Thursday Boonstra

Digital Signal Processing and
Beamforming
Tutorial Lecture
SKADS Marie Curie Technical Workshop

Design of Wideband Receiving Array Systems
ASTRON, November 26-29, The Netherlands
A.J. Boonstra, ASTRON, boonstra@astron.nl
1
Contents
Notation
Interferometry and imaging
Narrow band assumption
Discrete source data model
Subspace analysis
Beamforming
Spatial filtering
References
SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 2
Netherlands Foundation for Research in Astronomy November 29, 2007

Notation: signals and random variables
signal: s(t) or s, parameter t often omitted

random variable: x(t) or x, parameter t often omitted
probability density function: P (x)
x Efxg
e.g. white Gaussian noise (WGN): P (x) = p2 e
1 ( )2
2 2
R1
expected value: Efx(t)g = 1 x(t)P (x)dt
correlation: Rxy = Efx(t)y(t)g

Notation: scalars, vectors and matrices
scalar, lower case or upper case,

0 1 not bold: y or Y
x1
vector, bold lower case: x = B
..
.
C
A
xp
0 1
R1;1 R1;p
matrix, bold upper case: R = B
. ...
.. .. C
. A
Rp;1 Rp;p
estimated value: bg

Notation: operators
complex conjugate: y = Refyg Imfyg

transpose: xt = (x1; ; xp)
Hermitian or complex conjugate transpose: RH = Rt
conversion from vector d (or row) to diagonal matrix:
diag(d) = D

Spatial coherency and interferometry
Spatial coherence function: V ri ; rj ( )=

EfE ri; t E rj ; t g. ( ) ( )
Denote the intensity sky source distribution (continuous) by
Ib , then: Z 1 st (r
V (ri; rj ) = IB (s)e 2 | i rj ) d
elestial
sphere

Imaging
Van Cittert-Zernike theorem: the brightness distribution Ib (s)

and the complex visibility function V ri ; rj form a ( )
two-dimensional Fourier-transform pair:
2 D FT
I (s) ! V (r r )
b i j

Obtaining data
R
Recall: V (ri; rj ) = IB (s)e 2 | 1 st (r
i rj ) d
elestial
sphere
The covariance matrix R is a measure of the visibility,
( )
including (a.o.) antenna gains A s and electronic gains G:
Z
R(ri ; rj ) = Gi Gj Ai (s)Aj (s)IB (s)e 2| 1 st bij d
elestial
sphere

LOFAR ITS, configuration & beam shape
ITS, antenna locations ITS, snapshot uv configuration
100 200
W E
80
150
60
100
40
50
20
u (m)
y (m)
0 S N 0 S N
−20
−50
−40
−100
−60
−150
−80
E W
−100 −200
−100 −50 0 50 100 −200 −150 −100 −50 0 50 100 150 200
x (m) v (m)
0 locations:
antenna 1 PSF, or beam shape,
x1 y1 z1
R=B C spatial sample points
A of coherencies (uv-
zenith position.
imaging by beam-
xp yp zp points)
antenna signals: R Efxx g H = forming or FT
P Efkw H xk2 g = =
( )=[ ( )
x t xt t ; ; xp t t ( )℄ p p matrix H w Rw

LOFAR ITS, coherencies and sky maps
LOFAR ITS test station, ob-

served coherencies (left) and
corresponding sky map (upper
figure). Sky maps are obtained
by inverse FT or by beamforming
(as in this case: MVDR beam-
former).

LOFAR, beamforming and correlation
antenna beam
phased array beam
synthesized beam
Broadband vs narrowband processing:

in LOFAR both types, broadband mainly in analog domain.
Tradeoffs: cost, flexibility, narrowband limits (e.g. beam
squint).

Narrow band assumptions
()
If for a band-limited signal y t ,

then y (t) = m(t)e2| t , where m(t) is the complex envelope
(baseband signal) of the real-valued band-pass signal
z (t) = Refy (t)g.
Second narrow band condition: propagation delay
differences, ij = i j , along elements i and j of the
antenna/telescope array are inverse bandwidth.
(2ij ) 1
Then it can be shown that time delays across array can be
approximated by phase shifts of the baseband signal.

Narrow band assumptions
( ) = Refm(t)e2| tg, then
Recall z t
z (t ) = Refm(t )e 2| e2| t g
which implies m (t) = m(t )e 2| .

F
Define m(t) ! M( ), and let be the bandwidth of m(t):
Z 1
() = M( )e2|td
2
mt
1

Z
2
1

( ) = M( )e 2| e2|t d
2
mt
1
2

If (2ij ) 1 then, e 2 1, and m(t) (t )
, hence:
m (t) m(t)e 2| QED

Discrete source model
geometric main beam
delay
x3(t)
....
x1(t)
xp(t)
x2(t)
Consider an array of p antennas or telescopes with baselines

= ()
bij ri rj . Signals xi t are stacked in a vector, the spatially
()
white noise signals ni t (from LNAs, spill-over etc.) idem:
x(t) = [x1(t); ; xp(t)℄t

n(t) = [n1(t); ; np(t)℄t

Suppose there is one source, either astronomical or a
()
transmitter (RFI) with signal s t and from direction s
Then
x t as t n t ( )= ( )+ ( )
and h i
2 rs t
a= 1 e 2| r1 s ; ; p e
1 t | tp
1
Now define the sample estimate (observational data):

1 X
N
Rb = N xnxHn with xn = x(nTs)
n=1
()
Given i.i.d. noise vector n t with Efn t ( )n(t)H g = n2 I (I is the
() =
identity matrix), and Efs t 2 g 2 :
R = EfRb g = 2 aaH + n2 I

=
Recall for a single source: R 2 aaH n2 I, for multiple +
sources: X 2 H 2
R =
k ak ak n I +
k
Stacking q sources at directions s1 to sq in a matrix A:
A = [a1(s1); ; aq (sq )℄
and storing the source powers (brightness) is a diagonal
matrix B
=
B diag 12 ; ; q2 [ ℄
yields for the data model:
R = ABAH + n2 I

Recall the data model:
R = ABAH + n2 I
Assume:
there are astronomical sources "s" and interferers "r",
the noise powers are unequal: Efn(t)n(t)H g = D
(diagonal),
define a diagonal electronic gain matrix G,
then the discrete data model becomes:
R = G(Ar Br AHr + AsBsAHs )GH + D
Note: "whitening" can be done by pre and post multiplying R
with D or R 2 (yields correlation coefficients).
1 1
2

Subspace analysis
Given q sources, and an array of p antennas, then R can be
decomposed by means of an eigenvalue decomposition such
=
that R UUH , where U =[
u1; ; up t contains the ℄
eigenvectors ui (UH U I , and = ) = (
diag 1 ; ; p contains )
the eigenvalues. Also:
R = +
Ar Br Ar n2 Ip
= Ur r UHr"+ n2 (Ur UHr + UnU#"
n)
H
#
r + n Iq 0 U
2 H
= [Ur Un℄ 0 2 I UH
r
= U U H
n p q n
The sources are located in the "source subspace" Ur ,

separated from the "noise subspace" Un . Note that all
eigenvectors are perpendicular and UH r Un 0. =

Relation subspace structure and RFI
Eigenvalues of R Eigenvalues of R
Sky map 98 Sky map
105 95
98 1
1
96 94
96
100 94 93
0.5 0.5
eigenvalue [dB]
eigenvalue [dB]
94 92
92
m
0
m
95 0 91
92 90
90
−0.5 90 88 −0.5
90 89
86 88
−1 88 −1
85 −1 −0.5 0 0.5 1 84 −1 −0.5 0 0.5 1 87
0 20 40 60 l 0 20 40 60 l
eigenvalue number eigenvalue number
PN
Relation between multiple transmitters and eigenvalue
b 1
distribution of covariance matrix R N n=1 xn xH n . Data (not =
normalized/whitened) are obtained from LOFAR ITS test
station. One horizon transmitter (RFI) (left), and three
transmitters (right)(1) .
Note: for astronomical sources similar relation holds.
(1) A.J. Boonstra and S. van der Tol, 2004.

Detection based on subspace structure
For identical noise powers n2 , or
r p 2
for a whitened R matrix (1)
max;min 2 N 1
This can be used e.g. as detection threshold for intermittent
RFI, see figure below showing RFI at the WSRT (whitened
normalized dataset).
Eigenstructure, Iridium transmission
12
fc= 907.14 MHz, ∆ f = 156 kHz, ∆ t = 0.2 ms
10 An alternative, computa-
8 tionally less expensive, is
to use jjRjj2F (sum of ab-
λ1 ... λ8
4
solute squares of the ele-
2
ments) (2)
0
0.6 0.61 0.62 0.63 0.64 0.65
time (s)
(1) Edelman 1988. (2) Leshem and van der Veen, 2001.

Lagrange multipiers
Lagrange multipliers: find- 1.2
1
g(x,y)=c
ing extreme of a function

( )
0.8
y
f x; y subject to the con- 0.6
( ) =0
0.4
straint g x; y
f(x,y)=k2
. 0.2
f(x,y)=k
1
0
0 0.5 1 1.5 2
x
At extremum: gradient r? of f and g aligned (arrows in the

figure), and tangential component rk of derivative is zero:
( )= (( ) )
rf x; y r g x; y where is a scaling constant, the
Lagrange multiplier. Define L f x; y = ( ) (( )
g x; y . Then: )

x
L = 0 ;
y
L = 0 ;

L = 0
The latter equation is true as it reduces to the original
constraint.

Lagrange multipliers
Generalization to N-dimensional space.
In N -dimensional space, with x =[

x1 ; ; xN t , and with K ℄
constraints we define the Lagrangian L:
X
K
L(x; i; ; K ) = f (x) + k g (x)
k=1
where the previous constant of g x is now included in the ()

()
definition of g x . This yields N K equations and N K + +
unknowns:
rxL ; rL =0 =0
with = [1; ; K ℄t

MVDR beamformer
Given p zero-mean antenna element random signals
( )=[ ( ) ( )℄ = () ()
x t x1 t xp t t, with R Efx t x t H g. Let
y (t) = wH x(t)
be the beamformer output signal with beamformer weights w.

Minimize beamformer signal variance
P = min
w Efj y j2 g = min(wH Rw)
w
with constraint of unit gain in look-direction s, wH a = 1 where
a = e 2|Rsf 1
Now define the Lagrangian: L = wH Rw + (wH a 1)


MVDR beamformer
Recall L = wH Rw + (wH a 1)
Using rw (wH Rw) = Rt w, and rw (aH w) = a (1) yields:
rw L = Rtw + a = 0
This results in w = R 1a; inserting it in constraint equation
leads to
=
1
aH R 1a
and to the MVDR/Capon beamformer weights:
R 1a
w = aH R 1 a
A treatise on real vs complex derivative definitions and
relations can be found in (1) .
(1) cf. Kay 1998, Moon 2000

Conventional beamformer
Given p zero-mean antenna element random signals
( )=[ ( ) ( )℄ = () ()
x t x1 t xp t t, with R Efx t x t H g. Let
y (t) = wH x(t)
be the beamformer output signal with beamformer weights w.

Maximize beamformer output power
P = max
Efjyj2 g
w wH Rw = max
w ( )
with constraint of unit norm of w. Solving the gradient
equations yields: a w= p
aH a
The vector w can be interpreted as a spatial filter matched to
the impinging signal.

Beamformer spatial spectra
Assuming observational data is used, the spatial spectrum is
given by inserting the weight into the beamformer output
power formula:
a ba
HR
P onv (s) = H
a a
and
Pmvdr (s) =
1
aH Rb 1a
where the steering vector a, as before, is defined by
a = e 2|Rs 1
with R = [b1; ; bp℄t.


Music beamformer
Recall for the MVDR beamformer: Pmvdr (s) = aH Rb1 1 a
Recall also the data model:
=
R ABAH n2 I Ur r UHr + = + n2 UnUHn
As UHn a for all a in A, define the ad-hoc "MUSIC" beamformer
spatial spectrum by:
Pmusi (s) = H ?
a Ha
a a
b nUb Hn , with Ub n the noise subspace. This leads
with ? = U
(as with the MVDR) to an improved sidelobe suppression and
smaller beamwidth as compared to the conventional
beamformer. Drawback: these are data dependent and
computation scales with p3 , as compared to p for the
conventional beamformer.

Conventional vs subspace beamforming
Conventional beamforming (left) vs MUSIC beamforming (right), using LOFAR ITS data. The
sky map is dominated by a transmitter at the horizon, note the difference in sidelobe level and
the difference in beamwidth. The plusses indicate the location of astronomical 3C sources.

Conventional vs subspace beamforming
Conventional beamforming (left) vs MUSIC beamforming (right), using LOFAR ITS data. The
sky map is dominated by a transmitter at the horizon, note the difference in sidelobe level and
the difference in beamwidth. The plusses indicate the location of astronomical 3C sources.

Spatial projection filtering
Define the data model, using a separate covariance matrix for
the astronomical visibilities Rv , and assuming one RFI
source:
R Rv n2 I 2 aaH = + +
Projection filtering, define the projection matrix P:
P = I a(aH a) 1aH
e by
then Pa = 0. Define the filtered matrix R
Re = PRb P; then EfRe g = P(Rv + n2 I)P
The remaining distortions can be corrected, under certain
stationarity conditions(1) .
(1) Boonstra, 2005.

Spatial projection filtering
Recall the spatial projection filter
P = I a(aH a) 1aH
This can easily be extended to multiple (q) interferes:
P = I A(AH A) 1AH
where, as before,
A = [a1; aq ℄
The filter can be applied to covariance (correlation) data, but
also is a beamformer by multiplying it with the beamformer
weights.

Spatial subtraction filtering
Recall the data model:
R = Rv + n2 I + 2 aaH
The subtraction filtered covariance matrix R e is given by:
Re = Rb ^ 2a^a^H
where ^ 2 and a^ need to be estimated. Especially estimating
^ 2 is computationally expensive (needs short feedback loops).
Theoretical estimates of effectiveness of spatial filters have
been found1 .
(1) Boonstra, 2005.


Spatial filtering results
LOFAR ITS spatial filtering results.


References
Textbooks on array processing, detection, and estimation
techniques
D.G. Manolakis et al., Statistical and Adaptive Signal Processing, Mc. Graw Hill,
Boston, 2000.
S.M. Kay, Fundamentals of Statistical Signal Processing, Detection Theory, Prentice
Hall, New Jersey, 1998.
S.M. Kay, Fundamentals of Statistical Signal Processing, Estimation Theory, Prentice
Hall, New Jersey, 1993.
V.K. Madisetti and D.B. Williams, eds., The Digital Signal Processing Handbook, CRC
Press 1998.
H.L. van Trees, Optimum Array Processing. Part IV of Detection, Estimation and
Modulation theory, Wiley, New York, 2002.
T.K. Moon, and W.C. Srirling, Mathematical Methods amd Algorithms for Signal
Processing, Prentice Hall, NJ, 2000.

References
Overview articles on beamforming and on array signal
processing
H.Krim, and M.Viberg, "Two Decades of Array Signal Processing Research", IEEE
Signal Processing Magazine, July, 1996, p.67-94
B.D. van Veen and K. Buckley, "Beamforming: A Versatile Approach to Spatial

Filtering", IEEE ASSP Magazine, April 1988, pp.4-24.
Overview article and tutorials on (array) signal processing for

radio astronomy, and thesis on RFI mitigation
A.J. van der Veen, A. Leshem, and A.J. Boonstra, "Array Signal Processing for Radio
Astronomy", Experimental Astronomy (2004) 17: 231-240.
Ninth Synthesis Imaging Summer School, Socorro, USA, June 15-22,2004, lecture
notes, www.nrao.edu.
A.J. Boonstra, "RFI Mitigation in Radio Astronomy", PhD thesis, Delft University of
Technology, June 14, 2005.

Thursday Boonstra

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thursday Boonstra

Uploaded by

Copyright:

Available Formats

Digital Signal Processing and

SKADS Marie Curie Technical Workshop

A.J. Boonstra, ASTRON, boonstra@astron.nl

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 2

Netherlands Foundation for Research in Astronomy November 29, 2007

 signal: s(t) or s, parameter t often omitted

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 3

Netherlands Foundation for Research in Astronomy November 29, 2007

 scalar, lower case or upper case,

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 4

Netherlands Foundation for Research in Astronomy November 29, 2007

 complex conjugate: y = Refyg Imfyg

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 5

Netherlands Foundation for Research in Astronomy November 29, 2007

Spatial coherence function: V ri ; rj ( )=

Netherlands Foundation for Research in Astronomy November 29, 2007

Van Cittert-Zernike theorem: the brightness distribution Ib (s)

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 7

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

LOFAR ITS test station, ob-

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 10

Netherlands Foundation for Research in Astronomy November 29, 2007

phased array beam

Broadband vs narrowband processing:

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

which implies m (t) = m(t  )e 2|  .

Netherlands Foundation for Research in Astronomy November 29, 2007

Consider an array of p antennas or telescopes with baselines

x(t) = [x1(t); ; xp(t)℄t

Netherlands Foundation for Research in Astronomy November 29, 2007

Now define the sample estimate (observational data):

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 17

Netherlands Foundation for Research in Astronomy November 29, 2007

The sources are located in the "source subspace" Ur ,

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

ing extreme of a function

At extremum: gradient r? of f and g aligned (arrows in the

Netherlands Foundation for Research in Astronomy November 29, 2007

In N -dimensional space, with x =[

where the previous constant of g x is now included in the ()

Netherlands Foundation for Research in Astronomy November 29, 2007

be the beamformer output signal with beamformer weights w.

Now define the Lagrangian: L = wH Rw + (wH a 1)

Netherlands Foundation for Research in Astronomy November 29, 2007

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 24

Netherlands Foundation for Research in Astronomy November 29, 2007

be the beamformer output signal with beamformer weights w.

Netherlands Foundation for Research in Astronomy November 29, 2007

with R = [b1; ; bp℄t.

Netherlands Foundation for Research in Astronomy November 29, 2007

Netherlands Foundation for Research in Astronomy November 29, 2007

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 28

Netherlands Foundation for Research in Astronomy November 29, 2007

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 29

Netherlands Foundation for Research in Astronomy November 29, 2007

SKADS Marie Curie Workshop - Digital Signal Processing and Beamforming 30

signal: s(t) or s, parameter t often omitted

scalar, lower case or upper case,

complex conjugate: y = Refyg Imfyg

which implies m (t) = m(t )e 2| .

Now define the Lagrangian: L = wH Rw + (wH a 1)

B.D. van Veen and K. Buckley, "Beamforming: A Versatile Approach to Spatial