1374

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

Oblique Projections for Direction-of-Arrival Estimation With Prior Knowledge

Rmy Boyer and Guillaume Bouleux
AbstractEstimation of directions-of-arrival (DOA) is an important problem in various applications and a priori knowledge on the source location is sometimes available. To exploit this information, standard methods are based on the orthogonal projection of the steering manifold onto the noise subspace associated with the a priori known DOA. In this paper, we derive and analyze the Cramr-Rao bound associated with this model and in particular we point out the limitations of this approach when the known and unknown DOA are closely spaced and the associated sources are uncorrelated (block-diagonal source covariance). To ll this need, we propose to integrate a priori known locations of several sources into the MUSIC algorithm based on oblique projection of the steering manifold. Finally, we show that the proposed approach is able to almost completely cancel the inuence of the known DOA on the unknown ones for block-diagonal source covariance and for sufcient signal-to-noise ratio (SNR). Index TermsCramr-Rao bound, MUSIC algorithm, orthogonal and oblique projectors, prior knowledge of DOA.

I. INTRODUCTION

IRECTIONS-OF-ARRIVAL (DOA) of narrowband sources estimation is one of the central problems in passive radar, sensor sonar, radio-astronomy, and seismology. This problem has received considerable attention in the last 30 years, and a variety of techniques for its solution have been proposed. In practical situations, we have sometimes the knowledge of some a priori known subset of the DOA as for instance in a radar application where the emitted signal is backscattered by a number of stationary objects with known positions situated in the radars viewing eld [6], [11]. So, several methods have been proposed to incorporate this prior knowledge into an estimation algorithm. Prior knowledge of DOA can be classied into two families depending if we assume soft or hard constraints [11]. Soft constraints mean that we known approximatively all the DOA. This class of method is known under the name of beamspace methods [20] and has received attention as data reduction methods. The second class of approach incorporates the exact knowledge of a subset of the DOA. This constraint is somewhat more restricting but

Manuscript received September 19, 2006; revised April 27, 2007. This work was presented in part at the Fourth IEEE Workshop on Sensor Array and MultiChannel Processing (SAM-2006). The associate editor coordinating the review of this manuscript and approving it for publication was Prof. Simon J. Godsill. The authors are wiith the Universit Paris XI (UPS), CNRSLSS-Suplec, 91190 Gif-sur-Yvette, France (e-mail: [email protected]; guillaume. [email protected]). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TSP.2007.909348

more interesting gains can be expected. The exact knowledge of DOA allows the deation of the signal subspace and, thus, to mitigate the inuence of the known DOA on the unknown ones. It is well known that subspace-based methods are very sensitive to over or underestimation of the number of sources and a small error on this parameter leads to a poor accuracy. On the other hand, estimate all the DOA and extract from this set, only the unknown DOA is not easy, especially for closely spaced DOA or/and for low SNRs. So, a specic strategy must be carried out to set up an estimation scheme which extracts only the unknown DOA without bias. In [6] and [11], a constrainedMUSIC algorithm has been presented. The key idea is to orthogonally project the noisy array response onto the noise subspace spanned by the steering vectors associated with the known DOA. Note that, we can nd the same principle in some sequential MUSIC algorithms as in [3], [14], and [13]. But, in contrast with our framework, the prior knowledge is uncertain since it is constituted from the previously estimated DOA. In this paper, we derive and analyze the Cramr-Rao bound (CRB), named Prior-CRB (P-CRB), associated with the orthogonal deation of the signal subspace. In particular, we show that the prior knowledge of a subset of the DOA leads to a smaller variance for coherent (or highly correlated) sources associated with closely spaced DOA. Another result lies in the fact that if the known and unknown DOA are closely spaced and the associated sources are uncorrelated, the orthogonal deation cannot help. To ll this need, we advocate that a better scheme for the deation of the signal subspace is based not only on orthogonal projectors but also on oblique projectors. Based on this principle, we propose to rewrite the MUSIC [16], [19] Least-Squares (LS) criterion in the context of the oblique projector algebra. The resulting LS criterion can be decomposed into the sum of two contributions: the rst term is a MUSIC-like criterion and the second one is a corrective function which integrates the prior knowledge. Note that oblique projectors have received relatively little attention in the literature of signal processing. However, we can nd in [9] a rst application of oblique projectors to sensor arrays. In [1], Behrens and Scharf propose a very detailed review on this topic and several signal processing oriented applications. More recently, these projection operators have been exploited in [24] and [28] in the context of blind channel identication, in image restoration [27], in noise reduction [8] and in the context of DOA estimation by McCloud and Scharf [12]. Remark that the framework of this latter publication is different to the context of this paper since the authors propose a scaled version of the MUSIC algorithm based on oblique projection but they do not assume prior knowledge of DOA. Finally, our methodology is

1053-587X/$25.00 2008 IEEE

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1375

also valid for all the methods belonging to the class of subspace tting techniques [26]. In addition, some potential applications of this algorithm can be considered in biomedical signal analysis [5], [22] and harmonic retrieval [25]. After this brief introduction, Section II discusses the matrixbased model of the DOA problem and denes in particular the partitioned steering manifold. In Section III, we present our derivation and the analysis of the P-CRB. In Section IV, we explain how integrate into the MUSIC algorithm the prior knowledge of a subset of the DOA by means of oblique projectors. Section V is dedicated to the practical implementation of the spectral and the root versions of the Prior-MUSIC algorithm. Section VII presents some numerical simulations. Finally, we recall some important facts on oblique projectors and we present the demonstration of the theorems in the Appendix. We denote by bold font vectors and matrices. In addition, and mean, respectively, transpose, conjugated-transpose, Moore-Penrose pseudoinverse, trace, determinant, Kronecker product, Hadamard product, and the direct sum of two subspaces. II. MATRIX-BASED REPRESENTATION OF THE DOA ESTIMATION PROBLEM In this section, we introduce the standard matrix-based representation of the DOA estimation problem for an Uniform Linear Array (ULA) and in particular, we dene the notion of partitioned steering manifold. A. Parametric Multi-Input Multi-Output (MIMO) Model Assume there are narrowband plane waves simultaneously incident on an ULA with sensors. The complex array response for the th snapshot is given by (1) where is the noisy observation on the th sensor and is the complex amplitude of the th source. The noise vector is denoted by in which the noise on each , is assumed to be additive, white, and sensor, denoted by and is a positive real paramGaussian of parameter Vandermonde steering manifold eter. Matrix is the dened by (2) where is the steering vector parameterized by DOA radian), given by (in

where and

with is the noise matrix.

B. Partitioned Steering Manifold and Deated Signal Subspace Assume that we know DOA among . Without loss of generality, the steering manifold can be partitioned according to

(4)

where the submanifold is the matrix composed by the desired DOA and submanifold collects the a priori known DOA. We name the subspace of interest or the which is deated signal subspace as its dimension is . smaller than , the dimension of the signal subspace In addition, we assume that the DOA are all distinct ( for ) or equivalently the rank of the steering manifold is (here, we assume ) then and intersect . This implies that trivially, i.e., . C. Structure of the Spatial Covariance The spatial covariance matrix admits a Vandermondetype plus noise decomposition according to (5) where is the mathematical expectation and the noise-free spatial covariance is given by (6) If we assume that all the sources are correlated, the source cois a full matrix, otherwise in the sequel we will variance sometimes assume that this matrix is block-diagonal, i.e., the sources associated with the known and unknown parts of the steering manifold are uncorrelated. In that case, we have (7) with and where (re) is the covariance associated with the unknown spectively, (respectively, known) sources. III. DERIVATION AND ANALYSIS OF THE PRIOR-CRB A. Incorporate Prior Knowledge

in which is the intersensor distance and is the wavelength. is assumed to be known or previously estimated Parameter ([18, Appendix C]). So, the parametric MIMO model for snapshots can be written according to (3)

In [6] and [11], a prior knowledge MUSIC algorithm has been introduced and analyzed. This algorithm, called constrained MUSIC, is based on the projection of the noisy array response onto . In a view to derive the Prior-CRB, we vectorize model (3) according to (8)

1376

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

in which the noise is denoted by vector . Conseis an additive white Gaussian process of parameters quently, and

distribution according to

(9) . and To incorporate prior knowledge of the known DOA set: collected in submanifold , model (8) is modied according to (14) denotes the th block of vector for . where Finally, we can formulate the following result. Theorem 1: The P-CRB with respect to the signal parameter is given by

(10) and the notation prior indicates that where we have projected signal onto the noise subspace associated with the a priori known DOA. Following the same notation, we dene the noise-free signal according to (11) where . B. Prior-CRB Based on Model (10) (P-CRB) There are several ways to derive the P-CRB. A rst approach could be to derive this bound in the framework of Gaussian colored processes where the inverse of the noise error covariance, involved in the general CRB formula [21], is computed from . An equivalent and a truncated pseudoinverse of projector more handy way is based on the following proposition. Proposition 1: The compressed1 signal (12) where is a unitary basis of , follows a Gaussian distribution of parameters where . Proof: See Appendix B. Let dene the signal plus nuisance model parameter vector by where and . A standard result ([18], Appendix B) is that the Mean Squares Error (MSE) for any unbiased estimate, , of the parameter vector satises with (16) with (15) where . Proof: See Appendix C. C. CRB Without Prior Knowledge Based on Model (8) The derivation of these bounds is well known [18, p. 392, (B.6.32)]. Then, we have with and

(17) , and . Expressions (16) and (17) are, respectively, the CRB over the subspace of interest and over the whole . Remark that the P-CRB mixes the and the space in the sense that only the unknown derivative steering vectors are projected onto the orthogonal complement of the whole space. where D. Comparison of the Bounds The following theorem discusses some properties linking the derived bounds. Theorem 2: For any unbiased estimator of , we have the following relations. : For full (i) (ii) for . If is block-diagonal: (iii) (iv) for closely spaced DOA for closely spaced (v) decient. DOA and for widely spaced (vi) DOA. Proof: See Appendix D.

(13) where is the likelihood function. Consequently, the P-CRB is a lower bound on the minimal achievable variance. According to proposition 1, (13) can be rewritten for Gaussian
1In

the sense that (I

) is a fat matrix (more columns than rows).

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1377

Note that closely spaced DOA means that at least one known DOA is close to at least one unknown DOA.2 If the sources are correlated and according to property (i), we cannot expect perform better than the CRB over the subspace . This is a fundamental limit. Regarding propof interest, erty (ii), prior knowledge cannot help for a large number of sensors. If the known and unknown sources are uncorrelated, the known sources are not coherent and the DOA are widely spaced, then properties (iii) and (vi) mean that all the CRB are merged. Inversely [cf. property (v)], if the known sources are coherent or highly correlated, prior knowledge can be exploited. In conclusion, the use of orthogonal projector to introduce prior knowledge into an estimation algorithm is recommended for some limit situations as for closely spaced DOA associated with coherent (or possibly highly correlated) sources with small/moderate number of sensors. Some of these conclusions have been already obtained in [6] and [11] in the context of the constrained MUSIC algorithm but here we present a more general argumentation since we base our analysis on a statistical quantity which is independent of the specic choice of the estimation algorithm. An important point is that, for nite , property (iv) suggests that the orthogonal projector does not completely cancel the inuence of the known DOA on the estimation of the unknown for block-diagones and does not reach the CRB over onal source covariance associated with closely spaced DOA. In Section IV, we propose an estimation scheme which solves this problem.

where the cost function is dened by (20) in which is a complex amplitude vector. Let then the cost function can be rewritten according to (21) (22) since and . Consequently according to (22), minimizing is equivalent to look for in the subspace of interest and in the same vectors . time to discard the DOA belonging to the known space In contrast with orthogonal projectors, the steering manifold is not affected by projector [cf. (11)]. A. Prior-MUSIC (P-MUSIC) Algorithm A standard minimal-norm solution of parameter is given by with respect to

(23) This solution satises Consequently, the cost function of the P-MUSIC algorithm .

IV. PRIOR KNOWLEDGE-BASED MUSIC ALGORITHMS To be in line with property (iv), we assume that the sources associated with the known and with the unknown parts of the steering manifold are uncorrelated. Consequently, the spatial covariance matrix of the sources is block-diagonal and is dened in (7). In addition, we denote unknown quantities by the . hat symbol. According to this notation, we have It can be seen that the Constrained-MUSIC (CMUSIC) algorithm [6], [11] attempts to nd one component at a time which is most orthogonal to the noise subspace of the partially known steering manifold, . The CMUSIC optimization problem can where be described according to (18) (25) Note that this problem is different to the standard unconstrained problem associated with the MUSIC algorithm: with where both matrices and are unknown. Following the formalism introduced in [7], [16] for the MUSIC algorithm, the constrained Prior-MUSIC criterion is given by (19)
2More precisely, the inter-DOA distance must be under the Rayleigh resolution [23], i.e.,  (sin( ) sin( )) 2 6=(L 1).

(24) reaches its minimum value (wrt. ) for the known and unknown DOA. However, criterion (19) must be minimal only for the known DOA. So, we can expect that in some limit situations as for low SNR or for closely spaced DOA, this approach will be suboptimal. Consequently, in the next paragraph, we propose another approach which completely solves criterion (19). B. Weighted Prior-MUSIC (WP-MUSIC) Algorithm 1) A Second Resolution Based on the Obliquely Weighted Pseudoinverse: To solve criterion (19), consider the following minimal-norm solution:

where cording to

is the obliquely weighted pseudoinverse3 dened by . Then, the cost function can be rewritten ac-

(26) (27) (28)

3We do not confuse the obliquely weighted pseudoinverse with the standard oblique pseudoinverse dened in [1].

j

1378

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

The previous expressions are derived by using some basic properties of oblique projectors dened in (53) and (54) and the fact that projectors are idempotent. So, the nal WP-MUSIC . Note that criterion is given by , we have . So, and thus is minimal. , we have and and thus . So, is not minimal. To show that is a minimal-norm solution of criterion (19), of the cost function we consider the partial derivative wrt. according to

According to (23) and the fact that . In addition, it comes to show

, it is direct

(32) (33) (34) (35) where we have used the results given in proposition 2. Consequently, the P-MUSIC and WP-MUSIC cost functions become

(29) We have two cases, then . So, is a minimal norm solution of criterion (19) for the unknown DOA. then . Thus for the known DOA, is not a minimal-norm solution or in other words, the cost function does not reached its minimum wrt. for the known DOA. This fact is very desirable since criterion (19) must be minimum only for the unknown DOA. This property can be understood as a reinforcement of the rejection of the known DOA. In addition, for the . known DOA, it comes 2) Link To the MUSIC-Like Criterion: One can easily verify and, therefore, the cost function that can be decomposed into two contributions according to (30) has been dened in (18) and . As we can see, the above expression is a CMUSIC criterion with an additional corrective term which takes into account the prior knowledge. where C. Large Number of Sensors We begin by exposing an asymptotic result regarding oblique projectors. and in1) Proposition 2: For large and if tersect trivially, we have and . Proof: then As and , we have where and are mutually orthogonal. This implies . Using these properties together with the denition of an oblique projector given in (52), it is easy to show the proposition. It is straightforward to see that for a large number of sensors, we have This result means that asymptotically, the criterions of the P-MUSIC and WP-MUSIC are in fact the criterion of the MUSIC over the orthonormalized subspace of interest. In addition, the WP-MUSIC and P-MUSIC algorithms based, and are asymptotirespectively, on criterion cally equivalent. However, for more realistic situations where takes moderate values, we show in the simulation part that the two approaches are no longer equivalent, as expected in Sections IV-A and IV-B. V. IMPLEMENTATION OF THE WP-MUSIC CRITERION As claimed at the end of Section IV-B, we focus our analysis on the WP-MUSIC algorithm. To implement this algorithm, we have two possibilities. We can use (27) or (28). The latter is more readable since it explains how the WP-MUSIC algorithm works. However, it is preferable to implement (27) for the two following reasons. 1) More DOA can be estimated. Since the second expres, sion of the WP-MUSIC criterion involves projector to ensure that is we have to satisfy constraint a rank- matrix. For the rst expression, only matrices and are involved through projector . In that case, we have to satisfy the following constraints:

is of rankis of rankThe two last constraints can be reformulated as which is less restrictive than since we have . In fact, combining the constraints on the rank of matrices and , we obtain which allows possible values for greater than . 2) In a computational point of view, (28) involves the estimaand while (27) involves only tion of projectors the estimation of projector . A. Estimation of Oblique Projectors

where

(31)

1) Invariant to Change of Basis: The oblique projectors and are invariant to change of basis. Indeed

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1379

a basis of space is not unique, so consider another such as . We know that there exits basis such as . In that case, it an invertible matrix comes the two following equalities and . This invariance property for is a consequence of the fact that is essentially unique since . For projector this result in the following manner: , we can show

coincide with the The peaks in the pseudospectrum can be inunknown DOA. Note that the minimization of terpreted as a generic one-dimensional subspace tting problem [25]. C. Root WP-MUSIC The enumerative search procedure associated with the spectral WP-MUSIC criterion is a costly operation. Thanks to the ULA assumption, we expose the root version of the WP-MUSIC algorithm which has a lower complexity cost. In addition, it is well known that the root version of the MUSIC algorithm is superior to its spectral form [15]. . The cri1) Root-CMUSIC Principle: Let terion of the root-CMUSIC4 is given by (44) where denotes an unitary basis of the noise space obtained through the methodology introduced in [6] and [11]. has a Vandermonde strucDue to the ULA assumption, ture and the DOA estimation problem can be formulated in term of nding the zeros of the above conjugate centrosym. This symmetry is a consemetric polynomial of degree and the quence of the Hermitian character of projector denoted explicit computation of the coefcients of is given by summing along the diagonal of by and is the projector matrix. In addition, we have real and equals to . Moreover, one can is equal to its reciprocal polynoeasily verify that is a zero then is also a zero, mial [2] and, therefore, if occur in pairs. Note that for the desired DOA, i.e., , i.e., the DOA belong to the unit we have constraint circle. In presence of noise, the DOA may be extracted (among possible roots) based on their proximity to the unit circle. 2) Polynomial Form of the Corrective Function and Root WP-MUSIC Algorithm: Here, we follow the same methodology as for the root-CMUSIC approach, and we associate a polynosuch as for all unknown DOA, the folmial form to lowing polynomial: (45) (46) must be zero. By analogy to (42), we have Then, polynomial (46) is obtained by remarking that .

(36) : In criterion (27), we need 2) Estimation of Projector to partially estimate projector . So, knowing projector , we have to estimate a basis of . Consider the sample weighted spatial covariance matrix of the noise-free array response: (37) As admits a Vandermonde-type decomposition according to (6) and as we assume that is block-diagonal, it comes (38) is . Now, consider the Singular Value Deand the rank of composition (SVD, [18, p. 355]) of the sample weighted spatial covariance (39) is a unitary basis of and is a where unitary basis of the null-space of the sample weighted spatial . We use the right basis since covariance, denoted by destroys the Hermitian character of the sample projector spatial covariance. Projector can be computed according to (40) (41) (42) where . In presence of noise, it is preferable to consider where is dened in (55) since we have shown that does not destroy the statistical properties of the noise. Finally, we can formulate the spectral form of the WP-MUSIC algorithm. B. Spectral WP-MUSIC The spectral WP-MUSIC criterion is where (43)

(47) (48) (49)

4Here again, we expose the root version of the WP-MUSIC algorithm based on (28) since we consider that this expression clearly highlights the link between the root WP-MUSIC and the root-CMUSIC. However, for the two reasons explained in Section V, it is preferable to use (27) to really implement the root WP-MUSIC based on the resolution of polynomial p(1=z ) (I ^ ^ ^ E )(I E )p(z ) where projector E is given by (42).

1380

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

Fig. 2. CRB versus SNR [dB], Closely spaced DOA, (top) L = 10, (middle) L = 300, (bottom) L = 1000.

three derived bounds: the , the and the , we consider two situations. 1) One Known DOA and One Unknown: In this situation, is known and have to be estimated. The covariance matrix of the sources is given by with and . The coherent scenario . is considered for singular spatial covariance, i.e., for In Fig. 1(a), we consider uncorrelated, correlated and coherent sources with widely spaced DOA. In that case, the Prior-CRB is comparable with the CRB for a single or two DOA. So, in this situation, the knowledge of cannot help the estimation of . This situation conrms property (vi) and (iii) even if the sources are correlated. In Fig. 1(b), the DOA are closely spaced. In this case, the CRB for one DOA is much lower than the P-CRB and the CRB for two DOA, as expected in (iv). In addition, for block-diagonal source covariance, the P-CRB and the CRB for two DOA are merged according to property (iii). This illustrates in particular property (iv). The important point is that even if the sources are highly correlated, the gain associated with the P-CRB with respect to the CRB for two DOA is small. Inversely, for coherent sources with closely spaced DOA, the P-CRB is much lower than the CRB for two DOA. This observation illustrates property (v). In that case, prior knowledge is benecial. Finally, in Fig. 2, we vary the number of sensors for closely spaced DOA. We can see that the CRB are asymptotically merged which conrms property (ii). 2) Two Known DOA and One Unknown: Here, we know and and we want to estimate . The spatial covariance is given by (50) where and (51)

Fig. 1. CRB versus SNR [dB]. (a) Widely spaced DOA ( = [5 80 ]). (b) Closely spaced DOA ( = [5 5:2 ]).

Note that due to the fact that the coefcients of , noted

is Hermitian, , are conju-

gate centrosymmetry, i.e., and therefore occur in pairs. Consequently, the root WP-MUSIC is based on the following result. roots of polynomial Theorem 3: The where and has been dened in (44) and (45), respectively, are the set of the DOA without the subset of the known DOA. Proof: See Appendix E. VI. NUMERICAL SIMULATIONS A. Numerical Analysis of the Prior-CRB The geometry of the array is Uniform and Linear (ULA) of sensors and snapshots. So, the Rayleigh resolution is about 0.49. To illustrate the comparison between the

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1381

Fig. 3. CRB versus SNR [dB]. (a) Widely spaced DOA ( = [5 80 40 ]). (b) Closely spaced DOA ( = [5 5:2 7 ]).

with , and varies according to in the previous section. So, the known DOA are highly correlated and the correlation coefcient between the unknown DOA associated with the rst and third sources varies until to the coherent scenario. On Fig. 3, we have drawn the CRB for closely spaced and for widely spaced DOA. Like in the previous situation, the P-CRB indicates that the exploitation of prior knowledge is interesting only for coherent sources with closely spaced DOA. B. Illustration of the WP-MUSIC Algorithm Here, we consider a numerical example to illustrate the WP-MUSIC algorithm. On Fig. 4(a), we have drawn the pseudospectrums of the CMUSIC and the WP-MUSIC algorithms for three DOA where one is known and two others have to be estimated. First, note on the Prior-MUSIC pseudospectrum that the known DOA at 100 has been efciently cancelled from the CMUSIC pseudospectrum without altering the unknown
Fig. 4. (a) CMUSIC and WP-MUSIC pseudospectrums for three DOA ( ); f ( ) and C ( ) for (one known and two unknown). (b) f L = 18 sensors. (c) Zero location with respect to the unit circle.

one. In contrast with the CMUSIC algorithm, we can note on has only two null values at 50 and 150 . Fig. 4(b) that

1382

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

In Fig. 4(c), we have drawn the zero location with respect to the unit circle for the root-CMUSIC and root WP-MUSIC algorithms. Note that the zeros occur in pairs, as expected. However, in presence of noise, selecting the zeros (with unit modis a difcult task due to ulus constraint) based only on their proximity to the unit circle. So, a decision only based on seems ineffective. Inversely, note that a decision on criterion is a more practicable task.

C. Performances of the Algorithms 1) Accuracy of the Proposed Methods: We assume that the source covariance is block-diagonal. The tested methods are WP-MUSIC: The MUSIC algorithm with prior knowledge based on the obliquely weighted pseudoinverse (cf. Section IV-B). P-MUSIC: The MUSIC algorithm with prior knowledge based on standard LS resolution (cf. Section IV-A). P-MUSIC (SI): The MUSIC algorithm with prior knowledge based on the implementation of the oblique projector proposed by McCloud and Scharf [12]. MUSIC: The standard root version of the MUSIC algorithm. CMUSIC: The root version of the constrained MUSIC algorithm presented by DeGroat et al. [6]. The accuracy of the DOA of interest estimation is measured through the Standard Deviation (Std) which is dened as the root of the MSE. Each simulation is based on 1000 Monte Carlo trials. In Fig. 5(a), we consider widely spaced DOA, . In this situation, all the tested algorithms e.g., are equivalent. On Fig. 5(b), we choose closely spaced DOA, which corresponds to a distance inter-DOA e.g., much lower than the Rayleigh resolution for 10 sensors. In this scenario, the CRB for two DOA and the P-CRB for the DOA of interest are merged, as expected in property (iii) of Theorem 1 (cf. Section III-D) in the context of more than two DOA. We can note that for sufcient SNR, the CMUSIC and the MUSIC algorithms reach these bounds but they cannot outperform it for closely spaced DOA and for block-diagonal source covariance. The WP-MUSIC, P-MUSIC, and P-MUSIC (SI) show Std close to the CRB for only one DOA at high SNR. By the light of this example, we can say that most of the inuence of the known DOA has been efciently cancelled by the proposed algorithms. This is not the case for the CMUSIC algorithm. According to Fig. 5(b) and (c), the WP-MUSIC algorithm is slightly more efcient than the P-MUSIC algorithm at low dB) and for closely spaced DOA. This observation SNR ( conrms the discussion in Sections IV-A and IV.B. Finally, we perform in Fig. 5(d), (e), and (f), some experiments with two known DOA and one unknown. The conclusions are similar to the more simple case of one unknown and one known. 2) Robustness to a Small Error on the a Priori: The scenario is the same as for Fig. 5(a), i.e., the known DOA is and the unknown DOA is . We perturb according to and we compute the standard deviation for the DOA of interest, . This scenario is repeated for different SNR and number of sensors. The number of snapshots is equal to 100.

Fig. 6(a) shows that without noise, the P-MUSIC (SI) and the CMUSIC are very sensitive to a small error on the known DOA. Inversely, the P-MUSIC and WP-MUSIC algorithms are more robust. These observations are conrmed by Fig. 6(b) where the SNR is equal to 30 dB. Indeed, we can note the remarkable robustness of the P-MUSIC and WP-MUSIC algorithms since the Std associated with these methods follow a at curve. This is a clear advantage of these approaches. In Fig. 6(c), we can see that all the tested methods have similar robustness at low SNR (0 dB). In this situation, the error induced by the noise dominates the error associated with the error on the known DOA. In Fig. 6(d), we have drawn the Std for the tested methods . without noise and for a large number of sensor As expected, in this asymptotic regime, the CMUSIC, the P-MUSIC and the WP-MUSIC have the same efciency and robustness. Note that in this case, the bad results for the P-MUSIC (SI) algorithm. These observations are conrmed in Fig. 6(e) and (f) in the noisy situation. VII. CONCLUSION OF THE SIMULATION PART 1) For a small number of sensors, the WP-MUSIC shows the best accuracy and the best robustness to a small error on the known DOA. In particular, the accuracy of this algorithm is near the CRB associated with the subspace of interest. Consequently, the inuence of the known DOA is almost cancelled. 2) As explained in Sections IV-A, IV-B and in the simulation part, the P-MUSIC algorithm is slightly less efcient than the WP-MUSIC algorithm but its robustness is comparable. This is a valuable solution. 3) The P-MUSIC (SI) shows a comparable accuracy as the P-MUSIC and the WP-MUSIC algorithms at high SNR but this algorithm is less efcient in severely noisy situations. In addition, its robustness is weak. So, essentially for the latest reason mentioned, we prefer the implementation of the oblique projector introduced in Section V-A rather than the one presented by McCloud and Scharf.5 4) For a small number of sensors and for closely spaced DOA associated with block-diagonal source covariance, the P-MUSIC, the WP-MUSIC, and the P-MUSIC (SI) algorithms outperform the CMUSIC algorithms, in particular at high SNR where the CMUSIC algorithm is lower bounded by the CRB over the whole space. For a large number of sensors and/or widely spaced sources, this algorithm is equivalent to the ones based on oblique projectors. VIII. CONCLUSION In this paper, we have presented a subspace-based solution using the knowledge of to estimate DOA among known DOA. In a rst part of this paper, we have derived and analyzed the CRB associated with the orthogonal deation of the signal subspace and we have shown several limitations of this approach. Consequently in the second part, we have proposed alternative solutions based on an oblique deation of the
5Note that in this work, the proposed methodology also assumes the blockdiagonal structure of the spatial covariance of the sources.

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1383

Fig. 5. Std versus SNR for two sources. (a) Widely spaced DOA  = [80 5 ] with L = 10 sensors and T = 100 snapshots. Closely spaced DOA  = [8 5 ] with L = 10 sensors. (b) With T = 100 snapshots. (c) With T = 500 snapshots. Std versus SNR for three sources. (d) Closely spaced DOA  = [8 5 12 ] with L = 10 sensors and T = 100 snapshots. (e) Closely spaced DOA  = [8 5 80 ] L = 10 sensors and T = 100 snapshots. (f) Widely spaced DOA  = [8 50 80 ] L = 10 sensors and T = 100 snapshots.

1384

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

signal subspace. We show that the proposed algorithm, called Prior-MUSIC, mitigates almost the inuence of the known DOA on the DOA of interest in particular when the DOA are closely spaced and the source covariance is block-diagonal. Finally, the oblique projector framework provides a suitable way to integrate prior knowledge into subspace-based methods and more generally into subspace tting techniques. APPENDIX A. Brief Discussion of Oblique Projectors This appendix is dedicated to oblique projections [1]. In parto ticular, we recall that the only requirement of a matrix be a projector is is idempotent, i.e., . Let and be subspaces of that intersect trivially, . Then, the projector on along i.e., is the linear operator satisfying: ; ; . The geometric interpretation of the above properties is where and then . So, the complex Euclidean Space is decomposed according to . Let be a complex matrix having full column rank, obtaining by the concatenation of matrices and according . The orthogonal projector onto is then to and dened as (52) This property is important since it highlights the link between and oblique projectors and the orthogonal projector . In addition, the ranges for and are and , respectively, and the null spaces for and are and , respeccan be deduced from (VIII-A) tively. A useful rewritten of according to (53) Finally, note that (59) (54) (60) B. Demonstration of Proposition 1 First, note that an ordered eigendecomposition of any rankidempotent matrices is C. Demonstration of Theorem 1 We recall that the signal plus nuisance model parameter where vector by and . The rst term in (14) is associated with the noise and is given by (61)

Fig. 6. Std versus Error on the known DOA with L = 10 sensors. (a) Without noise. (b) With SNR = 30 dB. (c) With SNR = 0 dB. L = 100 sensors and T = 100 snapshots. (d) Without noise. (e) With L = 30 sensors and T = 100 snapshots and SNR = 30 dB. (f) With SNR = 0 dB.

(57) (58) The error covariance, noted , is given by the covariance of the following centered signal then

(55) where is the rst columns of the left eigenfactor. Thanks to the property that the noise is zero-mean and , it is not difcult to see that

(62) The second term which involved the partial derivatives of the noise-free model with respect to the parameter vector can be

(56)

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1385

expressed according to (63)

where to obtain (68) [respectively, (69)], we use (53) [respectively, (54)]. Consequently, the P-CRB can be simplied to (15). D. Demonstration of Theorem 2 (64) (65) To show property (i), it is equivalent to prove that is a positive semidenite , it is (psd) matrix. First, using is idempotent. Therefore, straightforward to see that the eigenvalues of is 1 or 0 and, thus, is is a nondecient matrix, psd. Next, as is also psd. Finally using that (1) the Hadamard product of two psd matrices is also a psd matrix (cf. result R19 in [18]) and (2) the real part of a psd matrix is psd itself, we prove (i). Property (ii) can be proved in the following manner. In [21], it has been shown that . For the P-CRB, we have and then

where . . . . . . (66)

with

. So, we obtain

The P-CRB is given by the matrix at the bottom of the page. As the nuisance and signal parameters are decoupled, the P-CRB has a block-diagonal structure. To obtain the P-CRB for , we follow the block-diagonalization method insubvector troduced in [18, p. 390]. Finally, we have

(71) Consequently, using the denition of the P-CRB, it comes

which proves property (ii). Next, note that the CRB over the whole space can be rewritten according to where . After some straightforward algebraic derivations and for sufcient number of snapshots, it comes

(73) . In addition, suppose that where is block-diagonal in (73) then it is easy to deduce property (iii) since

(respectively, ) be the oblique projector on (respectively, ) along (respectively, ) dened in (52). Based on these operators, we have the following property: (67)

Let

(74) Note that this relation holds for widely or closely spaced DOA. If the DOA are widely spaced, the inuence between the DOA . is weak then it is well known that Consequently, according to property (iii), the reaches its minimum near the by superior values [cf. relation (i)], which proves property (vi). Inis invariant versely, for closely spaced DOA, the

(68) (69) (70)

1386

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 4, APRIL 2008

while the inverse of is near the singularity (large condition number with respect to the inversion). Thus, . This fact together with property (iii) give property (iv). is decient (coherent sources), then Now suppose that is also decient.6 In addition, if we due to its block structure, takes a large consider closely spaced DOA then value as the inverse of the Hadamard product of two (near) sin, the gular matrices. In the same time and through projector P-CRB remains sensitive to the known DOA but insensitive to the correlation between the sources associated with the known . Consequently, propDOA, i.e., to the spatial covariance erty (v) holds. For correlated sources, the problem is more complicated and we have deferred the discussion to the simulation part. E. Demonstration of Theorem 3 As we know occur in pairs, we can give the factorand according ized forms of polynomials to

Clearly, has no trivial roots, i.e., any known or unknown . Inversely, we only have DOA are solution of for the unknown DOA. So, according to (77), zeros of are only the DOA which annulate , i.e., the unknown DOA. REFERENCES
[1] R. T. Behrens and L. L. Scharf, Signal processing applications of oblique projection operators, IEEE Trans. Signal Process., vol. 42, no. 6, pp. 14131424, Jun. 1994. [2] Y. Bistritz, Zero location with respect to the unit circle of discrete-time linear system polynomials, Proc. IEEE, vol. 72, no. 9, pp. 11311142, Sep. 1984. [3] G. Bouleux and R. Boyer, Zero-forcing based sequential MUSIC algorithm, in Proc. 32nd IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP07), 2007, vol. 3, pp. 10171020. [4] R. Boyer and G. Bouleux, Cramr-Rao lower bound for prior-subspace estimation, in Proc. IEEE Sens. Array Multi-channel Process. (SAM 2006), 2006, pp. 394398. [5] H. Chen, S. Van Huffel, D. van Ormondt, and R. de Beer, Parameter estimation with prior knowledge of known signal poles for the quantication of NMR spectroscopy data in the time domain, J. Magnet. Reson., ser. A, vol. 119, no. 2, pp. 225234, Apr. 1996. [6] R. D. DeGroat, E. M. Dowling, and D. A. Linebarger, The constrained MUSIC problem, IEEE Trans. Signal Process., vol. 41, no. 3, pp. 14451449, Mar. 1993. [7] Y. Ephraim, N. Merhav, and H. L. Van Trees, Min-norm interpretations and consistency of MUSIC, MODE and ML, IEEE Trans. Signal Process., vol. 43, no. 12, pp. 29372942, Dec. 1995. [8] P. C. Hansen and S. H. Jensen, Prewhitening for rank-decient noise in subspace methods for noise reduction, IEEE Trans. Signal Process., vol. 53, no. 10, pt. 1, pp. 37183726, Oct. 2005. [9] S. Kayalar and H. L. Weinert, Oblique projections: Formulas, algorithms and error bounds, Math. Contr. Signals Syst., vol. 2, no. 1, pp. 3345, 1989. [10] H. Krim and M. Viberg, Two decades of array signal processing research: The parametric approach, IEEE Signal Process. Mag., vol. 13, no. 4, pp. 6794, Jul. 1996. [11] D. A. Linebarger, R. D. DeGroat, E. M. Dowling, P. Stoica, and G. L. Fudge, Incorporating a priori information into MUSIC-algorithms and analysis, Signal Process., vol. 46, no. 1, pp. 84105, Sep. 1995. [12] M. L. McCloud and L. L. Scharf, A new subspace identication algorithm for high-resolution DOA estimation, IEEE Trans. Antennas Propag., vol. 50, no. 10, pp. 13821390, Oct. 2002. [13] J. C. Mosher and R. M. Leahy, Source localization using recursively applied and projected (RAP) MUSIC, IEEE Trans. Signal Process., vol. 47, no. 2, pp. 332340, Feb. 1999. [14] S. K. Oh and C. K. Un, A sequential estimation approach for performance improvement of eigenstructure-based methods in array processing, IEEE Trans. Signal Process., vol. 41, no. 1, pp. 457463, Jan. 1993. [15] B. D. Rao and K. V. S. Hari, Performance analysis of root-music, IEEE Trans. Signal Process., vol. 37, no. 12, pp. 19391949, Dec. 1989. [16] R. O. Schmidt, A signal subspace approach to multiple emitter location and spectral estimation, Ph.D. dissertation, Stanford Univ., Stanford, CA, 1981. [17] A. A. Shah and D. W. Tufts, Determination of the dimension of a signal subspace from short data records, IEEE Trans. Signal Process., vol. 42, no. 9, pp. 25312535, Sep. 1994. [18] P. Stoica and R. L. Moses, Spectral Analysis of Signals. Englewood Cliffs, NJ: Prentice-Hall, 2005. [19] P. Stoica and A. Nehorai, Performance comparison of subspace rotation and MUSIC methods for direction estimation, IEEE Trans. Signal Process., vol. 39, no. 2, pp. 446453, Feb. 1991. [20] P. Stoica and A. Nehorai, Comparative performance study of element-space and beamspace MUSIC estimators, Circuits, Syst. Signal Process., vol. 10, no. 3, pp. 285292, 1991. [21] P. Stoica and A. Nehorai, MUSIC, maximum likelihood, and Cramer-Rao bound, IEEE Trans. Signal Process., vol. 37, no. 5, pp. 720741, May 1989. [22] P. Stoica, Y. Selen, N. Sandgren, and S. Van Huffel, Using prior knowledge in SVD-based parameter estimation for magnetic resonance spectroscopythe ATP example, IEEE Trans. Biomed. Eng., vol. 51, no. 9, pp. 15681578, Sep. 2004.

(75) and

(76) where and (76), are the desired (known or unknown) zeros and are the extraneous zeros. Based on (75) and admits the following factorization: (77) where

and

6We

det(

have det(R ) = det(R ) = 0.

) det(

) = 0 since we assume that

BOYER AND BOULEUX: OBLIQUE PROJECTIONS FOR DOA ESTIMATION

1387

[23] D. Thubert and L. Kopp, Measurement accuracy and resolving power of high resolution passive methods, Proc. EURASIP, vol. 86, pp. 10371040, 1986. [24] P. Vandaele and M. Moonen, Two deterministic blind channel estimation algorithms based on oblique projections, Signal Process., vol. 80, no. 3, pp. 481495, Mar. 2000. [25] A.-J. Van Der Veen, E. F. Deprettere, and A. L. Swindlehurst, Subspace-based signal analysis using singular value decomposition, Proc. IEEE, vol. 81, no. 9, pp. 12771308, Sep. 1993. [26] M. Viberg and B. Ottersten, Sensor array processing based on subspace tting, IEEE Trans. Signal Process., vol. 39, no. 5, pp. 11101121, May 1991. [27] P. Wirawan, K. Abed-Meraim, H. Maitre, and P. Duhamel, Blind multichannel image restoration using oblique projections, in Proc. IEEE Sens. Array Multichannel Process. (SAM2002), 2002, pp. 125129. [28] X. Yu and L. Tong, Joint channel and symbol estimation by oblique projections, IEEE Trans. Signal Process., vol. 49, no. 12, pp. 30743083, Dec. 2001. Rmy Boyer received the M.Sc. and Ph.D. degrees in the eld of signal processing and communications from cole Nationale Suprieure des Tlcommunications (ENST), Paris, France, in 1999 and 2002, respectively.

From 2002 to 2003, he was a Visiting Researcher with the University of Sherbrooke, QC, Canada. He is currently an Associate Professor with the Laboratoire des Signaux et Systmes (L2S), Gif-sur-Yvette , with the CNRS-UPS-SUPELEC (National Center of Scientic Research and University of Paris-Sud). His research activities include multiway-arrays processing, sensor array, sinusoidal modeling, and watermarking.

Guillaume Bouleux received the engineering degree in electronic and signal processing from the Ecole Nationale Suprieure de lElectronique et de ses Applications (ENSEA), Cergy, France, in 2004 and the Master degree from Universite de Cergy-Pontoise, France, in 2005. He is currently working toward the Ph.D. degree in signal processing in collaboration with both Universite Jean Monnet (to LASPI laboratory), Saint Etienne, France and Universite de Paris-Sud, Orsay, France (to LSS-Supelec laboratory). His research interest includes array signal processing, statistical analysis, linear, and multilinear algebra.