Submitted to Signal Processing
Introduction
Spherically invariant random processes (SIRPs), also known as compound-Gaussian processes (CGPs), provide a flexible framework for modeling a broad class of non-Gaussian phenomena encountered in signal processing. This class includes, among others, ϵ \epsilon ϵ -contaminated Gaussian, Student-t t t , and K K K -distributed processes, as well as certain symmetric α \alpha α -stable processes, while also encompassing Gaussian processes as a particular case. The simplest models in this class are second-order (wide-sense) stationary parametric processes, for which the parameters of interest are estimated from a finite set of consecutive observations.
Evaluating the accuracy of parametric estimates in such models is fundamental to the analysis and design of signal processing systems. This accuracy is commonly characterized by the Cramér-Rao bound (CRB), which is determined by the inverse of the Fisher information matrix (FIM) and provides a lower bound on the covariance matrix of unbiased estimators. While the finite-sample FIM is given by the Slepian-Bangs formula [2 , 3 ] , evaluating it is computationally prohibitive for large data records. In contrast, Whittle’s formula [1 ] provides a closed-form expression for the asymptotic FIM rate of zero-mean, real-valued stationary Gaussian processes in terms of their power spectra. This drastically reduces computational complexity while offering a direct interpretation in terms of the spectral properties of the process. Recently, by exploiting the asymptotic properties of block Toeplitz matrices, this result was extended in [4 ] to real-valued, arbitrary-mean stationary SIRPs within a unified framework covering fully known, finite-dimensional parameterized, and completely unspecified density generators.
However, several important extensions are required to make such asymptotic FIM results applicable to a broader class of signal processing problems. First, many signals are frequency down-converted to baseband after filtering and are consequently represented by their complex-valued envelopes, whose distributions may be either circular or noncircular. Second, many man-made signals are not stationary but exhibit periodic statistical properties and are therefore naturally described by cyclostationary models. Such models arise in communication systems [5 , 6 ] , rotating machinery [7 , 8 ] , radar and sonar [9 , 10 ] , and power-line communications [11 , 12 ] . Third, practical communication signals may exhibit almost cyclostationary behavior, for which the relevant second-order statistics vary periodically with time but do not necessarily admit the strict periodicity of conventional cyclostationary models. A general asymptotic FIM framework encompassing these cases is therefore of considerable interest.
In this paper, we develop a unified extension of Whittle’s formula to multidimensional stationary, cyclostationary, and almost cyclostationary CGPs. The main contributions are as follows.
[4 ] to processes with periodic second-order statistics.
We further extend the results to purely and specifically almost cyclostationary complex-valued CGPs . In particular, we establish results that enable the treatment of linearly modulated digital communication signals exhibiting cyclostationarity and their extension to specific multidimensional almost cyclostationary models.
The proposed framework goes beyond the commonly used performance analysis based on independent and identically distributed Gaussian observations. In the considered setting, the observations are samples of continuous-time stochastic processes whose temporal and spectral characteristics can be arbitrary and whose distributions need not be Gaussian. Consequently, the resulting CRB explicitly incorporates information carried by the temporal spectrum, cyclic spectrum, and, for noncircular complex processes, complementary spectral characteristics. This provides a general analytical framework for studying how signal statistics influence parameter-estimation performance.
The remainder of this paper is organized as follows. Section II introduces the required preliminaries on real-valued multidimensional stationary CGPs and reviews the asymptotic frequency-domain FIM rate formula for these processes. Section III extends this formula to real-valued cyclostationary CGPs. Section IV further generalizes the results to circular and noncircular complex-valued multidimensional stationary CGPs, as well as to purely and specifically almost cyclostationary CGPs. Preliminary results for linearly modulated digital communication signals exhibiting cyclostationarity are also established. Section V applies the proposed asymptotic FIM rate formulas to investigate the influence of source spectral characteristics on the CRB for DOA estimation, with particular emphasis on measurements affected by frequency offsets. Finally, Section VI concludes the paper.
The following notations are used throughout this paper. Matrices and vectors are represented by bold upper-case and bold lower-case characters, respectively. Vectors are by default in column orientation, while the superscripts T T T , H H H and ∗ * ∗ stand for transpose, conjugate transpose and conjugate, respectively. E ( ⋅ ) E(\cdot) E ( ⋅ ) , ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ , and Tr ( ⋅ ) \operatorname{Tr}(\cdot) Tr ( ⋅ ) are the expectation, determinant, and trace operators, respectively. ∥ ⋅ ∥ F \|\cdot\|_{F} ∥ ⋅ ∥ F denotes the Frobenius norm. I \mathbf{I} I is the identity matrix with the appropriate dimension. ⊗ \otimes ⊗ is the Kronecker product of two matrices. Blkdiag ( B 1 , … , B n \operatorname{Blkdiag}(\mathbf{B}_{1},\ldots,\mathbf{B}_{n} Blkdiag ( B 1 , … , B n is the block diagonal matrix with B 1 , … , B n B_1,\ldots,B_n B 1 , … , B n along its diagonal. 1 C ( f ) = 1 \mathbb{1}_C(f)=1 1 C ( f ) = 1 if f ∈ C f\in C f ∈ C and 1 C ( f ) = 0 \mathbb{1}_C(f)=0 1 C ( f ) = 0 if f ∉ C f\notin C f ∈ / C and x ∼ P x x\sim P_x x ∼ P x means that the random variables x x x has the distribution P x P_x P x . The acronyms r.v., p.d.f., c.d.f. and i.f.f. are used to represent random variable, probability density function, cumulative distribution function, and ”if and only if”, respectively.
Statistical Background
Preliminaries on multidimensional CGP
Let ( x k ) k ∈ Z (\mathbf{x}_k)_{k\in\mathbb{Z}} ( x k ) k ∈ Z be a stochastic process with x k ∈ R m \mathbf{x}_k\in\mathbb{R}^m x k ∈ R m . This process is defined as an m m m -dimensional SIRP, or equivalently a CGP, if any finite collection of samples y : = ( x k 1 T , … , x k n T ) T \mathbf{y}\mathrel{\mathop:}= (\mathbf{x}_{k_1}^{T},\ldots,\mathbf{x}_{k_n}^{T})^{T} y : = ( x k 1 T , … , x k n T ) T with k 1 < … < k n k_1<\ldots<k_n k 1 < … < k n , follows a compound-Gaussian distribution. The p.d.f. is given by:
p ( y ) = ∣ Σ ∣ − 1 / 2 g n m [ ( y − μ ) T Σ − 1 ( y − μ ) ] . (1) p(\mathbf{y})=|\boldsymbol{\Sigma}|^{-1/2}g_{nm}\left[(\mathbf{y}-\boldsymbol{\mu})^{T}\boldsymbol{\Sigma}^{-1}(\mathbf{y}-\boldsymbol{\mu})\right]. \tag*{(1)} p ( y ) = ∣ Σ ∣ − 1/2 g nm [ ( y − μ ) T Σ − 1 ( y − μ ) ] . (1)
The first and second-order moments of y \mathbf{y} y are assumed to exist, satisfying E ( y ) = μ \mathbb{E}(\mathbf{y})=\boldsymbol{\mu} E ( y ) = μ and
Cov ( y ) = Σ (2) \operatorname{Cov}(\mathbf{y})=\boldsymbol{\Sigma} \tag*{(2)} Cov ( y ) = Σ (2)
respectively. This normalization in (2) is imposed to remove any scale ambiguity between Σ \boldsymbol{\Sigma} Σ and g n m ( ⋅ ) g_{nm}(\cdot) g nm ( ⋅ ) . The density generator g n m ( ⋅ ) g_{nm}(\cdot) g nm ( ⋅ ) is given by
g n m ( t ) = ( 2 π ) − n m / 2 ∫ 0 ∞ τ − n m / 2 exp ( − t / 2 τ ) d F τ ( τ ) , (3) g_{nm}(t)=(2\pi)^{-nm/2}\int_{0}^{\infty}\tau^{-nm/2}\exp(-t/2\tau)\,\mathrm{d}F_{\tau}(\tau), \tag*{(3)} g nm ( t ) = ( 2 π ) − nm /2 ∫ 0 ∞ τ − nm /2 exp ( − t /2 τ ) d F τ ( τ ) , (3)
where F τ F_{\tau} F τ denotes the c.d.f. of the r.v. τ \tau τ , commonly referred to as the texture variable. According to [13 Th. 2.2] , the p.d.f. in (1), with the density generator given by (3), is equivalently induced by the stochastic representation
y = μ + τ Σ 1 / 2 n 0 , (4) \mathbf{y}=\boldsymbol{\mu}+\sqrt{\tau}\boldsymbol{\Sigma}^{1/2}\mathbf{n}_0, \tag*{(4)} y = μ + τ Σ 1/2 n 0 , (4)
where τ \tau τ and n 0 \mathbf{n}_0 n 0 are independent, with n 0 ∼ N n m ( 0 , I ) \mathbf{n}_0\sim\mathcal{N}_{nm}(\mathbf{0},\mathbf{I}) n 0 ∼ N nm ( 0 , I ) . Under the normalization condition in (2), the texture variable satisfies E ( τ ) = 1 \mathbb{E}(\tau)=1 E ( τ ) = 1 . The density generator g n m ( ⋅ ) g_{nm}(\cdot) g nm ( ⋅ ) , or equivalently the c.d.f. F τ F_{\tau} F τ of τ \tau τ , may be fully known, parameterized by a finite-dimensional nuisance parameter [14 ] , or completely unspecified [15 ] .
In the following, the above definition is applied to stationary and cyclostationary processes, with the covariance matrix Σ \boldsymbol{\Sigma} Σ and the corresponding density generator determined by the temporal structure of the sampled observations.
Preliminaries on multidimensional stationary CGP
A CGP is said to be strictly stationary if the p.d.f. in (1) of any finite collection of samples y \mathbf{y} y is invariant under arbitrary translations of the sampling indices ( k 1 , … , k n ) (k_1,\ldots,k_n) ( k 1 , … , k n ) . Equivalently, the v.s ( x 1 + k , … , x n + k ) (\mathbf{x}_{1+k},\ldots,\mathbf{x}_{n+k}) ( x 1 + k , … , x n + k ) and ( x 1 , … , x n ) (\mathbf{x}_1,\ldots,\mathbf{x}_n) ( x 1 , … , x n ) share the same joint distribution for all n ∈ N ∗ n \in\mathbb{N}^* n ∈ N ∗ and k ∈ Z k \in\mathbb{Z} k ∈ Z . As for Gaussian processes, the compound-Gaussian structure implies that stationarity is equivalent to wide-sense stationarity. More precisely, a CGP is stationary if and only if E ( x ℓ ) \mathrm{E}(\mathbf{x}_{\ell}) E ( x ℓ ) and E ( x ℓ x ℓ + k T ) \mathrm{E}(\mathbf{x}_{\ell}\mathbf{x}_{\ell+k}^{T}) E ( x ℓ x ℓ + k T ) are independent of ℓ ∈ Z \ell\in\mathbb{Z} ℓ ∈ Z for every k ∈ Z k \in\mathbb{Z} k ∈ Z . Consequently, for a stationary CGP, the random vector formed by stacking n n n consecutive observations of x k \mathbf{x}_k x k is given by
y n ≡ d e f ( x 1 T , … , x n T ) T ∈ R m n ∀ n ∈ N ∗ , (5) \mathbf{y}_n \mathrel{\mathop{\equiv}^{\mathrm{def}}} (\mathbf{x}_1^T,\ldots,\mathbf{x}_n^T)^T \in\mathbb{R}^{mn}\quad\forall n \in\mathbb{N}^*, \tag*{(5)} y n ≡ def ( x 1 T , … , x n T ) T ∈ R mn ∀ n ∈ N ∗ , (5)
has mean μ y n ≡ d e f E ( y n ) = ( 1 n ⊗ I m ) μ x \boldsymbol{\mu}_{\mathbf{y}_n} \mathrel{\mathop{\equiv}^{\mathrm{def}}} \mathrm{E}(\mathbf{y}_n) = (\mathbf{1}_n \otimes\mathbf{I}_m)\boldsymbol{\mu}_{\mathbf{x}} μ y n ≡ def E ( y n ) = ( 1 n ⊗ I m ) μ x with μ x ≡ d e f E ( x ℓ ) \boldsymbol{\mu}_{\mathbf{x}} \mathrel{\mathop{\equiv}^{\mathrm{def}}} \mathrm{E}(\mathbf{x}_{\ell}) μ x ≡ def E ( x ℓ ) and covariance matrix Σ y n ≡ d e f Cov ( y n ) \boldsymbol{\Sigma}_{\mathbf{y}_n} \mathrel{\mathop{\equiv}^{\mathrm{def}}} \operatorname{Cov}(\mathbf{y}_n) Σ y n ≡ def Cov ( y n ) , which is symmetric positive definite and has a block Toeplitz structure with block size m m m . Its ( i , j ) (i,j) ( i , j ) -th block is given by R x ( j − i ) \mathbf{R}_{\mathbf{x}}(j-i) R x ( j − i ) with R x ( k ) ≡ d e f E [ ( x ℓ − μ x ) ( x ℓ + k − μ x ) T ] ∈ R m × m \mathbf{R}_{\mathbf{x}}(k) \mathrel{\mathop{\equiv}^{\mathrm{def}}} \mathrm{E}[(\mathbf{x}_{\ell}-\boldsymbol{\mu}_{\mathbf{x}})(\mathbf{x}_{\ell+k}-\boldsymbol{\mu}_{\mathbf{x}})^T] \in\mathbb{R}^{m \times m} R x ( k ) ≡ def E [( x ℓ − μ x ) ( x ℓ + k − μ x ) T ] ∈ R m × m . We assume that the sequence R x ( k ) \mathbf{R}_{\mathbf{x}}(k) R x ( k ) is absolutely summable, i.e., ∑ k = − ∞ ∞ ∥ R x ( k ) ∥ 2 < ∞ \sum_{k=-\infty}^{\infty}\|\mathbf{R}_{\mathbf{x}}(k)\|_2 < \infty ∑ k = − ∞ ∞ ∥ R x ( k ) ∥ 2 < ∞ , which guarantees that the spectrum1 R x ( f ) ≡ d e f ∑ k R x ( k ) e − i 2 π k f \mathbf{R}_{\mathbf{x}}(f) \mathrel{\mathop{\equiv}^{\mathrm{def}}} \sum_k \mathbf{R}_{\mathbf{x}}(k)e^{-i2\pi kf} R x ( f ) ≡ def ∑ k R x ( k ) e − i 2 π k f is well-defined.
Asymptotic Frequency-Domain FIM Rate for Stationary CGP
Suppose that the mean μ x \boldsymbol{\mu}_{\mathbf{x}} μ x and the sequence R x ( k ) \mathbf{R}_{\mathbf{x}}(k) R x ( k ) depend on a parameter θ = ( θ 1 , … , θ r ) T ∈ Θ ⊂ R r \boldsymbol{\theta} = (\theta_1,\ldots,\theta_r)^T \in\Theta\subset\mathbb{R}^r θ = ( θ 1 , … , θ r ) T ∈ Θ ⊂ R r , and consequently, so does the spectrum R x ( f ) \mathbf{R}_{\mathbf{x}}(f) R x ( f ) . For notational simplicity, this dependence on θ \boldsymbol{\theta} θ is omitted whenever no ambiguity arises. Assume that the mapping θ ↦ ( μ x , [ R x ( k ) ] k ∈ Z , R x ( f ) ) \boldsymbol{\theta} \mapsto(\boldsymbol{\mu}_{\mathbf{x}},[\mathbf{R}_{\mathbf{x}}(k)]_{k\in\mathbb{Z}},\mathbf{R}_{\mathbf{x}}(f)) θ ↦ ( μ x , [ R x ( k ) ] k ∈ Z , R x ( f )) is differentiable in a neighborhood of the true parameter value and that the regularity conditions stated in [4 ] hold, among which is E ( τ − 1 ) < ∞ \mathrm{E}(\tau^{-1}) < \infty E ( τ − 1 ) < ∞ . Then, combining the Slepian-Bangs expression for the finite-sample FIM with the asymptotic theory of block Toeplitz matrices yields the following asymptotic frequency domain FIM rate for a multidimensional real-valued stationary CGP:
lim n → ∞ 1 n ( FIM y n ( θ ) ) k , ℓ = E ( τ − 1 ) μ x , k ′ T R x − 1 ( 0 ) μ x , ℓ ′ + 1 2 ∫ 0 1 Tr ( R x − 1 ( f ) R x , k ′ ( f ) R x − 1 ( f ) R x , ℓ ′ ( f ) ) d f + c ( ∫ 0 1 Tr ( R x − 1 ( f ) R x , k ′ ( f ) ) d f ) ( ∫ 0 1 Tr ( R x − 1 ( f ) R x , ℓ ′ ( f ) ) d f ) , (6) \lim_{n\to\infty}\frac{1}{n}\left(\operatorname{FIM}_{\mathbf{y}_n}(\boldsymbol{\theta})\right)_{k,\ell}
=
\mathrm{E}(\tau^{-1})\,\boldsymbol{\mu}'_{\mathbf{x},k}{}^{T}\mathbf{R}_{\mathbf{x}}^{-1}(0)\boldsymbol{\mu}'_{\mathbf{x},\ell}
+\frac{1}{2}\int_0^1\operatorname{Tr}\left(\mathbf{R}_{\mathbf{x}}^{-1}(f)\mathbf{R}'_{\mathbf{x},k}(f)\mathbf{R}_{\mathbf{x}}^{-1}(f)\mathbf{R}'_{\mathbf{x},\ell}(f)\right)\mathrm{d}f
+c\left(\int_0^1\operatorname{Tr}\left(\mathbf{R}_{\mathbf{x}}^{-1}(f)\mathbf{R}'_{\mathbf{x},k}(f)\right)\mathrm{d}f\right)\left(\int_0^1\operatorname{Tr}\left(\mathbf{R}_{\mathbf{x}}^{-1}(f)\mathbf{R}'_{\mathbf{x},\ell}(f)\right)\mathrm{d}f\right), \tag*{(6)} n → ∞ lim n 1 ( FIM y n ( θ ) ) k , ℓ = E ( τ − 1 ) μ x , k ′ T R x − 1 ( 0 ) μ x , ℓ ′ + 2 1 ∫ 0 1 Tr ( R x − 1 ( f ) R x , k ′ ( f ) R x − 1 ( f ) R x , ℓ ′ ( f ) ) d f + c ( ∫ 0 1 Tr ( R x − 1 ( f ) R x , k ′ ( f ) ) d f ) ( ∫ 0 1 Tr ( R x − 1 ( f ) R x , ℓ ′ ( f ) ) d f ) , (6)
for k , ℓ = 1 , … , r k,\ell=1,\ldots,r k , ℓ = 1 , … , r , where μ x , k ′ ≡ d e f ∂ μ x ∂ θ k \boldsymbol{\mu}'_{\mathbf{x},k} \mathrel{\mathop{\equiv}^{\mathrm{def}}} \frac{\partial\boldsymbol{\mu}_{\mathbf{x}}}{\partial\theta_k} μ x , k ′ ≡ def ∂ θ k ∂ μ x , R x ( 0 ) = R x ( f ) ∣ f = 0 \mathbf{R}_{\mathbf{x}}(0)=\mathbf{R}_{\mathbf{x}}(f)|_{f=0} R x ( 0 ) = R x ( f ) ∣ f = 0 denotes the spectrum matrix evaluated at zero frequency, R x , k ′ ( f ) ≡ d e f ∂ R x ( f ) ∂ θ k \mathbf{R}'_{\mathbf{x},k}(f) \mathrel{\mathop{\equiv}^{\mathrm{def}}} \frac{\partial\mathbf{R}_{\mathbf{x}}(f)}{\partial\theta_k} R x , k ′ ( f ) ≡ def ∂ θ k ∂ R x ( f ) , and where c c c is given by (7) or (8) depending on whether the g n m ( ⋅ ) g_{nm}(\cdot) g nm ( ⋅ ) , or equivalently the c.d.f. F τ F_{\tau} F τ of τ \tau τ is fully known or completely unspecified, respectively. For the classical parametric case in which F τ F_{\tau} F τ is known, the coefficient c c c is
c C l a s = − 1 − β 2 m ≤ 0 , (7) c_{\mathrm{Clas}}=-\frac{1-\beta}{2m}\le0, \tag*{(7)} c Clas = − 2 m 1 − β ≤ 0 , (7)
where β \beta β denotes the total probability mass of the discrete component of the distribution of τ \tau τ , i.e., β ≔ ∑ t : P ( τ = t ) > 0 P ( τ = t ) \beta\coloneqq\sum_{t:P(\tau=t)>0}P(\tau=t) β : = ∑ t : P ( τ = t ) > 0 P ( τ = t ) . For the semiparametric case in which the density generator is completely unspecified,
c S e P a = − 1 2 m . (8) c_{\mathrm{SePa}}=-\frac{1}{2m}. \tag*{(8)} c SePa = − 2 m 1 . (8)
Eq. (6) extends Whittle’s frequency-domain FIM rate to multidimensional stationary CGPs with arbitrary mean. In the Gaussian case, τ = 1 \tau=1 τ = 1 and thus c C l a s = 0 c_{\mathrm{Clas}}=0 c Clas = 0 , recovering the multidimensional Gaussian Whittle formula with arbitrary mean.
Asymptotic FIM Rate for Cyclostationary CGP
Cyclostationary CGP with cycle period p p p
Analogous to strict stationarity, a CGP is strictly cyclostationary with cycle period p p p if its finite-dimensional distributions are invariant under time shifts of p p p . This means that r.v.s ( x 1 + p k , … , x n + p k ) (\mathbf{x}_{1+pk},\ldots,\mathbf{x}_{n+pk}) ( x 1 + p k , … , x n + p k ) and ( x 1 , … , x n ) (\mathbf{x}_1,\ldots,\mathbf{x}_n) ( x 1 , … , x n ) share the same joint distribution for all n ∈ N ∗ n\in\mathbb{N}^{*} n ∈ N ∗ , k ∈ Z k\in\mathbb{Z} k ∈ Z . Within the standard SIRP framework, a CGP is strictly cyclostationary if and only if it is wide-sense cyclostationary.2 Precisely, a CGP is cyclostationary with cycle period p p p i.f.f.
E ( x k + i p ) and E ( x k + i p x ℓ + i p T ) are independent of i ∈ Z , ∀ k , ℓ ∈ Z . (9) \mathbb{E}(\mathbf{x}_{k+ip})\text{ and }\mathbb{E}(\mathbf{x}_{k+ip}\mathbf{x}_{\ell+ip}^{T})\text{ are independent of }i\in\mathbb{Z},\ \forall k,\ell\in\mathbb{Z}. \tag*{(9)} E ( x k + i p ) and E ( x k + i p x ℓ + i p T ) are independent of i ∈ Z , ∀ k , ℓ ∈ Z . (9)
For a cyclostationary CGP with cycle period p p p , the blocked sequence ( x k ′ ) k ∈ Z ∈ R m p (\mathbf{x}'_k)_{k\in\mathbb{Z}}\in\mathbb{R}^{mp} ( x k ′ ) k ∈ Z ∈ R m p , k ∈ Z k\in\mathbb{Z} k ∈ Z with x k ′ ≔ [ x 1 + ( k − 1 ) p T , … , x p + ( k − 1 ) p T ] T \mathbf{x}'_k\coloneqq[\mathbf{x}_{1+(k-1)p}^{T},\ldots,\mathbf{x}_{p+(k-1)p}^{T}]^{T} x k ′ : = [ x 1 + ( k − 1 ) p T , … , x p + ( k − 1 ) p T ] T is stationary as a direct consequence of (9) (see also [18 ] , [19 ] ).
When the number of observations n n n is a multiple of p p p (i.e., n = p q n=pq n = pq for some integer q q q ) [20 ] , the vector y n \mathbf{y}_n y n defined in (5) can be partitioned into q q q blocks of size m p mp m p :
y n = ( x 1 T , … , x p T , … , x ( k − 1 ) p + 1 T , … , x k p T , … , x n − p + 1 T , … , x n T ) T ∈ R m n for n = p q . (10) \mathbf{y}_n=(\mathbf{x}_1^{T},\ldots,\mathbf{x}_p^{T},\ldots,\mathbf{x}_{(k-1)p+1}^{T},\ldots,\mathbf{x}_{kp}^{T},\ldots,\mathbf{x}_{n-p+1}^{T},\ldots,\mathbf{x}_n^{T})^{T}\in\mathbb{R}^{mn}\ \text{for }n=pq. \tag*{(10)} y n = ( x 1 T , … , x p T , … , x ( k − 1 ) p + 1 T , … , x k p T , … , x n − p + 1 T , … , x n T ) T ∈ R mn for n = pq . (10)
Consequently, the mean of y n \mathbf{y}_n y n is given by μ y n = ( 1 q ⊗ I m ) μ x ′ \boldsymbol{\mu}_{\mathbf{y}_n}=(\mathbf{1}_q\otimes\mathbf{I}_m)\boldsymbol{\mu}_{\mathbf{x}'} μ y n = ( 1 q ⊗ I m ) μ x ′ with μ x ′ ≔ ( μ x 1 ′ T , … , μ x p ′ T ) T \boldsymbol{\mu}_{\mathbf{x}'}\coloneqq(\boldsymbol{\mu}_{\mathbf{x}'_1}^{T},\ldots,\boldsymbol{\mu}_{\mathbf{x}'_p}^{T})^{T} μ x ′ : = ( μ x 1 ′ T , … , μ x p ′ T ) T . The covariance matrix Σ y n ≔ Cov ( y n ) \boldsymbol{\Sigma}_{\mathbf{y}_n}\coloneqq\operatorname{Cov}(\mathbf{y}_n) Σ y n : = Cov ( y n ) is also a symmetric positive definite and possesses a block-Toeplitz structure with block size m p mp m p . its ( k , ℓ ) (k,\ell) ( k , ℓ ) -th block is given by R x ′ ( ℓ − k ) \mathbf{R}_{\mathbf{x}'}(\ell-k) R x ′ ( ℓ − k ) for which the sequence $\mathbf{R}_{\mathbf{x}'}(k)\coloneqq E [ ( x ℓ ′ − μ x ′ ) ( x ′ ℓ + k − μ x ′ ) T ] ∈ R p m × p m \mathbb{E}[(\mathbf{x}'_{\ell}-\boldsymbol{\mu}_{\mathbf{x}'})({\mathbf{x}'}_{\ell+k}-\boldsymbol{\mu}_{\mathbf{x}'})^{T}] \in\mathbb{R}^{pm\times pm} E [( x ℓ ′ − μ x ′ ) ( x ′ ℓ + k − μ x ′ ) T ] ∈ R p m × p m is assumed to be absolutely summable. The corresponding spectral matrix of the blocked ( x ′ k ) k ∈ Z ∈ R m p ({\mathbf{x}'}_k)_{k\in\mathbb{Z}} \in\mathbb{R}^{mp} ( x ′ k ) k ∈ Z ∈ R m p is R x ′ ( f ) ≔ ∑ k ∈ Z R x ′ ( k ) e − i 2 π k f \mathbf{R}_{\mathbf{x}'}(f) \coloneqq\sum_{k\in\mathbb{Z}}\mathbf{R}_{\mathbf{x}'}(k)e^{-i2\pi kf} R x ′ ( f ) : = ∑ k ∈ Z R x ′ ( k ) e − i 2 π k f . Notably, the second-order frequency properties of the cyclostationary CGP are more advantageously characterized by its cyclic spectra3 . The cyclic spectral matrices are defined as: R x ( k / p ) ( f ) ≔ ∑ τ = − ∞ ∞ R x ( k / p ) ( τ ) e − i 2 π τ f \mathbf{R}_{\mathbf{x}}^{(k/p)}(f) \coloneqq\sum_{\tau=-\infty}^{\infty}\mathbf{R}_{\mathbf{x}}^{(k/p)}(\tau)e^{-i2\pi\tau f} R x ( k / p ) ( f ) : = ∑ τ = − ∞ ∞ R x ( k / p ) ( τ ) e − i 2 π τ f , k = 0 , 1 , … , p − 1 k=0,1,\ldots,p-1 k = 0 , 1 , … , p − 1 , where the cyclic autocovariance sequence R x ( k / p ) ( τ ) ≔ 1 p ∑ ℓ = 0 p − 1 E ( x ℓ x ℓ + τ T ) e − i 2 π k ℓ / p \mathbf{R}_{\mathbf{x}}^{(k/p)}(\tau) \coloneqq\frac{1}{p}\sum_{\ell=0}^{p-1}\mathbb{E}(\mathbf{x}_{\ell}\mathbf{x}_{\ell+\tau}^{T})e^{-i2\pi k\ell/p} R x ( k / p ) ( τ ) : = p 1 ∑ ℓ = 0 p − 1 E ( x ℓ x ℓ + τ T ) e − i 2 π k ℓ / p is the k k k -th Fourier series coefficient of the periodic covariance
E [ ( x ℓ − μ x ) ( x ℓ + τ − μ x ) T ] = ∑ k = 0 p − 1 R x ( k / p ) ( τ ) e i 2 π k ℓ / p , (11) \mathbb{E}[(\mathbf{x}_{\ell}-\boldsymbol{\mu}_{\mathbf{x}})(\mathbf{x}_{\ell+\tau}-\boldsymbol{\mu}_{\mathbf{x}})^{T}]
=\sum_{k=0}^{p-1}\mathbf{R}_{\mathbf{x}}^{(k/p)}(\tau)e^{i2\pi k\ell/p}, \tag*{(11)} E [( x ℓ − μ x ) ( x ℓ + τ − μ x ) T ] = k = 0 ∑ p − 1 R x ( k / p ) ( τ ) e i 2 π k ℓ / p , (11)
evaluated at the harmonic cyclic frequencies ( k p ) k = 0 , 1 , … , p − 1 (\frac{k}{p})_{k=0,1,\ldots,p-1} ( p k ) k = 0 , 1 , … , p − 1 . Specifically, the cyclic autocovariance is given by:
R x ( k / p ) ( τ ) ≔ 1 p ∑ ℓ = 0 p − 1 E [ ( x ℓ − μ x ( ℓ ) ) ( x ℓ + τ − μ x ( ℓ + τ ) ) T ] e − i 2 π k ℓ / p . \mathbf{R}_{\mathbf{x}}^{(k/p)}(\tau)\coloneqq\frac{1}{p}\sum_{\ell=0}^{p-1}\mathbb{E}[(\mathbf{x}_{\ell}-\boldsymbol{\mu}_{\mathbf{x}}(\ell))(\mathbf{x}_{\ell+\tau}-\boldsymbol{\mu}_{\mathbf{x}}(\ell+\tau))^{T}]e^{-i2\pi k\ell/p}. R x ( k / p ) ( τ ) : = p 1 ℓ = 0 ∑ p − 1 E [( x ℓ − μ x ( ℓ )) ( x ℓ + τ − μ x ( ℓ + τ ) ) T ] e − i 2 π k ℓ / p .
Furthermore, we assume that the spectral matrices R x ′ ( f ) \mathbf{R}_{\mathbf{x}'}(f) R x ′ ( f ) and R x ( k / p ) ( f ) \mathbf{R}_{\mathbf{x}}^{(k/p)}(f) R x ( k / p ) ( f ) depend on a parameter vector θ = ( θ 1 , … , θ q ) T ∈ R q \boldsymbol{\theta}=(\theta_1,\ldots,\theta_q)^T\in\mathbb{R}^{q} θ = ( θ 1 , … , θ q ) T ∈ R q . For notational simplicity, this dependence is omitted in the following. Finally, the mappings θ ↦ R x ′ ( f ) \boldsymbol{\theta}\mapsto\mathbf{R}_{\mathbf{x}'}(f) θ ↦ R x ′ ( f ) and θ ↦ R x ( k / p ) ( f ) \boldsymbol{\theta}\mapsto\mathbf{R}_{\mathbf{x}}^{(k/p)}(f) θ ↦ R x ( k / p ) ( f ) are assumed to be differentiable in a neighborhood of the true value of θ \boldsymbol{\theta} θ , and that the regularity conditions stated in [4 ] hold.
Asymptotic FIM rate for cyclostationary CGP with cycle period p p p
Since a cyclostationary CGP ( x k ) k ∈ Z ({\mathbf{x}}_k)_{k\in\mathbb{Z}} ( x k ) k ∈ Z with cycle period p p p gives rise to the stationary CGP ( x k ′ ) k ∈ Z ({\mathbf{x}}'_k)_{k\in\mathbb{Z}} ( x k ′ ) k ∈ Z defined above, the asymptotic frequency-domain FIM rate in (6) can be applied to x k ′ \mathbf{x}'_k x k ′ . However, because the blocked process consists of only q = n / p q=n/p q = n / p vector samples, the FIM rate for the original process scales by a factor of 1 / p 1/p 1/ p . By substituting the parameters of the blocked process and accounting for this scaling, the terms ( 1 2 , c ) (\frac{1}{2},c) ( 2 1 , c ) , R x ( f ) \mathbf{R}_{\mathbf{x}}(f) R x ( f ) , μ x , k ′ \boldsymbol{\mu}_{\mathbf{x},k}' μ x , k ′ , and R x , k ′ ( f ) \mathbf{R}_{\mathbf{x},k}'(f) R x , k ′ ( f ) are replaced by ( 1 2 p , c p 2 ) (\frac{1}{2p},\frac{c}{p^2}) ( 2 p 1 , p 2 c ) , μ x ′ , k ′ \boldsymbol{\mu}_{\mathbf{x}',k}' μ x ′ , k ′ , R x ′ ( f ) \mathbf{R}_{\mathbf{x}'}(f) R x ′ ( f ) , and R x ′ , k ′ ( f ) \mathbf{R}_{\mathbf{x}',k}'(f) R x ′ , k ′ ( f ) , respectively.
For a cyclostationary CGP, it is computationally and physically more insightful to express the resulting asymptotic frequency-domain FIM rate directly in terms of the cyclic spectra R x ( k / p ) ( f ) \mathbf{R}_{\mathbf{x}}^{(k/p)}(f) R x ( k / p ) ( f ) for k = 0 , … , p − 1 k=0,\ldots,p-1 k = 0 , … , p − 1 , rather than relying on the conventional spectral matrix R x ′ ( f ) \mathbf{R}_{\mathbf{x}'}(f) R x ′ ( f ) of the blocked process. Building upon the foundational work of Gladyshev [18 ] on periodically correlated random sequences (see also [19 eq. (2.7)] and [21 Proposition 6.9] ), this concept extends naturally to multidimensional cyclostationary processes. Specifically, the conventional spectral matrix R x ′ ( f ) \mathbf{R}_{\mathbf{x}'}(f) R x ′ ( f ) is unitarily similar to an m p × m p mp\times mp m p × m p symmetric positive definite polyspectral matrix, denoted R x c y c ( f ) \mathbf{R}_{\mathbf{x}}^{\mathrm{cyc}}(f) R x cyc ( f ) . Its ( k ′ , ℓ ′ ) (k',\ell') ( k ′ , ℓ ′ ) -th m × m m\times m m × m block of R x c y c ( f ) \mathbf{R}_{\mathbf{x}}^{\mathrm{cyc}}(f) R x cyc ( f ) is defined as function of the cyclic spectra R x ( ( ℓ ′ − k ′ ) / p ) ( f ) \mathbf{R}_{\mathbf{x}}^{((\ell'-k')/p)}(f) R x (( ℓ ′ − k ′ ) / p ) ( f ) by the following matrices [19 eq. (2.2)] :
1 p R x ( ( ℓ ′ − k ′ ) / p ) ( f − k ′ + 1 p ) , k ′ , ℓ ′ = 1 , … , p . (12) \frac{1}{p}\mathbf{R}_{\mathbf{x}}^{((\ell'-k')/p)}\left(\frac{f-k'+1}{p}\right),\quad k',\ell'=1,\ldots,p. \tag*{(12)} p 1 R x (( ℓ ′ − k ′ ) / p ) ( p f − k ′ + 1 ) , k ′ , ℓ ′ = 1 , … , p . (12)
This relationship yields the following block EVD of R x ′ ( f ) \mathbf{R}_{x'}(f) R x ′ ( f ) :
R x ′ ( f ) = p [ U p ( f ) ⊗ I m ] H R x c y c ( f ) [ U p ( f ) ⊗ I m ] , (13) \mathbf{R}_{x'}(f)=p\left[\mathbf{U}_p(f)\otimes\mathbf{I}_m\right]^{H}\mathbf{R}_{x}^{\mathrm{cyc}}(f)\left[\mathbf{U}_p(f)\otimes\mathbf{I}_m\right], \tag*{(13)} R x ′ ( f ) = p [ U p ( f ) ⊗ I m ] H R x cyc ( f ) [ U p ( f ) ⊗ I m ] , (13)
where U p ( f ) \mathbf{U}_p(f) U p ( f ) is the p × p p\times p p × p unitary DFT matrix with entries defined by [ U p ( f ) ] k ′ , ℓ ′ ≔ 1 p e i 2 π ( k ′ − 1 − f ) ( ℓ ′ − 1 ) / p [\mathbf{U}_p(f)]_{k',\ell'}\coloneqq\frac{1}{\sqrt{p}}e^{i2\pi(k'-1-f)(\ell'-1)/p} [ U p ( f ) ] k ′ , ℓ ′ : = p 1 e i 2 π ( k ′ − 1 − f ) ( ℓ ′ − 1 ) / p . Substitution of the block EVD (13) into (6) for the blocked process, and exploiting the unitary invariance of the trace operator, the factors of p p p inside the trace operations cancel out elegantly. This leads to the following fundamental result:
Result 1: For a multidimensional cyclostationary real-valued CGP with cycle period p p p , the asymptotic frequency-domain FIM rate is given by:
lim n → ∞ 1 n ( FIM y n ( θ ) ) k , ℓ = 1 p E ( τ − 1 ) ( μ x , k c y c ′ ) T [ R x c y c ( 0 ) ] − 1 μ x , ℓ c y c ′ \lim_{n\to\infty}\frac{1}{n}\left(\operatorname{FIM}_{y_n}(\theta)\right)_{k,\ell}
=
\frac{1}{p}\mathbb{E}\left(\tau^{-1}\right)\left(\boldsymbol{\mu}_{x,k}^{\mathrm{cyc}\prime}\right)^{T}\left[\mathbf{R}_{x}^{\mathrm{cyc}}(0)\right]^{-1}\boldsymbol{\mu}_{x,\ell}^{\mathrm{cyc}\prime} n → ∞ lim n 1 ( FIM y n ( θ ) ) k , ℓ = p 1 E ( τ − 1 ) ( μ x , k cyc ′ ) T [ R x cyc ( 0 ) ] − 1 μ x , ℓ cyc ′ + 1 2 p ∫ 0 1 Tr ( [ R x c y c ( f ) ] − 1 R x , k c y c ′ ( f ) [ R x c y c ( f ) ] − 1 R x , ℓ c y c ′ ( f ) ) d f \quad+\frac{1}{2p}\int_{0}^{1}\operatorname{Tr}\left(\left[\mathbf{R}_{x}^{\mathrm{cyc}}(f)\right]^{-1}\mathbf{R}_{x,k}^{\mathrm{cyc}\prime}(f)\left[\mathbf{R}_{x}^{\mathrm{cyc}}(f)\right]^{-1}\mathbf{R}_{x,\ell}^{\mathrm{cyc}\prime}(f)\right)\mathrm{d}f + 2 p 1 ∫ 0 1 Tr ( [ R x cyc ( f ) ] − 1 R x , k cyc ′ ( f ) [ R x cyc ( f ) ] − 1 R x , ℓ cyc ′ ( f ) ) d f + c p 2 ∫ 0 1 Tr ( [ R x c y c ( f ) ] − 1 R x , k c y c ′ ( f ) ) d f ∫ 0 1 Tr ( [ R x c y c ( f ) ] − 1 R x , ℓ c y c ′ ( f ) ) d f , (14) \quad+\frac{c}{p^{2}}\int_{0}^{1}\operatorname{Tr}\left(\left[\mathbf{R}_{x}^{\mathrm{cyc}}(f)\right]^{-1}\mathbf{R}_{x,k}^{\mathrm{cyc}\prime}(f)\right)\mathrm{d}f\int_{0}^{1}\operatorname{Tr}\left(\left[\mathbf{R}_{x}^{\mathrm{cyc}}(f)\right]^{-1}\mathbf{R}_{x,\ell}^{\mathrm{cyc}\prime}(f)\right)\mathrm{d}f, \tag*{(14)} + p 2 c ∫ 0 1 Tr ( [ R x cyc ( f ) ] − 1 R x , k cyc ′ ( f ) ) d f ∫ 0 1 Tr ( [ R x cyc ( f ) ] − 1 R x , ℓ cyc ′ ( f ) ) d f , (14)
where R x c y c ( 0 ) ≔ R x c y c ( f ) ∣ f = 0 \mathbf{R}_{x}^{\mathrm{cyc}}(0)\coloneqq\mathbf{R}_{x}^{\mathrm{cyc}}(f)\big|_{f=0} R x cyc ( 0 ) : = R x cyc ( f ) f = 0 , the derivatives are defined as R x , k c y c ′ ( f ) ≔ ∂ R x c y c ( f ) ∂ θ k \mathbf{R}_{x,k}^{\mathrm{cyc}\prime}(f)\coloneqq\frac{\partial\mathbf{R}_{x}^{\mathrm{cyc}}(f)}{\partial\theta_k} R x , k cyc ′ ( f ) : = ∂ θ k ∂ R x cyc ( f ) , and the transformed mean derivative vector is μ x , k c y c ′ ≔ ( W p ⊗ I m ) μ x ′ , k ′ \boldsymbol{\mu}_{x,k}^{\mathrm{cyc}\prime}\coloneqq(\mathbf{W}_p\otimes\mathbf{I}_m)\boldsymbol{\mu}_{x',k}^{\prime} μ x , k cyc ′ : = ( W p ⊗ I m ) μ x ′ , k ′ . Here, μ x ′ , k ′ = [ ( μ x 1 ′ , k ′ ) T , … , ( μ x p ′ , k ′ ) T ] T \boldsymbol{\mu}_{x',k}^{\prime}=[(\boldsymbol{\mu}_{x'_1,k}^{\prime})^{T},\ldots,(\boldsymbol{\mu}_{x'_p,k}^{\prime})^{T}]^{T} μ x ′ , k ′ = [( μ x 1 ′ , k ′ ) T , … , ( μ x p ′ , k ′ ) T ] T , and W p ∈ C p × p \mathbf{W}_p\in\mathbb{C}^{p\times p} W p ∈ C p × p is the standard DFT matrix with entries [ W p ] k , ℓ ≔ 1 p e i 2 π ( k − 1 ) ( ℓ − 1 ) / p [\mathbf{W}_p]_{k,\ell}\coloneqq\frac{1}{\sqrt{p}}e^{i2\pi(k-1)(\ell-1)/p} [ W p ] k , ℓ : = p 1 e i 2 π ( k − 1 ) ( ℓ − 1 ) / p .
Asymptotic FIM rate for stationary, cyclostationary and almost stationary complex-valued CGP
In numerous engineering applications, multidimensional complex-valued processes, whether circular or noncircular, arise naturally. In these instances, A complex-valued process ( x k ) k ∈ Z (x_k)_{k\in\mathbb{Z}} ( x k ) k ∈ Z with x k ∈ C m x_k\in\mathbb{C}^{m} x k ∈ C m is defined as a complex CGP if its associated real-valued process ( x ˉ k ) k ∈ Z (\bar{x}_k)_{k\in\mathbb{Z}} ( x ˉ k ) k ∈ Z with x ˉ k ≔ [ Re ( x k ) T , Im ( x k ) T ] T ∈ R 2 m \bar{x}_k\coloneqq[\operatorname{Re}(x_k)^{T},\operatorname{Im}(x_k)^{T}]^{T}\in\mathbb{R}^{2m} x ˉ k : = [ Re ( x k ) T , Im ( x k ) T ] T ∈ R 2 m is a real CGP. The one-to-one mapping x ˉ ↦ x ~ ≔ ( x T , x H ) T = 2 M x ˉ \bar{x}\mapsto\widetilde{x}\coloneqq(x^{T},x^{H})^{T}=\sqrt{2}\mathbf{M}\bar{x} x ˉ ↦ x : = ( x T , x H ) T = 2 M x ˉ , where M ≔ 1 2 ( I i I I − i I ) \mathbf{M}\coloneqq\frac{1}{\sqrt{2}}\begin{pmatrix}\mathbf{I}&i\mathbf{I}\\\mathbf{I}&-i\mathbf{I}\end{pmatrix} M : = 2 1 ( I I i I − i I ) is unitary, is subsequently referred to as a real-complex representation. It provides a convenient framework for extending the asymptotic frequency-domain FIM rates in (6) and (14) to both circular and noncircular complex-valued stationary and cyclostationary CGPs.
Asymptotic FIM rate for stationary and cyclostationary complex-valued CGP
Complex-valued extension of (6)
The real-to-complex representation of CGP is employed, and (6) is applied to the associated real-valued process ( x ˉ k ) k ∈ Z ∈ R 2 m (\bar{x}_k)_{k\in\mathbb{Z}}\in\mathbb{R}^{2m} ( x ˉ k ) k ∈ Z ∈ R 2 m . This yields a direct extension of (6) to stationary complex-valued CGP, with the corresponding parameters and functions summarized in Table 1 . In particular,
Table 1. Parameters and functions involved in the asymptotic FIM rate for real and complex-valued CGP.
C ~ x ( f ) = ( C x ( f ) R x ( f ) R x ∗ ( − f ) C x ∗ ( − f ) ) = ∑ k ∈ Z C ~ x ( k ) e − i 2 π k f , \widetilde{\mathbf C}_{x}(f)=
\begin{pmatrix}
\mathbf C_x(f) & \mathbf R_x(f)\\
\mathbf R_x^{*}(-f) & \mathbf C_x^{*}(-f)
\end{pmatrix}
=\sum_{k\in\mathbb Z}\widetilde{\mathbf C}_x(k)e^{-i2\pi kf}, C x ( f ) = ( C x ( f ) R x ∗ ( − f ) R x ( f ) C x ∗ ( − f ) ) = k ∈ Z ∑ C x ( k ) e − i 2 π k f ,
where the block matrices C x ( f ) ≔ ∑ k ∈ Z C x ( k ) e − i 2 π k f \mathbf C_x(f)\coloneqq\sum_{k\in\mathbb Z}\mathbf C_x(k)e^{-i2\pi kf} C x ( f ) : = ∑ k ∈ Z C x ( k ) e − i 2 π k f and R x ( f ) ≔ ∑ k ∈ Z R x ( k ) e − i 2 π k f \mathbf R_x(f)\coloneqq\sum_{k\in\mathbb Z}\mathbf R_x(k)e^{-i2\pi kf} R x ( f ) : = ∑ k ∈ Z R x ( k ) e − i 2 π k f represent the spectrum and pseudo-spectrum matrices, respectively, and where C ~ x ( k ) ≔ E [ ( x ~ ℓ − μ ~ x ) ( x ~ ℓ + k − μ ~ x ) H ] \widetilde{\mathbf C}_x(k)\coloneqq\mathbb E[(\widetilde{x}_{\ell}-\widetilde{\mu}_x)(\widetilde{x}_{\ell+k}-\widetilde{\mu}_x)^{H}] C x ( k ) : = E [( x ℓ − μ x ) ( x ℓ + k − μ x ) H ] with μ ~ x ≔ ( μ x T , μ x H ) T \widetilde{\mu}_x\coloneqq(\mu_x^{T},\mu_x^{H})^{T} μ x : = ( μ x T , μ x H ) T , C x ( k ) ≔ E [ ( x ℓ − μ x ) ( x ℓ + k − μ x ) H ] \mathbf C_x(k)\coloneqq\mathbb E[(x_{\ell}-\mu_x)(x_{\ell+k}-\mu_x)^{H}] C x ( k ) : = E [( x ℓ − μ x ) ( x ℓ + k − μ x ) H ] and R x ( k ) ≔ E [ ( x ℓ − μ x ) ( x ℓ + k − μ x ) T ] \mathbf R_x(k)\coloneqq\mathbb E[(x_{\ell}-\mu_x)(x_{\ell+k}-\mu_x)^{T}] R x ( k ) : = E [( x ℓ − μ x ) ( x ℓ + k − μ x ) T ] .
Complex-valued extension of Result 1
The noncircular complex-valued cyclostationary CGP with cycle period p p p are also characterized by their cyclic spectra, which are expressed as:
C x ~ ( k / p ) ( f ) ≔ ∑ τ ∈ Z C x ~ ( k / p ) ( τ ) e − i 2 π τ f , k = 0 , 1 , … , p − 1 , \mathbf C_{\widetilde{x}}^{(k/p)}(f)\coloneqq\sum_{\tau\in\mathbb Z}\mathbf C_{\widetilde{x}}^{(k/p)}(\tau)e^{-i2\pi\tau f},\quad k=0,1,\ldots,p-1, C x ( k / p ) ( f ) : = τ ∈ Z ∑ C x ( k / p ) ( τ ) e − i 2 π τ f , k = 0 , 1 , … , p − 1 ,
where
C x ~ ( k / p ) ( τ ) ≔ 1 p ∑ ℓ = 0 p − 1 E [ ( x ~ ℓ − μ ~ x ) ( x ~ ℓ + τ − μ ~ x ) H ] e − i 2 π k ℓ / p \mathbf C_{\widetilde{x}}^{(k/p)}(\tau)\coloneqq\frac{1}{p}\sum_{\ell=0}^{p-1}\mathbb E[(\widetilde{x}_{\ell}-\widetilde{\mu}_x)(\widetilde{x}_{\ell+\tau}-\widetilde{\mu}_x)^{H}]e^{-i2\pi k\ell/p} C x ( k / p ) ( τ ) : = p 1 ℓ = 0 ∑ p − 1 E [( x ℓ − μ x ) ( x ℓ + τ − μ x ) H ] e − i 2 π k ℓ / p
is the k k k -th Fourier series coefficient of the periodic augmented covariance:
E [ ( x ~ ℓ − μ ~ x ) ( x ~ ℓ + τ − μ ~ x ) H ] = ∑ k = 0 p − 1 C x ~ ( k / p ) ( τ ) e i 2 π k ℓ / p \mathbb E[(\widetilde{x}_{\ell}-\widetilde{\mu}_x)(\widetilde{x}_{\ell+\tau}-\widetilde{\mu}_x)^{H}]
=\sum_{k=0}^{p-1}\mathbf C_{\widetilde{x}}^{(k/p)}(\tau)e^{i2\pi k\ell/p} E [( x ℓ − μ x ) ( x ℓ + τ − μ x ) H ] = k = 0 ∑ p − 1 C x ( k / p ) ( τ ) e i 2 π k ℓ / p
evaluated at the harmonic cyclic frequencies ( k p ) k = 0 , 1 , … , p − 1 (\frac{k}{p})_{k=0,1,\ldots,p-1} ( p k ) k = 0 , 1 , … , p − 1 .
Using the one-to-one mapping x ˉ ↦ x ~ = 2 M x ˉ \bar{x}\mapsto\widetilde{x}=\sqrt{2}\mathbf M\bar{x} x ˉ ↦ x = 2 M x ˉ and the 2 m p × 2 m p 2mp\times2mp 2 m p × 2 m p unitary matrix Blkdiag ( M , … , M ) [ U p ( f ) ⊗ I 2 m ] \operatorname{Blkdiag}(\mathbf M,\ldots,\mathbf M)[\mathbf U_p(f)\otimes\mathbf I_{2m}] Blkdiag ( M , … , M ) [ U p ( f ) ⊗ I 2 m ] , we straightforwardly obtain the following result from Result 1:
Result 2. The asymptotic frequency-domain FIM rate for multidimensional cyclostationary noncircular complex-valued CGP is given by:
lim n → ∞ 1 n ( FIM y n ( θ ) ) k , ℓ = 1 p Tr ( τ − 1 ) ( μ ^ x , k c y c ) H [ C x ~ c y c ( 0 ) ] − 1 μ ^ x , ℓ c y c + 1 2 p ∫ 0 1 Tr ( [ C x ~ c y c ( f ) ] − 1 C x ~ , k c y c ′ ( f ) [ C x ~ c y c ( f ) ] − 1 C x ~ , ℓ c y c ′ ( f ) ) d f + c p 2 ∫ 0 1 Tr ( [ C x ~ c y c ( f ) ] − 1 C x ~ , k c y c ′ ( f ) ) d f ∫ 0 1 Tr ( [ C x ~ c y c ( f ) ] − 1 C x ~ , ℓ c y c ′ ( f ) ) d f . (15) \begin{aligned}
\lim_{n\to\infty}\frac{1}{n}\bigl(\operatorname{FIM}_{y_n}(\theta)\bigr)_{k,\ell}
={}&\frac{1}{p}\operatorname{Tr}\left(\tau^{-1}\right)\bigl(\widehat{\mu}_{x,k}^{\mathrm{cyc}}\bigr)^{H}
\left[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(0)\right]^{-1}
\widehat{\mu}_{x,\ell}^{\mathrm{cyc}}\\
&+\frac{1}{2p}\int_{0}^{1}\operatorname{Tr}\left(
\left[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)\right]^{-1}
\mathbf C_{\widetilde{x},k}^{\mathrm{cyc}\prime}(f)
\left[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)\right]^{-1}
\mathbf C_{\widetilde{x},\ell}^{\mathrm{cyc}\prime}(f)
\right)\,df\\
&+\frac{c}{p^{2}}\int_{0}^{1}\operatorname{Tr}\left(
\left[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)\right]^{-1}
\mathbf C_{\widetilde{x},k}^{\mathrm{cyc}\prime}(f)
\right)\,df
\int_{0}^{1}\operatorname{Tr}\left(
\left[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)\right]^{-1}
\mathbf C_{\widetilde{x},\ell}^{\mathrm{cyc}\prime}(f)
\right)\,df . \tag*{(15)}
\end{aligned} n → ∞ lim n 1 ( FIM y n ( θ ) ) k , ℓ = p 1 Tr ( τ − 1 ) ( μ x , k cyc ) H [ C x cyc ( 0 ) ] − 1 μ x , ℓ cyc + 2 p 1 ∫ 0 1 Tr ( [ C x cyc ( f ) ] − 1 C x , k cyc ′ ( f ) [ C x cyc ( f ) ] − 1 C x , ℓ cyc ′ ( f ) ) df + p 2 c ∫ 0 1 Tr ( [ C x cyc ( f ) ] − 1 C x , k cyc ′ ( f ) ) df ∫ 0 1 Tr ( [ C x cyc ( f ) ] − 1 C x , ℓ cyc ′ ( f ) ) df . (15) where μ ^ x , k c y c ≔ ( ( μ x , k c y c ′ ) T , ( μ x , k c y c ′ ) H ) T \widehat{\mu}_{x,k}^{\mathrm{cyc}}\coloneqq((\mu_{x,k}^{\mathrm{cyc}\prime})^{T},(\mu_{x,k}^{\mathrm{cyc}\prime})^{H})^{T} μ x , k cyc : = (( μ x , k cyc ′ ) T , ( μ x , k cyc ′ ) H ) T , C x ~ c y c ( 0 ) ≔ C x ~ c y c ( f ) ∣ f = 0 \mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(0)\coloneqq\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)|_{f=0} C x cyc ( 0 ) : = C x cyc ( f ) ∣ f = 0 and C x ~ , k c y c ′ ( f ) ≔ ∂ C x ~ c y c ( f ) ∂ θ k \mathbf C_{\widetilde{x},k}^{\mathrm{cyc}\prime}(f)\coloneqq\frac{\partial\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f)}{\partial\theta_k} C x , k cyc ′ ( f ) : = ∂ θ k ∂ C x cyc ( f ) . The ( k ′ , ℓ ′ ) (k^{\prime},\ell^{\prime}) ( k ′ , ℓ ′ ) th 2 m × 2 m 2m\times2m 2 m × 2 m block of the augmented poly-spectrum matrix C x ~ c y c ( f ) \mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f) C x cyc ( f ) is defined by 1 p C x ~ ( ( ℓ ′ − k ′ ) / p ) ( f − k ′ + 1 p ) \frac{1}{p}\mathbf C_{\widetilde{x}}^{((\ell^{\prime}-k^{\prime})/p)}\left(\frac{f-k^{\prime}+1}{p}\right) p 1 C x (( ℓ ′ − k ′ ) / p ) ( p f − k ′ + 1 ) , k ′ , ℓ ′ = 1 , … , p k^{\prime},\ell^{\prime}=1,\ldots,p k ′ , ℓ ′ = 1 , … , p .
where
C x ~ ( k / p ) ( f ) ≔ ( C x ( k / p ) ( f ) R x ( k / p ) ( f ) R x ( − k / p ) ∗ ( − f ) C x ( − k / p ) ∗ ( − f ) ) . \mathbf C_{\widetilde{x}}^{(k/p)}(f)\coloneqq
\begin{pmatrix}
\mathbf C_x^{(k/p)}(f) & \mathbf R_x^{(k/p)}(f)\\
\mathbf R_x^{(-k/p)*}(-f) & \mathbf C_x^{(-k/p)*}(-f)
\end{pmatrix}. C x ( k / p ) ( f ) : = ( C x ( k / p ) ( f ) R x ( − k / p ) ∗ ( − f ) R x ( k / p ) ( f ) C x ( − k / p ) ∗ ( − f ) ) . For multidimensional cyclostationary circular complex-valued CGPs, the pseudo-spectral matrices vanish (i.e., R x ( k / p ) ( f ) = 0 \mathbf R_x^{(k/p)}(f)=0 R x ( k / p ) ( f ) = 0 ), making C x ~ c y c ( f ) \mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f) C x cyc ( f ) block-diagonal. Consequently, (15) simplifies and can be expressed directly in terms of the m × m m\times m m × m spectral matrices C x c y c ( f ) \mathbf C_x^{\mathrm{cyc}}(f) C x cyc ( f ) by replacing ( 1 / 2 p , c / p 2 ) (1/2p,c/p^{2}) ( 1/2 p , c / p 2 ) , ( μ ^ x , k c y c ) H [ C x ~ c y c ( 0 ) ] − 1 μ ^ x , ℓ c y c (\widehat{\mu}_{x,k}^{\mathrm{cyc}})^{H}[\mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(0)]^{-1}\widehat{\mu}_{x,\ell}^{\mathrm{cyc}} ( μ x , k cyc ) H [ C x cyc ( 0 ) ] − 1 μ x , ℓ cyc , C x ~ c y c ( f ) \mathbf C_{\widetilde{x}}^{\mathrm{cyc}}(f) C x cyc ( f ) and C x ~ , k c y c ′ ( f ) \mathbf C_{\widetilde{x},k}^{\mathrm{cyc}\prime}(f) C x , k cyc ′ ( f ) by ( 1 / p , 2 c / p 2 ) (1/p,2c/p^{2}) ( 1/ p , 2 c / p 2 ) , 2 Re [ ( μ x , k c y c ′ ) H [ C x c y c ( 0 ) ] − 1 μ x , ℓ c y c ] 2\operatorname{Re}[(\mu_{x,k}^{\mathrm{cyc}\prime})^{H}[\mathbf C_x^{\mathrm{cyc}}(0)]^{-1}\mu_{x,\ell}^{\mathrm{cyc}}] 2 Re [( μ x , k cyc ′ ) H [ C x cyc ( 0 ) ] − 1 μ x , ℓ cyc ] , C x c y c ( f ) \mathbf C_x^{\mathrm{cyc}}(f) C x cyc ( f ) and C x , k c y c ′ ( f ) \mathbf C_{x,k}^{\mathrm{cyc}\prime}(f) C x , k cyc ′ ( f ) , respectively, where the ( k ′ , ℓ ′ ) (k^{\prime},\ell^{\prime}) ( k ′ , ℓ ′ ) th m × m m\times m m × m block of the poly-spectrum C x c y c ( f ) \mathbf C_x^{\mathrm{cyc}}(f) C x cyc ( f ) is defined by 1 p C x ( ( ℓ ′ − k ′ ) / p ) ( f − k ′ + 1 p ) \frac{1}{p}\mathbf C_x^{((\ell^{\prime}-k^{\prime})/p)}\left(\frac{f-k^{\prime}+1}{p}\right) p 1 C x (( ℓ ′ − k ′ ) / p ) ( p f − k ′ + 1 ) , k ′ , ℓ ′ = 1 , … , p k^{\prime},\ell^{\prime}=1,\ldots,p k ′ , ℓ ′ = 1 , … , p , where
C x ( k / p ) ( f ) ≔ ∑ τ ∈ Z C x ( k / p ) ( τ ) e − i 2 π τ f , k = 0 , 1 , … , p − 1. \mathbf C_x^{(k/p)}(f)\coloneqq\sum_{\tau\in\mathbb Z}\mathbf C_x^{(k/p)}(\tau)e^{-i2\pi\tau f},\quad k=0,1,\ldots,p-1. C x ( k / p ) ( f ) : = τ ∈ Z ∑ C x ( k / p ) ( τ ) e − i 2 π τ f , k = 0 , 1 , … , p − 1.
Finally, it is important to note that the conjecture proposed in [22 rels. (6), (7)] for circular complex Gaussian data, when expressed using our notation, takes the form: lim n → ∞ 1 n ( FIM y n ( θ ) ) k , ℓ ≈ 1 2 tr ( S − 1 ∂ S ∂ θ k S − 1 ∂ S ∂ θ ℓ ) \lim_{n\to\infty}\frac{1}{n}\left(\operatorname{FIM}_{y_n}(\boldsymbol{\theta})\right)_{k,\ell} \approx\frac{1}{2}\operatorname{tr}\left(\mathbf{S}^{-1}\frac{\partial\mathbf{S}}{\partial\theta_k}\mathbf{S}^{-1}\frac{\partial\mathbf{S}}{\partial\theta_\ell}\right) lim n → ∞ n 1 ( FIM y n ( θ ) ) k , ℓ ≈ 2 1 tr ( S − 1 ∂ θ k ∂ S S − 1 ∂ θ ℓ ∂ S ) , where the ( k ′ , ℓ ′ ) (k',\ell') ( k ′ , ℓ ′ ) -th block of S \mathbf{S} S is defined as [ S ] k ′ , ℓ ′ : = C x ~ ( k ′ − ℓ ′ p ) ( k ′ + ℓ ′ 2 p ) [\mathbf{S}]_{k',\ell'} \mathrel{\mathop:}= \mathbf{C}_{\tilde{x}}^{\left(\frac{k'-\ell'}{p}\right)}\left(\frac{k'+\ell'}{2p}\right) [ S ] k ′ , ℓ ′ : = C x ~ ( p k ′ − ℓ ′ ) ( 2 p k ′ + ℓ ′ ) where C x ( k p ) ( f ) \mathbf{C}_{x}^{\left(\frac{k}{p}\right)}(f) C x ( p k ) ( f ) represents the cyclic spectrum. This latter formula is grounded under any proof and differs significantly from Result 2 based on rigorous reasoning.
Application to linearly modulated digital communication signals
We give in this subsection meaningful interpretations of the various spectral matrices introduced previously, namely, C x ~ ′ ( f ) \mathbf{C}_{\tilde{x}}'(f) C x ~ ′ ( f ) , C x ′ ( f ) \mathbf{C}_{x'}(f) C x ′ ( f ) , C x ~ c y c ( f ) \mathbf{C}_{\tilde{x}}^{\mathrm{cyc}}(f) C x ~ cyc ( f ) and C x c y c ( f ) \mathbf{C}_{x}^{\mathrm{cyc}}(f) C x cyc ( f ) , by applying them to the case of noisy over-sampled pulse amplitude modulated signals. The continuous-time signal model is given by x ( t ) = ∑ ℓ ∈ Z b ℓ h ( t − ℓ T ) + w ( t ) ∈ C m \mathbf{x}(t)=\sum_{\ell\in\mathbb{Z}}b_\ell\mathbf{h}(t-\ell T)+\mathbf{w}(t)\in\mathbb{C}^{m} x ( t ) = ∑ ℓ ∈ Z b ℓ h ( t − ℓ T ) + w ( t ) ∈ C m , t ∈ R t\in\mathbb{R} t ∈ R where ( b ℓ ) ℓ ∈ Z ∈ C (b_\ell)_{\ell\in\mathbb{Z}}\in\mathbb{C} ( b ℓ ) ℓ ∈ Z ∈ C is a sequence of zero-mean independent identically distributed symbols with variance σ b 2 : = E ∣ b ℓ ∣ 2 \sigma_b^2\mathrel{\mathop:}=\mathbb{E}|b_\ell|^2 σ b 2 : = E ∣ b ℓ ∣ 2 , h ( t ) ∈ C m \mathbf{h}(t)\in\mathbb{C}^{m} h ( t ) ∈ C m is the overall impulse response and w ( t ) \mathbf{w}(t) w ( t ) is an additive, circular, temporally and spatially white noise with spectra C w ( f ) = σ w 2 I \mathbf{C}_{w}(f)=\sigma_w^2\mathbf{I} C w ( f ) = σ w 2 I . The over-sampled discrete-time process, denoted as x k : = x ( t ) ∣ t = k T / p \mathbf{x}_k\mathrel{\mathop:}=\mathbf{x}(t)|_{t=kT/p} x k : = x ( t ) ∣ t = k T / p , is cyclostationary with a cycle period equal to the oversampling rate p p p , and reduces to a stationary process when p = 1 p=1 p = 1 . Depending on whether the symbols b ℓ b_\ell b ℓ are real- or complex-valued, the process is noncircular or circular, respectively.
For this signal model, the conventional spectra of the blocked augmented process x ~ k ′ \tilde{\mathbf{x}}'_k x ~ k ′ and the blocked process x k ′ \mathbf{x}'_k x k ′ are respectively given by C x ~ ′ ( f ) = σ b 2 h ~ ′ ( f ) h ~ ′ H ( f ) + σ w 2 I \mathbf{C}_{\tilde{x}'}(f)=\sigma_b^2\tilde{\mathbf{h}}'(f)\tilde{\mathbf{h}}'^{H}(f)+\sigma_w^2\mathbf{I} C x ~ ′ ( f ) = σ b 2 h ~ ′ ( f ) h ~ ′ H ( f ) + σ w 2 I and C x ′ ( f ) = σ b 2 h ′ ( f ) h ′ H ( f ) + σ w 2 I \mathbf{C}_{x'}(f)=\sigma_b^2\mathbf{h}'(f)\mathbf{h}'^{H}(f)+\sigma_w^2\mathbf{I} C x ′ ( f ) = σ b 2 h ′ ( f ) h ′ H ( f ) + σ w 2 I where h ~ ′ ( f ) : = ∑ k ∈ Z h ~ k ′ e − i 2 π k f \tilde{\mathbf{h}}'(f)\mathrel{\mathop:}=\sum_{k\in\mathbb{Z}}\tilde{\mathbf{h}}'_k e^{-i2\pi kf} h ~ ′ ( f ) : = ∑ k ∈ Z h ~ k ′ e − i 2 π k f and h ′ ( f ) : = ∑ k ∈ Z h k ′ e − i 2 π k f \mathbf{h}'(f)\mathrel{\mathop:}=\sum_{k\in\mathbb{Z}}\mathbf{h}'_k e^{-i2\pi kf} h ′ ( f ) : = ∑ k ∈ Z h k ′ e − i 2 π k f . The blocked impulse response vectors are defined as h ~ k ′ : = [ h ~ T ( t ) ∣ t = k T + T / p , … , h ~ T ( t ) ∣ t = k T + p T / p ] T \tilde{\mathbf{h}}'_k\mathrel{\mathop:}=\left[\tilde{\mathbf{h}}^{T}(t)|_{t=kT+T/p},\ldots,\tilde{\mathbf{h}}^{T}(t)|_{t=kT+pT/p}\right]^{T} h ~ k ′ : = [ h ~ T ( t ) ∣ t = k T + T / p , … , h ~ T ( t ) ∣ t = k T + pT / p ] T with h ~ ( t ) : = [ h T ( t ) , h H ( t ) ] T \tilde{\mathbf{h}}(t)\mathrel{\mathop:}=\left[\mathbf{h}^{T}(t),\mathbf{h}^{H}(t)\right]^{T} h ~ ( t ) : = [ h T ( t ) , h H ( t ) ] T , and h k ′ : = [ h T ( t ) ∣ t = k T + T / p , … , h T ( t ) ∣ t = k T + p T / p ] T \mathbf{h}'_k\mathrel{\mathop:}=\left[\mathbf{h}^{T}(t)|_{t=kT+T/p},\ldots,\mathbf{h}^{T}(t)|_{t=kT+pT/p}\right]^{T} h k ′ : = [ h T ( t ) ∣ t = k T + T / p , … , h T ( t ) ∣ t = k T + pT / p ] T .
Under the simplifying assumption that x ~ k \tilde{\mathbf{x}}_k x ~ k and x k \mathbf{x}_k x k are zero-mean cyclostationary CGP, the asymptotic frequency-domain FIM rate extensions summarized in the last two columns of Table 1 apply directly. When applying these results, it is assumed that the channel responses h ~ ′ ( f ) \tilde{\mathbf{h}}'(f) h ~ ′ ( f ) and h ′ ( f ) \mathbf{h}'(f) h ′ ( f ) are known up to a finite number of unknown parameters (e.g., in specular multipath channels). It is worth noting that this formulation differs from standard blind channel estimation models for which (a) the number of consecutive non-zero time taps in h ~ k ′ \tilde{\mathbf{h}}'_k h ~ k ′ and h k ′ \mathbf{h}'_k h k ′ is finite and determined by the channel order; (b) the symbols b ℓ b_\ell b ℓ are deterministic, unknown parameters; and (c) specific constraints are imposed on the time taps to ensure identifiability (see, e.g., [23 ] and references therein).
The associated cyclic spectra are given by C x ~ ( k p ) ( f ) = σ b 2 p h ~ ( f ) h ~ H ( f − k p ) + δ k , 0 σ w 2 I \mathbf{C}_{\tilde{x}}^{\left(\frac{k}{p}\right)}(f)=\frac{\sigma_b^2}{p}\tilde{\mathbf{h}}(f)\tilde{\mathbf{h}}^{H}\left(f-\frac{k}{p}\right)+\delta_{k,0}\sigma_w^2\mathbf{I} C x ~ ( p k ) ( f ) = p σ b 2 h ~ ( f ) h ~ H ( f − p k ) + δ k , 0 σ w 2 I and C x ( k p ) ( f ) = σ b 2 p h ( f ) h H ( f − k p ) + δ k , 0 σ w 2 I \mathbf{C}_{x}^{\left(\frac{k}{p}\right)}(f)=\frac{\sigma_b^2}{p}\mathbf{h}(f)\mathbf{h}^{H}\left(f-\frac{k}{p}\right)+\delta_{k,0}\sigma_w^2\mathbf{I} C x ( p k ) ( f ) = p σ b 2 h ( f ) h H ( f − p k ) + δ k , 0 σ w 2 I for k = 0 , … , p − 1 k=0,\ldots,p-1 k = 0 , … , p − 1 , where h ~ ( f ) : = ∑ k h ~ k e − i 2 π k f \tilde{\mathbf{h}}(f)\mathrel{\mathop:}=\sum_k\tilde{\mathbf{h}}_k e^{-i2\pi kf} h ~ ( f ) : = ∑ k h ~ k e − i 2 π k f and h ( f ) : = ∑ k h k e − i 2 π k f \mathbf{h}(f)\mathrel{\mathop:}=\sum_k\mathbf{h}_k e^{-i2\pi kf} h ( f ) : = ∑ k h k e − i 2 π k f , with h ~ k : = ( h k T , h k H ) T \tilde{\mathbf{h}}_k\mathrel{\mathop:}=(\mathbf{h}^{T}_k,\mathbf{h}^{H}_k)^{T} h ~ k : = ( h k T , h k H ) T , and h k : = h ( t ) ∣ t = k T / p \mathbf{h}_k\mathrel{\mathop:}=\mathbf{h}(t)|_{t=kT/p} h k : = h ( t ) ∣ t = k T / p (see, e.g., [24 subsect. 17.4.3] for equivalent expressions with different notation). Consequently, it is straightforward to prove that the poly-spectral matrices defined in Result 2 exhibit the following rank-one structural characteristics:
C x ~ c y c ( f ) = σ b 2 p 2 h ~ c y c ( f ) ( h ~ c y c ( f ) ) H + σ w 2 p I and C x c y c ( f ) = σ b 2 p 2 h c y c ( f ) ( h c y c ( f ) ) H + σ w 2 p I (16) \mathbf{C}_{\tilde{x}}^{\mathrm{cyc}}(f)=\frac{\sigma_b^2}{p^2}\tilde{\mathbf{h}}^{\mathrm{cyc}}(f)\left(\tilde{\mathbf{h}}^{\mathrm{cyc}}(f)\right)^{H}+\frac{\sigma_w^2}{p}\mathbf{I}
\quad\text{and}\quad
\mathbf{C}_{x}^{\mathrm{cyc}}(f)=\frac{\sigma_b^2}{p^2}\mathbf{h}^{\mathrm{cyc}}(f)\left(\mathbf{h}^{\mathrm{cyc}}(f)\right)^{H}+\frac{\sigma_w^2}{p}\mathbf{I} \tag*{(16)} C x ~ cyc ( f ) = p 2 σ b 2 h ~ cyc ( f ) ( h ~ cyc ( f ) ) H + p σ w 2 I and C x cyc ( f ) = p 2 σ b 2 h cyc ( f ) ( h cyc ( f ) ) H + p σ w 2 I (16)
where the stacked frequency responses are defined as: h ~ c y c ( f ) ≔ ( h ~ T ( f p ) , h ~ T ( f − 1 p ) , … , h ~ T ( f − p + 1 p ) ) T \tilde{\mathbf{h}}^{\mathrm{cyc}}(f)\coloneqq\left(\tilde{\mathbf{h}}^{T}\left(\frac{f}{p}\right),\tilde{\mathbf{h}}^{T}\left(\frac{f-1}{p}\right),\ldots,\tilde{\mathbf{h}}^{T}\left(\frac{f-p+1}{p}\right)\right)^{T} h ~ cyc ( f ) : = ( h ~ T ( p f ) , h ~ T ( p f − 1 ) , … , h ~ T ( p f − p + 1 ) ) T and h c y c ( f ) ≔ ( h T ( f p ) , h T ( f − 1 p ) , … , h T ( f − p + 1 p ) ) T \mathbf{h}^{\mathrm{cyc}}(f)\coloneqq\left(\mathbf{h}^{T}\left(\frac{f}{p}\right),\mathbf{h}^{T}\left(\frac{f-1}{p}\right),\ldots,\mathbf{h}^{T}\left(\frac{f-p+1}{p}\right)\right)^{T} h cyc ( f ) : = ( h T ( p f ) , h T ( p f − 1 ) , … , h T ( p f − p + 1 ) ) T . The expressions in (16) for the poly-spectrum, which were introduced by Gladyshev in [18 ] , provide a meaningful interpretation and, to the best of our knowledge, represent a novel contribution to the field.
Finally, it is important to note that over-sampled offset QAM and, more generally, over-sampled quasi-rectilinear modulations [25 ] x ( t ) = ∑ ℓ b ℓ h ( t − ℓ T ) x(t)=\sum_{\ell}b_{\ell}h(t-\ell T) x ( t ) = ∑ ℓ b ℓ h ( t − ℓ T ) , where b ℓ b_{\ell} b ℓ are real-valued, are also noncircular cyclostationary but, with a period equal to twice the oversampling rate p p p because E ( x k + j p x ℓ + j p H ) = E ( x k x ℓ H ) \mathop{\mathrm{E}}(x_{k+jp}x_{\ell+jp}^{H})=\mathop{\mathrm{E}}(x_kx_{\ell}^{H}) E ( x k + j p x ℓ + j p H ) = E ( x k x ℓ H ) and E ( x k + j 2 p x ℓ + j 2 p T ) = E ( x k x ℓ T ) \mathop{\mathrm{E}}(x_{k+j2p}x_{\ell+j2p}^{T})=\mathop{\mathrm{E}}(x_kx_{\ell}^{T}) E ( x k + j 2 p x ℓ + j 2 p T ) = E ( x k x ℓ T ) for arbitrary j ∈ Z j\in\mathbb{Z} j ∈ Z .
Extension to particular multidimensional (wide-sense) almost cyclostationary complex-valued CGP
We notice that the methodology developed in this paper does not, in its present form, cover a general zero-mean wide-sense almost cyclostationary CGP, for which the covariance and pseudo-covariance sequences may contain arbitrary, not necessarily harmonically related, cycle frequencies. More precisely, a zero-mean complex-valued process is wide-sense almost cyclostationary if there exist sets of cycle frequencies ( α k ) k ∈ Z (\alpha_k)_{k\in\mathbb{Z}} ( α k ) k ∈ Z and ( β k ) k ∈ Z (\beta_k)_{k\in\mathbb{Z}} ( β k ) k ∈ Z such that4
E ( x ℓ x ℓ + τ T ) = ∑ k ∈ Z R x α k ( τ ) e i 2 π α k ℓ , E ( x ℓ x ℓ + τ H ) = ∑ k ∈ Z C x β k ( τ ) e i 2 π β k ℓ . (17) \mathop{\mathrm{E}}(x_{\ell}x_{\ell+\tau}^{T})=\sum_{k\in\mathbb{Z}}\mathbf{R}_{x}^{\alpha_k}(\tau)e^{i2\pi\alpha_k\ell},
\qquad
\mathop{\mathrm{E}}(x_{\ell}x_{\ell+\tau}^{H})=\sum_{k\in\mathbb{Z}}\mathbf{C}_{x}^{\beta_k}(\tau)e^{i2\pi\beta_k\ell}. \tag*{(17)} E ( x ℓ x ℓ + τ T ) = k ∈ Z ∑ R x α k ( τ ) e i 2 π α k ℓ , E ( x ℓ x ℓ + τ H ) = k ∈ Z ∑ C x β k ( τ ) e i 2 π β k ℓ . (17)
Although a general almost cyclostationary process is beyond the scope of the present methodology, an important practically relevant subclass can be treated. Consider
x k ≔ e i 2 π k θ 0 u k ∈ C m , k ∈ Z , (18) x_k\coloneqq e^{i2\pi k\theta_0}u_k\in\mathbb{C}^{m},\qquad k\in\mathbb{Z}, \tag*{(18)} x k : = e i 2 π k θ 0 u k ∈ C m , k ∈ Z , (18)
where ( u k ) k ∈ Z (u_k)_{k\in\mathbb{Z}} ( u k ) k ∈ Z is a zero-mean circular or noncircular complex-valued CGP that is either stationary or cyclostationary with period p p p . This model describes a common impairment in digital communications, namely, a carrier-frequency offset (CFO) θ 0 \theta_0 θ 0 resulting from a frequency mismatch between the transmitter and receiver local oscillators.
The effect of the frequency offset on the second-order statistics depends on the circularity and cyclostationarity of the underlying process. In particular, if u k u_k u k is circular and stationary, then x k x_k x k remains circular and stationary, with spectral density matrix C x ( f ) = C u ( f + θ 0 ) \mathbf{C}_x(f)=\mathbf{C}_u(f+\theta_0) C x ( f ) = C u ( f + θ 0 ) . The other cases are summarized in Table 2 .
Table 2. Table.2. Properties of ( u k , x k ) (u_k,x_k) ( u k , x k ) under a frequency offset.
When ( u k ) k ∈ Z (u_k)_{k\in\mathbb{Z}} ( u k ) k ∈ Z is an arbitrary (non necessarily stationary or cyclostationary) zero-mean CGP, the r.v. y n = ( x 1 T , … , x n T ) T ∈ C m n \mathbf{y}_n=(\mathbf{x}_1^T,\ldots,\mathbf{x}_n^T)^T\in\mathbb{C}^{mn} y n = ( x 1 T , … , x n T ) T ∈ C mn where x k \mathbf{x}_k x k is given by (18), is therefore also CG distributed. Applying the complex-valued extension of the Slepian-Bangs formula for CG distributions established in [26 ] to y n \mathbf{y}_n y n , yields the following result proven in the Appendix.
Result 3: For arbitrary circular [resp. noncircular] zero-mean complex-valued CGP ( u k ) k ∈ Z (\mathbf{u}_k)_{k\in\mathbb{Z}} ( u k ) k ∈ Z affected by the frequency offset θ 0 \theta_0 θ 0 , the parameter θ 0 \theta_0 θ 0 and the parameters ( θ 1 , … , θ q ) (\theta_1,\ldots,\theta_q) ( θ 1 , … , θ q ) of the covariance Σ z n : = E ( z n z n H ) \boldsymbol{\Sigma}_{\mathbf{z}_n}\mathrel{\mathop:}=\mathbb{E}(\mathbf{z}_n\mathbf{z}_n^H) Σ z n : = E ( z n z n H ) where z n : = ( u 1 T , … , u n T ) T \mathbf{z}_n\mathrel{\mathop:}=(\mathbf{u}_1^T,\ldots,\mathbf{u}_n^T)^T z n : = ( u 1 T , … , u n T ) T [resp. Σ ~ z n : = E ( z ~ n z ~ n H ) \widetilde{\boldsymbol{\Sigma}}_{\mathbf{z}_n}\mathrel{\mathop:}=\mathbb{E}(\widetilde{\mathbf{z}}_n\widetilde{\mathbf{z}}_n^H) Σ z n : = E ( z n z n H ) , where z ~ n : = ( u ~ 1 T , … , u ~ n T ) T \widetilde{\mathbf{z}}_n\mathrel{\mathop:}=(\widetilde{\mathbf{u}}_1^T,\ldots,\widetilde{\mathbf{u}}_n^T)^T z n : = ( u 1 T , … , u n T ) T ], are decoupled in the ( q + 1 ) × ( q + 1 ) (q+1)\times(q+1) ( q + 1 ) × ( q + 1 ) FIM y n ( θ ) \operatorname{FIM}_{\mathbf{y}_n}(\boldsymbol{\theta}) FIM y n ( θ ) (i.e., [ FIM y n ( θ ) ] 0 , k = 0 [\operatorname{FIM}_{\mathbf{y}_n}(\boldsymbol{\theta})]_{0,k}=0 [ FIM y n ( θ ) ] 0 , k = 0 for k = 1 , … , q k=1,\ldots,q k = 1 , … , q ) and furthermore [ FIM y n ( θ ) ] k , ℓ = [ FIM z n ( θ ) ] k , ℓ [\operatorname{FIM}_{\mathbf{y}_n}(\boldsymbol{\theta})]_{k,\ell}=[\operatorname{FIM}_{\mathbf{z}_n}(\boldsymbol{\theta})]_{k,\ell} [ FIM y n ( θ ) ] k , ℓ = [ FIM z n ( θ ) ] k , ℓ for k , ℓ = 1 , … , q k,\ell=1,\ldots,q k , ℓ = 1 , … , q .
Consequently, whenever the underlying process u k u_k u k satisfies the assumptions of the stationary or cyclostationary frequency-domain results developed above, the corresponding asymptotic frequency-domain FIM rate for the covariance parameters ( θ 1 , … , θ q ) (\theta_1,\ldots,\theta_q) ( θ 1 , … , θ q ) is obtained directly by applying the complex stationary formula (6) (as summarized in Table 1 ) and Result 2 to u k u_k u k , without explicitly incorporating the frequency offset θ 0 \theta_0 θ 0 . Thus, for this single-component model, the frequency offset does not modify the asymptotic frequency-domain FIM rate for the parameters governing the covariance and pseudo-covariance spectra of the underlying process.
Section V provides an illustrative application of this result, including a closed-form expression for the asymptotic CRB on θ 0 \theta_0 θ 0 for the model (18). It is also instructive to consider the multi-component model x k = ∑ ℓ = 1 q 0 e i 2 π k θ 0 , ℓ u k , ℓ x_k=\sum_{\ell=1}^{q_0}e^{i2\pi k\theta_{0,\ell}}u_{k,\ell} x k = ∑ ℓ = 1 q 0 e i 2 π k θ 0 , ℓ u k , ℓ , where the processes ( u k , ℓ ) ℓ = 1 , … , q 0 (u_{k,\ell})_{\ell=1,\ldots,q_0} ( u k , ℓ ) ℓ = 1 , … , q 0 are mutually uncorrelated, circular, and stationary, with spectra C u ℓ ( f ) C_{u_\ell}(f) C u ℓ ( f ) . In this case, x k x_k x k is circular and stationary and has spectrum C x ( f ) = ∑ ℓ = 1 q 0 C u ℓ ( f + θ 0 , ℓ ) C_x(f)=\sum_{\ell=1}^{q_0}C_{u_\ell}(f+\theta_{0,\ell}) C x ( f ) = ∑ ℓ = 1 q 0 C u ℓ ( f + θ 0 , ℓ ) . Therefore, the complex stationary extension of (6) applies directly, with parameter vector θ = ( θ 0 , 1 , … , θ 0 , q 0 , θ 1 , 1 , … , θ 1 , q 1 , … , θ q 0 , 1 , … , θ q 0 , q q 0 ) \boldsymbol{\theta}=(\theta_{0,1},\ldots,\theta_{0,q_0},\theta_{1,1},\ldots,\theta_{1,q_1},\ldots,\theta_{q_0,1},\ldots,\theta_{q_0,q_{q_0}}) θ = ( θ 0 , 1 , … , θ 0 , q 0 , θ 1 , 1 , … , θ 1 , q 1 , … , θ q 0 , 1 , … , θ q 0 , q q 0 ) , where ( θ ℓ , 1 , … , θ ℓ , q ℓ ) (\theta_{\ell,1},\ldots,\theta_{\ell,q_\ell}) ( θ ℓ , 1 , … , θ ℓ , q ℓ ) parameterizes C u ℓ ( f ) C_{u_\ell}(f) C u ℓ ( f ) .
In contrast to the single-component model, the frequency-offset parameters ( θ 0 , 1 , … , θ 0 , q 0 ) (\theta_{0,1},\ldots,\theta_{0,q_0}) ( θ 0 , 1 , … , θ 0 , q 0 ) and the spectral parameters ( θ 1 , 1 , … , θ 1 , q 1 , … , θ q 0 , 1 , … , θ q 0 , q q 0 ) (\theta_{1,1},\ldots,\theta_{1,q_1},\ldots,\theta_{q_0,1},\ldots,\theta_{q_0,q_{q_0}}) ( θ 1 , 1 , … , θ 1 , q 1 , … , θ q 0 , 1 , … , θ q 0 , q q 0 ) are generally no longer FIM-decoupled. Consequently, the different frequency offsets can affect the CRBs of the spectral parameters, and conversely; see, e.g., [27 ] . A subsequent section provides an illustrative example showing the impact of frequency offsets on the CRB for DOA estimation.
Application to DOA Estimation
This section illustrates the influence of the spectral characteristics of the source signals on the CRB for estimating DOA parameters, including scenarios in which the measurements are affected by a carrier frequency offset. More specifically, we consider one or two narrowband far-field signals impinge on an arbitrary array of m m m sensors. After frequency down-shifting the filtered sensor signals to baseband, the generated complex envelope, sampled at time k k k is modeled as
x k = A s k + n k ∈ C m . (19) \mathbf{x}_k=\mathbf{A}\mathbf{s}_k+\mathbf{n}_k\in\mathbb{C}^{m}. \tag*{(19)} x k = A s k + n k ∈ C m . (19)
The steering matrix A \mathbf{A} A is either A = a 1 \mathbf{A}=\mathbf{a}_1 A = a 1 or [ a 1 , a 2 ] [\mathbf{a}_1,\mathbf{a}_2] [ a 1 , a 2 ] where a i \mathbf{a}_i a i , i = 1 , 2 i=1,2 i = 1 , 2 are the steering vectors parameterized by the DOA ω i \omega_i ω i with ∥ a i ∥ 2 = m \lVert\mathbf{a}_i\rVert^2=m ∥ a i ∥ 2 = m . We define α i ≔ 2 d i H Π a i ⊥ d i \alpha_i\coloneqq2\mathbf{d}_i^H\boldsymbol{\Pi}_{\mathbf{a}_i}^{\perp}\mathbf{d}_i α i : = 2 d i H Π a i ⊥ d i with d i ≔ d a i d ω i \mathbf{d}_i\coloneqq\frac{\mathrm{d}\mathbf{a}_i}{\mathrm{d}\omega_i} d i : = d ω i d a i and where Π a i ⊥ \boldsymbol{\Pi}_{\mathbf{a}_i}^{\perp} Π a i ⊥ denotes the orthogonal projector onto the orthogonal complement of a i \mathbf{a}_i a i . The measurement noise ( n k ) k ∈ Z (\mathbf{n}_k)_{k\in\mathbb{Z}} ( n k ) k ∈ Z is assumed to be zero-mean, circular complex-valued, spatially uncorrelated with E ( n k n k H ) = σ n 2 I \mathbb{E}(\mathbf{n}_k\mathbf{n}_k^H)=\sigma_n^2\mathbf{I} E ( n k n k H ) = σ n 2 I and white stationary with spectrum C n ( f ) = σ n 2 I \mathbf{C}_n(f)=\sigma_n^2\mathbf{I} C n ( f ) = σ n 2 I . The source signals, denoted as s k = s k , 1 \mathbf{s}_k=s_{k,1} s k = s k , 1 or ( s k , 1 , s k , 2 ) T (s_{k,1},s_{k,2})^T ( s k , 1 , s k , 2 ) T are zero-mean complex-valued, and not correlated with n k \mathbf{n}_k n k . Furthermore, we assume that the process ( x k ) k ∈ Z (\mathbf{x}_k)_{k\in\mathbb{Z}} ( x k ) k ∈ Z is CG distributed5 . In contrast to all previously published performance studies, we assume here that the source signals ( s k ) k ∈ Z (\mathbf{s}_k)_{k\in\mathbb{Z}} ( s k ) k ∈ Z are not necessarily independent, but are instead stationary, cyclostationary or almost cyclostationary signals.
Stationary CGP
In this instance, the source signals ( s k ) k ∈ Z (\mathbf{s}_k)_{k\in\mathbb{Z}} ( s k ) k ∈ Z are stationary, and are either circular or rectilinear (i.e., strictly second-order noncircular). The SNRs are defined as r i ≔ σ s i 2 σ n 2 r_i\coloneqq\frac{\sigma_{s_i}^2}{\sigma_n^2} r i : = σ n 2 σ s i 2 with σ s i 2 ≔ E ( ∣ s k , i ∣ 2 ) \sigma_{s_i}^2\coloneqq\mathbb{E}(\lvert s_{k,i}\rvert^2) σ s i 2 : = E (∣ s k , i ∣ 2 ) .
For circular sources with temporal correlation C s ( k ) ≔ E ( s ℓ s ℓ + k ∗ ) C_s(k)\coloneqq\mathbb{E}(s_\ell s_{\ell+k}^*) C s ( k ) : = E ( s ℓ s ℓ + k ∗ ) and spectrum C s ( f ) C_s(f) C s ( f ) , model (19) implies that the observations have covariance C x ( k ) = A C s ( k ) A H + σ n 2 δ k , 0 I \mathbf{C}_x(k)=\mathbf{A}\mathbf{C}_s(k)\mathbf{A}^H+\sigma_n^2\delta_{k,0}\mathbf{I} C x ( k ) = A C s ( k ) A H + σ n 2 δ k , 0 I and spectrum C x ( f ) = A C s ( f ) A H + σ n 2 I \mathbf{C}_x(f)=\mathbf{A}\mathbf{C}_s(f)\mathbf{A}^H+\sigma_n^2\mathbf{I} C x ( f ) = A C s ( f ) A H + σ n 2 I .
For rectilinear sources, each source can be written as s k , i = ρ k , i e i ϕ i s_{k,i}=\rho_{k,i}e^{\mathrm{i}\phi_i} s k , i = ρ k , i e i ϕ i , i = 1 , 2 i=1,2 i = 1 , 2 where ϕ i \phi_i ϕ i is a fixed (deterministic) phase and ρ k , i ∈ R \rho_{k,i}\in\mathbb{R} ρ k , i ∈ R is a real-valued amplitude process. Let ρ k = ρ k , 1 \boldsymbol{\rho}_k=\rho_{k,1} ρ k = ρ k , 1 (single source) or ρ k = ( ρ k , 1 , ρ k , 2 ) T \boldsymbol{\rho}_k=(\rho_{k,1},\rho_{k,2})^T ρ k = ( ρ k , 1 , ρ k , 2 ) T (two sources), and define the temporal covariance R ρ ( k ) ≔ E ( ρ ℓ ρ ℓ + k T ) \mathbf{R}_{\rho}(k)\coloneqq\mathbb{E}(\boldsymbol{\rho}_\ell\boldsymbol{\rho}_{\ell+k}^T) R ρ ( k ) : = E ( ρ ℓ ρ ℓ + k T ) with associated spectrum R ρ ( f ) \mathbf{R}_{\rho}(f) R ρ ( f ) . This results in noncircular measurements x k \mathbf{x}_k x k , which therefore necessitates the use of the extended model of (19), as
x ~ k = A ~ ρ k + n ~ k ∈ C 2 m , (20) \widetilde{\mathbf{x}}_k=\widetilde{\mathbf{A}}\boldsymbol{\rho}_k+\widetilde{\mathbf{n}}_k\in\mathbb{C}^{2m}, \tag*{(20)} x k = A ρ k + n k ∈ C 2 m , (20)
with A ~ = a ~ 1 \widetilde{\mathbf{A}} = \widetilde{\mathbf{a}}_1 A = a 1 (single source) or [ a ~ 1 , a ~ 2 ] [\widetilde{\mathbf{a}}_1,\widetilde{\mathbf{a}}_2] [ a 1 , a 2 ] (two sources) where a ~ i = d e f ( a i T e − i ϕ i , a i H e − i ϕ i ) T \widetilde{\mathbf{a}}_i \mathrel{\mathop=^{\mathrm{def}}} (\mathbf{a}_i^{T}e^{-\mathrm{i}\phi_i},\mathbf{a}_i^{H}e^{-\mathrm{i}\phi_i})^{T} a i = def ( a i T e − i ϕ i , a i H e − i ϕ i ) T , i = 1 , 2 i=1,2 i = 1 , 2 . This model implies that C x ~ ( k ) = A ~ R ρ ( k ) A ~ H + σ n 2 δ k , 0 I \mathbf{C}_{\widetilde{x}}(k) = \widetilde{\mathbf{A}}\mathbf{R}_{\rho}(k)\widetilde{\mathbf{A}}^{H} + \sigma_n^2\delta_{k,0}\mathbf{I} C x ( k ) = A R ρ ( k ) A H + σ n 2 δ k , 0 I and associated spectrum C x ~ ( f ) = A ~ R ρ ( f ) A ~ H + σ n 2 I \mathbf{C}_{\widetilde{x}}(f) = \widetilde{\mathbf{A}}\mathbf{R}_{\rho}(f)\widetilde{\mathbf{A}}^{H} + \sigma_n^2\mathbf{I} C x ( f ) = A R ρ ( f ) A H + σ n 2 I .
In the case of a single source scenario, we assume that the spectrum C s 1 ( f ) \mathbf{C}_{s_1}(f) C s 1 ( f ) (for circular sources) and R ρ 1 ( f ) \mathbf{R}_{\rho_1}(f) R ρ 1 ( f ) (for rectilinear sources) are parameterized by the power σ s 1 2 \sigma_{s_1}^2 σ s 1 2 and a real-valued parameter6 {}^{6} 6 a 1 a_1 a 1 , such that C s 1 ( f ) = σ s 1 2 C a 1 ( f ) \mathbf{C}_{s_1}(f) = \sigma_{s_1}^2 C_{a_1}(f) C s 1 ( f ) = σ s 1 2 C a 1 ( f ) with ∫ 0 1 C a 1 ( f ) d f = 1 \int_0^1 C_{a_1}(f)df = 1 ∫ 0 1 C a 1 ( f ) df = 1 and R ρ 1 ( f ) = σ s 1 2 R a 1 ( f ) \mathbf{R}_{\rho_1}(f) = \sigma_{s_1}^2 R_{a_1}(f) R ρ 1 ( f ) = σ s 1 2 R a 1 ( f ) with ∫ 0 1 R a 1 ( f ) d f = 1 \int_0^1 R_{a_1}(f)df = 1 ∫ 0 1 R a 1 ( f ) df = 1 . The unknown parameters of our models are therefore θ = ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) T \boldsymbol{\theta} = (\omega_1,a_1,\sigma_{s_1}^2,\sigma_n^2)^T θ = ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) T for a single circular source, and θ = ( ω 1 , ϕ 1 , a 1 , σ s 1 2 , σ n 2 ) T \boldsymbol{\theta} = (\omega_1,\phi_1,a_1,\sigma_{s_1}^2,\sigma_n^2)^T θ = ( ω 1 , ϕ 1 , a 1 , σ s 1 2 , σ n 2 ) T for a single rectilinear source. Applying the inversion matrix lemma to the spectra C x ( f ) = σ s 1 2 C a 1 ( f ) a 1 a 1 H + σ n 2 I \mathbf{C}_x(f) = \sigma_{s_1}^2 C_{a_1}(f)\mathbf{a}_1\mathbf{a}_1^H + \sigma_n^2\mathbf{I} C x ( f ) = σ s 1 2 C a 1 ( f ) a 1 a 1 H + σ n 2 I and C x ~ ( f ) = σ s 1 2 R a 1 ( f ) a ~ 1 a ~ 1 H + σ n 2 I \mathbf{C}_{\widetilde{x}}(f) = \sigma_{s_1}^2 R_{a_1}(f)\widetilde{\mathbf{a}}_1\widetilde{\mathbf{a}}_1^H + \sigma_n^2\mathbf{I} C x ( f ) = σ s 1 2 R a 1 ( f ) a 1 a 1 H + σ n 2 I yields the following result, whose main steps of the proof are given in the Appendix.
Result 4: For a single circular [resp., rectilinear] source, the DOA parameter ω 1 \omega_1 ω 1 and the parameters ( a 1 , σ s 1 2 , σ n 2 ) (a_1,\sigma_{s_1}^2,\sigma_n^2) ( a 1 , σ s 1 2 , σ n 2 ) [resp., ( ϕ 1 , a 1 , σ s 1 2 , σ n 2 ) (\phi_1,a_1,\sigma_{s_1}^2,\sigma_n^2) ( ϕ 1 , a 1 , σ s 1 2 , σ n 2 ) ] are decoupled in the asymptotic frequency-domain FIM rate (6) . Moreover, the DOA-related FIM entries reduce solely on the second term of (6) . Consequently, the asymptotic CRB for ω 1 \omega_1 ω 1 is independent of the underlying CG distribution and is given by:
lim n → ∞ n CRB c ( ω 1 ) = 1 α 1 r 1 ∫ 0 1 m r 1 C a 1 2 ( f ) 1 + m r 1 C a 1 ( f ) d f . (21) \lim_{n\to\infty} n\operatorname{CRB}_{c}(\omega_1) = \frac{1}{\alpha_1 r_1\int_0^1 \frac{m_{r_1}C_{a_1}^2(f)}{1+m_{r_1}C_{a_1}(f)}df}. \tag*{(21)} n → ∞ lim n CRB c ( ω 1 ) = α 1 r 1 ∫ 0 1 1 + m r 1 C a 1 ( f ) m r 1 C a 1 2 ( f ) df 1 . (21)
and
lim n → ∞ n CRB r ( ω 1 ) = 1 2 α 1 r 1 ∫ 0 1 m r 1 R a 1 2 ( f ) 1 + 2 m r 1 R a 1 ( f ) d f , (22) \lim_{n\to\infty} n\operatorname{CRB}_{r}(\omega_1) = \frac{1}{2\alpha_1 r_1\int_0^1 \frac{m_{r_1}R_{a_1}^2(f)}{1+2m_{r_1}R_{a_1}(f)}df}, \tag*{(22)} n → ∞ lim n CRB r ( ω 1 ) = 2 α 1 r 1 ∫ 0 1 1 + 2 m r 1 R a 1 ( f ) m r 1 R a 1 2 ( f ) df 1 , (22)
respectively. Furthermore, these CRBs satisfy the following inequalities:
1 α 1 r 1 < lim n → ∞ n CRB c ( ω 1 ) ≤ 1 α 1 r 1 ( 1 + 1 m r 1 ) (23) \frac{1}{\alpha_1r_1} < \lim_{n\to\infty} n\operatorname{CRB}_{c}(\omega_1) \le\frac{1}{\alpha_1r_1}\left(1+\frac{1}{m r_1}\right) \tag*{(23)} α 1 r 1 1 < n → ∞ lim n CRB c ( ω 1 ) ≤ α 1 r 1 1 ( 1 + m r 1 1 ) (23)
and
1 α 1 r 1 < lim n → ∞ n CRB r ( ω 1 ) ≤ 1 2 α 1 r 1 ( 2 + 1 m r 1 ) , (24) \frac{1}{\alpha_1r_1} < \lim_{n\to\infty} n\operatorname{CRB}_{r}(\omega_1) \le\frac{1}{2\alpha_1r_1}\left(2+\frac{1}{m r_1}\right), \tag*{(24)} α 1 r 1 1 < n → ∞ lim n CRB r ( ω 1 ) ≤ 2 α 1 r 1 1 ( 2 + m r 1 1 ) , (24)
where the upper bound are achieved when C a 1 ( f ) = 1 C_{a_1}(f) = 1 C a 1 ( f ) = 1 and R a 1 ( f ) = 1 R_{a_1}(f) = 1 R a 1 ( f ) = 1 , respectively.
It is worth noting that (21) and (22) align with the classical expressions previously established for independent identically complex circular or rectilinear Gaussian signals [28 ] , In that case, C a 1 ( f ) = 1 C_{a_1}(f) = 1 C a 1 ( f ) = 1 and R a 1 ( f ) = 1 R_{a_1}(f) = 1 R a 1 ( f ) = 1 , yielding
CRB c ( ω 1 ) = 1 n α 1 r 1 ( 1 + 1 m r 1 ) (25) \operatorname{CRB}_{c}(\omega_1) = \frac{1}{n\alpha_1r_1}\left(1+\frac{1}{m r_1}\right) \tag*{(25)} CRB c ( ω 1 ) = n α 1 r 1 1 ( 1 + m r 1 1 ) (25)
and
CRB r ( ω 1 ) = 1 2 n α 1 r 1 ( 2 + 1 m r 1 ) . (26) \operatorname{CRB}_{r}(\omega_1)=\frac{1}{2n\alpha_1r_1}\left(2+\frac{1}{mr_1}\right). \tag*{(26)} CRB r ( ω 1 ) = 2 n α 1 r 1 1 ( 2 + m r 1 1 ) . (26)
Moreover, inequalities (23) and (24) are also consistent with the deterministic CRB (DCRB), obtained when the source signals are modeled as unknown deterministic parameters. In that case, DCRB c ( ω 1 ) = DCRB r ( ω 1 ) = 1 n α 1 r 1 \operatorname{DCRB}_{c}(\omega_1)=\operatorname{DCRB}_{r}(\omega_1)=\frac{1}{n\alpha_1r_1} DCRB c ( ω 1 ) = DCRB r ( ω 1 ) = n α 1 r 1 1 , which is upper bounded by the stochastic CRB. Furthermore, equalities (21) and (22) are not overly sensitive to the spectral shape C a 1 ( f ) C_{a_1}(f) C a 1 ( f ) [resp. R a 1 ( f ) R_{a_1}(f) R a 1 ( f ) ] when m r 1 ≫ 1 mr_1\gg1 m r 1 ≫ 1 .
We now illustrate Result 4 with two examples:
For a flat spectrum7 in [ b , b + a 1 ] ⊂ [ 0 , 1 ] [b,b+a_1]\subset[0,1] [ b , b + a 1 ] ⊂ [ 0 , 1 ] (21), substituting into (21) and (22) yields
lim n → ∞ n CRB c ( ω 1 ) = 1 α 1 r 1 ( 1 + a 1 m r 1 ) , (27) \lim_{n\to\infty}n\operatorname{CRB}_{c}(\omega_1)=\frac{1}{\alpha_1r_1}\left(1+\frac{a_1}{mr_1}\right), \tag*{(27)} n → ∞ lim n CRB c ( ω 1 ) = α 1 r 1 1 ( 1 + m r 1 a 1 ) , (27)
and
lim n → ∞ n CRB r ( ω 1 ) = 1 2 α 1 r 1 ( 2 + a 1 m r 1 ) . (28) \lim_{n\to\infty}n\operatorname{CRB}_{r}(\omega_1)=\frac{1}{2\alpha_1r_1}\left(2+\frac{a_1}{mr_1}\right). \tag*{(28)} n → ∞ lim n CRB r ( ω 1 ) = 2 α 1 r 1 1 ( 2 + m r 1 a 1 ) . (28)
For an AR(1) spectrum C a 1 ( f ) = R a 1 ( f ) = 1 − ∣ a 1 ∣ 2 ∣ 1 − a 1 e − i 2 π f ∣ 2 C_{a_1}(f)=R_{a_1}(f)=\frac{1-|a_1|^2}{|1-a_1e^{-i2\pi f}|^2} C a 1 ( f ) = R a 1 ( f ) = ∣1 − a 1 e − i 2 π f ∣ 2 1 − ∣ a 1 ∣ 2 with ∣ a 1 ∣ < 1 |a_1|<1 ∣ a 1 ∣ < 1 . After lengthy algebraic manipulations, (21) and (22) yield
lim n → ∞ n CRB c ( ω 1 ) = 1 α 1 r 1 ( 1 + 1 − ∣ a 1 ∣ 2 c − ( 1 − ∣ a 1 ∣ 2 ) ) , (29) \lim_{n\to\infty}n\operatorname{CRB}_{c}(\omega_1)=\frac{1}{\alpha_1r_1}\left(1+\frac{1-|a_1|^2}{c-(1-|a_1|^2)}\right), \tag*{(29)} n → ∞ lim n CRB c ( ω 1 ) = α 1 r 1 1 ( 1 + c − ( 1 − ∣ a 1 ∣ 2 ) 1 − ∣ a 1 ∣ 2 ) , (29)
and
lim n → ∞ n CRB r ( ω 1 ) = 1 α 1 r 1 ( 1 + 1 − ∣ a 1 ∣ 2 c − ( 1 − ∣ a 1 ∣ 2 ) ) , (30) \lim_{n\to\infty}n\operatorname{CRB}_{r}(\omega_1)=\frac{1}{\alpha_1r_1}\left(1+\frac{1-|a_1|^2}{c-(1-|a_1|^2)}\right), \tag*{(30)} n → ∞ lim n CRB r ( ω 1 ) = α 1 r 1 1 ( 1 + c − ( 1 − ∣ a 1 ∣ 2 ) 1 − ∣ a 1 ∣ 2 ) , (30)
where c : = { ( 1 − ∣ a 1 ∣ 2 ) [ ( 1 + m 1 r 1 ) 2 − ∣ a 1 ∣ 2 ( 1 − m 1 r 1 ) 2 ] } 1 / 2 c\mathrel{\mathop:}=\left\{(1-|a_1|^2)\left[(1+m_1r_1)^2-|a_1|^2(1-m_1r_1)^2\right]\right\}^{1/2} c : = { ( 1 − ∣ a 1 ∣ 2 ) [ ( 1 + m 1 r 1 ) 2 − ∣ a 1 ∣ 2 ( 1 − m 1 r 1 ) 2 ] } 1/2 , with m 1 = m m_1=m m 1 = m for the circular case and m 1 = 2 m m_1=2m m 1 = 2 m for the rectilinear case.
It can be verified from (27)–(28) and (29)–(30) that:
For the flat spectrum (27)–(28), as a 1 → 1 a_1\to1 a 1 → 1 (white spectrum), the CRBs tend to the Gaussian CRBs (25)–(26) (upper bounds). Conversely, as a 1 → 0 a_1\to0 a 1 → 0 (narrowband spectrum), the CRBs tend to the lower bound 1 α 1 r 1 \frac{1}{\alpha_1r_1} α 1 r 1 1 .
For the AR(1) spectrum (29)–(30), as ∣ a 1 ∣ → 0 |a_1|\to0 ∣ a 1 ∣ → 0 (white spectrum), the CRBs tend to the Gaussian CRBs (25)–(26) (upper bounds). Conversely, as ∣ a 1 ∣ → 1 |a_1|\to1 ∣ a 1 ∣ → 1 (narrowband/peak spectrum), the CRBs tend to the lower bound 1 α 1 r 1 \frac{1}{\alpha_1r_1} α 1 r 1 1 .
These results confirm that the lower bound 1 α 1 r 1 \frac{1}{\alpha_{1}r_{1}} α 1 r 1 1 in (23)–(24) is approached when the source spectrum becomes highly concentrated (narrowband), while the upper bound is achieved for spectrally white sources.
In the two-source scenario, we assume that the source processes ( s k , 1 ) k ∈ Z (s_{k,1})_{k\in\mathbb{Z}} ( s k , 1 ) k ∈ Z and ( s k , 2 ) k ∈ Z (s_{k,2})_{k\in\mathbb{Z}} ( s k , 2 ) k ∈ Z mutually uncorrelated and are either both circular or both rectilinear. This implies that the spectrum of x k \mathbf{x}_{k} x k [resp., x ~ k \widetilde{\mathbf{x}}_{k} x k ] is given by C x ( f ) = C s 1 ( f ) a 1 a 1 H + C s 2 ( f ) a 2 a 2 H + σ n 2 I \mathbf{C}_{x}(f)=C_{s_{1}}(f)\mathbf{a}_{1}\mathbf{a}_{1}^{H}+C_{s_{2}}(f)\mathbf{a}_{2}\mathbf{a}_{2}^{H}+\sigma_{n}^{2}\mathbf{I} C x ( f ) = C s 1 ( f ) a 1 a 1 H + C s 2 ( f ) a 2 a 2 H + σ n 2 I [resp., C ~ x ( f ) = R ρ 1 ( f ) a ~ 1 a ~ 1 H + R ρ 2 ( f ) a ~ 2 a ~ 2 H + σ n 2 I \widetilde{\mathbf{C}}_{x}(f)=R_{\rho_{1}}(f)\widetilde{\mathbf{a}}_{1}\widetilde{\mathbf{a}}_{1}^{H}+R_{\rho_{2}}(f)\widetilde{\mathbf{a}}_{2}\widetilde{\mathbf{a}}_{2}^{H}+\sigma_{n}^{2}\mathbf{I} C x ( f ) = R ρ 1 ( f ) a 1 a 1 H + R ρ 2 ( f ) a 2 a 2 H + σ n 2 I ] where the source spectra C s i ( f ) C_{s_{i}}(f) C s i ( f ) , [resp., R ρ i ( f ) R_{\rho_{i}}(f) R ρ i ( f ) ] i = 1 , 2 i=1,2 i = 1 , 2 , are parameterized by the powers σ s i 2 \sigma_{s_{i}}^{2} σ s i 2 and real-valued spectral shape parameters a i a_{i} a i , such that C s i ( f ) = σ s i 2 C a i ( f ) C_{s_{i}}(f)=\sigma_{s_{i}}^{2}C_{a_{i}}(f) C s i ( f ) = σ s i 2 C a i ( f ) [resp., R ρ i ( f ) = σ s i 2 R a i ( f ) R_{\rho_{i}}(f)=\sigma_{s_{i}}^{2}R_{a_{i}}(f) R ρ i ( f ) = σ s i 2 R a i ( f ) ] with ∫ 0 1 C a i ( f ) d f = 1 \int_{0}^{1}C_{a_{i}}(f)\mathrm{d}f=1 ∫ 0 1 C a i ( f ) d f = 1 [resp., ∫ 0 1 R a i ( f ) d f = 1 \int_{0}^{1}R_{a_{i}}(f)\mathrm{d}f=1 ∫ 0 1 R a i ( f ) d f = 1 ]. The unknown parameter vector is therefore θ = ( ω 1 , ω 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T \boldsymbol{\theta}=(\omega_{1},\omega_{2},a_{1},a_{2},\sigma_{s_{1}}^{2},\sigma_{s_{2}}^{2},\sigma_{n}^{2})^{T} θ = ( ω 1 , ω 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T for circular sources, and θ = ( ω 1 , ω 2 , ϕ 1 , ϕ 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T \boldsymbol{\theta}=(\omega_{1},\omega_{2},\phi_{1},\phi_{2},a_{1},a_{2},\sigma_{s_{1}}^{2},\sigma_{s_{2}}^{2},\sigma_{n}^{2})^{T} θ = ( ω 1 , ω 2 , ϕ 1 , ϕ 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T for rectilinear sources. Applying the matrix inversion lemma to the spectral matrices C x ( f ) \mathbf{C}_{x}(f) C x ( f ) and C ~ x ( f ) \widetilde{\mathbf{C}}_{x}(f) C x ( f ) , as detailed in the Appendix, yields the following result.
Result 5: In the circular [resp., rectilinear] two-source model previously described and for spectra without overlap (i.e., spectrum support B i ⊂ [ 0 , 1 ] B_{i}\subset[0,1] B i ⊂ [ 0 , 1 ] , such that B 1 ∩ B 2 = ∅ B_{1}\cap B_{2}=\varnothing B 1 ∩ B 2 = ∅ ), the DOA parameters ω 1 \omega_{1} ω 1 [resp., ( ω 1 , ϕ 1 ) (\omega_{1},\phi_{1}) ( ω 1 , ϕ 1 ) ] and ω 2 \omega_{2} ω 2 [resp., ( ω 2 , ϕ 2 ) (\omega_{2},\phi_{2}) ( ω 2 , ϕ 2 ) ] are decoupled from each other in the asymptotic frequency domain FIM rate (6). Moreover, the DOA parameters ( ω 1 , ω 2 ) (\omega_{1},\omega_{2}) ( ω 1 , ω 2 ) [resp., ( ω 1 , ϕ 1 , ω 2 , ϕ 2 ) (\omega_{1},\phi_{1},\omega_{2},\phi_{2}) ( ω 1 , ϕ 1 , ω 2 , ϕ 2 ) ] are FIM-decoupled from ( σ s 1 2 , σ s 2 2 , σ n 2 ) (\sigma_{s_{1}}^{2},\sigma_{s_{2}}^{2},\sigma_{n}^{2}) ( σ s 1 2 , σ s 2 2 , σ n 2 ) . Consequently, the FIM submatrix associated with the DOA parameters is determined solely by the second term of (6) and is therefore independent of the CG distribution. The asymptotic CRB for each DOA ω i \omega_{i} ω i , i = 1 , 2 i=1,2 i = 1 , 2 , is then given by
lim n → ∞ n CRB c ( ω i ) = 1 α i r i ∫ B i m r i C a i 2 ( f ) 1 + m r i C a i ( f ) d f (31) \lim_{n\to\infty}n\operatorname{CRB}_{c}(\omega_{i})=\frac{1}{\alpha_{i}r_{i}\int_{B_{i}}\frac{m r_{i}C_{a_{i}}^{2}(f)}{1+m r_{i}C_{a_{i}}(f)}\mathrm{d}f} \tag*{(31)} n → ∞ lim n CRB c ( ω i ) = α i r i ∫ B i 1 + m r i C a i ( f ) m r i C a i 2 ( f ) d f 1 (31)
and
lim n → ∞ n CRB r ( ω i ) = 1 2 α i r i ∫ B i m r i R a i 2 ( f ) 1 + 2 m r i R a i ( f ) d f . (32) \lim_{n\to\infty}n\operatorname{CRB}_{r}(\omega_{i})=\frac{1}{2\alpha_{i}r_{i}\int_{B_{i}}\frac{m r_{i}R_{a_{i}}^{2}(f)}{1+2m r_{i}R_{a_{i}}(f)}\mathrm{d}f}. \tag*{(32)} n → ∞ lim n CRB r ( ω i ) = 2 α i r i ∫ B i 1 + 2 m r i R a i ( f ) m r i R a i 2 ( f ) d f 1 . (32)
This result is illustrated in Figure 1 and Figure 2 for a circular and a rectilinear source, respectively. We consider two equipowered sources with flat spectra C s 1 ( f ) = R s 1 ( f ) = σ s 2 a 1 [ 0 , a ] ( f ) C_{s_{1}}(f)=R_{s_{1}}(f)=\frac{\sigma_{s}^{2}}{a}\mathbb{1}_{[0,a]}(f) C s 1 ( f ) = R s 1 ( f ) = a σ s 2 1 [ 0 , a ] ( f ) and C s 2 ( f ) = R s 2 ( f ) = σ s 2 a 1 [ 1 − a , 1 ] ( f ) C_{s_{2}}(f)=R_{s_{2}}(f)=\frac{\sigma_{s}^{2}}{a}\mathbb{1}_{[1-a,1]}(f) C s 2 ( f ) = R s 2 ( f ) = a σ s 2 1 [ 1 − a , 1 ] ( f ) where a ∈ ( 0 , 1 ] a\in(0,1] a ∈ ( 0 , 1 ] controls the degree of spectral overlap between the two sources. The sensor array is a uniform linear array (ULA) of m = 6 m=6 m = 6 elements with steering vectors a i = ( 1 , e i ω i , … , e i ( m − 1 ) ω i ) T \mathbf{a}_{i}=(1,e^{\mathrm{i}\omega_{i}},\ldots,e^{\mathrm{i}(m-1)\omega_{i}})^{T} a i = ( 1 , e i ω i , … , e i ( m − 1 ) ω i ) T , i = 1 , 2 i=1,2 i = 1 , 2 . For this symmetric configuration, lim n → ∞ n CRB ( ω 1 ) = lim n → ∞ n CRB ( ω 2 ) \lim_{n\to\infty}n\operatorname{CRB}(\omega_{1})=\lim_{n\to\infty}n\operatorname{CRB}(\omega_{2}) lim n → ∞ n CRB ( ω 1 ) = lim n → ∞ n CRB ( ω 2 ) , and we examine several values of the angular separation Δ ω : = ω 2 − ω 1 \Delta\omega\mathrel{\mathop:}=\omega_{2}-\omega_{1} Δ ω : = ω 2 − ω 1 and phase difference Δ ϕ : = ϕ 2 − ϕ 1 \Delta\phi\mathrel{\mathop:}=\phi_{2}-\phi_{1} Δ ϕ : = ϕ 2 − ϕ 1 .
Figure 1. lim n → ∞ n CRB c ( ω i ) \lim_{n \to\infty} n\operatorname{CRB}_{c}(\omega_i) lim n → ∞ n CRB c ( ω i ) as a function of a a a for different values of Δ ω \Delta\omega Δ ω for SNR = 0dB for two circular sources.
Figure 2. lim n → ∞ n C R B r ( ω i ) \lim_{n\to\infty} n\mathrm{CRB}_{r}(\omega_i) lim n → ∞ n CRB r ( ω i ) as a function of a a a for different values of Δ ω \Delta\omega Δ ω and Δ ϕ \Delta\phi Δ ϕ for SNR = 0 =0 = 0 dB for two rectilinear sources.
We verify that for non-overlapping spectra (i.e., a ∈ ( 0 , 1 / 2 ] a\in(0,1/2] a ∈ ( 0 , 1/2 ] ), the asymptotic CRB, lim n → ∞ n CRB ( ω i ) \lim_{n\to\infty}n\operatorname{CRB}(\omega_{i}) lim n → ∞ n CRB ( ω i ) , is exactly given by (27)-(28), confirming the decoupling property established in Result 5. Moreover, it can be observed that the CRB increases significantly as the spectral overlap increases (i.e., as a a a increases beyond 1 / 2 1/2 1/2 ) and as Δ ω \Delta\omega Δ ω decreases for circular sources [resp., as both Δ ω \Delta\omega Δ ω and Δ ϕ \Delta\phi Δ ϕ decrease for rectilinear sources]. This degradation in estimation performance is due to the increased difficulty in resolving closely spaced sources with overlapping spectral content.
Finally, for the case of complete spectral overlap (a = 1 a = 1 a = 1 ), we observe that the numerical value of lim n → ∞ n ⋅ CRB ( ω i ) \lim_{n \to\infty} n \cdot\operatorname{CRB}(\omega_i) lim n → ∞ n ⋅ CRB ( ω i ) computed from (6) coincides exactly with the closed-form expression n CRB c ( ω i ) n \operatorname{CRB}_{c}(\omega_i) n CRB c ( ω i ) given in [29 eq. (25)] , which was derived for the classical case of i.i.d. complex circular Gaussian signals x k \mathbf{x}_k x k .
Cyclostationary CGP
In this subsection, we consider a single source s k , 1 s_{k,1} s k , 1 that is cyclostationary with cycle period p p p , and is either circular or noncircular to deal with the linearly modulated digital communication signals presented in Subsection IV-B. As established in Subsections III-B and IV-A, the extension of the asymptotic frequency-domain FIM rate (6) to complex-valued cyclostationary signals applies when n n n is a multiple of p p p , the number of samples of the resulting stationary blocked process is n / p n/p n / p .
In the circular case, the blocked observation vector x k ′ \mathbf{x}'_k x k ′ defined in (10) can be written as x k ′ = A ′ s k , 1 ′ + n k ′ ∈ C m p \mathbf{x}'_k=\mathbf{A}'\mathbf{s}'_{k,1}+\mathbf{n}'_k\in\mathbb{C}^{mp} x k ′ = A ′ s k , 1 ′ + n k ′ ∈ C m p , where A ′ : = Blkdiag ( a 1 , … , a 1 ) ∈ C m p × p \mathbf{A}'\mathrel{\mathop:}=\operatorname{Blkdiag}(\mathbf{a}_1,\ldots,\mathbf{a}_1)\in\mathbb{C}^{mp\times p} A ′ : = Blkdiag ( a 1 , … , a 1 ) ∈ C m p × p (p p p identical blocks), s k , 1 ′ = ( s ( k − 1 ) p + 1 , 1 , … , s k p , 1 ) T ∈ C p \mathbf{s}'_{k,1}=(s_{(k-1)p+1,1},\ldots,s_{kp,1})^{T}\in\mathbb{C}^{p} s k , 1 ′ = ( s ( k − 1 ) p + 1 , 1 , … , s k p , 1 ) T ∈ C p , and n k ′ = ( n ( k − 1 ) p + 1 , … , n k p ) T ∈ C m p \mathbf{n}'_k=(n_{(k-1)p+1},\ldots,n_{kp})^{T}\in\mathbb{C}^{mp} n k ′ = ( n ( k − 1 ) p + 1 , … , n k p ) T ∈ C m p . Consequently, the covariance sequence and spectrum of the blocked process are respectively given by C x ′ ( ℓ ) : = E ( x k ′ x k + ℓ ′ H ) = A ′ C s ′ ( ℓ ) A ′ H + σ n 2 δ ℓ , 0 I m p \mathbf{C}_{\mathbf{x}'}(\ell)\mathrel{\mathop:}=\mathbb{E}(\mathbf{x}'_k\mathbf{x}'^{H}_{k+\ell})=\mathbf{A}'\mathbf{C}_{\mathbf{s}'}(\ell)\mathbf{A}'^{H}+\sigma_n^2\delta_{\ell,0}\mathbf{I}_{mp} C x ′ ( ℓ ) : = E ( x k ′ x k + ℓ ′ H ) = A ′ C s ′ ( ℓ ) A ′ H + σ n 2 δ ℓ , 0 I m p and C x ′ ( f ) = A ′ C s ′ ( f ) A ′ H + σ n 2 I m p \mathbf{C}_{\mathbf{x}'}(f)=\mathbf{A}'\mathbf{C}_{\mathbf{s}'}(f)\mathbf{A}'^{H}+\sigma_n^2\mathbf{I}_{mp} C x ′ ( f ) = A ′ C s ′ ( f ) A ′ H + σ n 2 I m p , where C s ′ ( ℓ ) \mathbf{C}_{\mathbf{s}'}(\ell) C s ′ ( ℓ ) and C s ′ ( f ) \mathbf{C}_{\mathbf{s}'}(f) C s ′ ( f ) are the p × p p\times p p × p covariance sequence and spectrum of the blocked source process ( s ~ k , 1 ) k ∈ Z (\tilde{s}_{k,1})_{k\in\mathbb{Z}} ( s ~ k , 1 ) k ∈ Z . The poly-spectral matrix of the source s k , 1 s_{k,1} s k , 1 as defined in (12) is denoted by C s 1 c y c ( f ) \mathbf{C}_{s_1}^{\mathrm{cyc}}(f) C s 1 cyc ( f ) .
In the noncircular case, the signal x ~ k ′ \tilde{\mathbf{x}}'_k x ~ k ′ defined from x ~ k \tilde{\mathbf{x}}_k x ~ k similarly as x k ′ \mathbf{x}'_k x k ′ from x k \mathbf{x}_k x k in (10) , is written in the form
x ~ k ′ = A ~ ′ s ~ k , 1 ′ + n ~ k ′ ∈ C 2 m p \tilde{\mathbf{x}}'_k=\tilde{\mathbf{A}}'\tilde{\mathbf{s}}'_{k,1}+\tilde{\mathbf{n}}'_k\in\mathbb{C}^{2mp} x ~ k ′ = A ~ ′ s ~ k , 1 ′ + n ~ k ′ ∈ C 2 m p
where A ~ ′ : = Blkdiag ( a 1 , a 1 ∗ , … , a 1 , a 1 ∗ ) ∈ C 2 m p × 2 p \tilde{\mathbf{A}}'\mathrel{\mathop{\mathrel{\mathop:}\mkern-1.2mu=}}\operatorname{Blkdiag}(\mathbf{a}_1,\mathbf{a}_1^*,\ldots,\mathbf{a}_1,\mathbf{a}_1^*)\in\mathbb{C}^{2mp\times2p} A ~ ′ : = Blkdiag ( a 1 , a 1 ∗ , … , a 1 , a 1 ∗ ) ∈ C 2 m p × 2 p , and s ~ k , 1 ′ \tilde{\mathbf{s}}'_{k,1} s ~ k , 1 ′ and n ~ k ′ \tilde{\mathbf{n}}'_k n ~ k ′ defined from s ~ k , 1 \tilde{\mathbf{s}}_{k,1} s ~ k , 1 and n ~ k \tilde{\mathbf{n}}_k n ~ k , respectively similarly as x k ′ \mathbf{x}'_k x k ′ from x k \mathbf{x}_k x k . This model also implies that C x ~ ′ ( k ) = A ~ ′ C s ~ ′ ( k ) A ~ ′ H + σ n 2 δ k , 0 I \mathbf{C}_{\tilde{x}'}(k)=\tilde{\mathbf{A}}'\mathbf{C}_{\tilde{s}'}(k)\tilde{\mathbf{A}}'^{H}+\sigma_n^2\delta_{k,0}\mathbf{I} C x ~ ′ ( k ) = A ~ ′ C s ~ ′ ( k ) A ~ ′ H + σ n 2 δ k , 0 I and C x ~ ′ ( f ) = A ~ ′ C s ~ ′ ( f ) A ~ ′ H + σ n 2 I \mathbf{C}_{\tilde{x}'}(f)=\tilde{\mathbf{A}}'\mathbf{C}_{\tilde{s}'}(f)\tilde{\mathbf{A}}'^{H}+\sigma_n^2\mathbf{I} C x ~ ′ ( f ) = A ~ ′ C s ~ ′ ( f ) A ~ ′ H + σ n 2 I , where C s ~ ′ ( k ) \mathbf{C}_{\tilde{s}'}(k) C s ~ ′ ( k ) and C s ~ ′ ( f ) \mathbf{C}_{\tilde{s}'}(f) C s ~ ′ ( f ) are the 2 p × 2 p 2p\times2p 2 p × 2 p augmented covariance sequence and spectrum of ( s ~ k , 1 ′ ) k ∈ Z (\tilde{\mathbf{s}}'_{k,1})_{k\in\mathbb{Z}} ( s ~ k , 1 ′ ) k ∈ Z . The poly-spectral matrix of the augmented source s ~ k , 1 \tilde{\mathbf{s}}_{k,1} s ~ k , 1 , as defined in (12) , is denoted by C s ~ 1 c y c ( f ) \mathbf{C}_{\tilde{s}_1}^{\mathrm{cyc}}(f) C s ~ 1 cyc ( f ) .
The source power is defined as the average power over one cycle period, σ s 1 2 : = 1 p ∑ k = 1 p E ( ∣ s k , 1 ∣ 2 ) \sigma_{s_1}^2\mathrel{\mathop{\mathrel{\mathop:}\mkern-1.2mu=}}\frac{1}{p}\sum_{k=1}^{p}\mathbb{E}(|s_{k,1}|^2) σ s 1 2 : = p 1 ∑ k = 1 p E ( ∣ s k , 1 ∣ 2 ) , and the SNR as r 1 : = σ s 1 2 / σ n 2 r_1\mathrel{\mathop{\mathrel{\mathop:}\mkern-1.2mu=}}\sigma_{s_1}^2/\sigma_n^2 r 1 : = σ s 1 2 / σ n 2 . Using the Wiener-Khinchin theorem for the blocked process, the instantaneous power can be expressed as E ( ∣ s k , 1 ∣ 2 ) = ∫ 0 1 [ C s 1 ( f ) ] k , k d f \mathbb{E}(|s_{k,1}|^2)=\int_0^1[\mathbf{C}_{s_1}(f)]_{k,k}\,\mathrm{d}f E ( ∣ s k , 1 ∣ 2 ) = ∫ 0 1 [ C s 1 ( f ) ] k , k d f . We assume that the source spectral matrices C s ′ ( f ) \mathbf{C}_{s'}(f) C s ′ ( f ) and C s ~ ′ ( f ) \mathbf{C}_{\tilde{s}'}(f) C s ~ ′ ( f ) are parameterized by the source power σ s 1 2 \sigma_{s_1}^2 σ s 1 2 and a real-valued spectral shape parameter a 1 a_1 a 1 . Specifically, we write C s ′ ( f ) = σ s 1 2 C a 1 ( f ) \mathbf{C}_{s'}(f)=\sigma_{s_1}^2\mathbf{C}_{a_1}(f) C s ′ ( f ) = σ s 1 2 C a 1 ( f ) and C s ~ ′ ( f ) = σ s 1 2 C a ~ 1 ( f ) \mathbf{C}_{\tilde{s}'}(f)=\sigma_{s_1}^2\mathbf{C}_{\tilde{a}_1}(f) C s ~ ′ ( f ) = σ s 1 2 C a ~ 1 ( f ) , where the normalized spectral shape matrices satisfy the trace conditions ∫ 0 1 tr ( C a 1 ′ ( f ) ) d f = p \int_0^1\operatorname{tr}(\mathbf{C}_{a_1'}(f))\,\mathrm{d}f=p ∫ 0 1 tr ( C a 1 ′ ( f )) d f = p and ∫ 0 1 tr ( C a ~ 1 ′ ( f ) ) d f = 2 p \int_0^1\operatorname{tr}(\mathbf{C}_{\tilde{a}_1'}(f))\,\mathrm{d}f=2p ∫ 0 1 tr ( C a ~ 1 ′ ( f )) d f = 2 p , respectively. Consequently, the corresponding poly-spectral (cyclic) matrices C s 1 c y c ( f ) \mathbf{C}_{s_1}^{\mathrm{cyc}}(f) C s 1 cyc ( f ) and C s ~ 1 c y c ( f ) \mathbf{C}_{\tilde{s}_1}^{\mathrm{cyc}}(f) C s ~ 1 cyc ( f ) inherit this parameterization, taking the form C s 1 c y c ( f ) = σ s 1 2 C a 1 c y c ( f ) \mathbf{C}_{s_1}^{\mathrm{cyc}}(f)=\sigma_{s_1}^2\mathbf{C}_{a_1}^{\mathrm{cyc}}(f) C s 1 cyc ( f ) = σ s 1 2 C a 1 cyc ( f ) and C s ~ 1 c y c ( f ) = σ s 1 2 C a ~ 1 c y c ( f ) \mathbf{C}_{\tilde{s}_1}^{\mathrm{cyc}}(f)=\sigma_{s_1}^2\mathbf{C}_{\tilde{a}_1}^{\mathrm{cyc}}(f) C s ~ 1 cyc ( f ) = σ s 1 2 C a ~ 1 cyc ( f ) . The unknown parameter vector for both the circular and noncircular models is therefore θ = ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) T \boldsymbol{\theta}=(\omega_1,a_1,\sigma_{s_1}^2,\sigma_n^2)^{T} θ = ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) T . Applying the matrix inversion lemma twice to the observation spectra C x ′ ( f ) \mathbf{C}_{x'}(f) C x ′ ( f ) and C x ~ ′ ( f ) \mathbf{C}_{\tilde{x}'}(f) C x ~ ′ ( f ) yields the following result, the main steps of which are detailed in the Appendix:
Result 6: For a single cyclostationary source of period p p p , either circular or noncircular, and for n n n a multiple of p p p , the DOA parameter ω 1 \omega_1 ω 1 is FIM-decoupled from the nuisance parameters ( a 1 , σ s 1 2 , σ n 2 ) (a_1,\sigma_{s_1}^2,\sigma_n^2) ( a 1 , σ s 1 2 , σ n 2 ) in the asymptotic frequency-domain FIM rate (6) . Moreover, the DOA-related FIM entry reduces to the second term of (6) . Consequently it does not depend on the CG distribution, and the asymptotic CRB on ω 1 \omega_1 ω 1 is given by
lim n → ∞ n CRB c ( ω 1 ) = 1 α 1 r 1 p ∫ 0 1 m r 1 Tr [ C a 1 c y c ( f ) ( I p + m r 1 p C a 1 c y c ( f ) ) − 1 C a 1 c y c ( f ) ] d f , (33) \lim_{n\to\infty}n\,\operatorname{CRB}_{c}(\omega_1)=\frac{1}{\alpha_1r_1p\int_0^1mr_1\operatorname{Tr}\left[\mathbf{C}_{a_1}^{\mathrm{cyc}}(f)\left(\mathbf{I}_p+mr_1p\mathbf{C}_{a_1}^{\mathrm{cyc}}(f)\right)^{-1}\mathbf{C}_{a_1}^{\mathrm{cyc}}(f)\right]\mathrm{d}f}, \tag*{(33)} n → ∞ lim n CRB c ( ω 1 ) = α 1 r 1 p ∫ 0 1 m r 1 Tr [ C a 1 cyc ( f ) ( I p + m r 1 p C a 1 cyc ( f ) ) − 1 C a 1 cyc ( f ) ] d f 1 , (33)
for the circular case, and
lim n → ∞ n CRB n c ( ω 1 ) = 2 α 1 r 1 p ∫ 0 1 m r 1 Tr [ C a ~ 1 c y c ( f ) ( I p + m r 1 p C a ~ 1 c y c ( f ) ) − 1 C a ~ 1 c y c ( f ) ] d f , (34) \lim_{n\to\infty}n\,\operatorname{CRB}_{nc}(\omega_1)=\frac{2}{\alpha_1r_1p\int_0^1mr_1\operatorname{Tr}\left[\mathbf{C}_{\tilde{a}_1}^{\mathrm{cyc}}(f)\left(\mathbf{I}_p+mr_1p\mathbf{C}_{\tilde{a}_1}^{\mathrm{cyc}}(f)\right)^{-1}\mathbf{C}_{\tilde{a}_1}^{\mathrm{cyc}}(f)\right]\mathrm{d}f}, \tag*{(34)} n → ∞ lim n CRB n c ( ω 1 ) = α 1 r 1 p ∫ 0 1 m r 1 Tr [ C a ~ 1 cyc ( f ) ( I p + m r 1 p C a ~ 1 cyc ( f ) ) − 1 C a ~ 1 cyc ( f ) ] d f 2 , (34)
for the noncircular case.
Finally, we verify that for the stationary case (p = 1 p=1 p = 1 ), we have C a 1 c y c ( f ) = C a 1 ( f ) \mathbf{C}_{a_1}^{\mathrm{cyc}}(f)=\mathbf{C}_{a_1}(f) C a 1 cyc ( f ) = C a 1 ( f ) . Consequently, (33) naturally reduces to (21) , recovering the classical result for a circular stationary source. But since noncircular sources are not necessarily rectilinear, (34) do not reduce to (22) for p = 1 p=1 p = 1 .
Stationary and cyclostationary CGP with frequency offset
For a single source s k , 1 s_{k,1} s k , 1 that is circular or noncircular, stationary or cyclostationary, and affected by a frequency offset f 1 ∈ ( 0 , 1 ) f_1 \in(0,1) f 1 ∈ ( 0 , 1 ) , model (19) becomes:
x k = e i 2 π k f 1 s k , 1 a 1 + n k . (35) \mathbf{x}_k = e^{i2\pi kf_1}s_{k,1}\mathbf{a}_1+\mathbf{n}_k. \tag*{(35)} x k = e i 2 π k f 1 s k , 1 a 1 + n k . (35)
This model is equivalent to the second-order model (18) with u k ≔ d e f s k , 1 a 1 + n k ′ \mathbf{u}_k \mathrel{\mathop{\coloneqq}^{\mathrm{def}}} s_{k,1}\mathbf{a}_1+\mathbf{n}'_k u k : = def s k , 1 a 1 + n k ′ where n k ′ ≔ d e f e − i 2 π k f 1 n k \mathbf{n}'_k \mathrel{\mathop{\coloneqq}^{\mathrm{def}}} e^{-i2\pi kf_1}\mathbf{n}_k n k ′ : = def e − i 2 π k f 1 n k . Since ( n k ) k ∈ Z (\mathbf{n}_k)_{k\in\mathbb{Z}} ( n k ) k ∈ Z is circular, the phase-rotated noise ( n k ′ ) k ∈ Z (\mathbf{n}'_k)_{k\in\mathbb{Z}} ( n k ′ ) k ∈ Z has identical second-order statistics, so ( u k ) k ∈ Z (\mathbf{u}_k)_{k\in\mathbb{Z}} ( u k ) k ∈ Z inherits the same covariance and pseudo-covariance structure as s k , 1 a 1 + n k s_{k,1}\mathbf{a}_1+\mathbf{n}_k s k , 1 a 1 + n k .
Consequently, Result 3 applies, and therefore, Results 4, 5 and 6 giving an interpretable closed-form expression of lim n → ∞ n C R B c ( ω 1 ) \lim_{n\to\infty} n\,\mathrm{CRB}_c(\omega_1) lim n → ∞ n CRB c ( ω 1 ) , lim n → ∞ n C R B r ( ω 1 ) \lim_{n\to\infty} n\,\mathrm{CRB}_r(\omega_1) lim n → ∞ n CRB r ( ω 1 ) and lim n → ∞ n C R B n c ( ω 1 ) \lim_{n\to\infty} n\,\mathrm{CRB}_{nc}(\omega_1) lim n → ∞ n CRB n c ( ω 1 ) , respectively, also apply. Furthermore, for the specific case of a circular stationary source, the following result is proven in the Appendix.
Result 7: For a circular stationary source following model (35) with frequency offset f 1 f_1 f 1 , the asymptotic CRB on f 1 f_1 f 1 is given by
lim n → ∞ n C R B c ( f 1 ) = 1 m 2 r 1 2 ∫ 0 1 ( ∂ C a 1 ( f ) ∂ f ) 2 ( 1 + m r 1 2 C a 1 ( f ) ) 2 d f (36) \lim_{n\to\infty} n\,\mathrm{CRB}_c(f_1)=\frac{1}{m^2r_1^2\int_0^1\frac{\left(\frac{\partial C_{a_1}(f)}{\partial f}\right)^2}{\left(1+m r_1^2 C_{a_1}(f)\right)^2}\,\mathrm{d}f} \tag*{(36)} n → ∞ lim n CRB c ( f 1 ) = m 2 r 1 2 ∫ 0 1 ( 1 + m r 1 2 C a 1 ( f ) ) 2 ( ∂ f ∂ C a 1 ( f ) ) 2 d f 1 (36)
Note that the integral in (36) is invariant to the frequency shift f 1 f_1 f 1 (since the integration is over a full period [ 0 , 1 ] [0,1] [ 0 , 1 ] ), so the CRB on f 1 f_1 f 1 does not depend on the actual value of the offset. Two important observations follow from (36):
The frequency offset f 1 f_1 f 1 is unidentifiable when the source spectrum is flat (C a 1 ( f ) C_{a_1}(f) C a 1 ( f ) constant), since ∂ C a 1 ( f ) / ∂ f = 0 \partial C_{a_1}(f)/\partial f=0 ∂ C a 1 ( f ) / ∂ f = 0 yields an infinite CRB.
The asymptotic CRB on f 1 f_1 f 1 is independent of the sensor array geometry, as the expression does not involve the geometric sensitivity parameter α 1 \alpha_1 α 1 or the steering vector a 1 \mathbf{a}_1 a 1 .
This behavior is illustrated in Figure 3 , which considers a circular stationary source with an AR(1) spectrum C a 1 ( f ) = 1 − ∣ a 1 ∣ 2 ∣ 1 − a 1 e − i 2 π f ∣ 2 C_{a_1}(f)=\frac{1-\lvert a_1\rvert^2}{\lvert1-a_1e^{-i2\pi f}\rvert^2} C a 1 ( f ) = ∣ 1 − a 1 e − i 2 π f ∣ 2 1 − ∣ a 1 ∣ 2 . As ∣ a 1 ∣ → 0 \lvert a_1\rvert\to0 ∣ a 1 ∣ → 0 , the spectrum becomes flat (C a 1 ( f ) → 1 C_{a_1}(f)\to1 C a 1 ( f ) → 1 ), and the asymptotic CRB on f 1 f_1 f 1 tends to infinity, confirming that frequency offset estimation is impossible for spectrally white sources.
Figure 3. lim n → ∞ n C R B c ( f 1 ) \lim_{n\to\infty} n\mathrm{CRB}_{c}(f_1) lim n → ∞ n CRB c ( f 1 ) as a function of a a a for different values of SNR and m = 6 m=6 m = 6 for a circular AR(1) source.
Note that in the case of two mutually uncorrelated circular stationary sources with spectra parameterized as C s 1 ( f ) = σ s 1 2 C a 1 ( f ) C_{s_1}(f)=\sigma_{s_1}^{2}C_{a_1}(f) C s 1 ( f ) = σ s 1 2 C a 1 ( f ) and C s 2 ( f ) = σ s 2 2 C a 2 ( f ) C_{s_2}(f)=\sigma_{s_2}^{2}C_{a_2}(f) C s 2 ( f ) = σ s 2 2 C a 2 ( f ) , and each affected by its own frequency offset, model (19) becomes:
x k = e i 2 π k f 1 s k , 1 a 1 + e i 2 π k f 2 s k , 2 a 2 + n k . (37) \mathbf{x}_k=e^{i2\pi kf_1}s_{k,1}\mathbf{a}_1+e^{i2\pi kf_2}s_{k,2}\mathbf{a}_2+\mathbf{n}_k. \tag*{(37)} x k = e i 2 π k f 1 s k , 1 a 1 + e i 2 π k f 2 s k , 2 a 2 + n k . (37)
The process ( x k ) k ∈ Z (\mathbf{x}_k)_{k\in\mathbb{Z}} ( x k ) k ∈ Z remains circular and stationary, with spectral density matrix:
C x ( f ) = σ s 1 2 C a 1 ( f + f 1 ) a 1 a 1 H + σ s 2 2 C a 2 ( f + f 2 ) a 2 a 2 H + σ n 2 I m . (38) \mathbf{C}_x(f)=\sigma_{s_1}^{2}C_{a_1}(f+f_1)\mathbf{a}_1\mathbf{a}_1^{H}+\sigma_{s_2}^{2}C_{a_2}(f+f_2)\mathbf{a}_2\mathbf{a}_2^{H}+\sigma_n^{2}\mathbf{I}_m. \tag*{(38)} C x ( f ) = σ s 1 2 C a 1 ( f + f 1 ) a 1 a 1 H + σ s 2 2 C a 2 ( f + f 2 ) a 2 a 2 H + σ n 2 I m . (38)
Thus, the complex circular extension of (6) (as summarized in Table 1 ) applies with the full parameter vector θ = ( f 1 , f 2 , ω 1 , ω 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T \boldsymbol{\theta}=(f_1,f_2,\omega_1,\omega_2,\mathbf{a}_1,\mathbf{a}_2,\sigma_{s_1}^{2},\sigma_{s_2}^{2},\sigma_n^{2})^{T} θ = ( f 1 , f 2 , ω 1 , ω 2 , a 1 , a 2 , σ s 1 2 , σ s 2 2 , σ n 2 ) T . Since these parameters are not decoupled in the FIM (both sources contribute to the same aggregate spectrum), it is not feasible to derive closed-form expressions for lim n → ∞ n C R B c ( ω i ) \lim_{n\to\infty}n\mathrm{CRB}_{c}(\omega_i) lim n → ∞ n CRB c ( ω i ) from (6) adapted to complex circular stationary processes.
Nevertheless, numerical evaluation of (6) permits investigation of the influence of the frequency offsets f 1 f_1 f 1 and f 2 f_2 f 2 on the DOA CRBs, which may be significantly affected by the source spectral characteristics. Consider two equipowered ( σ s 2 ≡ d e f σ s 1 2 = σ s 2 2 ) (\sigma_s^2 \mathrel{\mathop{\equiv}^{\mathrm{def}}}\sigma_{s_1}^2=\sigma_{s_2}^2) ( σ s 2 ≡ def σ s 1 2 = σ s 2 2 ) uncorrelated sources with identical AR(1) spectra, where a ≡ d e f a 1 = a 2 a \mathrel{\mathop{\equiv}^{\mathrm{def}}}a_1=a_2 a ≡ def a 1 = a 2 , impinging on a ULA of m = 6 m=6 m = 6 sensors. We show numerically from (6) that lim n → ∞ n C R B c ( ω 1 ) = lim n → ∞ n C R B c ( ω 2 ) \lim_{n\to\infty}n\mathrm{CRB}_{c}(\omega_1)=\lim_{n\to\infty}n\mathrm{CRB}_{c}(\omega_2) lim n → ∞ n CRB c ( ω 1 ) = lim n → ∞ n CRB c ( ω 2 ) , which depends only on the frequency offset difference Δ f ≡ d e f f 1 − f 2 \Delta f \mathrel{\mathop{\equiv}^{\mathrm{def}}}f_1-f_2 Δ f ≡ def f 1 − f 2 , the angular separation Δ ω ≡ d e f ω 1 − ω 2 \Delta\omega\mathrel{\mathop{\equiv}^{\mathrm{def}}}\omega_1-\omega_2 Δ ω ≡ def ω 1 − ω 2 , the AR(1) parameter a a a , and the SNR r ≡ d e f σ s 2 / σ n 2 r \mathrel{\mathop{\equiv}^{\mathrm{def}}}\sigma_s^2/\sigma_n^2 r ≡ def σ s 2 / σ n 2 . Figure 4 and Figure 5 display lim n → ∞ n C R B c ( ω i ) \lim_{n\to\infty}n\mathrm{CRB}_{c}(\omega_i) lim n → ∞ n CRB c ( ω i ) as a function of a a a and Δ ω \Delta\omega Δ ω , respectively, for varying values of Δ f \Delta f Δ f . These figures also show the numerical values of n C R B c ( ω i ) n\mathrm{CRB}_{c}(\omega_i) n CRB c ( ω i ) given by the closed-form expression [29 eq. (5)] for two independent circular Gaussian observations without frequency offset. For Δ f = 0 \Delta f=0 Δ f = 0 (i.e., identical frequency offsets), these figures clearly show that the CRB exhibits very low sensitivity to the spectral shape, reflecting the behavior observed in the single-source case. In contrast, for Δ f ≠ 0 \Delta f \ne0 Δ f = 0 , the CRB is markedly sensitive to the spectral shape, exhibiting a significant reduction for strongly correlated samples (i.e., highly peaked spectra with ∣ a ∣ → 1 |a| \to1 ∣ a ∣ → 1 ).
Figure 4. lim n → ∞ n C R B c ( ω i ) \lim_{n\to\infty} n\mathrm{CRB}_{c}(\omega_{i}) lim n → ∞ n CRB c ( ω i ) as a function of a a a for different values of Δ f \Delta f Δ f for Δ ω = 0.2 d \Delta\omega= 0.2\mathrm{d} Δ ω = 0.2 d and r = 15 d B r = 15\mathrm{dB} r = 15 dB .
Figure 5. lim n → ∞ n C R B c ( ω i ) \lim_{n\to\infty} n\mathrm{CRB}_{c}(\omega_{i}) lim n → ∞ n CRB c ( ω i ) as a function of Δ ω \Delta\omega Δ ω for different values of Δ f \Delta f Δ f for a = 0.5 a = 0.5 a = 0.5 and r = 15 d B r = 15\mathrm{dB} r = 15 dB .
Furthermore, it is important to note that these asymptotic CRBs do not depend numerically on the CG distribution of the stationary observations. This is because the integral ∫ 0 1 tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω i ) d f ≈ 0 \int_{0}^{1}\operatorname{tr}\left(\mathbf{C}_{x}^{-1}(f)\frac{\partial\mathbf{C}_{x}(f)}{\partial\omega_{i}}\right)\,\mathrm{d}f \approx0 ∫ 0 1 tr ( C x − 1 ( f ) ∂ ω i ∂ C x ( f ) ) d f ≈ 0 , so the CG-dependent correction term c c c in (6) has negligible effect on the asymptotic CRB. This property contrasts with the case of independent identically CG distributed observations, where C R B c ( ω i ) \mathrm{CRB}_{c}(\omega_{i}) CRB c ( ω i ) is inversely proportional to n ξ 1 , m c n\xi_{1,m}^{c} n ξ 1 , m c with ξ 1 , m c ≔ E [ Q m 2 ϕ m 2 ( Q m ) ] m ( m + 1 ) ≤ 1 \xi_{1,m}^{c} \coloneqq\frac{\mathbb{E}[Q_{m}^{2}\phi_{m}^{2}(Q_{m})]}{m(m+1)} \le1 ξ 1 , m c : = m ( m + 1 ) E [ Q m 2 ϕ m 2 ( Q m )] ≤ 1 [29 ] , and thus depends explicitly on the CG distribution even asymptotically.
Finally, it is worth noting that the numerical CRBs analyzed in [30 ] were derived under the assumption of stationary, rectilinear Gaussian sources with flat spectra. In that specific context, the effect of a carrier frequency offset manifests simply as a reduction in the effective source power across the full receiver bandwidth.
Conclusion
This paper has presented comprehensive extensions of the asymptotic frequency-domain FIM rate, originally introduced in [4 ] for multidimensional stationary real-valued CGPs. First, we extended this result to purely cyclostationary real-valued CGPs. Subsequently, we generalized the formulation to encompass circular and noncircular complex-valued multidimensional stationary CGPs, as well as both purely and almost cyclostationary regimes. These theoretical advancements establish a unified framework for evaluating the asymptotic FIM rate across a broad and practically relevant class of stationary and cyclostationary CGPs. A primary advantage of these generalizations is the significant reduction in computational complexity when evaluating the CRB for parameters characterizing the mean and spectral structure of such processes. While the derived expressions are broadly applicable to diverse parameter-estimation problems in signal processing, we specifically applied them to investigate the influence of source spectral characteristics on the CRB for DOA estimation. Our analysis reveals that while the temporal correlation (spectral shape) of the sources has a bounded impact on estimation accuracy, the presence of spectral offsets exerts a pronounced and systematic effect on the resulting DOA CRB. These findings underscore the critical importance of accurately modeling the spectral characteristics and potential frequency offsets of source signals when assessing the fundamental performance limits of DOA estimators. Future work will focus on extending this theoretical framework to multidimensional, spatiotemporally stationary spherically invariant random fields, thereby broadening its applicability to advanced signal and image processing applications.
Appendix
Proof of Result 3. Let us firstly consider the case of a circularly complex-valued u k u_k u k . From (18), we obtain: y n = Δ n ( θ 0 ) z n ∈ C m n y_n = \Delta_n(\theta_0)z_n \in\mathbb{C}^{mn} y n = Δ n ( θ 0 ) z n ∈ C mn where Δ n ( θ 0 ) ≔ Blkdiag ( e i 2 π θ 0 I m , … , e i 2 π n θ 0 I m ) \Delta_n(\theta_0) \coloneqq\operatorname{Blkdiag}(e^{\mathrm{i}2\pi\theta_0 I_m}, \ldots, e^{\mathrm{i}2\pi n\theta_0 I_m}) Δ n ( θ 0 ) : = Blkdiag ( e i 2 π θ 0 I m , … , e i 2 π n θ 0 I m ) is unitary. It then follows that Σ y n ≔ E ( y n y n H ) = Δ n ( θ 0 ) Σ z n Δ n H ( θ 0 ) \Sigma_{y_n} \coloneqq\mathbb{E}(y_n y_n^{H}) = \Delta_n(\theta_0)\Sigma_{z_n}\Delta_n^{H}(\theta_0) Σ y n : = E ( y n y n H ) = Δ n ( θ 0 ) Σ z n Δ n H ( θ 0 ) , Σ y n − 1 = Δ n ( θ 0 ) Σ z n − 1 Δ n H ( θ 0 ) \Sigma_{y_n}^{-1} = \Delta_n(\theta_0)\Sigma_{z_n}^{-1}\Delta_n^{H}(\theta_0) Σ y n − 1 = Δ n ( θ 0 ) Σ z n − 1 Δ n H ( θ 0 ) , Σ y n , 0 ′ = Δ n ′ ( θ 0 ) Σ z n Δ n H ( θ 0 ) + Δ n ( θ 0 ) Σ z n Δ n ′ H ( θ 0 ) \Sigma'_{y_n,0} = \Delta'_n(\theta_0)\Sigma_{z_n}\Delta_n^{H}(\theta_0) + \Delta_n(\theta_0)\Sigma_{z_n}\Delta_n^{\prime H}(\theta_0) Σ y n , 0 ′ = Δ n ′ ( θ 0 ) Σ z n Δ n H ( θ 0 ) + Δ n ( θ 0 ) Σ z n Δ n ′ H ( θ 0 ) and Σ y n , ℓ ′ = Δ n ( θ 0 ) Σ z n , ℓ ′ Δ n H ( θ 0 ) \Sigma'_{y_n,\ell} = \Delta_n(\theta_0)\Sigma'_{z_n,\ell}\Delta_n^{H}(\theta_0) Σ y n , ℓ ′ = Δ n ( θ 0 ) Σ z n , ℓ ′ Δ n H ( θ 0 ) , ℓ = 1 , … , q \ell= 1,\ldots,q ℓ = 1 , … , q where Δ n ′ ( θ 0 ) ≔ d Δ n ( θ 0 ) d θ 0 = i 2 π Δ n Δ n ( θ 0 ) \Delta'_n(\theta_0) \coloneqq\frac{\mathrm{d}\Delta_n(\theta_0)}{\mathrm{d}\theta_0} = \mathrm{i}2\pi\Delta_n\Delta_n(\theta_0) Δ n ′ ( θ 0 ) : = d θ 0 d Δ n ( θ 0 ) = i 2 π Δ n Δ n ( θ 0 ) with Δ n ≔ Blkdiag ( I m , 2 I m , … , n I m ) \Delta_n \coloneqq\operatorname{Blkdiag}(I_m,2I_m,\ldots,nI_m) Δ n : = Blkdiag ( I m , 2 I m , … , n I m ) . Plugging the expressions of Σ y n − 1 \Sigma_{y_n}^{-1} Σ y n − 1 , Σ y n , k ′ \Sigma'_{y_n,k} Σ y n , k ′ into Tr ( Σ y n − 1 Σ y n , k ′ ) \operatorname{Tr}(\Sigma_{y_n}^{-1}\Sigma'_{y_n,k}) Tr ( Σ y n − 1 Σ y n , k ′ ) and Tr ( Σ y n − 1 Σ y n , k ′ Σ y n − 1 Σ y n , ℓ ′ ) \operatorname{Tr}(\Sigma_{y_n}^{-1}\Sigma'_{y_n,k}\Sigma_{y_n}^{-1}\Sigma'_{y_n,\ell}) Tr ( Σ y n − 1 Σ y n , k ′ Σ y n − 1 Σ y n , ℓ ′ ) , we get after simple algebraic manipulations, for k , ℓ = 1 , … , q k,\ell= 1,\ldots,q k , ℓ = 1 , … , q :
Tr ( Σ y n − 1 Σ y n , 0 ′ ) = i 2 π Tr ( Δ n ) − i 2 π Tr ( Δ n ) = 0 , (39) \operatorname{Tr}\left(\Sigma_{y_n}^{-1}\Sigma'_{y_n,0}\right) = \mathrm{i}2\pi\operatorname{Tr}(\Delta_n) - \mathrm{i}2\pi\operatorname{Tr}(\Delta_n) = 0, \tag*{(39)} Tr ( Σ y n − 1 Σ y n , 0 ′ ) = i 2 π Tr ( Δ n ) − i 2 π Tr ( Δ n ) = 0 , (39)
Tr ( Σ y n − 1 Σ y n , 0 ′ Σ y n − 1 Σ y n , ℓ ′ ) = i 2 π Tr ( Δ n Σ z n − 1 Σ z n , ℓ ′ ) − i 2 π Tr ( Δ n Σ z n − 1 Σ z n , ℓ ′ ) = 0. (40) \operatorname{Tr}\left(\Sigma_{y_n}^{-1}\Sigma'_{y_n,0}\Sigma_{y_n}^{-1}\Sigma'_{y_n,\ell}\right) = \mathrm{i}2\pi\operatorname{Tr}\left(\Delta_n\Sigma_{z_n}^{-1}\Sigma'_{z_n,\ell}\right) - \mathrm{i}2\pi\operatorname{Tr}\left(\Delta_n\Sigma_{z_n}^{-1}\Sigma'_{z_n,\ell}\right) = 0. \tag*{(40)} Tr ( Σ y n − 1 Σ y n , 0 ′ Σ y n − 1 Σ y n , ℓ ′ ) = i 2 π Tr ( Δ n Σ z n − 1 Σ z n , ℓ ′ ) − i 2 π Tr ( Δ n Σ z n − 1 Σ z n , ℓ ′ ) = 0. (40)
This allows us to conclude from the expression of FIM that θ 0 \theta_0 θ 0 is decoupled from { θ k } k = 1 q \{\theta_k\}_{k=1}^{q} { θ k } k = 1 q . Furthermore, the remaining terms utilized in the calculation of the FIM elements associated with { θ k } k = 1 q \{\theta_k\}_{k=1}^{q} { θ k } k = 1 q can be expressed as follows:
Tr ( Σ y n − 1 Σ y n , ℓ ′ ) = Tr ( Σ z n − 1 Σ z n , ℓ ′ ) (41) \operatorname{Tr}\left(\boldsymbol{\Sigma}_{y_n}^{-1}\boldsymbol{\Sigma}_{y_n,\ell}^{\prime}\right)=\operatorname{Tr}\left(\boldsymbol{\Sigma}_{z_n}^{-1}\boldsymbol{\Sigma}_{z_n,\ell}^{\prime}\right) \tag*{(41)} Tr ( Σ y n − 1 Σ y n , ℓ ′ ) = Tr ( Σ z n − 1 Σ z n , ℓ ′ ) (41)
Tr ( Σ y n − 1 Σ y n , k ′ Σ y n − 1 Σ y n , ℓ ′ ) = Tr ( Σ z n − 1 Σ z n , k ′ Σ z n − 1 Σ z n , ℓ ′ ) . (42) \operatorname{Tr}\left(\boldsymbol{\Sigma}_{y_n}^{-1}\boldsymbol{\Sigma}_{y_n,k}^{\prime}\boldsymbol{\Sigma}_{y_n}^{-1}\boldsymbol{\Sigma}_{y_n,\ell}^{\prime}\right)=\operatorname{Tr}\left(\boldsymbol{\Sigma}_{z_n}^{-1}\boldsymbol{\Sigma}_{z_n,k}^{\prime}\boldsymbol{\Sigma}_{z_n}^{-1}\boldsymbol{\Sigma}_{z_n,\ell}^{\prime}\right). \tag*{(42)} Tr ( Σ y n − 1 Σ y n , k ′ Σ y n − 1 Σ y n , ℓ ′ ) = Tr ( Σ z n − 1 Σ z n , k ′ Σ z n − 1 Σ z n , ℓ ′ ) . (42)
For noncircular complex-valued u k \mathbf{u}_k u k , the proof follows the same steps, using y ~ n = Δ ~ n ( θ 0 ) z ~ n ∈ C 2 m n \widetilde{\mathbf{y}}_n=\widetilde{\boldsymbol{\Delta}}_n(\theta_0)\widetilde{\mathbf{z}}_n\in\mathbb{C}^{2mn} y n = Δ n ( θ 0 ) z n ∈ C 2 mn where Δ ~ n ( θ 0 ) ≡ d e f Blkdiag ( e i 2 π θ 0 I m , e − i 2 π θ 0 I m , … , e i 2 π n θ 0 I m , e − i 2 π n θ 0 I m ) \widetilde{\boldsymbol{\Delta}}_n(\theta_0)\mathrel{\mathop{\equiv}^{\mathrm{def}}}\operatorname{Blkdiag}(e^{i2\pi\theta_0}\mathbf{I}_m,e^{-i2\pi\theta_0}\mathbf{I}_m,\ldots,e^{i2\pi n\theta_0}\mathbf{I}_m,e^{-i2\pi n\theta_0}\mathbf{I}_m) Δ n ( θ 0 ) ≡ def Blkdiag ( e i 2 π θ 0 I m , e − i 2 π θ 0 I m , … , e i 2 π n θ 0 I m , e − i 2 π n θ 0 I m ) and z ~ n ≡ d e f ( u ~ 1 T , … , u ~ n T ) T \widetilde{\mathbf{z}}_n\mathrel{\mathop{\equiv}^{\mathrm{def}}}(\widetilde{\mathbf{u}}_1^{\mathsf{T}},\ldots,\widetilde{\mathbf{u}}_n^{\mathsf{T}})^{\mathsf{T}} z n ≡ def ( u 1 T , … , u n T ) T .
Proof of Result 4
In order to compute the extended version of (6) for complex circular-valued signals, it is necessary to evaluate the elements of FIM y n ( θ ) \operatorname{FIM}_{y_n}(\theta) FIM y n ( θ ) . This can be achieved by employing the following readily derived identities:
C x − 1 ( f ) = 1 σ 1 2 I − r 1 C a 1 ( f ) m σ 1 2 C a 1 ( f ) + σ n 2 a 1 a 1 H \mathbf{C}_x^{-1}(f)=\frac{1}{\sigma_1^2}\mathbf{I}-\frac{r_1 C_{a_1}(f)}{m\sigma_1^2 C_{a_1}(f)+\sigma_n^2}\mathbf{a}_1\mathbf{a}_1^{\mathsf{H}} C x − 1 ( f ) = σ 1 2 1 I − m σ 1 2 C a 1 ( f ) + σ n 2 r 1 C a 1 ( f ) a 1 a 1 H , ∂ C x ( f ) ∂ ω 1 = σ 1 2 C a 1 ( f ) [ a 1 d 1 H + d 1 a 1 H ] \frac{\partial\mathbf{C}_x(f)}{\partial\omega_1}=\sigma_1^2 C_{a_1}(f)[\mathbf{a}_1\mathbf{d}_1^{\mathsf{H}}+\mathbf{d}_1\mathbf{a}_1^{\mathsf{H}}] ∂ ω 1 ∂ C x ( f ) = σ 1 2 C a 1 ( f ) [ a 1 d 1 H + d 1 a 1 H ] , ∂ C x ( f ) ∂ a 1 = σ 1 2 ∂ C a 1 ( f ) ∂ a 1 a 1 a 1 H \frac{\partial\mathbf{C}_x(f)}{\partial a_1}=\sigma_1^2\frac{\partial C_{a_1}(f)}{\partial a_1}\mathbf{a}_1\mathbf{a}_1^{\mathsf{H}} ∂ a 1 ∂ C x ( f ) = σ 1 2 ∂ a 1 ∂ C a 1 ( f ) a 1 a 1 H , ∂ C x ( f ) ∂ σ 1 2 = C a 1 ( f ) a 1 a 1 H \frac{\partial\mathbf{C}_x(f)}{\partial\sigma_1^2}=C_{a_1}(f)\mathbf{a}_1\mathbf{a}_1^{\mathsf{H}} ∂ σ 1 2 ∂ C x ( f ) = C a 1 ( f ) a 1 a 1 H , ∂ C x ( f ) ∂ σ n 2 = I \frac{\partial\mathbf{C}_x(f)}{\partial\sigma_n^2}=\mathbf{I} ∂ σ n 2 ∂ C x ( f ) = I , ∥ a 1 ∥ 2 = m \lVert\mathbf{a}_1\rVert^2=m ∥ a 1 ∥ 2 = m and a 1 H d 1 + d 1 H a 1 = 0 \mathbf{a}_1^{\mathsf{H}}\mathbf{d}_1+\mathbf{d}_1^{\mathsf{H}}\mathbf{a}_1=0 a 1 H d 1 + d 1 H a 1 = 0 , which, after simple algebraic manipulation, leads to [ FIM y n ( θ ) ] ω 1 , a 1 = [ FIM y n ( θ ) ] ω 1 , σ 1 2 = [ FIM y n ( θ ) ] ω 1 , σ n 2 = 0 [\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,a_1}=[\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,\sigma_1^2}=[\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,\sigma_n^2}=0 [ FIM y n ( θ ) ] ω 1 , a 1 = [ FIM y n ( θ ) ] ω 1 , σ 1 2 = [ FIM y n ( θ ) ] ω 1 , σ n 2 = 0 (the parameter ω 1 \omega_1 ω 1 and ( a 1 , σ 1 2 , σ n 2 ) (a_1,\sigma_1^2,\sigma_n^2) ( a 1 , σ 1 2 , σ n 2 ) are decoupled in the FIM y n ( θ ) \operatorname{FIM}_{y_n}(\theta) FIM y n ( θ ) ) with Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω 1 ) = 0 \operatorname{Tr}\left(\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial\omega_1}\right)=0 Tr ( C x − 1 ( f ) ∂ ω 1 ∂ C x ( f ) ) = 0 . This implies:
lim n → ∞ [ n CRB ( ω 1 ) ] − 1 = ∫ 0 1 Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω 1 C x − 1 ( f ) ∂ C x ( f ) ∂ ω 1 ) d f . (43) \lim_{n\to\infty}[n\operatorname{CRB}(\omega_1)]^{-1}=\int_0^1\operatorname{Tr}\left(\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial\omega_1}\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial\omega_1}\right)\,df. \tag*{(43)} n → ∞ lim [ n CRB ( ω 1 ) ] − 1 = ∫ 0 1 Tr ( C x − 1 ( f ) ∂ ω 1 ∂ C x ( f ) C x − 1 ( f ) ∂ ω 1 ∂ C x ( f ) ) df . (43)
Upon evaluating the trace of the product matrices in (43) that depend on the aforementioned identities, we obtain (21) through straightforward algebraic manipulation using the properties Tr ( [ a 1 d 1 H + d 1 a 1 H ] [ a 1 d 1 H + d 1 a 1 H ] ) = m α 1 \operatorname{Tr}([\mathbf{a}_1\mathbf{d}_1^{\mathsf{H}}+\mathbf{d}_1\mathbf{a}_1^{\mathsf{H}}][\mathbf{a}_1\mathbf{d}_1^{\mathsf{H}}+\mathbf{d}_1\mathbf{a}_1^{\mathsf{H}}])=m\alpha_1 Tr ([ a 1 d 1 H + d 1 a 1 H ] [ a 1 d 1 H + d 1 a 1 H ]) = m α 1 and Tr ( a 1 a 1 H [ a 1 d 1 H + d 1 a 1 H ] a 1 a 1 H [ a 1 d 1 H + d 1 a 1 H ] ) = 0 \operatorname{Tr}(\mathbf{a}_1\mathbf{a}_1^{\mathsf{H}}[\mathbf{a}_1\mathbf{d}_1^{\mathsf{H}}+\mathbf{d}_1\mathbf{a}_1^{\mathsf{H}}]\mathbf{a}_1\mathbf{a}_1^{\mathsf{H}}[\mathbf{a}_1\mathbf{d}_1^{\mathsf{H}}+\mathbf{d}_1\mathbf{a}_1^{\mathsf{H}}])=0 Tr ( a 1 a 1 H [ a 1 d 1 H + d 1 a 1 H ] a 1 a 1 H [ a 1 d 1 H + d 1 a 1 H ]) = 0 .
For the lower bound of CRB c ( ω 1 ) \operatorname{CRB}_c(\omega_1) CRB c ( ω 1 ) , (23) ⇔ \Leftrightarrow ⇔ ∫ 0 1 m r 1 C a 1 2 ( f ) 1 + m r 1 C a 1 ( f ) d f < 1 ⇔ ∫ 0 1 ( C a 1 ( f ) − C a 1 ( f ) 1 + m r 1 C a 1 ( f ) ) d f < 1 \int_0^1\frac{mr_1C_{a_1}^2(f)}{1+mr_1C_{a_1}(f)}df<1\Leftrightarrow\int_0^1\left(C_{a_1}(f)-\frac{C_{a_1}(f)}{1+mr_1C_{a_1}(f)}\right)df<1 ∫ 0 1 1 + m r 1 C a 1 ( f ) m r 1 C a 1 2 ( f ) df < 1 ⇔ ∫ 0 1 ( C a 1 ( f ) − 1 + m r 1 C a 1 ( f ) C a 1 ( f ) ) df < 1 which is true because ∫ 0 1 C a 1 ( f ) d f = 1 \int_0^1C_{a_1}(f)df=1 ∫ 0 1 C a 1 ( f ) df = 1 and C a 1 ( f ) ≥ 0 ∀ f C_{a_1}(f)\geq0\ \forall f C a 1 ( f ) ≥ 0 ∀ f .
For the upper bound of CRB c ( ω 1 ) \operatorname{CRB}_c(\omega_1) CRB c ( ω 1 ) , (23) ⇔ \Leftrightarrow ⇔ ∫ 0 1 m r 1 C a 1 2 ( f ) 1 + m r 1 C a 1 ( f ) d f ≥ m r 1 1 + m r 1 ⇔ 1 − ∫ 0 1 C a 1 ( f ) 1 + m r 1 C a 1 ( f ) d f ≥ m r 1 1 + m r 1 ⇔ 1 − [ 1 m r 1 − 1 m r 1 ∫ 0 1 1 1 + m r 1 C a 1 ( f ) d f ] ≥ m r 1 1 + m r 1 ⇔ ∫ 0 1 1 1 + m r 1 C a 1 ( f ) d f ≥ 1 1 + m r 1 \int_0^1\frac{mr_1C_{a_1}^2(f)}{1+mr_1C_{a_1}(f)}df\geq\frac{mr_1}{1+mr_1}\Leftrightarrow1-\int_0^1\frac{C_{a_1}(f)}{1+mr_1C_{a_1}(f)}df\geq\frac{mr_1}{1+mr_1}\Leftrightarrow1-\left[\frac{1}{mr_1}-\frac{1}{mr_1}\int_0^1\frac{1}{1+mr_1C_{a_1}(f)}df\right]\geq\frac{mr_1}{1+mr_1}\Leftrightarrow\int_0^1\frac{1}{1+mr_1C_{a_1}(f)}df\geq\frac{1}{1+mr_1} ∫ 0 1 1 + m r 1 C a 1 ( f ) m r 1 C a 1 2 ( f ) df ≥ 1 + m r 1 m r 1 ⇔ 1 − ∫ 0 1 1 + m r 1 C a 1 ( f ) C a 1 ( f ) df ≥ 1 + m r 1 m r 1 ⇔ 1 − [ m r 1 1 − m r 1 1 ∫ 0 1 1 + m r 1 C a 1 ( f ) 1 df ] ≥ 1 + m r 1 m r 1 ⇔ ∫ 0 1 1 + m r 1 C a 1 ( f ) 1 df ≥ 1 + m r 1 1 . To prove this last inequality, we return to the minimization of an integral under constraint. This is achieved through the discretization of the Riemann integral with c k ≡ d e f C a 1 ( k n ) c_k\mathrel{\mathop{\equiv}^{\mathrm{def}}}C_{a_1}\left(\frac{k}{n}\right) c k ≡ def C a 1 ( n k ) , k = 1 , … , n k=1,\ldots,n k = 1 , … , n , after which we consider the following optimization problem for arbitrary n n n :
min c 1 , … , c n ∑ k = 1 n 1 1 + m r 1 c k subject to ∑ k = 1 n c k = n , (44) \min_{c_1,\ldots,c_n}\sum_{k=1}^{n}\frac{1}{1+mr_1c_k}\quad\text{subject to}\quad\sum_{k=1}^{n}c_k=n, \tag*{(44)} c 1 , … , c n min k = 1 ∑ n 1 + m r 1 c k 1 subject to k = 1 ∑ n c k = n , (44)
which is a standard convex optimization problem that can be solved by the Lagrange multiplier method [31 chap. 5] . Using the Lagrange function L ( c 1 , … , c n , λ ) = ∑ k = 1 n 1 1 + m r 1 c k + λ ( ∑ k = 1 n c k − n ) L(c_1,\ldots,c_n,\lambda)=\sum_{k=1}^{n}\frac{1}{1+mr_1c_k}+\lambda\left(\sum_{k=1}^{n}c_k-n\right) L ( c 1 , … , c n , λ ) = ∑ k = 1 n 1 + m r 1 c k 1 + λ ( ∑ k = 1 n c k − n ) , it is straightforward to see that L ( c 1 , … , c n , λ ) L(c_1,\ldots,c_n,\lambda) L ( c 1 , … , c n , λ ) is minimum for c k = 1 c_k=1 c k = 1 , i = 1 , … , n i=1,\ldots,n i = 1 , … , n . Consequently, ∫ 0 1 1 1 + m r 1 C a 1 ( f ) d f \int_0^1\frac{1}{1+mr_1C_{a_1}(f)}df ∫ 0 1 1 + m r 1 C a 1 ( f ) 1 df under the constraint ∫ 0 1 C a 1 ( f ) d f = 1 \int_0^1C_{a_1}(f)df=1 ∫ 0 1 C a 1 ( f ) df = 1 is minimum for C a 1 ( f ) = 1 C_{a_1}(f)=1 C a 1 ( f ) = 1 . The proofs for CRB r ( ω 1 ) \operatorname{CRB}_r(\omega_1) CRB r ( ω 1 ) (22),(24) proceed in a similar manner.
Proof of Result 5
For non overlapping spectra C s 1 ( f ) C_{s_1}(f) C s 1 ( f ) and C s 2 ( f ) C_{s_2}(f) C s 2 ( f ) , the inversion matrix lemma still applies and leads to
C x − 1 ( f ) = { 1 σ n 2 I − r i C a i ( f ) m σ s i 2 C a i ( f ) + σ n 2 a i a i H for f ∈ B i , i = 1 , 2 , 1 σ n 2 I for f ∈ [ 0 , 1 ] , f ∉ B 1 , f ∉ B 2 . (45) C_x^{-1}(f)=
\begin{cases}
\frac{1}{\sigma_n^2}\mathbf{I}-\frac{r_iC_{a_i}(f)}{m\sigma_{s_i}^2C_{a_i}(f)+\sigma_n^2}\mathbf{a}_i\mathbf{a}_i^H & \text{for } f\in B_i, i=1,2,\\
\frac{1}{\sigma_n^2}\mathbf{I} & \text{for } f\in[0,1], f\notin B_1, f\notin B_2.
\end{cases} \tag*{(45)} C x − 1 ( f ) = ⎩ ⎨ ⎧ σ n 2 1 I − m σ s i 2 C a i ( f ) + σ n 2 r i C a i ( f ) a i a i H σ n 2 1 I for f ∈ B i , i = 1 , 2 , for f ∈ [ 0 , 1 ] , f ∈ / B 1 , f ∈ / B 2 . (45)
Using ∂ C x ( f ) ∂ ω i = σ s i 2 C a i ( f ) [ a i d i H + d i a i H ] \frac{\partial C_x(f)}{\partial\omega_i}=\sigma_{s_i}^2C_{a_i}(f)[\mathbf{a}_i\mathbf{d}_i^H+\mathbf{d}_i\mathbf{a}_i^H] ∂ ω i ∂ C x ( f ) = σ s i 2 C a i ( f ) [ a i d i H + d i a i H ] for f ∈ B i f\in B_i f ∈ B i and following the derivation steps in the proof of Result 4, we arrive at Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω 1 ) Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω 2 ) = 0 \operatorname{Tr}(C_x^{-1}(f)\frac{\partial C_x(f)}{\partial\omega_1})\operatorname{Tr}(C_x^{-1}(f)\frac{\partial C_x(f)}{\partial\omega_2})=0 Tr ( C x − 1 ( f ) ∂ ω 1 ∂ C x ( f ) ) Tr ( C x − 1 ( f ) ∂ ω 2 ∂ C x ( f ) ) = 0 for f ∈ [ 0 , 1 ] f\in[0,1] f ∈ [ 0 , 1 ] and Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ ω 1 C x − 1 ( f ) ∂ C x ( f ) ∂ ω 2 ) = 0 \operatorname{Tr}(C_x^{-1}(f)\frac{\partial C_x(f)}{\partial\omega_1}C_x^{-1}(f)\frac{\partial C_x(f)}{\partial\omega_2})=0 Tr ( C x − 1 ( f ) ∂ ω 1 ∂ C x ( f ) C x − 1 ( f ) ∂ ω 2 ∂ C x ( f ) ) = 0 for f ∈ [ 0 , 1 ] f\in[0,1] f ∈ [ 0 , 1 ] . Thus, [ FIM y n ( θ ) ] ω 1 , ω 2 = 0 [\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,\omega_2}=0 [ FIM y n ( θ ) ] ω 1 , ω 2 = 0 . Similarly to the single source case, the decoupling between ( ω 1 , ω 2 ) (\omega_1,\omega_2) ( ω 1 , ω 2 ) and ( σ s 1 2 , σ s 2 2 , σ n 2 ) (\sigma_{s_1}^2,\sigma_{s_2}^2,\sigma_n^2) ( σ s 1 2 , σ s 2 2 , σ n 2 ) is proven, and the application of (43) to ω 1 \omega_1 ω 1 and ω 2 \omega_2 ω 2 applies. This leads to the deduction of (31) . The proofs for non-overlapping rectilinear sources follow the same steps.
Proof of Result 6
The proof proceeds by expressing the frequency-domain FIM rate in terms of the spectrum of the stationary stacked source process s k , 1 ′ \mathbf{s}'_{k,1} s k , 1 ′ . Applying the matrix inversion lemma to C x ′ ( f ) = A ′ C s ′ ( f ) A ′ H + σ n 2 I m p C_{x'}(f)=\mathbf{A}'C_{s'}(f)\mathbf{A}'^H+\sigma_n^2\mathbf{I}_{mp} C x ′ ( f ) = A ′ C s ′ ( f ) A ′ H + σ n 2 I m p gives:
C x ′ − 1 ( f ) = 1 σ n 2 ( I m p − A ′ B A ′ H ) with B = d e f [ σ n 2 C s ′ − 1 ( f ) + A ′ H A ′ ] − 1 = r 1 [ I p + m r 1 C a 1 ′ ( f ) ] − 1 C a 1 ′ ( f ) . C_{x'}^{-1}(f)=\frac{1}{\sigma_n^2}(\mathbf{I}_{mp}-\mathbf{A}'\mathbf{B}\mathbf{A}'^H)\quad\text{with}\quad\mathbf{B}\overset{\mathrm{def}}{=}\left[\sigma_n^2C_{s'}^{-1}(f)+\mathbf{A}'^H\mathbf{A}'\right]^{-1}=r_1\left[\mathbf{I}_p+mr_1C_{a_1}'(f)\right]^{-1}C_{a_1}'(f). C x ′ − 1 ( f ) = σ n 2 1 ( I m p − A ′ B A ′ H ) with B = def [ σ n 2 C s ′ − 1 ( f ) + A ′ H A ′ ] − 1 = r 1 [ I p + m r 1 C a 1 ′ ( f ) ] − 1 C a 1 ′ ( f ) .
Then using the derivatives ∂ C x ′ ( f ) ∂ ω 1 = D ′ C s ′ ( f ) A ′ H + A ′ C s ′ ( f ) D ′ H \frac{\partial C_{x'}(f)}{\partial\omega_1}=\mathbf{D}'C_{s'}(f)\mathbf{A}'^H+\mathbf{A}'C_{s'}(f)\mathbf{D}'^H ∂ ω 1 ∂ C x ′ ( f ) = D ′ C s ′ ( f ) A ′ H + A ′ C s ′ ( f ) D ′ H , ∂ C x ′ ( f ) ∂ a 1 = σ s 1 2 A ′ ∂ C a 1 ′ ( f ) ∂ a 1 A ′ H \frac{\partial C_{x'}(f)}{\partial a_1}=\sigma_{s_1}^2\mathbf{A}'\frac{\partial C_{a_1}'(f)}{\partial a_1}\mathbf{A}'^H ∂ a 1 ∂ C x ′ ( f ) = σ s 1 2 A ′ ∂ a 1 ∂ C a 1 ′ ( f ) A ′ H , ∂ C x ′ ( f ) ∂ σ s 1 2 = A ′ C a 1 ′ ( f ) A ′ H \frac{\partial C_{x'}(f)}{\partial\sigma_{s_1}^2}=\mathbf{A}'C_{a_1}'(f)\mathbf{A}'^H ∂ σ s 1 2 ∂ C x ′ ( f ) = A ′ C a 1 ′ ( f ) A ′ H and ∂ C x ′ ( f ) ∂ σ n 2 = I m p \frac{\partial C_{x'}(f)}{\partial\sigma_n^2}=\mathbf{I}_{mp} ∂ σ n 2 ∂ C x ′ ( f ) = I m p with D ′ = Blkdiag ( d 1 , … , d 1 ) \mathbf{D}'=\operatorname{Blkdiag}(d_1,\ldots,d_1) D ′ = Blkdiag ( d 1 , … , d 1 ) , and the identities A ′ H A ′ = m I p \mathbf{A}'^H\mathbf{A}'=m\mathbf{I}_p A ′ H A ′ = m I p and A ′ H D ′ + D ′ H A ′ = ( a 1 H d 1 + d 1 H a 1 ) I p = 0 \mathbf{A}'^H\mathbf{D}'+\mathbf{D}'^H\mathbf{A}'=(a_1^Hd_1+d_1^Ha_1)\mathbf{I}_p=0 A ′ H D ′ + D ′ H A ′ = ( a 1 H d 1 + d 1 H a 1 ) I p = 0 leads after simple algebraic manipulations to [ FIM y n ( θ ) ] ω 1 , a 1 = [ FIM y n ( θ ) ] ω 1 , σ s 1 2 = [ FIM y n ( θ ) ] ω 1 , σ n 2 = 0 [\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,a_1}=[\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,\sigma_{s_1}^2}=[\operatorname{FIM}_{y_n}(\theta)]_{\omega_1,\sigma_n^2}=0 [ FIM y n ( θ ) ] ω 1 , a 1 = [ FIM y n ( θ ) ] ω 1 , σ s 1 2 = [ FIM y n ( θ ) ] ω 1 , σ n 2 = 0 (the parameter ω 1 \omega_1 ω 1 and ( a 1 , σ s 1 2 , σ n 2 ) (a_1,\sigma_{s_1}^2,\sigma_n^2) ( a 1 , σ s 1 2 , σ n 2 ) are decoupled in the FIM y n ( θ ) \operatorname{FIM}_{y_n}(\theta) FIM y n ( θ ) ) with Tr ( C x ′ − 1 ( f ) ∂ C x ′ ( f ) ∂ ω 1 ) = 0 \operatorname{Tr}(C_{x'}^{-1}(f)\frac{\partial C_{x'}(f)}{\partial\omega_1})=0 Tr ( C x ′ − 1 ( f ) ∂ ω 1 ∂ C x ′ ( f ) ) = 0 . This implies from the extended version of (6) for complex circular-valued cyclostationary signals:
lim n → ∞ [ n CRB ( ω 1 ) ] − 1 = ∫ 0 1 Tr ( C x ′ − 1 ( f ) ∂ C x ′ ( f ) ∂ ω 1 C x ′ − 1 ( f ) ∂ C x ′ ( f ) ∂ ω 1 ) d f . (46) \lim_{n\to\infty}[n\operatorname{CRB}(\omega_1)]^{-1}=\int_0^1\operatorname{Tr}\left(C_{x'}^{-1}(f)\frac{\partial C_{x'}(f)}{\partial\omega_1}C_{x'}^{-1}(f)\frac{\partial C_{x'}(f)}{\partial\omega_1}\right)\,df. \tag*{(46)} n → ∞ lim [ n CRB ( ω 1 ) ] − 1 = ∫ 0 1 Tr ( C x ′ − 1 ( f ) ∂ ω 1 ∂ C x ′ ( f ) C x ′ − 1 ( f ) ∂ ω 1 ∂ C x ′ ( f ) ) df . (46)
Upon evaluating the trace of the product matrices in (46) , using the same calculations as for the stationary circular case in Result 4, we obtain
lim n → ∞ n CRB c ( ω 1 ) = p α 1 r 1 ∫ 0 1 m r 1 Tr [ C a 1 ′ ( f ) ( I p + m r 1 C a 1 ′ ( f ) ) − 1 C a 1 ′ ( f ) ] d f . (47) \lim_{n\to\infty}n\operatorname{CRB}_{c}(\omega_1)=\frac{p}{\alpha_1r_1\int_0^1mr_1\operatorname{Tr}[C_{a_1}'(f)(\mathbf{I}_p+mr_1C_{a_1}'(f))^{-1}C_{a_1}'(f)]\,df}. \tag*{(47)} n → ∞ lim n CRB c ( ω 1 ) = α 1 r 1 ∫ 0 1 m r 1 Tr [ C a 1 ′ ( f ) ( I p + m r 1 C a 1 ′ ( f ) ) − 1 C a 1 ′ ( f )] df p . (47)
Finally, substituting the unitary EVD C a 1 ′ ( f ) = p U p H ( f ) C a 1 c y c ( f ) U p ( f ) C_{a_1}'(f)=p\mathbf{U}_p^H(f)C_{a_1}^{\mathrm{cyc}}(f)\mathbf{U}_p(f) C a 1 ′ ( f ) = p U p H ( f ) C a 1 cyc ( f ) U p ( f ) from (13) and exploiting the unitary invariance of the trace (Tr [ U p H ( f ) X U p ( f ) ] = Tr ( X ) \operatorname{Tr}[\mathbf{U}_p^H(f)\mathbf{X}\mathbf{U}_p(f)]=\operatorname{Tr}(\mathbf{X}) Tr [ U p H ( f ) X U p ( f )] = Tr ( X ) ) directly yields (33) .
The same methodology is employed to prove (34) for cyclostationary noncircular sources, using the augmented matrices A ~ ′ = d e f Blkdiag ( a 1 , a 1 ∗ , … , a 1 , a 1 ∗ ) \widetilde{\mathbf{A}}' \overset{\mathrm{def}}{=}\operatorname{Blkdiag}(a_1,a_1^*,\ldots,a_1,a_1^*) A ′ = def Blkdiag ( a 1 , a 1 ∗ , … , a 1 , a 1 ∗ ) and D ~ ′ = d e f Blkdiag ( d 1 , d 1 ∗ , … , d 1 , d 1 ∗ ) \widetilde{\mathbf{D}}' \overset{\mathrm{def}}{=}\operatorname{Blkdiag}(d_1,d_1^*,\ldots,d_1,d_1^*) D ′ = def Blkdiag ( d 1 , d 1 ∗ , … , d 1 , d 1 ∗ ) .
Proof of Result 7.
Using the decoupling between f 1 f_1 f 1 and ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) (\omega_1,a_1,\sigma_{s_1}^{2},\sigma_{n}^{2}) ( ω 1 , a 1 , σ s 1 2 , σ n 2 ) in the FIM, as outlined in Result 3 and (6) extended to circular complex-valued x k \mathbf{x}_k x k , we obtain:
lim n → ∞ [ n CRB ( f 1 ) ] − 1 = ∫ 0 1 Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ f 1 C x − 1 ( f ) ∂ C x ( f ) ∂ f 1 ) d f + 2 c [ ∫ 0 1 Tr ( C x − 1 ( f ) ∂ C x ( f ) ∂ f 1 ) d f ] 2 (48) \lim_{n\to\infty}[n\operatorname{CRB}(f_1)]^{-1}=\int_{0}^{1}\operatorname{Tr}\left(\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial f_1}\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial f_1}\right)\mathrm{d}f+2c\left[\int_{0}^{1}\operatorname{Tr}\left(\mathbf{C}_x^{-1}(f)\frac{\partial\mathbf{C}_x(f)}{\partial f_1}\right)\mathrm{d}f\right]^2 \tag*{(48)} n → ∞ lim [ n CRB ( f 1 ) ] − 1 = ∫ 0 1 Tr ( C x − 1 ( f ) ∂ f 1 ∂ C x ( f ) C x − 1 ( f ) ∂ f 1 ∂ C x ( f ) ) d f + 2 c [ ∫ 0 1 Tr ( C x − 1 ( f ) ∂ f 1 ∂ C x ( f ) ) d f ] 2 (48)
where C x ( f ) = σ s 1 2 C a 1 ( f + f 1 ) a 1 a 1 H + σ n 2 I \mathbf{C}_x(f)=\sigma_{s_1}^{2}C_{a_1}(f+f_1)\mathbf{a}_1\mathbf{a}_1^{H}+\sigma_n^2\mathbf{I} C x ( f ) = σ s 1 2 C a 1 ( f + f 1 ) a 1 a 1 H + σ n 2 I , C x − 1 ( f ) = 1 σ n 2 I − σ s 1 2 C a 1 ( f + f 1 ) m σ s 1 2 C a 1 ( f + f 1 ) + σ n 2 a 1 a 1 H \mathbf{C}_x^{-1}(f)=\frac{1}{\sigma_n^2}\mathbf{I}-\frac{\sigma_{s_1}^{2}C_{a_1}(f+f_1)}{m\sigma_{s_1}^{2}C_{a_1}(f+f_1)+\sigma_n^2}\mathbf{a}_1\mathbf{a}_1^{H} C x − 1 ( f ) = σ n 2 1 I − m σ s 1 2 C a 1 ( f + f 1 ) + σ n 2 σ s 1 2 C a 1 ( f + f 1 ) a 1 a 1 H and ∂ C x ( f ) ∂ f 1 = σ s 1 2 ∂ C a 1 ( f + f 1 ) ∂ f a 1 a 1 H \frac{\partial\mathbf{C}_x(f)}{\partial f_1}=\sigma_{s_1}^{2}\frac{\partial C_{a_1}(f+f_1)}{\partial f}\mathbf{a}_1\mathbf{a}_1^{H} ∂ f 1 ∂ C x ( f ) = σ s 1 2 ∂ f ∂ C a 1 ( f + f 1 ) a 1 a 1 H because ∂ C a 1 ( f + f 1 ) ∂ f 1 = ∂ C a 1 ( f + f 1 ) ∂ f \frac{\partial C_{a_1}(f+f_1)}{\partial f_1}=\frac{\partial C_{a_1}(f+f_1)}{\partial f} ∂ f 1 ∂ C a 1 ( f + f 1 ) = ∂ f ∂ C a 1 ( f + f 1 ) . Upon inserting the values of C x − 1 ( f ) \mathbf{C}_x^{-1}(f) C x − 1 ( f ) and ∂ C x ( f ) ∂ f 1 \frac{\partial\mathbf{C}_x(f)}{\partial f_1} ∂ f 1 ∂ C x ( f ) into (48) and performing the requisite algebraic manipulations, we obtain
lim n → ∞ [ n CRB c ( f 1 ) ] − 1 = m 2 r 1 2 ∫ 0 1 ( ∂ C a 1 ( f + f 1 ) ∂ f ) 2 ( 1 + m r 1 C a 1 ( f + f 1 ) ) 2 d f + 2 c [ ∫ 0 1 m r 1 2 1 + m r 1 C a 1 ( f + f 1 ) ∂ C a 1 ( f + f 1 ) ∂ f d f ] 2 (49) \lim_{n\to\infty}[n\operatorname{CRB}^{\mathrm{c}}(f_1)]^{-1}=m^2r_1^2\int_{0}^{1}\frac{\left(\frac{\partial C_{a_1}(f+f_1)}{\partial f}\right)^2}{\left(1+mr_1C_{a_1}(f+f_1)\right)^2}\mathrm{d}f+2c\left[\int_{0}^{1}\frac{mr_1^2}{1+mr_1C_{a_1}(f+f_1)}\frac{\partial C_{a_1}(f+f_1)}{\partial f}\mathrm{d}f\right]^2 \tag*{(49)} n → ∞ lim [ n CRB c ( f 1 ) ] − 1 = m 2 r 1 2 ∫ 0 1 ( 1 + m r 1 C a 1 ( f + f 1 ) ) 2 ( ∂ f ∂ C a 1 ( f + f 1 ) ) 2 d f + 2 c [ ∫ 0 1 1 + m r 1 C a 1 ( f + f 1 ) m r 1 2 ∂ f ∂ C a 1 ( f + f 1 ) d f ] 2 (49)
Using the periodicity of C a 1 ( f + f 1 ) C_{a_1}(f+f_1) C a 1 ( f + f 1 ) in f f f of period unity, the first term of (49) reduces to (36) and the second integral to r 1 [ ln ( 1 + m r 1 C a 1 ( f + f 1 ) ) ] 0 1 = 0 r_1[\ln(1+mr_1C_{a_1}(f+f_1))]_0^1=0 r 1 [ ln ( 1 + m r 1 C a 1 ( f + f 1 )) ] 0 1 = 0 .
Acknowledgement
Habti Abeida would like to acknowledge the Deanship of Graduate Studies and Scientific Research, Taif University, for funding this work.