1110 IEEE SIGNAL PROCESSING LETTERS, VOL. 20, NO.

11, NOVEMBER 2013

Wideband Gaussian Source Processing

Using a Linear Nested Array
Keyong Han, Student Member, IEEE, and Arye Nehorai, Fellow, IEEE

Abstract—Based on a recently proposed linear nested array, we

consider the problem of source number detection and direction of
arrival estimation for wideband Gaussian sources. This array pro-
vides degrees of freedom with sensors, enabling
us to estimate sources with sensors. To employ the
nested array for the wideband case, we propose effective strategies
to apply nested-array processing to each frequency component, Fig. 1. A 2-level nested array with sensors in the inner ULA, and sen-
and combine all the spectral information of various frequencies sors in the outer ULA, with intersensor spacings and respectively.
to conduct the detection and estimation. In particular, for source
detection, we propose a novel approach employing the idea of en-
semble, used in machine learning and statistics. Numerical simu- property of nonuniformity, the resulting difference co-array has
lations demonstrate the advantages of our strategies. significantly more DOF than the original sparse array, which
makes it possible for the nested array to detect more sources
Index Terms—Direction of arrival estimation, ensemble, nested
array, source number detection, wideband source.
than sensors [9]. Similar strategies for identifying more sources
than sensors were proposed by using co-prime arrays [10].
Note that the application of nested microphone array is not new
I. INTRODUCTION in the processing of acoustic and speech signals [11]. However,
they usually do not examine the analytical aspects of the nested

S OURCE number detection and direction of arrival (DOA)

estimation are two major applications of antenna arrays.
For narrowband sources, theories are well established and a
array.
In this letter, we propose two algorithms for wideband source
number detection based on SORTE. Further we propose a com-
large body of literature exists [1]. Owing to the narrowband bined method, employing the idea of ensemble [12], used in ma-
property, the array model can be greatly simplified. Numerous chine learning and statistics. In addition, we construct a com-
methods exist for source number detection and DOA estimation, bined MUSIC spectrum to exploit all the spectral information
including the eigenvalue-based method second order statistic of from different frequency analyses. Simulations are provided to
eigenvalues (SORTE) [2] and the multiple signal classification demonstrate the advantages of our strategies.
(MUSIC) [3], [4].
For wideband sources, however, the literature is less abun-
dant. Wax et al. are among the earlier researchers in this field [5], II. SIGNAL MODEL
decomposing the incoherent wideband signal into many narrow- We assume there is a 2-level nested array with sensors,
band signals, and using discrete Fourier transform (DFT) along including two concatenated ULAs. Suppose the inner ULA has
the temporal domain. We consider uncorrelated sources in this sensors with spacing , and the outer ULA has sensors
letter. with spacing , where . More
Using conventional subspace-based methods such as precisely, a 2-level nested array is a linear array with sensor
MUSIC, a uniform linear array (ULA) with sensors can locations given by the union of the sets
resolve at most sources. In earlier works, the problem and , as
of detecting more sources than sensors has been addressed in shown in Fig. 1.
different ways [6], [7]. Based on a linear nested array, a sys- We further assume wideband sources are in the surveil-
tematic approach to achieve degrees of freedom (DOF) lance region, impinging on this linear array from directions
using sensors was recently proposed in [8]. Owing to the . Suppose that the incident wideband
signals have a common bandwidth with center frequency .
Manuscript received July 30, 2013; revised September 02, 2013; accepted Let denote the th baseband signal. Then the observed
September 07, 2013. Date of publication September 11, 2013; date of current bandpass signal at a reference point can be written as
version September 20, 2013. This work was supported by the Air Force Office of
Scientific Research under Grant FA9550-11-1-0210 and by the Office of Naval
Research under Grant N000141310050. The associate editor coordinating the (1)
review of this manuscript and approving it for publication was Prof. Augusto
Sarti.
The authors are with the Preston M. Green Department of Electrical and Sys-
If we observe the signal over the time interval , then the
tems Engineering, Washington University in St. Louis, St. Louis, MO, 63130 baseband signal can be written as [13]
USA (e-mail: kyhannah@ese.wustl.edu; nehorai@ese.wustl.edu).
Color versions of one or more of the figures in this letter are available online
at http://ieeexplore.ieee.org. (2)
Digital Object Identifier 10.1109/LSP.2013.2281514

1070-9908 © 2013 IEEE

HAN AND NEHORAI: WIDEBAND GAUSSIAN SOURCE PROCESSING 1111

where are the Fourier coefficients III. ARRAY PROCESSING FOR WIDEBAND SOURCES
We will use the nested array mentioned above to conduct
(3) source number detection and direction of arrival estimation.
First, we will briefly introduce spatial smoothing [8], which
with . denotes is used to exploit the increased degrees of freedom. Then we
the lowest frequency included in the bandwidth , and is the propose two algorithms based on SORTE, and further propose
number of frequency components. We choose and so that a novel strategy employing the idea of ensemble for source
the frequencies are symmetric about 0 Hz. By considering the number detection. Third, we will provide a novel strategy for
propagation delay of the th signal at the th sensor, the wideband source estimation using MUSIC.
modulated bandpass signal at the reference point can be pre-
sented as A. Spatial Smoothing
To exploit the increased degrees of freedom provided by the
(4) co-array, we need to apply spatial smoothing. We remove the
repeated rows from and also sort them so that the th
row corresponds to the sensor location
where , , and in the difference co-array of the 2-level nested array, giving a
, with being the propagation speed.
new vector: , where
The demodulated signal can be expressed as
is a vector of all zeros except a 1 at the center position. The
(5) difference co-array of this 2-level nested array has sensors lo-
cated at
Stacking according to sensor number, we get
the 1 vector . Let denote the 1
steering vector of the th source and the th frequency compo-
nent:
We now divide these sensors into
(6) overlapping subarrays, where the th subarray has sensors
located at .
Then the received data vector has the form The th subarray corresponds to the th
to th rows of , denoted as .
We can check that where
(7)
.
Viewing as a newly received vector, we get the equivalent
where , covariance matrix Taking the average of yields
is the 1 signal vector, and
is the 1 noise Fourier
coefficient vector. Define (10)

(8)
The spatially smoothed matrix enables us to identify up to
Then are by definition the 1 Fourier coefficient vec- sources with sensors.
tors of .
We assume the source signals follow Gaussian distribu- B. Source Number Detection
tions, , and that they are all independent As mentioned in the introduction, we consider a narrowband
of each other. The noise is assumed to be white decomposition for the wideband case. Considering the spatial
Gaussian and uncorrelated with sources. Based on our as- smoothing matrix for the th frequency, , we do eigen-
sumption, the source autocorrelation matrix is diagonal: value decomposition: , where
. We use to represent are the eigenvalues, and
for brevity. Then the autocorrelation matrix of
is the corresponding eigenvector ma-
is , where is the noise power, and
trix. We suppose the eigenvalues are sorted decreasingly:
is the identity matrix. Vectorizing [8], we get

(9)

where , and , We denote . According to the SORTE

with being a vector of all zeros except a 1 at the th position. method, we define a gap:
We can view vector in (9) as some new longer received sig-
nals with the new manifold matrix , and the new source
signals . The symbol denotes conjugation without transpose,
and denotes the Khatri-Rao product.
1112
where , , and
Then the source number is . For each segment , we employ DFT to get the
Note that the number is based on
the information of frequency . To exploit all the frequency
information, we propose the following two algorithms.
• SORTEC
Based on the detected source number for each fre-
quency, we count the occurrence of distinct numbers
. Then we find the source number:
.
• SORTES
Based on the gaps for each frequency, we take the
summation first rather than conduct detection. Then
we decide the source number using the summation:
. ingly, we can conduct source detection and DOA estimation
Existing literature investigates source number detection by
using various methods singly. The idea of ensemble, used in
machine learning and statistics, inspires us to conduct source
detection by combining multiple methods. Typically, more com-
putation is required to evaluate the ensemble’s performance than
for a single method.
• Ensemble
Suppose we have detection methods in total:
D D D , each with detection accuracy . Then
we obtain the detected source number
based on each method. Next, we count the occurrence of
each distinct number, denoted as , there are
with summation . The final source number is chosen as
the one that occurs most frequently: . us to use a 6-sensor ULA to detect seven sources. However,
When the detection accuracy for each method is greater than or
equal to 50%, the improvement of the ensemble is guaranteed.
C. Direction of Arrival Estimation
For this section, we assume the source number has
been correctly detected. We write the noise subspace as
, which consists of the last
eigenvectors corresponding to the smallest
eigenvalues. The estimated DOA can be found through
an exhaustive search over all the direction space of the
MUSIC spectrum: , where
, with . components via DFT. The selection of proper value of will be
Combining the resulting measurements for all the different
frequencies, we consider the new combined MUSIC spectrum:
(11)
Then the estimated DOAs correspond to the largest values of
the spectrum . and the combined ensemble method. It describes the detection
D. Wideband Sample Covariance for Array Processing
According to Section II, our observed data is in (7), and
our problem of interest is to detect the source number and esti-
mate the DOAs from the Fourier coefficients
in (8). Suppose our total observation time is , and we divide
IEEE SIGNAL PROCESSING LETTERS, VOL. 20, NO. 11, NOVEMBER 2013
it into segments, with each segment . We as-
sume that there are samples within each segment. Therefore,
we have samples: .
We will investigate the impact of and on the performance
through numerical examples.
corre-
sponding frequency coefficient matrix:
(12)
Considering all the segments, we can get the coefficient
matrix for each frequency index :
(13)
The resulting sample covariance matrix for frequency index
can be written as .
Following the spatial smoothing technique in subsection ,
we can get the sample spatial smoothing matrix . Accord-
based on this sample covariance matrix.
IV. NUMERICAL EXAMPLES
In this section, we use numerical examples to show the effec-
tiveness of our proposed strategies for wideband source detec-
tion and DOA estimation with a linear nested array.
In the examples, we consider a 2-level nested array with
sensors, with both the inner and outer ULAs having
three sensors. The interspacing is chosen as half of the
shortest wavelength of the wideband signals, which ensures
that there is no spatial aliasing. is equal to . Suppose
wideband sources impinging from directions
. It is impossible for
the spatial smoothing matrix in (10) helps a nested array
obtain this goal. Suppose the wideband sources have the same
center frequency, , and the same bandwidth,
. Further suppose the sources follow zero mean
Gaussian random processes with equal power, independent of
each other. We choose equal power for simplicity here, but our
method works also for different powers. The noises are white
Gaussian, and uncorrelated with the sources.
A. Source Number Detection
The array output is decomposed into narrowband
explained next. We choose the segment number to be .
Therefore we use a total of .
To employ the ensemble strategy, we considered three
methods: SORTEC, SORTES, and one that utilizes the variance
of transformed rotational submatrix (VTRS) [14]. The VTRS
method as used here applies a strategy similar to SORTEC.
Fig. 2 shows the results of the aformentioned three methods
accuracy with respect to the signal-to-noise ratio (SNR). The
detection accuracy is defined as , where is the trial
number, and is the number of times that the true source
number is detected. We can see that the ensemble method
outperforms all of the three separate methods, and achieves
great improvement. Note that the detection accuracy is almost
HAN AND NEHORAI: WIDEBAND GAUSSIAN SOURCE PROCESSING
Fig. 2. Detection accuracy comparison of SORTES, VTRS, SORTEC, and the
combined ensemble method with a 6-sensor nested array, , ,
. performance. Simulations demonstrate the effectiveness of
Fig. 3. MUSIC spectrum using the spatial smoothing technique, as a function
of the DOA, , , , , . The red dash
lines are the true DOAs.
always above 0.5 for different SNRs, which guarantees the
improvement of the ensemble, as discussed in the previous
section. SORTES performs much better than SORTEC, which
is reasonable because SORTES determines the source number
based on a combined gap of different frequencies.
We also investigated the performance with respect to various
numbers of snapshots: with , the effectiveness of the
ensemble can be guaranteed with over 2500 snapshots.
B. Music Spectra for DOA Estimation
Fig. 3 shows the representative MUSIC spectra using the spa-
tial smoothing technique, with respect to various angles at a
SNR of 0 dB. We can see that the proposed method can resolve
the seven wideband sources sufficiently well.
C. Impact of the Choice of
To investigate the impact on the estimation performance of
the choice of , we fixed the sample number at 4000. For dif-
1113
TABLE I
MSE VERSUS DIFFERENT NUMBERS OF
ferent numbers of , Table I shows the MSE results for esti-
mation of a wideband source with . We can see that a
moderate guarantees good performance. When is too small,
it will lose information on most frequencies. On the other hand,
when is too large, the fusion process will perform badly.
V. CONCLUSION
In this letter, we proposed novel strategies for wideband
source detection and DOA estimation with a nested array. This
approach can estimate a number of wideband sources that is
greater than the number of sensors, and obtain good estimation
our strategy. One thing to note is that the assumptions of this
strategy are relatively restrictive, and the strategy is not suitable
for correlated sources. For future work, we will investigate
the nested-array processing for more complex scenarios, like
correlated or coherent wideband sources, sources with different
bandwidth and different carrier frequency.
REFERENCES
[1] H. Krim and M. Viberg, “Two decades of array signal processing re-
search: The parametric approach,” IEEE Signal Process. Mag., vol. 13,
pp. 67–94, Jul. 1996.
[2] Z. He, A. Cichocke, S. Xie, and K. Choi, “Detecting the number of
clusters in n-way probabilistic clustering,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 32, pp. 2006–2021, Nov. 2010.
[3] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
tion,” IEEE Trans. Antennas Propag., vol. 34, pp. 276–280, Mar. 1986.
[4] P. Stoica and A. Nehorai, “MUSIC, maximum likelihood and
Cramér-Rao bound,” IEEE Trans. Acoust., Speech, Signal Process.,
vol. 37, pp. 720–741, May 1989.
[5] M. Wax, T.-J. Shan, and T. Kailath, “Spatio-temporal spectral anal-
ysis by eigenstructure methods,” IEEE Trans. Acoust., Speech, Signal
Process., vol. 32, pp. 817–827, Aug. 1984.
[6] S. Pillai, F. Haber, and Y. Bar-Ness, “A new approach to array geom-
etry for improved spatial spectrum estimation,” in Proc. ICASSP, Oct.
1985, pp. 1522–1524.
[7] Q.-K. Ma, T.-H. Hsieh, and C.-Y. Chi, “DOA estimation of quasista-
tionary signals via Khatri-Rao subspace,” in Proc. ICASSP, Apr. 2009,
pp. 2165–2168.
[8] P. Pal and P. P. Vaidyanathan, “Nested array: A novel approach to array
processing with enhanced degrees of freedom,” IEEE Trans. Signal
Process., vol. 58, pp. 4167–4181, Aug. 2010.
[9] K. Han and A. Nehorai, “Improved source number detection and direc-
tion estimation with nested arrays and ULAs using jackknifing,” IEEE
Trans. Signal Process., to be published.
[10] P. P. Vaidyanathan and P. Pal, “Sparse sensing with co-prime samplers
and arrays,” IEEE Trans. Signal Process., vol. 59, pp. 573–586, Feb.
2011.
[11] Y. R. Zheng, R. A. Goubran, and M. El-Tanany, “Experimental eval-
uation of a nested microphone array with adaptive noise cancellers,”
IEEE Trans. Instrum. Meas., vol. 53, pp. 777–786, Jun. 2004.
[12] D. Opitz and R. Maclin, “Popular ensemble methods: An empirical
study,” J. Artif. Intell. Res., vol. 11, pp. 169–198, 1999.
[13] J. Li, D. Zheng, and P. Stoica, “Angle and waveform estimation via
RELAX,” IEEE Trans. Aerosp. Electron. Syst., vol. 33, pp. 1077–1087,
Jul. 1997.
[14] J. S. Jiang and M. A. Ingram, “Robust detection of number of sources
using the transformed rotational matrix,” in Wireless Commun. and
Networking Conf., Mar. 2004, pp. 501–506.