*{(26)} Λ ( 2 , 1 ) , p ( V ) = v H B v , (26)
where v ‾ \overline{\mathbf{v}} v is given in (11) . We now formalize our analysis for Case (2,1) in the following theorem.
Theorem 2. The detector Λ ( 2 , 1 ) , p ( V ) \Lambda_{(2,1),\mathrm{p}}(\mathbf{V}) Λ ( 2 , 1 ) , p ( V ) in (26) under H 1 \mathcal{H}_{1} H 1 and H 0 \mathcal{H}_{0} H 0 is respectively a linear combination of non-central chi-squares and a linear combination of chi-squares, as
Λ ( 2 , 1 ) , p ∣ H 1 = ∑ j ′ = 1 K T N R λ ^ j ′ χ 2 , j ′ 2 ( Re { μ ^ j ′ } 2 + Im { μ ^ j ′ } 2 ) , (27) \Lambda_{(2,1),\mathrm{p}}|\mathcal{H}_{1}=\sum_{j'=1}^{KTN_{\mathrm{R}}}\widehat{\lambda}_{j'}\chi^{2}_{2,j'}\left(\operatorname{Re}\{\widehat{\mu}_{j'}\}^{2}+\operatorname{Im}\{\widehat{\mu}_{j'}\}^{2}\right), \tag*{(27)} Λ ( 2 , 1 ) , p ∣ H 1 = j ′ = 1 ∑ K T N R λ j ′ χ 2 , j ′ 2 ( Re { μ j ′ } 2 + Im { μ j ′ } 2 ) , (27) Λ ( 2 , 1 ) , p ∣ H 0 = ∑ j ′ = 1 K T N R λ ^ j ′ χ 2 , j ′ 2 , (28) \Lambda_{(2,1),\mathrm{p}}|\mathcal{H}_{0}=\sum_{j'=1}^{KTN_{\mathrm{R}}}\widehat{\lambda}_{j'}\chi^{2}_{2,j'}, \tag*{(28)} Λ ( 2 , 1 ) , p ∣ H 0 = j ′ = 1 ∑ K T N R λ j ′ χ 2 , j ′ 2 , (28) where λ ^ j ′ \widehat{\lambda}_{j'} λ j ′ and μ ^ j ′ \widehat{\mu}_{j'} μ j ′ are defined below (29) .
Proof. Our proposed detector for Case (2,1) can be used in a general scenario where the signal is time-varying. Thus, for the purpose of generality we consider a time-varying signal. Thus, we use α k , t \alpha_{k,t} α k , t in the analysis.
To calculate Λ ( 2 , 1 ) , p ∣ H 1 \Lambda_{(2,1),\mathrm{p}}|\mathcal{H}_{1} Λ ( 2 , 1 ) , p ∣ H 1 we substitute α k , t + R ‾ k 1 2 v ~ k , t \alpha_{k,t}+\overline{\mathbf{R}}_{k}^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v k , t for v k , t \mathbf{v}_{k,t} v k , t in (26) . Therefore, as in Section III-A, (26) can be expressed as
Λ ( 2 , 1 ) , p ( V ) = b ‾ H R ‾ B ‾ R ‾ b ‾ , (29) \Lambda_{(2,1),\mathrm{p}}(\mathbf{V})=\overline{\mathbf{b}}^{H}\overline{\mathbf{R}}\overline{\mathbf{B}}\overline{\mathbf{R}}\overline{\mathbf{b}}, \tag*{(29)} Λ ( 2 , 1 ) , p ( V ) = b H R B R b , (29)
where vector b ‾ \overline{\mathbf{b}} b is given in (15) and matrix R ‾ \overline{\mathbf{R}} R is given in (16) . We use the eigenvalue decomposition of matrix R ‾ B ‾ R ‾ \overline{\mathbf{R}}\overline{\mathbf{B}}\overline{\mathbf{R}} R B R in (29) as Φ ^ Δ ^ Φ ^ H \widehat{\mathbf{\Phi}}\widehat{\mathbf{\Delta}}\widehat{\mathbf{\Phi}}^{H} Φ Δ Φ H where Φ ^ \widehat{\mathbf{\Phi}} Φ denotes a unitary matrix and Δ ^ \widehat{\mathbf{\Delta}} Δ denotes a diagonal matrix containing the eigenvalues, λ ^ j ′ \widehat{\lambda}_{j'} λ j ′ . Define c ^ \widehat{\mathbf{c}} c and μ ^ \widehat{\boldsymbol{\mu}} μ by replacing Φ ‾ \overline{\mathbf{\Phi}} Φ with Φ ^ \widehat{\mathbf{\Phi}} Φ in (18) and (19) respectively. Also let μ ^ j ′ \widehat{\mu}_{j'} μ j ′ and c ^ j ′ \widehat{c}_{j'} c j ′ be elements of vectors μ ^ \widehat{\boldsymbol{\mu}} μ and c ^ \widehat{\mathbf{c}} c respectively. By substituting the eigenvalue decomposition into (29) we get
Λ ( 2 , 1 ) , p ( V ) = ( μ ^ H + c ^ H ) Δ ^ ( μ ^ + c ^ ) = ∑ j ′ = 1 K T N R λ ^ j ′ ∣ μ ^ j ′ + c ^ j ′ ∣ 2 . (30) \Lambda_{(2,1),\mathrm{p}}(\mathbf{V})=(\widehat{\boldsymbol{\mu}}^{H}+\widehat{\mathbf{c}}^{H})\widehat{\mathbf{\Delta}}(\widehat{\boldsymbol{\mu}}+\widehat{\mathbf{c}})=\sum_{j'=1}^{KTN_{\mathrm{R}}}\widehat{\lambda}_{j'}\left|\widehat{\mu}_{j'}+\widehat{c}_{j'}\right|^{2}. \tag*{(30)} Λ ( 2 , 1 ) , p ( V ) = ( μ H + c H ) Δ ( μ + c ) = j ′ = 1 ∑ K T N R λ j ′ ∣ μ j ′ + c j ′ ∣ 2 . (30)
It is observed that the right hand side of the expression in (30) is a linear combination of K T N R KTN_{\mathrm{R}} K T N R non-central chi-squares with 2 DoFs and non-centrality parameter Re { μ ^ j ′ } 2 + Im { μ ^ j ′ } 2 \operatorname{Re}\{\widehat{\mu}_{j'}\}^{2}+\operatorname{Im}\{\widehat{\mu}_{j'}\}^{2} Re { μ j ′ } 2 + Im { μ j ′ } 2 , given in (27) . For the null hypothesis, (28) can be similarly proved by replacing v k , t \mathbf{v}_{k,t} v k , t in (26) with R ‾ k 1 2 v ~ k , t \overline{\mathbf{R}}_{k}^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} R k 2 1 v k , t .
Proposed Detector for Case (2,2): Slowly varying in Both Domains
We make use of slow variation in both the frequency and time domains and propose a detector using a rectangular two-dimensional (2-D) window. The window moves in both frequency and time. Therefore we propose to use an estimate of α k , t \alpha_{k,t} α k , t as α k , t ( 2 , 2 ) \alpha_{k,t}^{(2,2)} α k , t ( 2 , 2 ) that is a weighted average of a 2-D window consisting of the received voltage vectors as
α k , t ( 2 , 2 ) = ∑ m = − M M ∑ q = − Q Q v k − m , t − q w k , m , t , q . (31) \alpha_{k,t}^{(2,2)}=\sum_{m=-M}^{M}\sum_{q=-Q}^{Q}\mathbf{v}_{k-m,t-q}w_{k,m,t,q}. \tag*{(31)} α k , t ( 2 , 2 ) = m = − M ∑ M q = − Q ∑ Q v k − m , t − q w k , m , t , q . (31)
with w k , m , t , q w_{k,m,t,q} w k , m , t , q being the weights of the window. Parameters M M M and Q Q Q control the length of the window in the frequency domain, i.e., 2 M + 1 2M+1 2 M + 1 , and in the time domain, i.e., 2 Q + 1 2Q+1 2 Q + 1 respectively. Also, in the weighted average expression in (31) , the center of the 2-D window with m = 0 m=0 m = 0 and q = 0 q=0 q = 0 coincides with the received voltage vector for the pair ( k , t ) (k,t) ( k , t ) . Moreover, frequency and time indices are in the ranges 1 1 1 to K K K and 1 1 1 to T T T respectively. Thus, to avoid edge effects, w k , m , t , q w_{k,m,t,q} w k , m , t , q is set to zero for k − m < 1 k-m<1 k − m < 1 or k − m > K k-m>K k − m > K or t − q < 1 t-q<1 t − q < 1 or t − q > T t-q>T t − q > T . Also w k , m , t , q w_{k,m,t,q} w k , m , t , q is set to zero for m < − M m<-M m < − M or m > M m>M m > M or q < − Q q<-Q q < − Q or q > Q q>Q q > Q .
There are 9 possible conditions on the weights depending on where the 2-D window lies with respect to the frequency and time domains, including edge cases which were elaborated on in Section III-A. The weights are calculated by uniform averaging in both time and frequency and sum to one. Calculation of the weights is straightforward and similar to that of previous subsections. Hence, we omit it for brevity.
Referring to (4) and (31) , our proposed detector for Case (2,2) is written as
Λ ( 2 , 2 ) , p ( V ) = ∑ k ∑ t [ v k , t H R k − 1 ( ∑ m ∑ q v k − m , t − q w k , m , t , q ) + ( ∑ m ∑ q v k − m , t − q H w k , m , t , q ) R k − 1 v k , t − ( ∑ m ∑ q v k − m , t − q H w k , m , t , q ) R k − 1 × ( ∑ m ′ ∑ q ′ v k − m ′ , t − q ′ w k , m ′ , t , q ′ ) ] . (32) \begin{aligned}
\Lambda_{(2,2),\mathrm{p}}(\mathbf{V})={}&\sum_{k}\sum_{t}\Bigg[\mathbf{v}_{k,t}^{H}\mathbf{R}_{k}^{-1}\left(\sum_{m}\sum_{q}\mathbf{v}_{k-m,t-q}w_{k,m,t,q}\right)\\
&+\left(\sum_{m}\sum_{q}\mathbf{v}_{k-m,t-q}^{H}w_{k,m,t,q}\right)\mathbf{R}_{k}^{-1}\mathbf{v}_{k,t}\\
&-\left(\sum_{m}\sum_{q}\mathbf{v}_{k-m,t-q}^{H}w_{k,m,t,q}\right)\mathbf{R}_{k}^{-1}\\
&\qquad\times\left(\sum_{m'}\sum_{q'}\mathbf{v}_{k-m',t-q'}w_{k,m',t,q'}\right)\Bigg]. \tag*{(32)}
\end{aligned} Λ ( 2 , 2 ) , p ( V ) = k ∑ t ∑ [ v k , t H R k − 1 ( m ∑ q ∑ v k − m , t − q w k , m , t , q ) + ( m ∑ q ∑ v k − m , t − q H w k , m , t , q ) R k − 1 v k , t − ( m ∑ q ∑ v k − m , t − q H w k , m , t , q ) R k − 1 × m ′ ∑ q ′ ∑ v k − m ′ , t − q ′ w k , m ′ , t , q ′ ] . (32)
The detector in (32) can be written as a quadratic form:
Λ ( 2 , 2 ) , p ( V ) = ∑ k 1 = 1 K ∑ t 1 = 1 T ∑ k 2 = 1 K ∑ t 2 = 1 T v k 1 , t 1 H C k 1 t 1 k 2 t 2 v k 2 , t 2 , (33) \Lambda_{(2,2),\mathrm{p}}(\mathbf{V})=\sum_{k_{1}=1}^{K}\sum_{t_{1}=1}^{T}\sum_{k_{2}=1}^{K}\sum_{t_{2}=1}^{T}\mathbf{v}_{k_{1},t_{1}}^{H}\mathbf{C}_{k_{1}t_{1}k_{2}t_{2}}\mathbf{v}_{k_{2},t_{2}}, \tag*{(33)} Λ ( 2 , 2 ) , p ( V ) = k 1 = 1 ∑ K t 1 = 1 ∑ T k 2 = 1 ∑ K t 2 = 1 ∑ T v k 1 , t 1 H C k 1 t 1 k 2 t 2 v k 2 , t 2 , (33)
where we obtain matrix C k 1 t 1 k 2 t 2 \mathbf{C}_{k_{1}t_{1}k_{2}t_{2}} C k 1 t 1 k 2 t 2 as
C k 1 t 1 k 2 t 2 = R k 1 − 1 w k 1 , k 1 − k 2 , t 1 , t 1 − t 2 + w k 2 , k 2 − k 1 , t 2 , t 2 − t 1 R k 2 − 1 − ∑ k ∑ t w k , k − 1 , t , t − t 1 R k − 1 w k , k − k 2 , t , t − t 2 . (34) \mathbf{C}_{k_{1}t_{1}k_{2}t_{2}}=\mathbf{R}_{k_{1}}^{-1}w_{k_{1},k_{1}-k_{2},t_{1},t_{1}-t_{2}}+w_{k_{2},k_{2}-k_{1},t_{2},t_{2}-t_{1}}\mathbf{R}_{k_{2}}^{-1}-\sum_{k}\sum_{t}w_{k,k-1,t,t-t_{1}}\mathbf{R}_{k}^{-1}w_{k,k-k_{2},t,t-t_{2}}. \tag*{(34)} C k 1 t 1 k 2 t 2 = R k 1 − 1 w k 1 , k 1 − k 2 , t 1 , t 1 − t 2 + w k 2 , k 2 − k 1 , t 2 , t 2 − t 1 R k 2 − 1 − k ∑ t ∑ w k , k − 1 , t , t − t 1 R k − 1 w k , k − k 2 , t , t − t 2 . (34)
Define a K T N R KTN_{\mathrm{R}} K T N R by K T N R KTN_{\mathrm{R}} K T N R block matrix C ‾ \overline{\mathbf{C}} C where block ( i , j ) (i,j) ( i , j ) is C k 1 t 1 k 2 t 2 \mathbf{C}_{k_{1}t_{1}k_{2}t_{2}} C k 1 t 1 k 2 t 2 and the indices k 1 , t 1 , k 2 k_{1},t_{1},k_{2} k 1 , t 1 , k 2 and t 2 t_{2} t 2 are obtained using (10) , with i = 1 , 2 , … , K T i=1,2,\ldots,KT i = 1 , 2 , … , K T and j = 1 , 2 , … , K T j=1,2,\ldots,KT j = 1 , 2 , … , K T . As such, referring to (33) , our proposed detector is expressed as
Λ ( 2 , 2 ) , p ( V ) = v ‾ H C ‾ v ‾ , (35) \Lambda_{(2,2),\mathrm{p}}(\mathbf{V})=\overline{\mathbf{v}}^{H}\overline{\mathbf{C}}\overline{\mathbf{v}}, \tag*{(35)} Λ ( 2 , 2 ) , p ( V ) = v H C v , (35)
where v ‾ \overline{\mathbf{v}} v is given in (11) .
In the following theorem we present an analysis for our proposed detector for Case (2,2).
Theorem 3. The detector Λ ( 2 , 2 ) , p ( V ) \Lambda_{(2,2),\mathrm{p}}(\mathbf{V}) Λ ( 2 , 2 ) , p ( V ) in (35) under H 1 \mathcal{H}_{1} H 1 and H 0 \mathcal{H}_{0} H 0 is respectively a linear combination of non-central chi-squares and a linear combination of chi-squares, as
Λ ( 2 , 2 ) , p ∣ H 1 = ∑ j ′ = 1 K T N R ω j ′ χ 2 , j ′ 2 ( Re { ζ j ′ } 2 + Im { ζ j ′ } 2 ) , (36) \Lambda_{(2,2),\mathrm{p}}|\mathcal{H}_{1}=\sum_{j'=1}^{KTN_{\mathrm{R}}}\omega_{j'}\chi^{2}_{2,j'}\left(\operatorname{Re}\{\zeta_{j'}\}^{2}+\operatorname{Im}\{\zeta_{j'}\}^{2}\right), \tag*{(36)} Λ ( 2 , 2 ) , p ∣ H 1 = j ′ = 1 ∑ K T N R ω j ′ χ 2 , j ′ 2 ( Re { ζ j ′ } 2 + Im { ζ j ′ } 2 ) , (36)
Λ ( 2 , 2 ) , p ∣ H 0 = ∑ j ′ = 1 K T N R ω j ′ ′ χ 2 , j ′ 2 (37) \Lambda_{(2,2),p}\mid\mathcal{H}_0=\sum_{j'=1}^{KTN_R}\omega_{j'}'\chi_{2,j'}^2 \tag*{(37)} Λ ( 2 , 2 ) , p ∣ H 0 = j ′ = 1 ∑ K T N R ω j ′ ′ χ 2 , j ′ 2 (37)
where ω j ′ ′ \omega_{j'}' ω j ′ ′ and ζ j ′ ′ \zeta_{j'}' ζ j ′ ′ are defined above (39).
Proof. Similar to the proofs of Theorems 1 and 2, (35) can be expressed as
Λ ( 2 , 2 ) , p ( V ) = b ˉ H R ˉ C ˉ R ˉ b ˉ , (38) \Lambda_{(2,2),p}(\mathbf{V})=\bar{\mathbf{b}}^{H}\bar{\mathbf{R}}\bar{\mathbf{C}}\bar{\mathbf{R}}\bar{\mathbf{b}}, \tag*{(38)} Λ ( 2 , 2 ) , p ( V ) = b ˉ H R ˉ C ˉ R ˉ b ˉ , (38)
where b ˉ \bar{\mathbf{b}} b ˉ and R ˉ \bar{\mathbf{R}} R ˉ are given in (15) and (16). Matrix R ˉ C ˉ R ˉ \bar{\mathbf{R}}\bar{\mathbf{C}}\bar{\mathbf{R}} R ˉ C ˉ R ˉ in (38) can be replaced by its eigenvalue decomposition as Ψ Ω Ψ H \mathbf{\Psi}\mathbf{\Omega}\mathbf{\Psi}^{H} ΨΩ Ψ H where Ψ \mathbf{\Psi} Ψ denotes a unitary matrix and Ω \mathbf{\Omega} Ω denotes a diagonal matrix containing the eigenvalues, ω j ′ ′ \omega_{j'}' ω j ′ ′ . Define d \mathbf{d} d and ζ \boldsymbol{\zeta} ζ by replacing Φ \mathbf{\Phi} Φ with Ψ \mathbf{\Psi} Ψ in (18) and (19) respectively. In addition, denote the elements of vectors ζ \boldsymbol{\zeta} ζ and d \mathbf{d} d by ζ j ′ ′ \zeta_{j'}' ζ j ′ ′ and d j ′ ′ d_{j'}' d j ′ ′ respectively. Thus we obtain
Λ ( 2 , 2 ) , p ( V ) = ( ζ H + d H ) Ω ( ζ + d ) = ∑ j ′ = 1 K T N R ω j ′ ′ ∣ ζ j ′ ′ + d j ′ ′ ∣ 2 . (39) \Lambda_{(2,2),p}(\mathbf{V})=(\boldsymbol{\zeta}^{H}+\mathbf{d}^{H})\mathbf{\Omega}(\boldsymbol{\zeta}+\mathbf{d})=\sum_{j'=1}^{KTN_R}\omega_{j'}'\left|\zeta_{j'}'+d_{j'}'\right|^2. \tag*{(39)} Λ ( 2 , 2 ) , p ( V ) = ( ζ H + d H ) Ω ( ζ + d ) = j ′ = 1 ∑ K T N R ω j ′ ′ ζ j ′ ′ + d j ′ ′ 2 . (39)
Consequently the proof of (36) is complete. We also refer the reader to the proofs of the previous theorems to prove (37).
Proposed Detector for Case (2,3): Slowly Varying in Frequency and Rapidly Varying in Time
We design a detector that uses a weighted average of received voltage vectors in the frequency domain as an estimate of the signal. We estimate α k , t \alpha_{k,t} α k , t with α k , t ( 2 , 3 ) \alpha_{k,t}^{(2,3)} α k , t ( 2 , 3 ) which is, for a frequency sample k k k , a weighted average of vectors v k , t \mathbf{v}_{k,t} v k , t in the frequency domain as
α k , t ( 2 , 3 ) = ∑ ℓ 3 = − L 3 L 3 v k − ℓ 3 , t w k , ℓ 3 , (40) \alpha_{k,t}^{(2,3)}=\sum_{\ell_3=-L_3}^{L_3}\mathbf{v}_{k-\ell_3,t}w_{k,\ell_3}, \tag*{(40)} α k , t ( 2 , 3 ) = ℓ 3 = − L 3 ∑ L 3 v k − ℓ 3 , t w k , ℓ 3 , (40)
where w k , ℓ 3 w_{k,\ell_3} w k , ℓ 3 denote the weights. Replacing α k , t \alpha_{k,t} α k , t by the weighted average in (40) implements smoothing via a window of surrounding frequency samples. The center of the window corresponding to ℓ 3 = 0 \ell_3=0 ℓ 3 = 0 matches to the k k k -th frequency sample. Parameter L 3 L_3 L 3 determines the window length, i.e., 2 L 3 + 1 2L_3+1 2 L 3 + 1 , where 2 L 3 + 1 < K 2L_3+1<K 2 L 3 + 1 < K . Weights w k , ℓ 3 w_{k,\ell_3} w k , ℓ 3 are set to zero when ℓ 3 < − L 3 \ell_3<-L_3 ℓ 3 < − L 3 or ℓ 3 > L 3 \ell_3>L_3 ℓ 3 > L 3 . Also, w k , ℓ 3 w_{k,\ell_3} w k , ℓ 3 are set to zero for k − ℓ 3 < 1 k-\ell_3<1 k − ℓ 3 < 1 or k − ℓ 3 > K k-\ell_3>K k − ℓ 3 > K due to edge effects. As explained above in Section III-A, we obtain three conditions for the weights as
w k , ℓ 3 = { 1 / ( 2 L 3 + 1 ) k ≥ L 3 + 1 , k ≤ K − L 3 1 / ( k + L 3 ) k ≤ L 3 1 / ( K − k + 1 + L 3 ) k ≥ K − L 3 + 1 , (41) w_{k,\ell_3}=
\begin{cases}
1/(2L_3+1) & k\ge L_3+1,\ k\le K-L_3\\
1/(k+L_3) & k\le L_3\\
1/(K-k+1+L_3) & k\ge K-L_3+1,
\end{cases} \tag*{(41)} w k , ℓ 3 = ⎩ ⎨ ⎧ 1/ ( 2 L 3 + 1 ) 1/ ( k + L 3 ) 1/ ( K − k + 1 + L 3 ) k ≥ L 3 + 1 , k ≤ K − L 3 k ≤ L 3 k ≥ K − L 3 + 1 , (41)
which we calculate based on uniform averaging across frequency. Referring to (4) and (40) we write our proposed detector for Case (2,3) as
Λ ( 2 , 3 ) , p ( V ) = ∑ t [ ∑ k ∑ ℓ 3 = − L 3 L 3 v k , t H R k − 1 v k − ℓ 3 , t w k , ℓ 3 + ∑ k ( ∑ ℓ 3 = − L 3 L 3 w k , ℓ 3 v k − ℓ 3 , t H ) R k − 1 v k , t − ∑ k ( ∑ ℓ 3 = − L 3 L 3 w k , ℓ 3 v k − ℓ 3 , t H ) R k − 1 ( ∑ ℓ 3 ′ = − L 3 L 3 v k − ℓ 3 ′ , t w k , ℓ 3 ′ ) ] . (42) \begin{aligned}
\Lambda_{(2,3),p}(\mathbf{V})={}&\sum_t\left[\sum_k\sum_{\ell_3=-L_3}^{L_3}\mathbf{v}_{k,t}^{H}\mathbf{R}_k^{-1}\mathbf{v}_{k-\ell_3,t}w_{k,\ell_3}\right.\\
&+\sum_k\left(\sum_{\ell_3=-L_3}^{L_3}w_{k,\ell_3}\mathbf{v}_{k-\ell_3,t}^{H}\right)\mathbf{R}_k^{-1}\mathbf{v}_{k,t}\\
&\left.-\sum_k\left(\sum_{\ell_3=-L_3}^{L_3}w_{k,\ell_3}\mathbf{v}_{k-\ell_3,t}^{H}\right)\mathbf{R}_k^{-1}\left(\sum_{\ell_3'=-L_3}^{L_3}\mathbf{v}_{k-\ell_3',t}w_{k,\ell_3'}\right)\right]. \tag*{(42)}
\end{aligned} Λ ( 2 , 3 ) , p ( V ) = t ∑ [ k ∑ ℓ 3 = − L 3 ∑ L 3 v k , t H R k − 1 v k − ℓ 3 , t w k , ℓ 3 + k ∑ ( ℓ 3 = − L 3 ∑ L 3 w k , ℓ 3 v k − ℓ 3 , t H ) R k − 1 v k , t − k ∑ ( ℓ 3 = − L 3 ∑ L 3 w k , ℓ 3 v k − ℓ 3 , t H ) R k − 1 ℓ 3 ′ = − L 3 ∑ L 3 v k − ℓ 3 ′ , t w k , ℓ 3 ′ . (42)
We then rewrite Λ ( 2 , 3 ) , p ( V ) \Lambda_{(2,3),p}(\mathbf{V}) Λ ( 2 , 3 ) , p ( V ) in (42) in a quadratic form as
Λ ( 2 , 3 ) , p ( V ) = ∑ t [ v 1 , t H v 2 , t H … v K , t H ] A [ v 1 , t T v 2 , t T … v K , t T ] T , (43) \Lambda_{(2,3),p}(\mathbf{V})=\sum_t\left[\mathbf{v}_{1,t}^{H}\ \mathbf{v}_{2,t}^{H}\ldots\mathbf{v}_{K,t}^{H}\right]\mathbf{A}\left[\mathbf{v}_{1,t}^{T}\ \mathbf{v}_{2,t}^{T}\ldots\mathbf{v}_{K,t}^{T}\right]^{T}, \tag*{(43)} Λ ( 2 , 3 ) , p ( V ) = t ∑ [ v 1 , t H v 2 , t H … v K , t H ] A [ v 1 , t T v 2 , t T … v K , t T ] T , (43)
where matrix A K N R × K N R \mathbf{A}_{K N_R\times K N_R} A K N R × K N R is a block matrix with N R N_R N R by N R N_R N R sub-matrices A r s \mathbf{A}_{rs} A r s , for 1 ≤ r ≤ K 1\le r\le K 1 ≤ r ≤ K and 1 ≤ s ≤ K 1\le s\le K 1 ≤ s ≤ K . We calculate the sub-matrices as
A r s = R r − 1 w r , r − s + w s , s − r R s − 1 − ∑ k w k , k − r R k − 1 w k , k − s . (44) \mathbf{A}_{rs}=\mathbf{R}_r^{-1}w_{r,r-s}+w_{s,s-r}\mathbf{R}_s^{-1}-\sum_k w_{k,k-r}\mathbf{R}_k^{-1}w_{k,k-s}. \tag*{(44)} A r s = R r − 1 w r , r − s + w s , s − r R s − 1 − k ∑ w k , k − r R k − 1 w k , k − s . (44)
We now proceed to provide a performance analysis for our proposed detector with the following theorem.
Theorem 4. The detector Λ ( 2 , 3 ) , p ( V ) \Lambda_{(2,3),p}(\mathbf{V}) Λ ( 2 , 3 ) , p ( V ) in (43) under H 1 \mathcal{H}_1 H 1 and H 0 \mathcal{H}_0 H 0 is respectively a linear combination of non-central chi-squares and chi-squares, as
Λ ( 2 , 3 ) , p ∣ H 1 = ∑ i = 1 K N R λ i χ 2 T , i 2 ( ∑ t Re { μ i , t } 2 + Im { μ i , t } 2 ) , (45) \Lambda_{(2,3),p}\mid\mathcal{H}_1=\sum_{i=1}^{K N_R}\lambda_i\chi_{2T,i}^{2}\left(\sum_t\operatorname{Re}\{\mu_{i,t}\}^{2}+\operatorname{Im}\{\mu_{i,t}\}^{2}\right), \tag*{(45)} Λ ( 2 , 3 ) , p ∣ H 1 = i = 1 ∑ K N R λ i χ 2 T , i 2 ( t ∑ Re { μ i , t } 2 + Im { μ i , t } 2 ) , (45) Λ ( 2 , 3 ) , p ∣ H 0 = ∑ i = 1 K N R λ i χ 2 T , i 2 , (46) \Lambda_{(2,3),p}\mid\mathcal{H}_0=\sum_{i=1}^{K N_R}\lambda_i\chi_{2T,i}^{2}, \tag*{(46)} Λ ( 2 , 3 ) , p ∣ H 0 = i = 1 ∑ K N R λ i χ 2 T , i 2 , (46) where λ i \lambda_i λ i and μ i , t \mu_{i,t} μ i , t are defined above and below (48) respectively.
Proof. Similar to the previous proofs, under H 1 \mathcal{H}_1 H 1 we replace v k , t \mathbf{v}_{k,t} v k , t in (43) with α k , t + R k 1 2 v ~ k , t \boldsymbol{\alpha}_{k,t}+\mathbf{R}_k^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v ~ k , t . Let b ~ k , t = R k − 1 2 α k , t + v ~ k , t \tilde{\mathbf{b}}_{k,t}=\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,t}+\tilde{\mathbf{v}}_{k,t} b ~ k , t = R k − 2 1 α k , t + v ~ k , t and define R ~ K N R × K N R \tilde{\mathbf{R}}_{K N_R\times K N_R} R ~ K N R × K N R as a block diagonal matrix with sub-matrices R 1 1 2 , R 2 1 2 , … , R K 1 2 \mathbf{R}_1^{\frac{1}{2}},\mathbf{R}_2^{\frac{1}{2}},\ldots,\mathbf{R}_K^{\frac{1}{2}} R 1 2 1 , R 2 2 1 , … , R K 2 1 on the diagonal. Thus, we write Λ ( 2 , 3 ) , p ( V ) \Lambda_{(2,3),p}(\mathbf{V}) Λ ( 2 , 3 ) , p ( V ) in (43) under H 1 \mathcal{H}_1 H 1 as
Λ ( 2 , 3 ) , p ∣ H 1 = ∑ t [ b ~ 1 , t H b ~ 2 , t H … b ~ K , t H ] R ~ A R ~ [ b ~ 1 , t T b ~ 2 , t T … b ~ K , t T ] T . (47) \Lambda_{(2,3),p}\mid\mathcal{H}_1=\sum_t\left[\tilde{\mathbf{b}}_{1,t}^{H}\ \tilde{\mathbf{b}}_{2,t}^{H}\ldots\tilde{\mathbf{b}}_{K,t}^{H}\right]\tilde{\mathbf{R}}\mathbf{A}\tilde{\mathbf{R}}\left[\tilde{\mathbf{b}}_{1,t}^{T}\ \tilde{\mathbf{b}}_{2,t}^{T}\ldots\tilde{\mathbf{b}}_{K,t}^{T}\right]^{T}. \tag*{(47)} Λ ( 2 , 3 ) , p ∣ H 1 = t ∑ [ b ~ 1 , t H b ~ 2 , t H … b ~ K , t H ] R ~ A R ~ [ b ~ 1 , t T b ~ 2 , t T … b ~ K , t T ] T . (47) We replace matrix R ~ A R ~ \tilde{\mathbf{R}}\mathbf{A}\tilde{\mathbf{R}} R ~ A R ~ in (47) by its eigenvalue decomposition Φ Δ Φ H \mathbf{\Phi}\mathbf{\Delta}\mathbf{\Phi}^{H} ΦΔ Φ H , where Φ \mathbf{\Phi} Φ is a unitary matrix and Δ \mathbf{\Delta} Δ is a diagonal matrix containing the eigenvalues λ i \lambda_i λ i for i = 1 , 2 , … , K N R i=1,2,\ldots,K N_R i = 1 , 2 , … , K N R . Define
μ t = Φ H [ ( R 1 − 1 2 α 1 , t ) T ( R 2 − 1 2 α 2 , t ) T … ( R K − 1 2 α K , t ) T ] T , (48) \boldsymbol{\mu}_t=\mathbf{\Phi}^{H}\left[\left(\mathbf{R}_1^{-\frac{1}{2}}\boldsymbol{\alpha}_{1,t}\right)^{T}\ \left(\mathbf{R}_2^{-\frac{1}{2}}\boldsymbol{\alpha}_{2,t}\right)^{T}\ldots\left(\mathbf{R}_K^{-\frac{1}{2}}\boldsymbol{\alpha}_{K,t}\right)^{T}\right]^{T}, \tag*{(48)} μ t = Φ H [ ( R 1 − 2 1 α 1 , t ) T ( R 2 − 2 1 α 2 , t ) T … ( R K − 2 1 α K , t ) T ] T , (48) and c t = Φ H [ v ~ 1 , t T v ~ 2 , t T … v ~ K , t T ] T \mathbf{c}_t=\mathbf{\Phi}^{H}\left[\tilde{\mathbf{v}}_{1,t}^{T}\ \tilde{\mathbf{v}}_{2,t}^{T}\ldots\tilde{\mathbf{v}}_{K,t}^{T}\right]^{T} c t = Φ H [ v ~ 1 , t T v ~ 2 , t T … v ~ K , t T ] T . Note that c t ∼ C N ( 0 , I ) \mathbf{c}_t\sim\mathcal{CN}(\mathbf{0},\mathbf{I}) c t ∼ C N ( 0 , I ) as Φ \mathbf{\Phi} Φ is unitary. Also denote by μ i , t \mu_{i,t} μ i , t and c i , t c_{i,t} c i , t the i i i -th elements of μ t \boldsymbol{\mu}_t μ t and c t \mathbf{c}_t c t respectively. Thus, (47) is given as
Λ ( 2 , 3 ) , p ∣ H 1 = ∑ t ( μ t H + c t H ) Δ ( μ t + c t ) = ∑ i = 1 K N R λ i ∑ t ∣ μ i , t + c i , t ∣ 2 . (49) \Lambda_{(2,3),p}\mid\mathcal{H}_1=\sum_t(\boldsymbol{\mu}_t^{H}+\mathbf{c}_t^{H})\mathbf{\Delta}(\boldsymbol{\mu}_t+\mathbf{c}_t)=\sum_{i=1}^{K N_R}\lambda_i\sum_t\left|\mu_{i,t}+c_{i,t}\right|^2. \tag*{(49)} Λ ( 2 , 3 ) , p ∣ H 1 = t ∑ ( μ t H + c t H ) Δ ( μ t + c t ) = i = 1 ∑ K N R λ i t ∑ ∣ μ i , t + c i , t ∣ 2 . (49) Finally, the summation over t t t in (49) is a chi-square with 2 T 2T 2 T DoFs and non-centrality parameter ∑ t ( Re { μ i , t } 2 + Im { μ i , t } 2 ) \sum_t(\operatorname{Re}\{\mu_{i,t}\}^{2}+\operatorname{Im}\{\mu_{i,t}\}^{2}) ∑ t ( Re { μ i , t } 2 + Im { μ i , t } 2 ) . Hence, (49) is a linear combination of non-central chi-squares, which results in (45). Note that for the null hypothesis there is no signal component. Thus, to calculate Λ ( 2 , 3 ) , p ∣ H 0 \Lambda_{(2,3),p}\mid\mathcal{H}_0 Λ ( 2 , 3 ) , p ∣ H 0 we replace v k , t \mathbf{v}_{k,t} v k , t in (43) with R k 1 2 v ~ k , t \mathbf{R}_k^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t and repeat the same process, which gives the expression in (46).
Proposed Detector for Case (3,2): Rapidly Varying in Frequency and Slowly Varying in Time
For Case (3,2), considering a slowly varying signal in time we propose a detector using a weighted average of received voltage vectors in the time domain. We estimate α k , t \alpha_{k,t} α k , t by α k , t ( 3 , 2 ) \alpha_{k,t}^{(3,2)} α k , t ( 3 , 2 ) which is, for time sample t t t , a weighted average taken in the time domain as
α k , t ( 3 , 2 ) = ∑ ℓ 4 = − L 4 L 4 v k , t − ℓ 4 w t , ℓ 4 , (50) \alpha_{k,t}^{(3,2)}=\sum_{\ell_4=-L_4}^{L_4}\mathbf{v}_{k,t-\ell_4}w_{t,\ell_4}, \tag*{(50)} α k , t ( 3 , 2 ) = ℓ 4 = − L 4 ∑ L 4 v k , t − ℓ 4 w t , ℓ 4 , (50)
where w t , ℓ 4 w_{t,\ell_4} w t , ℓ 4 denotes the weights. We use (50) to replace α k , t \alpha_{k,t} α k , t by a window in the time domain whose center, corresponding to ℓ 4 = 0 \ell_4=0 ℓ 4 = 0 , matches the t t t -th time sample. The length of the window is 2 L 4 + 1 2L_4+1 2 L 4 + 1 which is assumed to be smaller than T T T . To avoid edge cases, we set weights w t , ℓ 4 w_{t,\ell_4} w t , ℓ 4 to zero when t − ℓ 4 < 1 t-\ell_4<1 t − ℓ 4 < 1 or t − ℓ 4 > T t-\ell_4>T t − ℓ 4 > T . Also w t , ℓ 4 = 0 w_{t,\ell_4}=0 w t , ℓ 4 = 0 for ℓ 4 < − L 4 \ell_4<-L_4 ℓ 4 < − L 4 or ℓ 4 > L 4 \ell_4>L_4 ℓ 4 > L 4 . We calculate the weights by uniform averaging in time as
w t , ℓ 4 = { 1 / ( 2 L 4 + 1 ) t ≥ L 4 + 1 , t ≤ T − L 4 1 / ( t + L 4 ) t ≤ L 4 1 / ( T − t + 1 + L 4 ) t ≥ T − L 4 + 1. (51) w_{t,\ell_4}=
\begin{cases}
1/(2L_4+1) & t\ge L_4+1,t\le T-L_4\\
1/(t+L_4) & t\le L_4\\
1/(T-t+1+L_4) & t\ge T-L_4+1.
\end{cases} \tag*{(51)} w t , ℓ 4 = ⎩ ⎨ ⎧ 1/ ( 2 L 4 + 1 ) 1/ ( t + L 4 ) 1/ ( T − t + 1 + L 4 ) t ≥ L 4 + 1 , t ≤ T − L 4 t ≤ L 4 t ≥ T − L 4 + 1. (51)
We write our proposed detector for Case (3,2), referring to (4) and (50), as
Λ ( 3 , 2 ) , p ( V ) = ∑ k [ ∑ t v k , t H R k − 1 ( ∑ ℓ 4 = − L 4 L 4 v k , t − ℓ 4 w t , ℓ 4 ) + ∑ t ( ∑ ℓ 4 = − L 4 L 4 w t , ℓ 4 v k , t − ℓ 4 H ) R k − 1 v k , t − ( ∑ ℓ 4 = − L 4 L 4 w t , ℓ 4 v k , t − ℓ 4 H ) R k − 1 ( ∑ ℓ 4 ′ = − L 4 L 4 v k , t − ℓ 4 ′ w t , ℓ 4 ′ ) ] . (52) \begin{aligned}
\Lambda_{(3,2),p}(\mathbf{V})={}&\sum_k\left[\sum_t\mathbf{v}_{k,t}^{H}\mathbf{R}_k^{-1}\left(\sum_{\ell_4=-L_4}^{L_4}\mathbf{v}_{k,t-\ell_4}w_{t,\ell_4}\right)\right.\\
&+\sum_t\left(\sum_{\ell_4=-L_4}^{L_4}w_{t,\ell_4}\mathbf{v}_{k,t-\ell_4}^{H}\right)\mathbf{R}_k^{-1}\mathbf{v}_{k,t}\\
&\left.-\left(\sum_{\ell_4=-L_4}^{L_4}w_{t,\ell_4}\mathbf{v}_{k,t-\ell_4}^{H}\right)\mathbf{R}_k^{-1}\left(\sum_{\ell_4'=-L_4}^{L_4}\mathbf{v}_{k,t-\ell_4'}w_{t,\ell_4'}\right)\right]. \tag*{(52)}
\end{aligned} Λ ( 3 , 2 ) , p ( V ) = k ∑ [ t ∑ v k , t H R k − 1 ( ℓ 4 = − L 4 ∑ L 4 v k , t − ℓ 4 w t , ℓ 4 ) + t ∑ ( ℓ 4 = − L 4 ∑ L 4 w t , ℓ 4 v k , t − ℓ 4 H ) R k − 1 v k , t − ( ℓ 4 = − L 4 ∑ L 4 w t , ℓ 4 v k , t − ℓ 4 H ) R k − 1 ℓ 4 ′ = − L 4 ∑ L 4 v k , t − ℓ 4 ′ w t , ℓ 4 ′ . (52)
We then rewrite Λ ( 3 , 2 ) , p ( V ) \Lambda_{(3,2),p}(\mathbf{V}) Λ ( 3 , 2 ) , p ( V ) in (52) in a quadratic form as
Λ ( 3 , 2 ) , p ( V ) = ∑ k [ v k , 1 H v k , 2 H ⋯ v k , T H ] A k [ v k , 1 T v k , 2 T ⋯ v k , T T ] T , (53) \Lambda_{(3,2),p}(\mathbf{V})=\sum_k\begin{bmatrix}\mathbf{v}_{k,1}^{H}&\mathbf{v}_{k,2}^{H}&\cdots&\mathbf{v}_{k,T}^{H}\end{bmatrix}\mathbf{A}_k\begin{bmatrix}\mathbf{v}_{k,1}^{T}&\mathbf{v}_{k,2}^{T}&\cdots&\mathbf{v}_{k,T}^{T}\end{bmatrix}^{T}, \tag*{(53)} Λ ( 3 , 2 ) , p ( V ) = k ∑ [ v k , 1 H v k , 2 H ⋯ v k , T H ] A k [ v k , 1 T v k , 2 T ⋯ v k , T T ] T , (53)
where A k \mathbf{A}_k A k is a T N R TN_R T N R by T N R TN_R T N R block matrix whose block ( r ′ , s ′ ) (r',s') ( r ′ , s ′ ) is an N R N_R N R by N R N_R N R matrix denoted by A r ′ , s ′ , k \mathbf{A}_{r',s',k} A r ′ , s ′ , k , for 1 ≤ r ′ ≤ T 1\le r'\le T 1 ≤ r ′ ≤ T and 1 ≤ s ′ ≤ T 1\le s'\le T 1 ≤ s ′ ≤ T . We calculate the blocks as
A r ′ s ′ , k = z r ′ s ′ R k − 1 , (54) \mathbf{A}_{r's',k}=z_{r's'}\mathbf{R}_k^{-1}, \tag*{(54)} A r ′ s ′ , k = z r ′ s ′ R k − 1 , (54)
where the scalar z r ′ s ′ z_{r's'} z r ′ s ′ is defined as
z r ′ s ′ = w r ′ , r ′ − s ′ + w s ′ , s ′ − r ′ − ∑ t w t , t − r ′ w t , t − s ′ . (55) z_{r's'}=w_{r',r'-s'}+w_{s',s'-r'}-\sum_t w_{t,t-r'}w_{t,t-s'}. \tag*{(55)} z r ′ s ′ = w r ′ , r ′ − s ′ + w s ′ , s ′ − r ′ − t ∑ w t , t − r ′ w t , t − s ′ . (55)
We now formalize our performance analysis for our proposed detector, Λ ( 3 , 2 ) , p \Lambda_{(3,2),p} Λ ( 3 , 2 ) , p , in the following theorem.
Theorem 5. The detector Λ ( 3 , 2 ) , p ( V ) \Lambda_{(3,2),p}(\mathbf{V}) Λ ( 3 , 2 ) , p ( V ) in (53) under H 1 \mathcal{H}_1 H 1 and H 0 \mathcal{H}_0 H 0 is respectively a linear combination of non-central chi-squares and chi-squares, as
Λ ( 3 , 2 ) , p ∣ H 1 = ∑ i ′ = 1 T N R λ i ′ χ 2 K , i ′ 2 ( ∑ k { Re { μ i ′ , k } 2 + Im { μ i ′ , k } 2 } ) , (56) \Lambda_{(3,2),p}\mid\mathcal{H}_1=\sum_{i'=1}^{TN_R}\lambda_{i'}\chi_{2K,i'}^2\left(\sum_k\left\{\operatorname{Re}\{\mu_{i',k}\}^2+\operatorname{Im}\{\mu_{i',k}\}^2\right\}\right), \tag*{(56)} Λ ( 3 , 2 ) , p ∣ H 1 = i ′ = 1 ∑ T N R λ i ′ χ 2 K , i ′ 2 ( k ∑ { Re { μ i ′ , k } 2 + Im { μ i ′ , k } 2 } ) , (56) Λ ( 3 , 2 ) , p ∣ H 0 = ∑ i ′ = 1 T N R λ i ′ χ 2 K , i ′ 2 , (57) \Lambda_{(3,2),p}\mid\mathcal{H}_0=\sum_{i'=1}^{TN_R}\lambda_{i'}\chi_{2K,i'}^2, \tag*{(57)} Λ ( 3 , 2 ) , p ∣ H 0 = i ′ = 1 ∑ T N R λ i ′ χ 2 K , i ′ 2 , (57) where λ i ′ \lambda_{i'} λ i ′ and μ i ′ , k \mu_{i',k} μ i ′ , k are defined above and below (59) respectively.
Proof. Define b ~ k , t = R k − 1 2 α k , t + v ~ k , t \widetilde{\mathbf{b}}_{k,t}=\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,t}+\widetilde{\mathbf{v}}_{k,t} b k , t = R k − 2 1 α k , t + v k , t and matrix G ~ \widetilde{\mathbf{G}} G as G ~ = R ~ k A k R ~ k \widetilde{\mathbf{G}}=\widetilde{\mathbf{R}}_k\mathbf{A}_k\widetilde{\mathbf{R}}_k G = R k A k R k with R ~ k \widetilde{\mathbf{R}}_k R k being the T N R TN_R T N R by T N R TN_R T N R block diagonal matrix, R ~ k = blockdiag { R k − 1 2 , … , R k − 1 2 } \widetilde{\mathbf{R}}_k=\operatorname{blockdiag}\{\mathbf{R}_k^{-\frac{1}{2}},\ldots,\mathbf{R}_k^{-\frac{1}{2}}\} R k = blockdiag { R k − 2 1 , … , R k − 2 1 } . As in previous proofs, we substitute α k , t + R k 1 2 v ~ k , t \boldsymbol{\alpha}_{k,t}+\mathbf{R}_k^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v k , t for v k , t \mathbf{v}_{k,t} v k , t in (53) to obtain Λ ( 3 , 2 ) , p ∣ H 1 \Lambda_{(3,2),p}\mid\mathcal{H}_1 Λ ( 3 , 2 ) , p ∣ H 1 as
Λ ( 3 , 2 ) , p ∣ H 1 = ∑ k [ b ~ k , 1 H b ~ k , 2 H ⋯ b ~ k , T H ] G ~ [ b ~ k , 1 T b ~ k , 2 T ⋯ b ~ k , T T ] T . (58) \Lambda_{(3,2),p}\mid\mathcal{H}_1=\sum_k\begin{bmatrix}\widetilde{\mathbf{b}}_{k,1}^{H}&\widetilde{\mathbf{b}}_{k,2}^{H}&\cdots&\widetilde{\mathbf{b}}_{k,T}^{H}\end{bmatrix}\widetilde{\mathbf{G}}\begin{bmatrix}\widetilde{\mathbf{b}}_{k,1}^{T}&\widetilde{\mathbf{b}}_{k,2}^{T}&\cdots&\widetilde{\mathbf{b}}_{k,T}^{T}\end{bmatrix}^{T}. \tag*{(58)} Λ ( 3 , 2 ) , p ∣ H 1 = k ∑ [ b k , 1 H b k , 2 H ⋯ b k , T H ] G [ b k , 1 T b k , 2 T ⋯ b k , T T ] T . (58)
Matrix G ~ \widetilde{\mathbf{G}} G does not depend on k k k since, referring to (54), it can be simplified as G ~ = Z ′ ⊗ I N R \widetilde{\mathbf{G}}=\mathbf{Z}'\otimes\mathbf{I}_{N_R} G = Z ′ ⊗ I N R , where Z ′ \mathbf{Z}' Z ′ denote a T T T by T T T matrix including z r ′ s ′ z_{r's'} z r ′ s ′ . Also ⊗ \otimes ⊗ indicates the Kronecker product. We replace matrix G ~ \widetilde{\mathbf{G}} G in (58) by its eigenvalue decomposition Φ ~ Δ ~ Φ ~ H \widetilde{\mathbf{\Phi}}\widetilde{\mathbf{\Delta}}\widetilde{\mathbf{\Phi}}^{H} Φ Δ Φ H , where Φ ~ \widetilde{\mathbf{\Phi}} Φ is a unitary matrix and Δ ~ \widetilde{\mathbf{\Delta}} Δ is a diagonal matrix containing the eigenvalues, λ i ′ \lambda_{i'} λ i ′ , where i ′ = 1 , 2 , … , T N R i'=1,2,\ldots,TN_R i ′ = 1 , 2 , … , T N R . Define
μ k = Φ ~ H [ ( R k − 1 2 α k , 1 ) T ( R k − 1 2 α k , 2 ) T ⋯ ( R k − 1 2 α k , T ) T ] T , (59) \boldsymbol{\mu}_k=\widetilde{\mathbf{\Phi}}^{H}\begin{bmatrix}\left(\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,1}\right)^{T}&\left(\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,2}\right)^{T}&\cdots&\left(\mathbf{R}_k^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,T}\right)^{T}\end{bmatrix}^{T}, \tag*{(59)} μ k = Φ H [ ( R k − 2 1 α k , 1 ) T ( R k − 2 1 α k , 2 ) T ⋯ ( R k − 2 1 α k , T ) T ] T , (59)
and c k = Φ ~ H [ v ~ k , 1 T v ~ k , 2 T ⋯ v ~ k , T T ] T \mathbf{c}_k=\widetilde{\mathbf{\Phi}}^{H}\begin{bmatrix}\widetilde{\mathbf{v}}_{k,1}^{T}&\widetilde{\mathbf{v}}_{k,2}^{T}&\cdots&\widetilde{\mathbf{v}}_{k,T}^{T}\end{bmatrix}^{T} c k = Φ H [ v k , 1 T v k , 2 T ⋯ v k , T T ] T . Note that c k ∼ C N ( 0 , I ) \mathbf{c}_k\sim\mathcal{CN}(\mathbf{0},\mathbf{I}) c k ∼ C N ( 0 , I ) since Φ ~ \widetilde{\mathbf{\Phi}} Φ is a unitary matrix. Also, denote the i ′ i' i ′ -th elements of μ k \boldsymbol{\mu}_k μ k and c k \mathbf{c}_k c k by μ i ′ , k \mu_{i',k} μ i ′ , k and c i ′ , k c_{i',k} c i ′ , k respectively. Thus, (58) is given
Λ ( 3 , 2 ) , p ∣ H 1 = ∑ k ( μ k H + c k H ) Δ ~ ( μ k + c k ) = ∑ i ′ = 1 T N R λ i ′ ∑ k ∣ μ i ′ , k + c i ′ , k ∣ 2 . (60) \Lambda_{(3,2),p}\mid\mathcal{H}_1=\sum_k\left(\boldsymbol{\mu}_k^{H}+\mathbf{c}_k^{H}\right)\widetilde{\mathbf{\Delta}}\left(\boldsymbol{\mu}_k+\mathbf{c}_k\right)=\sum_{i'=1}^{TN_R}\lambda_{i'}\sum_k\left|\mu_{i',k}+c_{i',k}\right|^2. \tag*{(60)} Λ ( 3 , 2 ) , p ∣ H 1 = k ∑ ( μ k H + c k H ) Δ ( μ k + c k ) = i ′ = 1 ∑ T N R λ i ′ k ∑ ∣ μ i ′ , k + c i ′ , k ∣ 2 . (60)
Finally, the expression in (60) is a linear combination of chi-squares with 2 K 2K 2 K DoFs and non-centrality parameter ∑ k { Re { μ i ′ , k } 2 + Im { μ i ′ , k } 2 } \sum_k\left\{\operatorname{Re}\{\mu_{i',k}\}^2+\operatorname{Im}\{\mu_{i',k}\}^2\right\} ∑ k { Re { μ i ′ , k } 2 + Im { μ i ′ , k } 2 } , resulting in (56). To calculate Λ ( 3 , 2 ) , p ∣ H 0 \Lambda_{(3,2),p}\mid\mathcal{H}_0 Λ ( 3 , 2 ) , p ∣ H 0 , note that there is no signal component under null hypothesis H 0 \mathcal{H}_0 H 0 . Thus, we replace v k , t \mathbf{v}_{k,t} v k , t in (53) with R k 1 2 v ~ k , t \mathbf{R}_k^{\frac{1}{2}}\widetilde{\mathbf{v}}_{k,t} R k 2 1 v k , t and repeat the same process, giving the expression in (57). □ \square □
Analysis for a Signal Without Slow Variations
We now proceed to derive detectors and provide our analysis for a signal without slow variations, considering frequency-dependent noise correlation. These are the cases where the signal is constant or rapidly varying in a domain, i.e., Case (1,1), Case (1,3), Case (3,1) and Case (3,3). For these detectors, the subscript p p p is omitted, as we derive them based on the MLE method, rather than our proposed approach for signals with slow variations.
Analysis for Case (1,1): Constant in Both Domains
This is the simplest case where the signal is constant in both frequency and time domains so that we can drop the indices k k k and t t t from α k , t \alpha_{k,t} α k , t , i.e., α k , t = α \alpha_{k,t}=\alpha α k , t = α . The LLR of (2) for this case is written as
Λ ( 1 , 1 ) ( V ) = − α H ( T ∑ k R k − 1 ) α + 2 Re { α H ∑ k ( R k − 1 ∑ t v k , t ) } . (61) \Lambda_{(1,1)}(\mathbf{V})=-\alpha^{H}\left(T\sum_{k}\mathbf{R}_{k}^{-1}\right)\alpha+2\operatorname{Re}\left\{\alpha^{H}\sum_{k}\left(\mathbf{R}_{k}^{-1}\sum_{t}\mathbf{v}_{k,t}\right)\right\}. \tag*{(61)} Λ ( 1 , 1 ) ( V ) = − α H ( T k ∑ R k − 1 ) α + 2 Re { α H k ∑ ( R k − 1 t ∑ v k , t ) } . (61)
As the signal α \alpha α is unknown, referring to (3) , we use the MLE of α \alpha α which is denoted by α ∗ \alpha^{*} α ∗ . The expression in (61) has a quadratic structure. Thus, we compute the maximizer of (61) as α ∗ = ( T ∑ k R k − 1 ) − 1 ∑ k ( R k − 1 ∑ t v k , t ) \alpha^{*}=\left(T\sum_{k}\mathbf{R}_{k}^{-1}\right)^{-1}\sum_{k}\left(\mathbf{R}_{k}^{-1}\sum_{t}\mathbf{v}_{k,t}\right) α ∗ = ( T ∑ k R k − 1 ) − 1 ∑ k ( R k − 1 ∑ t v k , t ) , giving the detector for Case ( 1 , 1 ) (1,1) ( 1 , 1 ) as
Λ ( 1 , 1 ) ∗ ( V ) = ( ∑ k ∑ t v k , t H R k − 1 ) G ( ∑ k R k − 1 ∑ t v k , t ) , (62) \Lambda_{(1,1)}^{*}(\mathbf{V})=\left(\sum_{k}\sum_{t}\mathbf{v}_{k,t}^{H}\mathbf{R}_{k}^{-1}\right)\mathbf{G}\left(\sum_{k}\mathbf{R}_{k}^{-1}\sum_{t}\mathbf{v}_{k,t}\right), \tag*{(62)} Λ ( 1 , 1 ) ∗ ( V ) = ( k ∑ t ∑ v k , t H R k − 1 ) G ( k ∑ R k − 1 t ∑ v k , t ) , (62)
where G = ( T ∑ k R k − 1 ) − 1 \mathbf{G}=\left(T\sum_{k}\mathbf{R}_{k}^{-1}\right)^{-1} G = ( T ∑ k R k − 1 ) − 1 . Next, we provide analysis for Case ( 1 , 1 ) (1,1) ( 1 , 1 ) . The detector for this case can be used for time and frequency varying signals, hence we conduct analysis for a general signal model indexed by both k k k and t t t as α k , t \alpha_{k,t} α k , t .
As in Section III, under H 1 \mathcal{H}_{1} H 1 , we replace v k , t \mathbf{v}_{k,t} v k , t in (62) with α k , t + R k 1 2 v ~ k , t \alpha_{k,t}+\mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v ~ k , t which gives
Λ ( 1 , 1 ) ∗ ∣ H 1 = [ ∑ k ( ∑ t α k , t H ) R k − 1 + a H ] G [ ∑ k ( R k − 1 ∑ t α k , t ) + a ] , (63) \Lambda_{(1,1)}^{*} \mid\mathcal{H}_{1}=\left[\sum_{k}\left(\sum_{t}\alpha_{k,t}^{H}\right)\mathbf{R}_{k}^{-1}+\mathbf{a}^{H}\right]\mathbf{G}\left[\sum_{k}\left(\mathbf{R}_{k}^{-1}\sum_{t}\alpha_{k,t}\right)+\mathbf{a}\right], \tag*{(63)} Λ ( 1 , 1 ) ∗ ∣ H 1 = [ k ∑ ( t ∑ α k , t H ) R k − 1 + a H ] G [ k ∑ ( R k − 1 t ∑ α k , t ) + a ] , (63)
where a = ∑ k R k − 1 2 ∑ t v ~ k , t \mathbf{a}=\sum_{k}\mathbf{R}_{k}^{-\frac{1}{2}}\sum_{t}\tilde{\mathbf{v}}_{k,t} a = ∑ k R k − 2 1 ∑ t v ~ k , t . Note that a \mathbf{a} a is a zero-mean Gaussian vector with covariance matrix T ∑ k R k − 1 T\sum_{k}\mathbf{R}_{k}^{-1} T ∑ k R k − 1 which is equal to G − 1 \mathbf{G}^{-1} G − 1 . Therefore we can rewrite a \mathbf{a} a as a = G − 1 2 w ~ \mathbf{a}=\mathbf{G}^{-\frac{1}{2}}\tilde{\mathbf{w}} a = G − 2 1 w ~ where w ~ ∼ C N ( 0 , I ) \tilde{\mathbf{w}}\sim\mathcal{CN}(0,\mathbf{I}) w ~ ∼ C N ( 0 , I ) . Also we can factor G \mathbf{G} G in (63) as G 1 2 G 1 2 \mathbf{G}^{\frac{1}{2}}\mathbf{G}^{\frac{1}{2}} G 2 1 G 2 1 to rewrite (63) as
Λ ( 1 , 1 ) ∗ ∣ H 1 = ( x ^ + w ~ ) H ( x ^ + w ~ ) , (64) \Lambda_{(1,1)}^{*}\mid\mathcal{H}_{1}=(\hat{\mathbf{x}}+\tilde{\mathbf{w}})^{H}(\hat{\mathbf{x}}+\tilde{\mathbf{w}}), \tag*{(64)} Λ ( 1 , 1 ) ∗ ∣ H 1 = ( x ^ + w ~ ) H ( x ^ + w ~ ) , (64)
where x ^ = G 1 2 ∑ k ( R k − 1 ∑ t α k , t ) \hat{\mathbf{x}}=\mathbf{G}^{\frac{1}{2}}\sum_{k}(\mathbf{R}_{k}^{-1}\sum_{t}\alpha_{k,t}) x ^ = G 2 1 ∑ k ( R k − 1 ∑ t α k , t ) . As w ~ ∼ C N ( 0 , I ) \tilde{\mathbf{w}}\sim\mathcal{CN}(0,\mathbf{I}) w ~ ∼ C N ( 0 , I ) , the expression in (64) is a non-central chi-square. By defining x ^ j \hat{x}_{j} x ^ j as the j j j -th element of x ^ \hat{\mathbf{x}} x ^ , the detector Λ ( 1 , 1 ) ∗ ( V ) \Lambda_{(1,1)}^{*}(\mathbf{V}) Λ ( 1 , 1 ) ∗ ( V ) under H 1 \mathcal{H}_{1} H 1 is a non-central chi-square as
Λ ( 1 , 1 ) ∗ ∣ H 1 = χ 2 N R 2 ( ∑ j = 1 N R ( Re { x ^ j } 2 + Im { x ^ j } 2 ) ) . (65) \Lambda_{(1,1)}^{*}\mid\mathcal{H}_{1}=\chi_{2N_{R}}^{2}\left(\sum_{j=1}^{N_{R}}\left(\operatorname{Re}\{\hat{x}_{j}\}^{2}+\operatorname{Im}\{\hat{x}_{j}\}^{2}\right)\right). \tag*{(65)} Λ ( 1 , 1 ) ∗ ∣ H 1 = χ 2 N R 2 ( j = 1 ∑ N R ( Re { x ^ j } 2 + Im { x ^ j } 2 ) ) . (65)
For the null hypothesis there is no signal component. Thus, to calculate Λ ( 1 , 1 ) ∗ ∣ H 0 \Lambda_{(1,1)}^{*}\mid\mathcal{H}_{0} Λ ( 1 , 1 ) ∗ ∣ H 0 we replace v k , t \mathbf{v}_{k,t} v k , t in (62) with R k 1 2 v ~ k , t \mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t and repeat the above process, resulting in a chi-square as
Λ ( 1 , 1 ) ∗ ∣ H 0 = χ 2 N R 2 . (66) \Lambda_{(1,1)}^{*}\mid\mathcal{H}_{0}=\chi_{2N_{R}}^{2}. \tag*{(66)} Λ ( 1 , 1 ) ∗ ∣ H 0 = χ 2 N R 2 . (66)
Analysis for Case (1,3): Constant in Frequency and Rapidly Varying in Time
We drop k k k from α k , t \alpha_{k,t} α k , t to have a constant signal in the frequency domain but varying in time. Define β = [ α 1 T α 2 T … α T T ] T \boldsymbol{\beta}=[\alpha_{1}^{T}\ \alpha_{2}^{T}\ \ldots\ \alpha_{T}^{T}]^{T} β = [ α 1 T α 2 T … α T T ] T , M \mathbf{M} M as a block diagonal matrix whose all diagonal blocks are ∑ k R k − 1 \sum_{k}\mathbf{R}_{k}^{-1} ∑ k R k − 1 or M = blockdiag { ∑ k R k − 1 , … , ∑ k R k − 1 } \mathbf{M}=\operatorname{blockdiag}\{\sum_{k}\mathbf{R}_{k}^{-1},\ldots,\sum_{k}\mathbf{R}_{k}^{-1}\} M = blockdiag { ∑ k R k − 1 , … , ∑ k R k − 1 } , and u = [ ( ∑ k R k − 1 v k , 1 ) T ( ∑ k R k − 1 v k , 2 ) T … ( ∑ k R k − 1 v k , T ) T ] T \mathbf{u}=[(\sum_{k}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,1})^{T}\ (\sum_{k}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,2})^{T}\ldots(\sum_{k}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,T})^{T}]^{T} u = [( ∑ k R k − 1 v k , 1 ) T ( ∑ k R k − 1 v k , 2 ) T … ( ∑ k R k − 1 v k , T ) T ] T . Hence, the LLR of (2) is given as
Λ ( 1 , 3 ) ( V ) = 2 Re { ∑ t α t H ∑ k R k − 1 v k , t − ∑ t α t H ( ∑ k R k − 1 ) α t } = 2 Re { β H u } − β H M β . (67) \Lambda_{(1,3)}(\mathbf{V})=2\operatorname{Re}\left\{\sum_{t}\alpha_{t}^{H}\sum_{k}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,t}-\sum_{t}\alpha_{t}^{H}\left(\sum_{k}\mathbf{R}_{k}^{-1}\right)\alpha_{t}\right\}
=2\operatorname{Re}\{\boldsymbol{\beta}^{H}\mathbf{u}\}-\boldsymbol{\beta}^{H}\mathbf{M}\boldsymbol{\beta}. \tag*{(67)} Λ ( 1 , 3 ) ( V ) = 2 Re { t ∑ α t H k ∑ R k − 1 v k , t − t ∑ α t H ( k ∑ R k − 1 ) α t } = 2 Re { β H u } − β H M β . (67)
Since the signal is unknown we use the MLE of β \boldsymbol{\beta} β as β ∗ = M − 1 u \boldsymbol{\beta}^{*}=\mathbf{M}^{-1}\mathbf{u} β ∗ = M − 1 u , which is equivalent to replacing the MLE of α t \alpha_{t} α t in (3) . This gives the detector as
Λ ( 1 , 3 ) ∗ ( V ) = u H M − 1 u = ∑ t ( ∑ k v k , t H R k − 1 ) S ( ∑ k R k − 1 v k , t ) , (68) \Lambda_{(1,3)}^{*}(\mathbf{V})=\mathbf{u}^{H}\mathbf{M}^{-1}\mathbf{u}=\sum_{t}\left(\sum_{k}\mathbf{v}_{k,t}^{H}\mathbf{R}_{k}^{-1}\right)\mathbf{S}\left(\sum_{k}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,t}\right), \tag*{(68)} Λ ( 1 , 3 ) ∗ ( V ) = u H M − 1 u = t ∑ ( k ∑ v k , t H R k − 1 ) S ( k ∑ R k − 1 v k , t ) , (68)
where S = ( ∑ k R k − 1 ) − 1 \mathbf{S}=\left(\sum_{k}\mathbf{R}_{k}^{-1}\right)^{-1} S = ( ∑ k R k − 1 ) − 1 . We now present our analysis for Case ( 1 , 3 ) (1,3) ( 1 , 3 ) .
The detector can be used for frequency-varying signals, thus we perform the analysis for a general signal model indexed by both k k k and t t t as α k , t \alpha_{k,t} α k , t . Similar to previous analysis under H 1 \mathcal{H}_{1} H 1 , (68) is rewritten as
Λ ( 1 , 3 ) ∗ ∣ H 1 = ∑ t ( ∑ k α k , t H R k − 1 + f t H ) S ( ∑ k R k − 1 α k , t + f t ) , (69) \Lambda_{(1,3)}^{*}\mid\mathcal{H}_{1}=\sum_{t}\left(\sum_{k}\alpha_{k,t}^{H}\mathbf{R}_{k}^{-1}+\mathbf{f}_{t}^{H}\right)\mathbf{S}\left(\sum_{k}\mathbf{R}_{k}^{-1}\alpha_{k,t}+\mathbf{f}_{t}\right), \tag*{(69)} Λ ( 1 , 3 ) ∗ ∣ H 1 = t ∑ ( k ∑ α k , t H R k − 1 + f t H ) S ( k ∑ R k − 1 α k , t + f t ) , (69)
where f t = ∑ k R k − 1 2 v ~ k , t \mathbf{f}_{t}=\sum_{k}\mathbf{R}_{k}^{-\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} f t = ∑ k R k − 2 1 v ~ k , t and thus f t \mathbf{f}_{t} f t is a zero-mean Gaussian vector with covariance matrix S − 1 \mathbf{S}^{-1} S − 1 . Replacing f t \mathbf{f}_{t} f t with S − 1 2 w t \mathbf{S}^{-\frac{1}{2}}\mathbf{w}_{t} S − 2 1 w t where w t ∼ C N ( 0 , I ) \mathbf{w}_{t}\sim\mathcal{CN}(0,\mathbf{I}) w t ∼ C N ( 0 , I ) and replacing S \mathbf{S} S in (69) with S 1 2 S 1 2 \mathbf{S}^{\frac{1}{2}}\mathbf{S}^{\frac{1}{2}} S 2 1 S 2 1 we rewrite (69) as
Λ ( 1 , 3 ) ∗ ∣ H 1 = ∑ t ( q t + w t ) H ( q t + w t ) , (70) \Lambda_{(1,3)}^{*}\mid\mathcal{H}_{1}=\sum_{t}(\mathbf{q}_{t}+\mathbf{w}_{t})^{H}(\mathbf{q}_{t}+\mathbf{w}_{t}), \tag*{(70)} Λ ( 1 , 3 ) ∗ ∣ H 1 = t ∑ ( q t + w t ) H ( q t + w t ) , (70)
where q t = S 1 2 ∑ k R k − 1 α k , t \mathbf{q}_{t}=\mathbf{S}^{\frac{1}{2}}\sum_{k}\mathbf{R}_{k}^{-1}\alpha_{k,t} q t = S 2 1 ∑ k R k − 1 α k , t , whose j j j -th element is q j , t q_{j,t} q j , t . The n n n -th term in (70) is a non-central chi-square χ 2 N R 2 ( ∑ j = 1 N R { Re { q j , t } 2 + Im { q j , t } 2 } ) \chi_{2N_{R}}^{2}\left(\sum_{j=1}^{N_{R}}\{\operatorname{Re}\{q_{j,t}\}^{2}+\operatorname{Im}\{q_{j,t}\}^{2}\}\right) χ 2 N R 2 ( ∑ j = 1 N R { Re { q j , t } 2 + Im { q j , t } 2 } ) and therefore (70) is an addition of non-central chi-squares which is also a non-central chi-square as
Λ ( 1 , 3 ) ∗ ∣ H 1 = χ 2 N R T 2 ( ∑ t ∑ j = 1 N R { Re { q j , t } 2 + Im { q j , t } 2 } ) . (71) \Lambda_{(1,3)}^{*}\mid\mathcal{H}_{1}=\chi_{2N_{R}T}^{2}\left(\sum_{t}\sum_{j=1}^{N_{R}}\left\{\operatorname{Re}\{q_{j,t}\}^{2}+\operatorname{Im}\{q_{j,t}\}^{2}\right\}\right). \tag*{(71)} Λ ( 1 , 3 ) ∗ ∣ H 1 = χ 2 N R T 2 ( t ∑ j = 1 ∑ N R { Re { q j , t } 2 + Im { q j , t } 2 } ) . (71)
Following the same approach and replacing v k , t \mathbf{v}_{k,t} v k , t in (68) with R k 1 2 v ~ k , t \mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t , the detector Λ ( 1 , 3 ) ∗ ( V ) \Lambda_{(1,3)}^{*}(\mathbf{V}) Λ ( 1 , 3 ) ∗ ( V ) under H 0 \mathcal{H}_{0} H 0 is also obtained, which is a chi-square as follows
Λ ( 1 , 3 ) ∗ ∣ H 0 = χ 2 N R T 2 . (72) \Lambda_{(1,3)}^{*}\mid\mathcal{H}_{0}=\chi_{2N_{R}T}^{2}. \tag*{(72)} Λ ( 1 , 3 ) ∗ ∣ H 0 = χ 2 N R T 2 . (72)
Analysis for Case (3,1): Rapidly Varying in Frequency and Constant in Time
For case ( 3 , 1 ) (3,1) ( 3 , 1 ) , the time index t t t is dropped from the signal and only the frequency index is considered, i.e., α k \alpha_{k} α k . Define α ˉ = [ α 1 T α 2 T … α K T ] T \bar{\boldsymbol{\alpha}}=[\alpha_{1}^{T}\ \alpha_{2}^{T}\ldots\alpha_{K}^{T}]^{T} α ˉ = [ α 1 T α 2 T … α K T ] T , M ˉ \bar{\mathbf{M}} M ˉ as a block diagonal matrix with within blocks T R 1 − 1 , T R 2 − 1 , … , T R K − 1 \mathbf{T}\mathbf{R}_{1}^{-1},\mathbf{T}\mathbf{R}_{2}^{-1},\ldots,\mathbf{T}\mathbf{R}_{K}^{-1} T R 1 − 1 , T R 2 − 1 , … , T R K − 1 on its diagonal, i.e., M ˉ = blockdiag { T R 1 − 1 , T R 2 − 1 , … , T R K − 1 } \bar{\mathbf{M}}=\operatorname{blockdiag}\{\mathbf{T}\mathbf{R}_{1}^{-1},\mathbf{T}\mathbf{R}_{2}^{-1},\ldots,\mathbf{T}\mathbf{R}_{K}^{-1}\} M ˉ = blockdiag { T R 1 − 1 , T R 2 − 1 , … , T R K − 1 } , and u ˉ = [ R 1 − 1 ∑ t v 1 , t , R 2 − 1 ∑ t v 2 , t … R K − 1 ∑ t v K , t ] T \bar{\mathbf{u}}=[\mathbf{R}_{1}^{-1}\sum_{t}\mathbf{v}_{1,t},\mathbf{R}_{2}^{-1}\sum_{t}\mathbf{v}_{2,t}\ldots\mathbf{R}_{K}^{-1}\sum_{t}\mathbf{v}_{K,t}]^{T} u ˉ = [ R 1 − 1 ∑ t v 1 , t , R 2 − 1 ∑ t v 2 , t … R K − 1 ∑ t v K , t ] T . Using the LLR of (2) , we write the detector for this case as
Λ ( 3 , 1 ) ( V ) = 2 Re { ∑ k α k H R k − 1 ∑ t v k , t } − T ∑ k α k H R k − 1 α k = 2 Re { α ˉ H u ˉ } − α ˉ H M ˉ α . ˉ (73) \Lambda_{(3,1)}(\mathbf{V})=2\operatorname{Re}\left\{\sum_{k}\alpha_{k}^{H}\mathbf{R}_{k}^{-1}\sum_{t}\mathbf{v}_{k,t}\right\}-T\sum_{k}\alpha_{k}^{H}\mathbf{R}_{k}^{-1}\alpha_{k}
=2\operatorname{Re}\{\bar{\boldsymbol{\alpha}}^{H}\bar{\mathbf{u}}\}-\bar{\boldsymbol{\alpha}}^{H}\bar{\mathbf{M}}\bar{\boldsymbol{\alpha}.} \tag*{(73)} Λ ( 3 , 1 ) ( V ) = 2 Re { k ∑ α k H R k − 1 t ∑ v k , t } − T k ∑ α k H R k − 1 α k = 2 Re { α ˉ H u ˉ } − α ˉ H M ˉ α . ˉ (73)
The signal is unknown thus we use the MLE of α ~ \tilde{\boldsymbol{\alpha}} α ~ as α ~ ∗ = M ~ − 1 u ~ \tilde{\boldsymbol{\alpha}}^{*}=\tilde{\mathbf{M}}^{-1}\tilde{\mathbf{u}} α ~ ∗ = M ~ − 1 u ~ , which is equivalent to substituting the MLE of α k \boldsymbol{\alpha}_{k} α k in (3), resulting in a detector as follows:
Λ ( 3 , 1 ) ∗ ( V ) = u ~ H M ~ − 1 u ~ = 1 T ∑ k ( ∑ t v k , t H R k − 1 ) ( ∑ t v k , t ) . (74) \Lambda^{*}_{(3,1)}(\mathbf{V})=\tilde{\mathbf{u}}^{H}\tilde{\mathbf{M}}^{-1}\tilde{\mathbf{u}}=\frac{1}{T}\sum_{k}\left(\sum_{t}\mathbf{v}_{k,t}^{H}\mathbf{R}_{k}^{-1}\right)\left(\sum_{t}\mathbf{v}_{k,t}\right). \tag*{(74)} Λ ( 3 , 1 ) ∗ ( V ) = u ~ H M ~ − 1 u ~ = T 1 k ∑ ( t ∑ v k , t H R k − 1 ) ( t ∑ v k , t ) . (74)
We now proceed to provide analysis for the detector for Case (3,1). Note that since the detector Λ ( 3 , 1 ) ∗ ( V ) \Lambda^{*}_{(3,1)}(\mathbf{V}) Λ ( 3 , 1 ) ∗ ( V ) can be used in a general detection scenario where the signal is varying in time, we use both time and frequency indices for the signal in our analysis. Hence, similar to previous analysis, the detector under H 1 \mathcal{H}_{1} H 1 is obtained as
Λ ( 3 , 1 ) ∗ ∣ H 1 = ∑ k [ 1 T ∑ t α k , t H R k − 1 2 + 1 T ∑ t v ~ k , t H ] [ 1 T R k − 1 2 ∑ t α k , t + 1 T ∑ t v ~ k , t ] . (75) \Lambda^{*}_{(3,1)}\mid\mathcal{H}_{1}=\sum_{k}\left[\frac{1}{\sqrt{T}}\sum_{t}\boldsymbol{\alpha}_{k,t}^{H}\mathbf{R}_{k}^{-\frac{1}{2}}+\frac{1}{\sqrt{T}}\sum_{t}\tilde{\mathbf{v}}_{k,t}^{H}\right]\left[\frac{1}{\sqrt{T}}\mathbf{R}_{k}^{-\frac{1}{2}}\sum_{t}\boldsymbol{\alpha}_{k,t}+\frac{1}{\sqrt{T}}\sum_{t}\tilde{\mathbf{v}}_{k,t}\right]. \tag*{(75)} Λ ( 3 , 1 ) ∗ ∣ H 1 = k ∑ [ T 1 t ∑ α k , t H R k − 2 1 + T 1 t ∑ v ~ k , t H ] [ T 1 R k − 2 1 t ∑ α k , t + T 1 t ∑ v ~ k , t ] . (75)
We rewrite (75) as
Λ ( 3 , 1 ) ∗ ∣ H 1 = ∑ k ( ℓ k + ℓ ~ k ) H ( ℓ k + ℓ ~ k ) , (76) \Lambda^{*}_{(3,1)}\mid\mathcal{H}_{1}=\sum_{k}(\boldsymbol{\ell}_{k}+\tilde{\boldsymbol{\ell}}_{k})^{H}(\boldsymbol{\ell}_{k}+\tilde{\boldsymbol{\ell}}_{k}), \tag*{(76)} Λ ( 3 , 1 ) ∗ ∣ H 1 = k ∑ ( ℓ k + ℓ ~ k ) H ( ℓ k + ℓ ~ k ) , (76)
where ℓ k = 1 T R k − 1 2 ∑ t α k , t \boldsymbol{\ell}_{k}=\frac{1}{\sqrt{T}}\mathbf{R}_{k}^{-\frac{1}{2}}\sum_{t}\boldsymbol{\alpha}_{k,t} ℓ k = T 1 R k − 2 1 ∑ t α k , t and ℓ ~ k = 1 T ∑ t v ~ k , t \tilde{\boldsymbol{\ell}}_{k}=\frac{1}{\sqrt{T}}\sum_{t}\tilde{\mathbf{v}}_{k,t} ℓ ~ k = T 1 ∑ t v ~ k , t . The k k k -th term in the summation in (76) is a non-central chi-square as χ 2 N R 2 ( ∑ j = 1 N R Re { l j , k } 2 + Im { l j , k } 2 ) \chi^{2}_{2N_{R}}\left(\sum_{j=1}^{N_{R}}\operatorname{Re}\{l_{j,k}\}^{2}+\operatorname{Im}\{l_{j,k}\}^{2}\right) χ 2 N R 2 ( ∑ j = 1 N R Re { l j , k } 2 + Im { l j , k } 2 ) where l j , k l_{j,k} l j , k is the j j j -th component of vector ℓ k \boldsymbol{\ell}_{k} ℓ k . Therefore (76) is an addition of non-central chi-squares which is another non-central chi-square as
Λ ( 3 , 1 ) ∗ ∣ H 1 = χ 2 N R K 2 ( ∑ k ∑ j = 1 N R Re { l j , k } 2 + Im { l j , k } 2 ) . (77) \Lambda^{*}_{(3,1)}\mid\mathcal{H}_{1}=\chi^{2}_{2N_{R}K}\left(\sum_{k}\sum_{j=1}^{N_{R}}\operatorname{Re}\{l_{j,k}\}^{2}+\operatorname{Im}\{l_{j,k}\}^{2}\right). \tag*{(77)} Λ ( 3 , 1 ) ∗ ∣ H 1 = χ 2 N R K 2 ( k ∑ j = 1 ∑ N R Re { l j , k } 2 + Im { l j , k } 2 ) . (77)
To compute Λ ( 3 , 1 ) ∗ ∣ H 0 \Lambda^{*}_{(3,1)}\mid\mathcal{H}_{0} Λ ( 3 , 1 ) ∗ ∣ H 0 , similarly we substitute R k 1 2 v ~ k , t \mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t for v k , t \mathbf{v}_{k,t} v k , t in (74) and repeat a similar process, giving a chi-square as
Λ ( 3 , 1 ) ∗ ∣ H 0 = χ 2 N R K 2 . (78) \Lambda^{*}_{(3,1)}\mid\mathcal{H}_{0}=\chi^{2}_{2N_{R}K}. \tag*{(78)} Λ ( 3 , 1 ) ∗ ∣ H 0 = χ 2 N R K 2 . (78)
Extended Energy Detector for Case (3,3): Rapidly Varying in Both Frequency and Time
We derive a detector for a signal rapidly varying in both frequency and time. The detector that we derive is an extended energy detector for noise with a frequency-dependent covariance matrix as well as both frequency and time samples. Define two vectors
α ^ = [ α 1 , 1 T , α 2 , 1 T … α K , 1 T , α 1 , 2 T , α 2 , 2 T … α K , 2 T … α 1 , T T , α 2 , T T … α K , T T ] T , v ^ = [ v 1 , 1 T , v 2 , 1 T … v K , 1 T , v 1 , 2 T , v 2 , 2 T … v K , 2 T … v 1 , T T , v 2 , T T … v K , T T ] T . (79) \hat{\boldsymbol{\alpha}}=\left[\boldsymbol{\alpha}_{1,1}^{T},\boldsymbol{\alpha}_{2,1}^{T}\ldots\boldsymbol{\alpha}_{K,1}^{T},\boldsymbol{\alpha}_{1,2}^{T},\boldsymbol{\alpha}_{2,2}^{T}\ldots\boldsymbol{\alpha}_{K,2}^{T}\ldots\boldsymbol{\alpha}_{1,T}^{T},\boldsymbol{\alpha}_{2,T}^{T}\ldots\boldsymbol{\alpha}_{K,T}^{T}\right]^{T},
\quad
\hat{\mathbf{v}}=\left[\mathbf{v}_{1,1}^{T},\mathbf{v}_{2,1}^{T}\ldots\mathbf{v}_{K,1}^{T},\mathbf{v}_{1,2}^{T},\mathbf{v}_{2,2}^{T}\ldots\mathbf{v}_{K,2}^{T}\ldots\mathbf{v}_{1,T}^{T},\mathbf{v}_{2,T}^{T}\ldots\mathbf{v}_{K,T}^{T}\right]^{T}. \tag*{(79)} α ^ = [ α 1 , 1 T , α 2 , 1 T … α K , 1 T , α 1 , 2 T , α 2 , 2 T … α K , 2 T … α 1 , T T , α 2 , T T … α K , T T ] T , v ^ = [ v 1 , 1 T , v 2 , 1 T … v K , 1 T , v 1 , 2 T , v 2 , 2 T … v K , 2 T … v 1 , T T , v 2 , T T … v K , T T ] T . (79)
Also define R ^ N R K T × N R K T \hat{\mathbf{R}}_{N_{R}KT\times N_{R}KT} R ^ N R K T × N R K T as a block diagonal matrix where blocks R 1 , R 2 , … , R K \mathbf{R}_{1},\mathbf{R}_{2},\ldots,\mathbf{R}_{K} R 1 , R 2 , … , R K repeat on its diagonal for T T T times, i.e.,
R ^ = blockdiag { R 1 , R 2 , … , R K , … , R 1 , R 2 , … , R K } . (80) \hat{\mathbf{R}}=\operatorname{blockdiag}\{\mathbf{R}_{1},\mathbf{R}_{2},\ldots,\mathbf{R}_{K},\ldots,\mathbf{R}_{1},\mathbf{R}_{2},\ldots,\mathbf{R}_{K}\}. \tag*{(80)} R ^ = blockdiag { R 1 , R 2 , … , R K , … , R 1 , R 2 , … , R K } . (80)
We use the LLR of (2) to get the detector for Case (3,3) as
Λ ( 3 , 3 ) ( V ) = 2 Re { α ^ H R ^ − 1 v ^ } − α ^ H R ^ − 1 α ^ . (81) \Lambda_{(3,3)}(\mathbf{V})=2\operatorname{Re}\{\hat{\boldsymbol{\alpha}}^{H}\hat{\mathbf{R}}^{-1}\hat{\mathbf{v}}\}-\hat{\boldsymbol{\alpha}}^{H}\hat{\mathbf{R}}^{-1}\hat{\boldsymbol{\alpha}}. \tag*{(81)} Λ ( 3 , 3 ) ( V ) = 2 Re { α ^ H R ^ − 1 v ^ } − α ^ H R ^ − 1 α ^ . (81)
Since the signal is unknown we use the MLE of α ^ \hat{\boldsymbol{\alpha}} α ^ which is α ^ ∗ = R ^ R ^ − 1 v ^ = v ^ \hat{\boldsymbol{\alpha}}^{*}=\hat{\mathbf{R}}\hat{\mathbf{R}}^{-1}\hat{\mathbf{v}}=\hat{\mathbf{v}} α ^ ∗ = R ^ R ^ − 1 v ^ = v ^ . This is equivalent to replacing the MLE of α k , t \boldsymbol{\alpha}_{k,t} α k , t into (3). Hence, the detector is written as
Λ ( 3 , 3 ) ∗ ( V ) = v ^ H R ^ − 1 v ^ = ∑ k ∑ t v k , t H R k − 1 v k , t . (82) \Lambda^{*}_{(3,3)}(\mathbf{V})=\hat{\mathbf{v}}^{H}\hat{\mathbf{R}}^{-1}\hat{\mathbf{v}}=\sum_{k}\sum_{t}\mathbf{v}_{k,t}^{H}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,t}. \tag*{(82)} Λ ( 3 , 3 ) ∗ ( V ) = v ^ H R ^ − 1 v ^ = k ∑ t ∑ v k , t H R k − 1 v k , t . (82)
We now present our analysis for Λ ( 3 , 3 ) ∗ ( V ) \Lambda^{*}_{(3,3)}(\mathbf{V}) Λ ( 3 , 3 ) ∗ ( V ) . We write R k − 1 \mathbf{R}_{k}^{-1} R k − 1 in (82) as R k − 1 2 R k − 1 2 \mathbf{R}_{k}^{-\frac{1}{2}}\mathbf{R}_{k}^{-\frac{1}{2}} R k − 2 1 R k − 2 1 . Then, the detector under H 1 \mathcal{H}_{1} H 1 , as in previous analysis, is given as
Λ ( 3 , 3 ) ∗ ∣ H 1 = ∑ t ∑ k ( x k , t + v ~ k , t ) H ( x k , t + v ~ k , t ) , (83) \Lambda^{*}_{(3,3)}\mid\mathcal{H}_{1}=\sum_{t}\sum_{k}(\mathbf{x}_{k,t}+\tilde{\mathbf{v}}_{k,t})^{H}(\mathbf{x}_{k,t}+\tilde{\mathbf{v}}_{k,t}), \tag*{(83)} Λ ( 3 , 3 ) ∗ ∣ H 1 = t ∑ k ∑ ( x k , t + v ~ k , t ) H ( x k , t + v ~ k , t ) , (83)
where x k , t = R k − 1 2 α k , t \mathbf{x}_{k,t}=\mathbf{R}_{k}^{-\frac{1}{2}}\boldsymbol{\alpha}_{k,t} x k , t = R k − 2 1 α k , t is a vector whose j j j -th element is denoted by x j , k , t x_{j,k,t} x j , k , t . For fixed indices k k k and t t t , each term in the summation in (83) is a non-central chi-square χ 2 N R 2 ( ∑ j = 1 N R Re { x j , k , t } 2 + Im { x j , k , t } 2 ) \chi^{2}_{2N_{R}}\left(\sum_{j=1}^{N_{R}}\operatorname{Re}\{x_{j,k,t}\}^{2}+\operatorname{Im}\{x_{j,k,t}\}^{2}\right) χ 2 N R 2 ( ∑ j = 1 N R Re { x j , k , t } 2 + Im { x j , k , t } 2 ) . Thus, (83) is an addition of non-central chi-squares, resulting in
Λ ( 3 , 3 ) ∗ ∣ H 1 = χ 2 N R K T 2 ( ∑ t ∑ k ∑ j = 1 N R Re { x j , k , t } 2 + Im { x j , k , t } 2 ) . (84) \Lambda^{*}_{(3,3)}\mid\mathcal{H}_{1}=\chi^{2}_{2N_{R}KT}\left(\sum_{t}\sum_{k}\sum_{j=1}^{N_{R}}\operatorname{Re}\{x_{j,k,t}\}^{2}+\operatorname{Im}\{x_{j,k,t}\}^{2}\right). \tag*{(84)} Λ ( 3 , 3 ) ∗ ∣ H 1 = χ 2 N R K T 2 ( t ∑ k ∑ j = 1 ∑ N R Re { x j , k , t } 2 + Im { x j , k , t } 2 ) . (84)
Similarly, Λ ( 3 , 3 ) ∗ ∣ H 0 \Lambda^{*}_{(3,3)}\mid\mathcal{H}_{0} Λ ( 3 , 3 ) ∗ ∣ H 0 is obtained by replacing v k , t \mathbf{v}_{k,t} v k , t in (82) with R k 1 2 v ~ k , t \mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t and following a similar process as above. As a result, Λ ( 3 , 3 ) ∗ ( V ) \Lambda^{*}_{(3,3)}(\mathbf{V}) Λ ( 3 , 3 ) ∗ ( V ) under H 0 \mathcal{H}_{0} H 0 is a chi-square as
Λ ( 3 , 3 ) ∗ ∣ H 0 = χ 2 N R K T 2 . (85) \Lambda^{*}_{(3,3)}\mid\mathcal{H}_{0}=\chi^{2}_{2N_{R}KT}. \tag*{(85)} Λ ( 3 , 3 ) ∗ ∣ H 0 = χ 2 N R K T 2 . (85)
Upper Bound on Detection Performance
We have assumed an unknown signal throughout our derivations and analysis thus far. In this section, we derive a detector that gives an upper bound on the target detection performance. To obtain such a detector, we assume that the signal α k , t \boldsymbol{\alpha}_{k,t} α k , t is known. Hence, using the LLR of (2), the upper bound detector is given as Λ U p . B . ( V ) = Re { ∑ k ∑ t α k , t H R k − 1 v k , t } \Lambda_{\mathrm{Up.B.}}(\mathbf{V})=\operatorname{Re}\{\sum_{k}\sum_{t}\boldsymbol{\alpha}_{k,t}^{H}\mathbf{R}_{k}^{-1}\mathbf{v}_{k,t}\} Λ Up.B. ( V ) = Re { ∑ k ∑ t α k , t H R k − 1 v k , t } , where the subscript Up.B. indicates the upper bound detector. To calculate Λ U p . B . ( V ) \Lambda_{\mathrm{Up.B.}}(\mathbf{V}) Λ Up.B. ( V ) under H 1 \mathcal{H}_{1} H 1 , similar to the previous proofs, we replace v k , t \mathbf{v}_{k,t} v k , t in Λ U p . B . ( V ) \Lambda_{\mathrm{Up.B.}}(\mathbf{V}) Λ Up.B. ( V ) with α k , t + R k 1 2 v ~ k , t \boldsymbol{\alpha}_{k,t}+\mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} α k , t + R k 2 1 v ~ k , t where v ~ k , t ∼ C N ( 0 , I ) \tilde{\mathbf{v}}_{k,t}\sim\mathcal{CN}(0,\mathbf{I}) v ~ k , t ∼ C N ( 0 , I ) . This results in
Λ U p . B . ∣ H 1 = m u + Re { ∑ k ∑ t α k , t H R k − 1 2 v ~ k , t } , (86) \Lambda_{\mathrm{Up.B.}}\mid\mathcal{H}_{1}=m_{u}+\operatorname{Re}\left\{\sum_{k}\sum_{t}\boldsymbol{\alpha}_{k,t}^{H}\mathbf{R}_{k}^{-\frac{1}{2}}\tilde{\mathbf{v}}_{k,t}\right\}, \tag*{(86)} Λ Up.B. ∣ H 1 = m u + Re { k ∑ t ∑ α k , t H R k − 2 1 v ~ k , t } , (86)
where m u = Re { ∑ k ∑ t α k , t H R k − 1 α k , t } m_{u}=\operatorname{Re}\{\sum_{k}\sum_{t}\boldsymbol{\alpha}_{k,t}^{H}\mathbf{R}_{k}^{-1}\boldsymbol{\alpha}_{k,t}\} m u = Re { ∑ k ∑ t α k , t H R k − 1 α k , t } is a known scalar. The second term in (86) is a zero-mean real Gaussian variable with variance ν u = 1 2 ∑ k ∑ t α k , t H R k − 1 α k , t \nu_{u}=\frac{1}{2}\sum_{k}\sum_{t}\boldsymbol{\alpha}_{k,t}^{H}\mathbf{R}_{k}^{-1}\boldsymbol{\alpha}_{k,t} ν u = 2 1 ∑ k ∑ t α k , t H R k − 1 α k , t . Hence, Λ U p . B . ∣ H 1 \Lambda_{\mathrm{Up.B.}}\mid\mathcal{H}_{1} Λ Up.B. ∣ H 1 is a real Gaussian variable with mean m u m_{u} m u and variance ν u \nu_{u} ν u . The upper bound detector under H 0 \mathcal{H}_{0} H 0 is similarly calculated by replacing v k , t \mathbf{v}_{k,t} v k , t in Λ U p . B . ( V ) \Lambda_{\mathrm{Up.B.}}(\mathbf{V}) Λ Up.B. ( V ) with R k 1 2 v ~ k , t \mathbf{R}_{k}^{\frac{1}{2}}\tilde{\mathbf{v}}_{k,t} R k 2 1 v ~ k , t . As a result, Λ U p . B . ∣ H 0 \Lambda_{\mathrm{Up.B.}}\mid\mathcal{H}_{0} Λ Up.B. ∣ H 0 is a zero-mean real Gaussian variable with variance ν u \nu_{u} ν u .
Denote by P D , u P_{D,u} P D , u and P F A , u P_{FA,u} P F A , u the probabilities of detection and false alarm for the upper bound detector. Referring to Section II-A, we calculate P D , u P_{D,u} P D , u in terms of P F A , u P_{FA,u} P F A , u as P D , u = Q ( Q − 1 ( P F A , u ) − m u ν u ) P_{D,u}=Q\left(Q^{-1}(P_{FA,u})-\frac{m_{u}}{\sqrt{\nu_{u}}}\right) P D , u = Q ( Q − 1 ( P F A , u ) − ν u m u ) , where Q ( ⋅ ) Q(\cdot) Q ( ⋅ ) indicates the Q-function.
Numerical Results
In this section, we validate our analysis through simulations and present a performance comparison of the detectors. We consider a collinear configuration with either tight coupling (TC) or weak coupling (WC) for the receiver antenna array for which N R = 32 N_{R}=32 N R = 32 . TC and WC respectively refer to large and small mutual coupling among the antennas, which determines the ratio δ a R \frac{\delta}{a_R} a R δ where δ \delta δ is set to 0.5 cm [1 , 13 ] . For arrays with TC, δ a R \frac{\delta}{a_R} a R δ is set to the optimal value in terms of the achievable rate, i.e., ( δ a R ) ∗ \left(\frac{\delta}{a_R}\right)^* ( a R δ ) ∗ which is 1.932, and for arrays with WC δ a R \frac{\delta}{a_R} a R δ is set to twice the optimal value [1 ] . Note that superscript * here indicates the optimal value. The array size is fixed and only a R a_R a R changes for the two coupling cases. Our plots show detection performance for echo signals received by the sensing receiver from the end-fire direction. Note that end-fire gives the largest bandwidth gain for TCAs with a colinear configuration [14 ] . Other directions give reduced performance, with the lowest performance at broadside. All parameters related to the RLC circuit network including the resistance of the source voltage generator, the noise figure, the internal resistance of the low-noise amplifier and its gain are set as in [1 ] . Also, the mutual impedance among the antennas and the self-impedance of the equivalent RLC circuit for the Chu’s CMS antenna are computed using [1 Eq. (5)] and [1 Eq. (7)] respectively. For one time sample, the average received signal-to-noise ratio (SNR) across frequency for K K K frequency samples is denoted by γ ‾ K \overline{\gamma}_K γ K .
We adopt a simple signal model that captures both frequency and time variations via two easily interpretable parameters. This simple model is chosen as the purpose of this paper is to compare the performance of different detectors in different environments rather than to investigate a particular sensing scenario. Also, we assume that the signal, which models the echo signal from the target, has the same bandwidth as the transmit signal. The bandwidth of the signal is restricted to a certain range we specify for our results. The unknown signal is modeled as
α k , t = h k , t g k , t e j ( k − 1 ) υ e j ( t − 1 ) τ , (87) \alpha_{k,t} = h_{k,t}g_{k,t}e^{j(k-1)\upsilon}e^{j(t-1)\tau}, \tag*{(87)} α k , t = h k , t g k , t e j ( k − 1 ) υ e j ( t − 1 ) τ , (87)
where h k , t h_{k,t} h k , t denotes the physically-consistent channel model between the target and the receiver and g k , t g_{k,t} g k , t that is the echo signal from the target is modeled as a source voltage. We consider a physically-consistent line-of-sight (LoS) channel for h k h_k h k given in [1 ] and a fixed source voltage g k , t = g g_{k,t}=g g k , t = g obtained by equal power allocation among the frequency samples for one time sample. Also, in (87) υ \upsilon υ and τ \tau τ control the rate of phase variation across frequency and time respectively. Our model has no amplitude variation except amplitude variation due to the coupling effect caused by h k , t h_{k,t} h k , t , as variation over frequency and time is likely to be more pronounced than amplitude variation especially in a LoS channel.
Increasing υ \upsilon υ induces larger frequency variations, however υ = 0 \upsilon=0 υ = 0 does not necessarily indicate that the signal is fixed in the frequency domain as h k , t h_{k,t} h k , t in (87) is a frequency-dependent channel model. It is also assumed that the samples are collected in a short span of time, which essentially means that h k , t h_{k,t} h k , t is constant in time. Hence, τ = 0 \tau=0 τ = 0 does correspond to a constant signal in the time domain and τ \tau τ larger than 0 causes time variations.
Figs. 1 (a) and 1 (b) show a comparison between our analysis and simulations for the CDF of the detectors under null and alternative hypotheses. In the figure, blue markers are used to show curves corresponding to analysis. Also, red solid lines with the markers correspond to simulations. For the upper bound detector we use the abbreviation Up.B. in figure
Figure 1a.
Figure 1b.
caption: Distributions of the detectors for (a) the null hypothesis and (b) the alternative hypothesis for TC, K = T = 5 K=T=5 K = T = 5 , bandwidth [20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 ] GHz, υ = 0.2 \upsilon=0.2 υ = 0.2 , τ = 0.2 \tau=0.2 τ = 0.2 and γ ‾ K = 1.4 \overline{\gamma}_K=1.4 γ K = 1.4 dB. All window lengths are 3.
legends. For simplicity, K K K and T T T are set to 5. The figure shows that our analytical distributions perfectly match the simulated distributions.
Next, we show the performance of the detectors. In practice, signals are not precisely divided into the 9 possible cases in frequency and time. Thus, the assumptions that lead to the 9 detectors cannot be satisfied exactly and the detectors can be employed in scenarios for which they are not designed. Hence, it is necessary to compare the detectors in terms of robustness to spectral and temporal variation of the signal. Figure 2 (a)-(Figure 2 c) show the probability of detection P D P_D P D for a fixed probability of false alarm P F A = 10 − 6 P_{FA}=10^{-6} P F A = 1 0 − 6 and a large bandwidth of 2 GHz. Values of P D P_D P D are computed as explained in Section II-A. The letter ‘p’ in the legend indicates our proposed detectors. In Figure 2 (a)-(Figure 2 c), the x-axis shows the rate of variation in frequency, i.e. υ \upsilon υ . Figure 2 (a)-(Figure 2 c) each have a different τ \tau τ value to show the impact of the time variation rate, with τ \tau τ set to 0, 0.5 and 1 respectively. Also, we set the total power such that the upper bound detector approaches P D = 1 P_D=1 P D = 1 , ensuring the other detectors are not restricted by limited power. Note that the window lengths in frequency and time for our proposed detectors are fixed across Figure 2 (a)-(Figure 2 c). This corresponds to the worst case performance of our detectors as the performance of our proposed detectors can be improved by optimizing the window lengths based on the specific scenario. We show the impact of the window length later in this section. As seen, even in the worst case, our detectors are robust across different scenarios and show the best overall performance. This means that, as illustrated, other detectors work only in specific scenarios but in all scenarios a suitable proposed detector works equally well or better. The only exception is the detector for Case (3,3) for υ = 1 \upsilon= 1 υ = 1 in Figure 2 (c) that shows higher P D P_D P D , but we emphasize that the figure shows the worst case performance of our detectors. By optimizing the window length, our proposed detectors result in performance superior to that of the detector for Case (3,3). In Figure 2 (a)-(Figure 2 c), detectors that are designed for rapid
Figure 2. Probability of detection VS. rate of variation in frequency for an array with tight coupling (TC) with (a) τ = 0 \tau= 0 τ = 0 (b) τ = 0.5 \tau= 0.5 τ = 0.5 and (c) τ = 1 \tau= 1 τ = 1 , for bandwidth [28 , 29 , 30 ] GHz. Average of γ ˉ K \bar{\gamma}_K γ ˉ K is 3.7 dB, P F A = 10 − 6 P_{FA} = 10^{-6} P F A = 1 0 − 6 , K = 15 K = 15 K = 15 and T = 8 T = 8 T = 8 . All window lengths are 5.
variations in the frequency domain, i.e., detectors for Cases (3,1), (3,2) and (3,3), have constant P D P_D P D . This is because they work similarly to an energy detector in the frequency domain, with phase changes or changing υ \upsilon υ having no effect on their performance. Also, phase changes in the time domain has no impact on the detection performance of the detectors designed for rapid variations in the time domain, that is the detectors for Cases (1,3), (2,3) and (3,3). These detectors show the same performance for all values of τ \tau τ . This is evident by comparing Figure 2 (a)-(Figure 2 c). In addition, for the bandwidth considered, the channel h k , t h_{k,t} h k , t varies slightly in frequency. As a result, the detectors for Cases (1,1), (1,2) and (1,3) perform better for very small υ \upsilon υ compared to larger υ \upsilon υ .
Figure 3 shows detection performance for an array with WC for τ = 0 \tau= 0 τ = 0 . The merit of TC in detection is clear by comparing Figure 3 with Figure 2 (a) plotted for an array with TC. Even though for both the arrays the same total power of the echoes is used, the average SNR over the considered wide bandwidth for the array with TC is larger than that of for the array with WC. The larger SNR value across the bandwidth for arrays with TC is due to the coupling effect as shown in [1 ] , [13 ] . This contributes to better detection performance of the array with TC. In fact, the detection performance for every detector for the array with WC degrades, except for the upper bound detector. For brevity we omit showing results for τ \tau τ larger than 0, since the same observation applies.
Figure 3. Detection performance for an array with weak coupling (WC) for τ = 0 \tau= 0 τ = 0 and bandwidth [28 , 29 , 30 ] GHz. Average of γ ˉ K \bar{\gamma}_K γ ˉ K is 0.8 dB, P F A = 10 − 6 P_{FA} = 10^{-6} P F A = 1 0 − 6 , K = 15 K = 15 K = 15 and T = 8 T = 8 T = 8 . All window lengths are 5.
Figure 4 (a) illustrates the impact of window length on the detection performance of our proposed approach for rapid variations in the time domain with τ = 1 \tau= 1 τ = 1 . To have a concise legend for the figure, we define W L k \mathrm{WL}_k WL k and W L t \mathrm{WL}_t WL t as the window lengths in the frequency and time domains that are related to the window parameters in Section III as L 2 = L 3 = M = ( W L k − 1 ) / 2 L_2 = L_3 = M = (\mathrm{WL}_k - 1)/2 L 2 = L 3 = M = ( WL k − 1 ) /2 and L 1 = L 4 = Q = ( W L t − 1 ) / 2 L_1 = L_4 = Q = (\mathrm{WL}_t - 1)/2 L 1 = L 4 = Q = ( WL t − 1 ) /2 . We also define the pair W L = ( W L k , W L t ) \mathrm{WL} = (\mathrm{WL}_k,\mathrm{WL}_t) WL = ( WL k , WL t ) which is shown in the legend.
Figure 4. Impact of the window lengths for an array with tight coupling (TC). (a) For rapid variation in the time domain with τ = 1 \tau= 1 τ = 1 and average of γ ˉ K \bar{\gamma}_K γ ˉ K is 3.7 dB. (b) For slow variation in the time domain with τ = 0.2 \tau= 0.2 τ = 0.2 and average of γ ˉ K \bar{\gamma}_K γ ˉ K is 0.02 dB. The bandwidth is [28 , 29 , 30 ] GHz, P F A = 10 − 6 P_{FA} = 10^{-6} P F A = 1 0 − 6 , K = 15 K = 15 K = 15 and T = 8 T = 8 T = 8 .
Figure 4 (a) shows that when there is large variation in either frequency or time domains smaller window lengths yield a better performance. The figure is plotted for τ = 1 \tau= 1 τ = 1 indicating rapid variation across time. As can be seen, the curves with W L = ( 5 , 3 ) \mathrm{WL} = (5,3) WL = ( 5 , 3 ) perform better than the curves with W L = ( 5 , 5 ) \mathrm{WL} = (5,5) WL = ( 5 , 5 ) . Also, it can be seen that curves with W L = ( 3 , 5 ) \mathrm{WL} = (3,5) WL = ( 3 , 5 ) shows better performance than the curves with W L = ( 5 , 5 ) \mathrm{WL} = (5,5) WL = ( 5 , 5 ) when υ \upsilon υ is increased.
Figure 4 (b) shows how window length affects the detection performance of our proposed approach for slow variations in the time domain with τ = 0.2 \tau= 0.2 τ = 0.2 . For this figure, a smaller total power is used to show the distinction between the detectors when υ \upsilon υ is small. The figure shows when there is slow variation in either frequency or time domains, larger window lengths contribute to a better performance. With the slow variation in time corresponding to τ = 0.2 \tau= 0.2 τ = 0.2 , it can be seen that the curves with W L = ( 5 , 7 ) \mathrm{WL} = (5, 7) WL = ( 5 , 7 ) show superior performance compared to the curves with W L = ( 5 , 5 ) \mathrm{WL} = (5, 5) WL = ( 5 , 5 ) . In addition, it is evident from the figure that for small υ \upsilon υ , the curves for W L = ( 7 , 5 ) \mathrm{WL} = (7, 5) WL = ( 7 , 5 ) show better performance compared to the curves for W L = ( 5 , 5 ) \mathrm{WL} = (5, 5) WL = ( 5 , 5 ) . When the signal is slowly varying, increasing the window lengths results in better averaging and consequently better estimation of the signal. As seen, even a small change in W L k \mathrm{WL}_k WL k and W L t \mathrm{WL}_t WL t has a notable impact on detection performance. Of course, W L k \mathrm{WL}_k WL k and W L t \mathrm{WL}_t WL t can be optimized to make our detectors work very efficiently, however this is left for future work.
Conclusion
We conducted joint time and frequency target detection assuming unknown wideband signals and using TCAs. We considered correlated noise with a frequency-dependent covariance matrix. We provided detectors and analysis for signals with various temporal/spectral dynamics and proposed a novel detection approach for slowly-varying signals. We also derived a detector with upper bound performance. Our results demonstrate the notable benefits of TCAs over weakly-coupled antennas in target detection, caused by the higher SNR TCAs provide over a large bandwidth. We examined different detectors in practical scenarios where assumptions used to design them are not exact and highlighted the superior performance of our proposed approach.