Você está na página 1de 28

Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.

com
Extended Fourier analysis of signals
Abstract. The extended summary of Dr.Sc.Comp. thesis [6] is created to emphasis the tight
connection of the proposed spectral analysis method with the Discrete Fourier Transform
(DFT ! the most extensively studied and fre"uently used approach in the history of signal
processing. #t is shown that in a typical application case$ where uniform data readings are
transformed to the same num%er of uniformly spaced fre"uencies$ the results of the classical
DFT and proposed approach coincide. The difference in performance appears when the length
of the DFT is selected greater than the length of the data. The DFT solves the un&nown data
pro%lem %y padding readings with 'eros up to the length of the DFT$ while the proposed
(xtended DFT ((DFT deals with this situation in a different way$ it uses the Fourier integral
transform as a target and optimi'es the transform %asis in the extended fre"uency range
without putting such restrictions on the time domain. Thus$ the #nverse DFT (#DFT applied to
the result of (DFT returns not only &nown readings %ut also the extrapolated data$ where
classical DFT is a%le to give %ac& )ust 'eros. The (DFT significantly extends the usa%ility of
the DFT %ased methods$ where previously these approaches were considered inapplica%le [8!32].
The (DFT founds the solution in an iterative way and re"uires repeated calculations to get the
adaptive %asis$ and this ma&es its numerical complexity much higher compared to DFT. This
disadvantage was a serious pro%lem in *++,s$ when the method has %een proposed.
Fortunately$ since then the power of computers has increased so much that nowadays (DFT
application could %e considered as a real alternative.
Table of Contents
(xtended Fourier analysis of signals...............................................................................*
* #ntroduction......................................................................................................-
- .ro%lem formulation........................................................................................-
-.* /asic expressions of classical Fourier analysis.................................-
-.- /asic expressions of extended Fourier analysis................................0
0 .ro%lem solution..............................................................................................1
0.* (xtended Fourier transform of continuous time signals...................1
0.- (xtended Discrete Time Fourier Transform......................................2
0.-.* 3 particular solution for discrete time signals....................2
0.-.- 4enerali'ed solution for discrete time signals...................5
0.0.0 #terative (DTFT algorithm.................................................6
7 (xtended DFT algorithm..................................................................................6
1 (DFT and other nonparametric approaches...................................................*,
1.* Capon filter approach......................................................................*,
1.- 489S solution................................................................................*-
1.0 :igh!;esolution DFT......................................................................*0
2 Computer simulations....................................................................................*0
5 (DFT algorithm in <3T93/ code...............................................................-,
6 ;eferences......................................................................................................-2
Extended summary of Dr.Sc.Comp. thesis 1
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
1 Introduction
3 Fourier transform is a powerful tool of signal analysis and representation of a real or complex!
valued function of time x(t (hereinafter referred to as the signal in the fre"uency domain
}

= dt e t x F
t je
e ( ( $ (*.*
e e
t
e
d
e
F = t x
t j
(
-
*
(
}


. (*.-
The Fourier transforms orthogonality property
( -
,
,
e e to
e e
=
}

dt e e
t j t j
(-
providing a %asis for the signal selective fre"uency analysis$ where e, e
,
are cyclic fre"uencies
and o(ee
,
is the Dirac delta function. =nfortunately$ the Fourier transforms calculation
according to (*.* re"uiring &nowledge of the signal x(t as well as performing of integration
operation in infinite time interval. Therefore$ for practical evaluation of (*.* numerically$ the
signal o%servation period and the interval of integration is always limited %y some finite value O,
-O>-?t?O>-. The same applies to the Fourier analysis of the signal x(t sampled versions:
nonuniformly sampled signal x(t

or uniformly sampled signal x(!$ @-$A$-*$,$*$A$B.


Cnly a finite length se"uence x(t

or x(!$ @,$*$-$A$"!*$ are su%)ect of Fourier analysis$


where " is a discrete se"uence length$ ! is sampling period and the signal o%servation period
O@t
"#$
-t
%
or O@"!. To avoid aliasing and satisfy the Dy"uist limit$ uniform sampling of
continuous time signal should %e performed with the sampling period !?t>O$ where O is upper
cyclic fre"uency of signal x(t. 3lthough nonuniform sampling has no such strict limitation on
the mean sampling period !
s
@O>"$ the following analysis we suppose that %oth se"uences$ x(t


and x(!$ are derived from the %and!limited in O signal x(t. 9et write the %asic expressions of
the classical and the proposed extended Fourier analysis of continuous time signal x(t and its
sampled versions x(t

and x(!.
2 Problem formulation
E!he formulation of a pro&lem is often more essential than its solution 'hich may &e merely a
matter of mathematical or experimental sill.( Albert Einstein
2.1 Basic expressions of classical Fourier analysis
The classical Fourier analysis dealing with the following finite time Fourier transforms
}
O
O

O
=
- >
- >
( ( dt e t x F
t je
e $ (0.*a

O
*
,
( (
"

t j

e t x = F
e
e
$ (0.*%

O
*
,
( (
"

! j
e ! x = F
e
e
$ (0.*c
e e
t
e
d
e
F = t x
t j
(
-
*
(
O
O
O
O
}
. (0.-
where (0.- is the inverse Fourier transform o%tained from (*.- for %and!limited in O signal.
Transforms (0.*% and (0.*c are &nown as Discrete Time Fourier Transforms (DTFT of
Extended summary of Dr.Sc.Comp. thesis 2
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
nonuniformly and uniformly sampled signals. The values of reconstructed signal x
O
(t outside
the o%servation period O are 'eros or vanishes depending on whether (0.- applies to the results
(0.*a or (0.*% and (0.*c.
The signal amplitude spectrum is the Fourier transform (0.* divided %y the o%servation
period O$
(
*
( e e
O O
O
F = S . (7
The fre"uency resolution of the classical Fourier analysis is inversely proportional to the signal
o%servation period O.
C%viously$ one can get the formula (0.*a %y truncation of infinite integration limits in (*.* and
the DTFT (0.*a and (0.*% as result of replacement of infinite sums %y finite ones. This mean$
the classical Fourier analysis supposed that the signal outside O is 'eros. #n other words$ the
Fourier transform calculation %y formulas (0.* is well )ustified if applied to time!limited within
O signals. Cn the other hand$ a %and!limited in O signal cannot %e also time!limited and
o%viously have non'ero values outside O. 4enerally$ the Fourier analysis results o%tained %y
using the exponential %asis tend to the Fourier transform$ if O, while in any finite O there
may exist another transform %asis providing a more accurate estimation of (*.*.
2.2 Basic expressions of extended Fourier analysis
The idea of extended Fourier analysis is finding the transform %asis$ applica%le for a %and!limited
signals registered in finite time interval O and providing the results as close as possi%le to the
Fourier transform (*.* defined in infinite time interval. The formulas for proposed extended
Fourier analysis could %e written as
dt t t x = F $ ( ( (
- >
- >
e o e
o
}
O
O
$ (1.*a

=
*
,
$ ( ( (
"


t t x = F e o e
o
$ (1.*%

=
*
,
$ ( ( (
"

! ! x = F e o e
o
$ (1.*c
e e
t
e
o o
d e F = t x
t j
(
-
*
(
}
O
O
$ (1.-
where in general case the transform %asis o(e$t$ o(e$t

and o(e$! are not e"ual to the


classical ones (0.*. Dote that the inverse Fourier transform (1.- still holds the exponential %asis.
To ensure that the results of transforms (1.* are close to the result of the Fourier transform (*.*
for the signal x(t$ the following minimum least s"uares expression will %e composed and solved
min ( (
-
e e
o
F F . (2
=nfortunately$ as already stated a%ove$ the calculation of F(e for a %and!limited signal cannot
%e performed directly. So$ in order to compose (2$ we should find an ade"uate su%stitution. 9etFs
recall that a complex exponent$ at cyclic fre"uency e
,
and with a complex amplitude S(e
,
$ is
defined in infinite time interval as
< < = t e S t x
t j
$ ( $ (
,
, ,
e
e e
. (5
The Fourier transform of a signal (5 can %e expressed %y the Dirac delta function (-
( ( - $ (
, , ,
e e o e t e
e
=


}
S dt
e
t x
t j
. (6
Extended summary of Dr.Sc.Comp. thesis 3
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
Dow$ letFs use (5 as a signal model with &nown amplitude spectrum S(e
,
for fre"uencies in
range -O?e
,
?O and$ in the minimum least s"uare expression (2$ su%stitute F(e %y the signal
model Fourier transform (6 and the signals x(t$ x(t

and x(! in (1.* %y the signal models (5$


correspondingly. Finally$ the integral least s"uare error estimators for all the three signal cases get
the following form
,
- >
- >
-
, , ,
( ( ( ( -
,
e e o e e e o e t
e
d dt t )
e
S S =
t j
} }
O
O
O
O
A
$ (+a

}

=
O
O
A
*
,
,
-
, , ,
$ ( ( ( ( -
,
"

t j
d t e S S =

e e o e e e o e t
e
$ (+%

}

=
O
O
A
*
,
,
-
, , ,
$ ( ( ( ( -
,
"

! j
d ! e S S = e e o e e e o e t
e
. (+c
The solutions of (+ for a definite signal model (5 provide the %asis o(e$t$ o(e$t

and o(e$!
for the extended Fourier transforms (1.*. To control how close the selected signal model
amplitudes S(e
,
are to the signals x(t$ x(t

and x(! amplitude spectrum$ we will find the


formulas for estimate signal amplitude spectrum S
*
(e in the extended Fourier %asis o(e$t$
o(e$t

and o(e$!.
The formula (6 is showing the connection %etween the signal model Fourier transform and its
amplitude spectrum$ from where S(e
,
could %e expressed as signal model Fourier transform
divided %y -to(ee
,
. Ta&ing (6 into account$ S
*
(e is calculated as the transforms (1.*
divided %y the estimate of -to(ee
,
in the extended Fourier %asis$ which is determined from
(+ in the case A=0 and e
,
=e$
dt t
e
dt t t x
= S
t j
$ (
$ ( (
(
- >
- >
- >
- >
e o
e o
e
e
o
}
}
O
O
O
O
$ (*,a

=
*
,
*
,
$ (
$ ( (
(
"

t j
"


t e
t t x
= S

e o
e o
e
e
o
$ (*,%

=
*
,
*
,
$ (
$ ( (
(
"

! j
"

! e
! ! x
= S
e o
e o
e
e
o
$ (*,c
and showing that the amplitude spectrum on the fre"uency e is estimated as ratio of the signal
extended Fourier transform to the transform of exponent with a unit amplitude in the same %asis.
This is true also for classical Fourier transform. For example$ after su%stituting exponential %asis
t j
e = t
e
e o

$ ( in (*,a$ the denominator %ecomes e"ual to O as in formula (7 for the classical
Fourier analysis.
Galues of the denominator in formulas (*, are in inverse ratio to the fre"uency resolution of the
extended Fourier transform.
/efore finding the the extended %asis functions for ar%itrary S(e
,
$ it is reasona%le to consider a
simple signal model having a rectangular form$ S(e
,
@* for -O?e
,
?O and 'eros outside. Then
the estimators (+ reduces to
Extended summary of Dr.Sc.Comp. thesis 4
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
,
- >
- >
-
,
( ( -
,
e e o e e to
e
d dt t )
e
=
t j
} }
O
O
O
O
A
$ (**a

}

=
O
O
A
*
,
,
-
,
$ ( ( -
,
"

t j
#
d t e =

e e o e e to
e
$ (**%

}

=
O
O
A
*
,
,
-
,
$ ( ( -
,
"

! j
#
d ! e = e e o e e to
e
. (**c
The solution of (** allows to esta%lish relationship %etween the classical and extended Fourier
analysis.
Problem solution
#n this section the integral least s"uare error estimators (+ and (** are solved and su%se"uent
analysis of the o%tained results are performed to find out the only those solutions that can lead to
practically reali'a%le algorithms.
.1 Extended Fourier transform of continuous time signals
The solution of (**a for continuous time signal x(t is found as a partial derivation
- - $ ,
$ (
+ + = O s s O
c
A c
t
t e o
$ and leads to the linear integral e"uation
( )
et
e o
t t
t
j
e dt t )
t
t

O
O
}
=

O
- >
- >
(
(
( sin
. (*-
Step %y step solution of (*- is given in [2]. Finally$ the %asis o(e$t are o%tained %y applying a
specific functions system ! a prolate spheroidal wave functions

(t$ @,$*$-$... and are written


as series expansion

=
, @
(
(
$ (

t
,
t

e
e o
. (*0
The extended Fourier Transform of continuous time signal x(t are given %y
O s s O =

e e e
o
! $ ( (
, @

a , F
$ (*7.*
< < =

t a t t x


! $ ( (
, @

o
$ (*7.-
-
,
,
(
(
(
e
e
e
o
,
a
,
S

=
=
$ (*7.0
where
t t t

d x =
a

+
+

( (
*
-
-
}
O
O
$
dt t =

+
+

(
-
-
-

}
O
O
and

j =
,
(
-
(
|
.
|

\
|
O
O
O
O
e

t
e .
The extended Fourier transform in accordance with (*7.* re"uesting a calculations of infinite
sums$ this mean$ an infinite "uantity of mathematical operations$ therefore itFs impossi%le for real
world applications. Theoretically$ the value of denominator
-
,
(e
,
"

=
in amplitude spectrum
formula (*7.0 tends to infinite for "$ and the extended Fourier transform (*7.* provide a
supper!resolution ! an a%ility to determine the Fourier transform for the sum of sinusoids or
Extended summary of Dr.Sc.Comp. thesis 5
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
complex exponents$ if fre"uencies of them differ %y ar%itrary small finite value.
.2 Extended !iscrete Time Fourier Transform
#n this su%section the minimum least s"uare error estimators (+%$c and (**%$c are solved and the
extended Fourier transforms for uniformly and nonuniformly sampled complex!valued signals
are o%tained. The proposed approaches have %een developed in articles [3] and [4]$ where the
derivations for real!valued discrete signals are given.
.lease note that the following notations are used in the matrix e"uations:
superscripts X
!*
$X
!
$X
H
$X
-
denote inverse$ transpose$ complex con)ugate$ :ermitian
(complex con)ugate transpose of the matrix XI
.> represents element!%y!element division of two matrices with the same si'eI
sum(X means addition of all matrix X elementsI
dia.(X forms the row vector %y extracting the main diagonal elements from "uadratic
matrix X or it puts the elements of vector X on the main diagonal to form a diagonal
matrix.
.2.1 A particular solution for discrete time signals
The solutions of (**%$c can %e o%tained similarly to (**a as partial derivatives of
,
$ (
=
t
l
e o c
A c

and
,
$ (
=
l! e o c
A c
$ l@,$*$-$..."!*$ and leads to the systems of linear e"uations
( )
l
t j
"

l
l
e t )
t t
t t
e
e o
t

=
=

*
,
(
(
( sin
$ (*1.*
( )
l! j
"

e ! )
! l
! l
e
e o
t

=
=

*
,
(
(
( sin
. (*1.-
The solution of (*1 in the matrix form is expressed as
e e
E R A
*
= $ (*2
where A
e
("x* and E
e
("x* are the extended Fourier and the exponential %asis.
The formulas of (xtended Discrete Time Fourier Transform ((DTFT for signal model S(e
,
@*$
-O?e
,
?O, are derived %y su%stituting of transform %asis (*2 into expressions (1 and (*,
$ $ (
*
O s s O =

e e
e o
E xR F (*5.*
$ $ (
*
< < =

t t x
t
E xR
o
(*5.-
e e
e
o
e
E R E
E xR
*
*
(

-
= S
. (*5.0
The matrices for nonuniformly sampled signal case are composed as follows
x (*x" !
(

t x
$ E
e
("x* !
l
t j
e
e
$ R ("x" !
(
( sin
$
l
l
l
t t
t t
r

O
=
t
and E
t
("x* !
(
( sin
l
l
t t
t t

O
t
.
=niformly sampled se"uence x(! can %e considered as a special case of nonuniform
sampling at time moments t

@!$ @,$*$-$A$"!*. Then the matrices elements in (*2$ *5 are


x (*x" !
(! x
$ E
e
("x* !
l! j
e
e
$ R ("x" !
! l
! l
r
l
(
( sin
$

O
=
t
$ E
t
("x* !
(
( sin
l! t
l! t

O
t
.
#n particular$ if sampling of signal x(! is done with Dy"uist rate$ !=t/O$ the matrix R
%ecomes a unit matrix I and the formula (*5.* coincide with classical DTFT (0.*c$ %ut the
formula (*5.0 reduces to well &nown relationship %etween discrete signal Fourier transform and
Extended summary of Dr.Sc.Comp. thesis 6
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
its amplitude spectrum
e o
e e xE = =
O
( ( F F
$ (*6.*
e o
e xE
"
S
*
( = . (*6.-
8hereas for nonuniformly sampled signal x(t

the matrix R=I$ even if mean sampling period


!
s
=t/O and formulas (*5 give results superior to those that o%tained %y the classical nonuniform
DTFT (0.*%. For oversampled signals$ !(or !
s
)<t/O, the (DTFT approach can provide a high
fre"uency resolution and improved spectral estimation "uality. =nfortunate an achievement of
such results is limited %y finite precision in the mathematical calculations and %y restrictions on
fre"uency range in the process of signal sampling. Theoretical value of denominator in (*5.0
"
-
=

e e
E R E
*
and the fre"uency resolution should increase proportionally to the num%er of
samples in the signal o%servation period O. #n the %order!case$ if num%er of samples within O
increasing infinitely$ "$ and the discrete time signal tends to the continuous time signal x(t$
the (DTFT (*5.* gives the same results as (*7.*.
.2.2 "enerali#ed solution for discrete time signals
Dow$ let consider the solution of the minimum least s"uare error estimators (+%$c for ar%itrary
selected signal model S(e
,
. The derivation formulas for %oth estimators are similar to ones
given in previous section. For example$ a partial derivation of (+% %y %asis functions
$ * $...$ - $ * $ , for $ ,
$ (
=
c
A c
" l =
t
l
e o
provide the least s"uare solution
, ( $ ( ( ( ( -
, ,
H
*
,
, , ,
, ,
= |
.
|

\
|

=
O
O

}
e e e o e e e o e t
e e
d e S t e S S
l
t j
"

t j
$ (*+
("uation (*+ can %e rewritten as
, ,
-
,
*
,
,
(
-
,
( ( - $ ( (
, ,
e e e o e t e o e e
e e
d e S t d e S
l l
t j

"

t t j
=
|
|
.
|

\
|

O
O

O
O
}

}
. (-,
The filtering feature of Dirac delta function ( ( (
, ,
x f dx x x x f =
}


o applied to the right part
of (-, gives the final form of the system of linear e"uations
l l
t j

"

t t j
e S t d e S
e e
e e o e e
t

O
O
=
|
|
.
|

\
|

}
-
*
,
,
(
-
,
( $ ( (
-
*
,
$ (-*.*
l! j
"

! l j
e S ! d e S
e e
e e o e e
t

O
O
=
|
|
.
|

\
|

}
-
*
,
,
(
-
,
( $ ( (
-
*
,
$ (-*.-
for l@,$*$-$...$"!*$ where
-
(e S is the signal model power at e
,
=e. The e"uations (-*.- are
applica%le for uniformly sampled signal x(! and can %e derived from (+c in a similar way as
(-*.*. The (DTFT %asis A
e
("x* ! o(e$t

or o(e$! are found as a solution of (-*



e e
e E R A
*
-
(

S = . (--
Su%stituting of transform %asis (-- into expressions (1 and (*,$ yields the formulas for
calculation of the (DTFT:
$ J $ ( (
*
-
O s s = =

e e e
e e o
E xR xA S F (-0.*
$ $ (
*
< <

t = t x
t
E xR
o
(-0.-
Extended summary of Dr.Sc.Comp. thesis 7
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
e e
e
e e
e
e e
e
o
e
e
e
E R E
E xR
E R E
E R x
A E
xA
*
*
*
-
*
-
(
(
(

= =
-
-
-
S
S
= S
(-0.0
The elements of matrix R ("x" in the formulas (--$ -0 are expressed %y integrals
}
O
O

=
,
(
-
, $
,
(
-
*
e e
t
e
d e S r
l
t t j
l
$ (-7.*
}
O
O

=
,
(
-
, $
,
(
-
*
e e
t
e
d e S r
! l j
l
$ (-7.-
for nonuniformly and uniformly sampled signal x (*x" cases$ correspondingly. #f the signal and
its model power spectra are close$
- -
( ( e e
o
S S ~ $ then the matrix R elements (-7 are also an
estimate of the autocorrelation function for the se"uence x. Similarly$ the elements of matrix E
t
("x* in (-0.- ac"uire integral form

}
O
O

= e e
t
e
d e S e
l
t t j
l
(
-
(
-
*
or
}
O
O

= e e
t
e
d e S e
l! t j
l
(
-
(
-
*
.
The inverse transform (-0.- calculated on time moments t=t

or t=!$ @,$*$-$A$"!*$ returns
%ac& the input se"uence x undistorted. Case signal model S(e
,
@* the formulas (-- and (-0
reduces to (*2 and (*5.
The fre"uency resolution of the (DTFT is in inverse ration to
e e
e E R E
*
-
(
-
S and varied in
the fre"uency range -O?e?O.
.. Iterati$e E!TFT algorit%m
Calculation of the (DTFT %y formulas (-0 re"uires &nowledge of the signal model spectrum
which generally is not &nown. 3t the same time$ the amplitude spectrum o%tained in the previous
section %y the formula (*5.0 can %e used as a source of such information. This suggests the
following iterative algorithm$ where the signal model spectrum S(e
,
tends to the signal
spectrum S
*
(e:
/teration $. Calculate (
* (
e
a
S (*5.0 applying default signal model S(e
,
@*.
/teration 0. Calculate (
- (
e
a
S (-0.0 %y using the signal model (
,
* (
e
a
S .
/teration 1. Calculate (
0 (
e
a
S (-0.0 %y using the signal model (
,
- (
e
a
S .
2
/teration i. Calculate (
(
e
i
a
S (-0.0 %y using the signal model (
,
* (
e
i
a
S .
The iterations are repeated until the given maximum iteration num%er is reached or the power
spectrum do not alter from iteration to iteration$
-
* (
-
(
( ( e e

~
i
a
i
a
S S .
The (DTFT output F
*
(e (-0.* is calculated for the last performed iteration /.
/y default the signal model S(e
,
@* is used as input of the (DTFT algorithm. :owever$
additional information a%out the signal to %e analy'ed can %e applied to create a more realistic
signal model for the (DTFT input and to reduce the num%er of iterations re"uired to reach the
stopping iteration criteria.
& Extended !FT algorit%m
The (DTFT considered in the previous section is a function of continuous fre"uency(-O?e?O$
while descri%ed %elow the (DFT algorithm calculate the (DTFT on a discrete fre"uency set
-Ose
n
<O for n@,$*$-$A$3!*. The num%er of fre"uency points 3>" and it should %e selected
sufficiently great to su%stitute the integrals (-7 used for calculation of matrix R ("x" in the
Extended summary of Dr.Sc.Comp. thesis 8
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
expressions (--$ -0 %y the finite sums

}

=

O
O

O
~ =
*
,
(
-
,
(
-
, $
( (
-
*
,
3
n
t t j
n
t t j
l
l n l
e S
3
d e S r
e e
e
t
e e
t
$ (-1.*

}

=

O
O

O
~ =
*
,
(
-
,
(
-
, $
( (
-
*
,
3
n
! l j
n
! l j
l
n
e S
3
d e S r
e e
e
t
e e
t
$ (-1.-
l$@,$*$-$A$"!*. The matrix composed of (-1.* and (-1.-$
(
(
(
(
(
(





=




, ( ... ( ( (
... ... ... ... ...
( ... , ( ( (
( ... ( , ( (
( ... ( ( , (
* $ * * - - $ * * * * $ * * , , $ *
- * * $ - - $ - - * * $ - - , , $ -
* * * $ * * - - $ * * $ * * , , $ *
, * * $ , , - - $ , , * * $ , , $ ,
" " " " " " " "
" "
" "
" "
r t t r t t r t t r
t t r r t t r t t r
t t r t t r r t t r
t t r t t r t t r r
R $ (-2.*
( )
( )
( )
( ) ( ) ( )
(
(
(
(
(
(

, ( ... 0 ( - ( * (
... ... ... ... ...
0 ( ... , ( ( - (
- ( ... ( , ( (
* ( ... - ( ( , (
* $ * - $ * * $ * , $ *
* $ - - $ - * $ - , $ -
* $ * - $ * * $ * , $ *
* $ , - $ , * $ , , $ ,
" " " " "
"
"
"
r ! " r ! " r ! " r
! " r r ! r ! r
! " r ! r r ! r
! " r ! r ! r r
R
$ (-2.-
possesses :ermitian symmetry$
H
$ $ l l
r r =
$ %ut (-2.- for uniformly sampled signal has also a
Toeplit' structure. The matrix elements r
l)
representing the autocorrelation function of the
selected signal model and can %e calculated %y applying the #DFT to the signal model power
spectrum
-
(
n
S e . The fre"uency O/t=2f
u
=f
3
in (-1$ where f
u
is the signal upper fre"uency and
f
3
is the Dy"uist rate of a %and!limited signal$ and it is assumed to %e normali'ed (e"ual to * in
DFT calculations. The choice of the fre"uencies Ke
n
L@K-tf
n
L depends on the num%er of
fre"uencies needed for accurate estimation of (-1 as well as for detailed signal spectrum
representation$ and the limitations on the total amount of computation. (ventually$ the uniform
set of fre"uencies is prefera%le in most application cases.
The (DFT can %e expressed %y the following iterative algorithm
- i i
3
E EW R
( (
*
=
$ (-5.*
( * ( ( (
(
i i i i
EW R x xA F

= = $ (-5.-
( (
. (
* (
* (
(
E R E
E R x
S

=
i -
i
i
dia.
$ (-5.0
M (M
- ( * ( i i
dia. S W =
+
$ (-5.7
for the iteration num%er i@*$-$0$A/$ where (-5.* is (-1 expressed in the matrix form. The
exponents matrix E ("x3 has elements
n
t f j
e
t -
or
! f j
n
e
t -
if the sampling is uniform. /y default
the diagonal weight matrix W
(i
(3x3 for the first iteration is a unit matrix W
(*
@I. #f other
diagonal matrix is used as input of the (DFT algorithm then it must have at least " non'ero
elements. For the su%se"uent iterations W
(i
is filled with power spectrum values calculated %y
(-5.7. There could %e additional criteria for stopping the iterations %efore the maximum num%er
of iterations / is reached$ for example$ the iterations could %e interrupted$ if the relative change of
the power spectrum |sum(W
(i
-sum(W
(i#*
|>sum(W
(-
$ for iN-$ is smaller than a given threshold.
The #DFT can %e applied to output F and return %ac& original "#samples of uniform or
nonuniform se"uence
-
3
FE x
*
=
. (-6
Extended summary of Dr.Sc.Comp. thesis 9
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
Since the length of the fre"uency set 3>"$ then (-6 can %e modified to o%tain a se"uence x
*
(*x3 ! x
*
(t
m
$ m@,$*$-$A$3!*$
-
3
3
FE x
*
=
o
$ (-+
where exponents matrix E
3
(3x3 has elements
m n
t f j
e
t -
or
m! f j
n
e
t -
for uniform sampling case.
The reconstructed %y the formula (-+ se"uence is the original se"uence plus forward and
%ac&ward extrapolation of x to length 3 and>or interpolation if there are gaps inside of x. The
maximum fre"uency resolution of the iterative algorithm is limited %y the length 3 of fre"uency
set$ not %y the length " of se"uence as in application of classical DFT. This mean$ the (DFT is
a%le to increase the fre"uency resolution 3>" times in comparison with the classical DFT. This
can %e verified %y comparing the diagonal elements of the product of #DFT and DFT %asis$
* >
*
( < = 3 "
3
dia.
-
E E $ with the relationship$ * > .
*

*
( , s = < S F A E
3 3
dia.
-
$
corresponding to the #DFT and (DFT %asis A (-5.-. 3t the same time there is a restriction on
the fre"uency resolution sum(F.>S@3"$ which is satisfied %y iteration$ and in order to achieve
a high resolution at certain fre"uencies$ the (DFT must decrease the resolution on other
fre"uencies. The deviation |sum(F.>S-3"| also could %e used as an additional criteria for
stopping of the (DFT iterations$ %ecause of indicate the possi%le inaccuracy in the o%tained
result$ mainly caused %y the finite precision in calculations. #f this happens$ the result of the
previous (DFT iteration should %e considered as a final one.
#n a %order!case 3@"$ the iterative algorithm output do not depend on weight matrix W and
the optimal (DFT %asis can %e found in a non!iterative way (as result of the first (DFT
iteration.
' E!FT and ot%er nonparametric approac%es
#n the previous sections$ starting with the Fourier integral (* and using its orthogonality
property (-$ %y esta%lishing and solving the minimum least s"uare error estimators (+$ the
(xtended DFT is o%tained analytically. Dow letFs ma&e comparison with other nonparametric
methods ! Capon filter$ 4enerali'ed (8eighted 9east S"uares (489S solution and :igh!
;esolution Discrete Fourier Transform introduced %y Sacchi$ =lrych and 8al&er in *++6$ and
try to analy'e the ways and opportunities of derivation of an iterative (DFT algorithm %ased
on these approaches.
'.1 Capon filter approac%
The Capon filter also &nown as <inimum Gariance spectrum estimate (see [8$ 9$ 19$ 22] can
%e viewed as the output of a %an& of filters with each filter centered at one of the analysis
fre"uencies
( ) $... - $ * $ , $
O
( ( (
*
,
= = =

=
n ! h ! n x n! y
"

e e e
h x
. (0,
#n the matrix notation ( ) ( ) ( ) | | ! " n x ! n x ! n x n! x * ( $...$ - ( $ * ( $ (
O
+ = x is the filter input
signal and h
e
@[h
e
(,$h
e
(!$h
e
(-!$...$h
e
(("!*!]
!
is the filter coefficients. :ere the su%script
4 indicate a dependence on the filterPs center fre"uency.
The Capon filter is designed to minimi'e the variance on the filter output
{ } { } { } { }
e e e e e e e e e
c c c c o h R h h x x h h x x h
x
- - - - - -
y
n! y n! y n! y = = = = =
O O O O
( ( (
-
-
$ (0*
su%)ect to the constraint that its fre"uency response at the fre"uency of interest 4 has unity
gain
Extended summary of Dr.Sc.Comp. thesis 10
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
* ( (
*
,
= = =

=
e e
e
e
e h E
! ! j
"

e ! h -
$ (0-.*
* ( (
H
*
,
H
= = =

=
e e
e
e
e E h
- ! j
"

e ! h -
$ (0-.-
where {} . c denotes the expectation operator and the matrix E
e
("x* has elements
! j
e
e
. The
constraints (0-.* and (0-.- must %e satisfied %y the filter (0, and %y the :ermitian transpose
filter
- - -
n! y x h
O
(
e e
= $ correspondingly. The matrix { } x x R
O O-
x
c = ("x" is the sample
autocorrelation matrix and it can %e composed from the values of the signal autocorrelation
function. For example$ so called %iased estimate is calculated %y
( ) * $...$ - $ * $ , $ ( (
*
(
*
,
H
= + =


=
" l ! x ! l x
"
l! r
l "

xx
(00
and$ ta&ing into account that ( (
H
l! r l! r
xx xx
= $ the sample autocorrelation matrix is filled as
( )
( )
( )
( ) ( ) ( )
(
(
(
(
(
(

, ( ... 0 ( - ( * (
... ... ... ... ...
0 ( ... , ( ( - (
- ( ... ( , ( (
* ( ... - ( ( , (
* $ * - $ * * $ * , $ *
* $ - - $ - * $ - , $ -
* $ * - $ * * $ * , $ *
* $ , - $ , * $ , , $ ,
" " " " "
"
"
"
x
r ! " r ! " r ! " r
! " r r ! r ! r
! " r ! r r ! r
! " r ! r ! r r
R
. (07
<athematically$ the Capon filter coefficients can %e o%tained %y minimi'ing the variance (0*
under the constrains given %y (0-.* and (0-.-
min * ( * (
H
= =
e e e e e e
E h h E h R h
- !
x
-
5 $ (01
where $ are 9agrange multipliers. The conditions
, =
c
c
e
h
5
and
, =
c
c
-
5
e
h
have to %e fulfilled
to determine the minimum of (01. /oth re"uirements lead to the same solution
H *
H *
e e
e
e
E R E
E R
h

=
x
!
x
. (02
and$ traditionally$ the Capon power spectrum is computed as
H *
*
(
e e
e e
e
E R E
h R h

= =
x
! x
-
Capon
6
. (05
#n order to o%tain an iterative (DFT algorithm from the original Capon filter approach$ the
sample autocorrelation matrix R
x
(07 has to %e su%stituted %y R
!
@E
H
WE
!
. The matrix R
!
("x"
can also %e o%tained as a transpose of the (DFT matrix R defined %y (-2. The elements of
"uadratic diagonal matrix W (3x3 represent an estimate of power at time moment n!@,$
determined from one sample at the output of each Capon filter

( )
( )
-
H
*
H
*
- -
O
O
, (
e e
e
e e
E R E
E R x
h x

= =
! !
!
y $ (06
where the filter input se"uence x
O
(0, is related to the (DFT input se"uence x as
( ) ! " x ! x * ( (
O
+ = or
( (
O
* +
=
"
t x t x
$ @,$-*$--$..$-("-*$ for uniformly or
nonuniformly sampled se"uence cases$ respectively.
Finally$ an iterative algorithm$ with the initial condition for W
(*
@I$ can %e formed as follows
! i i !
E W E R
( H (
=
$ (0+.*
( )
( ) ( )
H
*
(
H
*
(
(
.
O
E R E
E R x
S

=
i ! !
i !
i
Capon
dia.
$ (0+.-
Extended summary of Dr.Sc.Comp. thesis 11
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
M (M
- ( * ( i
Capon
i
dia. S W =
+
$ (0+.0
with the iteration num%er i@*$-$0$A/. The estimate of the power spectrum
-
M M
Capon
S
coincides
with the results of the (DFT$ while the phase spectrum$ definitely$ is different. #t should %e noted
that the calculation of the Capon filter output power %y formula (05 is theoretically well
)ustified$ whereas the derivation of (0+ re"uires ad hoc assumptions and su%stitutions$ and
actually is a measurement of power o%tained from )ust a one sample at the output of filter. This
leads to conclusion that the approach (0+ is simply a filter!%an& interpretation of the (DFT$
similarly to the DFT which can also %e considered as %an& of filters. #n addition$ an iterative
algorithm derived on the %asis of the filter!%an& can not reveal all the (DFT capacity such as the
a%ility to estimate the DFT (-5.- and restore the signal (-6$ -+.
'.2 "()* solution
The 4enerali'ed (8eighted 9east S"uares approach (see [13$ 16, 23, 32] in the fre"uency
analysis is %ased on the following data model
7 89LS
!
S e E x + = (
H
e
e
$ (7,
with e
7
denoting the noise and interference (signals at fre"uencies other than 4 component$
and
(
H
e
e 89LS
S E
representing the signal component on the fre"uency of interest with
un&nown complex amplitude S
89LS
(4. The 489S minimi'es
Q ( R Q ( R
H * H
e e
e e 89LS
! -
89LS
!
S S E x Q E x

$ (7*
which is solved %y
H *
*
(
e e
e
e
E Q E
x Q E

=
!
! !
89LS
S
$ (7-
where Q ("x" is the covariance matrix of the data model component e
7
. There are two
special cases of 489S called 8eighted 9east S"uares (89S and ordinary 9east S"uares
(9S. 89S occurs when all the off!diagonal entries of Q are ,$ while 9S solution is o%tained
from the 489S under assumption that e
7
in (7, is a white noise$ hence Q@I.
The pro%lem of 489S estimator is that$ in general$ the covariance matrix Q is not &nown$
and must %e estimated from the data along with the S
89LS
(4. The initial estimate (the *
st
iteration could %e e"ual to 9S solution$ it is (7- with Q@I. Dext$ to ensure that the 489S
solution wor&s in an iterative way as (DFT do$ covariance matrix Q should %e replaced %y
R
!
@E
H
WE
!
. 3s a result$ 489S solution (7- coincides with the (DTFT formula (-0.0
( )
( )
( (
*
*
H
*
*
e e
o
e e
e
e e
e
S S
-
! !
! ! !
89LS
= = =

E R E
E xR
E R E
x R E
(70
and$ as shown in the Section 0.0.0$ can %e successfully used for calculation of the amplitude
spectrum iteratively. 3lthough su%stitution of a noise matrix %y R
!
would %e easy done$ it is
not supported %y 489S data model (7,$ from where the matrix Q represents the data model
component e
7
only and the signal component (
H
e
e 89LS
S E must %e excluded from it$ whereas
the matrix R
!
is calculated for the entire signal x
!
$ including e
7
and (
H
e
e 89LS
S E . Furthermore$
the derivation of (DFT shows that the signal can %e restored %y formula (-6$ which fits
perfectly to the iterative update of the matrix R. =sing estimates
( ( e e
o
S S
89LS
=
in the
data model (7, leads to a predetermined split of overall energy at the fre"uency 4 in %etween
components (
H
e
e 89LS
S E and e
7
. The conclusion reached is that ma&ing the derivation of the
(xtended DFT algorithm possi%le$ invalidates 489S minimi'ation expression (7* which
re"uire separation of %oth data model components.
Extended summary of Dr.Sc.Comp. thesis 12
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
'. +ig%,-esolution !FT
The third method considered here is the :igh!;esolution DFT (:;DFT [7]. The authors
presented an iterative nonparametric approach of spectral estimation$ which minimi'es the cost
function deduced from /ayesP theorem and$ as well as the (xtended DFT$ ma&es it possi%le to
o%tain high!resolution Fourier spectrum. The :;DFT algorithm can %e reduced to the following
iterative procedure:
- i i
3
E EW R
( (
*
= $ (77.*
( * ( (
(
i i i
-:DF!
EW R x F

= $ (77.-
|
.
|

\
|
=
+
-
( * (
*
i
-:DF!
i
3
dia. F W
$ (77.0
for iteration num%er i@*$-$0$A/ and with the initial condition W
(*
@I.
The #DFT (-6 applied for any iteration output (77.-$ return %ac& the se"uence x undistorted.
The main difference %etween approaches is that the :;DFT algorithm lac& of formula for
estimate of amplitude spectrum (-5.0. #nstead$ as input for the next iteration$ it uses the Fourier
spectrum estimated in previous iteration (77.0. Therefore$ the results of the :;DFT differ from
output of the (DFT significantly.
. Computer simulations
The computer modeling of the (DFT algorithm are performed for the complex!value test
signal used in [6]. True spectrum of the test signal consisting of a %and!limited noise in fre"uency
range R-,.1...-,.-1Q :'$ a rectangular pulse in range R,...,.-1Q :' and unit power complex
exponent at fre"uency ,.01 :'. The signal upper fre"uency is f
u
@,.1 :'. =niform and
nonuniform test se"uences of length "@27 samples are derived %y simulating *,!%it 3nalog!to!
Digital Converter (3DC. Sampling and mean sampling periods of %oth se"uences are e"ual$
!=!
s
@*s. Sampling time moments for the nonuniform se"uence are generated as$ t

@!Bt

$
@,$-$...$"-*$ where Kt

L are uniformly distri%uted random values in range R,...,.6sQ. Thus$ the


true spectrum of %oth test se"uences consisting of three non!overlapping components and 3DC
added floor noise (~!2,d/$ and it is sym%oli'ed %y red color lines in the Figures *!7.
The plots in Figures * and - shows the performance of (DFT (%lac& line for uniform and
nonuniform se"uences and allows to compare it with the classical DFT (%lue line. The num%er
of fre"uencies (the length of DFT here is chosen to %e 3@*,,,$ which gives spectral estimate
with step %y fre"uencies -f
u
>3@,,,* :'. The range R-,.1...,.1R :' is uniformly covered %y
fre"uencies and used for the calculations of (-1$ -5 and for the signal representation in the
fre"uency domain (spectrum plots.
Figures *a and -a display the power spectrum of the (DFT calculated as *,lg(MSM
-
in a non!
iterative way. The input matrix W for this case is composed from the values of true spectrum (red
line in the plots$ therefore there is no need for iterations. The o%tained non!iterative (DFT
estimate is very close to the (DFT *,
th
iteration result depicted in Fig.*% and -%$ where the input
matrix W@I is used for the first iteration. The Figure *c (-c shows the .ower Spectral Density
(.SD calculated %y the (DFT as *,lg(MFM
-
>D and proves the expectations$ that the .SD estimate
of complex exponent (,.01 :' should increase in a value in comparison with the classical DFT$
if the proposed method achieves a high resolution around this fre"uency.
Figures *d and -d plot the relative fre"uency resolution for the (DFT *,
th
iteration$ calculated as
S F > .
-
*
!" f
u
case of uniform or
S F > .
-
*
" ! f
s u
case of nonuniform se"uence$ in respect to the
DFT for which$ according to (*6.-$ it is simply e"ual to * at all fre"uencies. The value of
Extended summary of Dr.Sc.Comp. thesis 13
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
2f
u
!@2f
u
!
s
@* and this means that the signal is processed in one Dy"uist 'one. The DFT is
showing a normal fre"uency resolution$ whereas the (DFT have a%ility to increase the resolution
(in plot appears values N* around the powerful signal components and decrease the resolution
(in plot appears values S* at fre"uencies where the signal have wea& power components. The
(DFT is called as high!resolution method and thatFs true$ %ut with the following remar& ! it still
&eeps the same FsummaryF resolution as the traditional DFT or$ in other words$ s"uares under
Figure *. =niform complex!value test se"uence. The estimate of:
(a .ower spectrum ! True (red$ DFT (%lue and non!iterative (DFT (%lac&$
(% .ower spectrum ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
(c .ower Spectral Density ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
(d ;elative fre"uency resolution ! DFT (%lue and (DFT (*,
th
iteration.
Extended summary of Dr.Sc.Comp. thesis 14
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
%lac& and %lue curves in the plots *d (-d are e"ual. The maximum fre"uency resolution is
limited %y value of division 3>". For example$ if "@27 and 3@*,,,$ then the (DFT can
potentially improve the fre"uency resolution *,,,>27~*2 times. <aximum resolution is achieved
on narrow!%and signal components$ for test signal ! at fre"uency ,.01 :'. The rectangular pulse
is processed %y the (DFT with approximately the same resolution as the DFT (~*$ normal
fre"uency resolution$ the relative resolution for %and!limited noise R!,.1...!,.-1Q :' fluctuates
Figure -. Donuniform complex!value test se"uence. The estimate of:
(a .ower spectrum ! True (red$ DFT (%lue and non!iterative (DFT (%lac&$
(% .ower spectrum ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
(c .ower Spectral Density ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
(d ;elative fre"uency resolution ! DFT (%lue and (DFT (*,
th
iteration.
Extended summary of Dr.Sc.Comp. thesis 15
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
around *$ while in the range where )ust 3DC noise can %e found$ the (DFT decreases the
fre"uency resolution %ellow the normal. The difference %etween uniform and nonuniform (DFT
is explained in Figure 0$ where the same uniform and nonuniform test se"uences are analy'ed in
extended fre"uency range$ R!*...*R :'. The num%er of fre"uency points and the upper fre"uency
are increased two times$ 3@-,,, and f
u
@* :'. This means that the step %y fre"uency remains the
same as in the previous plots. The true spectrum of test se"uences at fre"uencies a%ove ,.1 :'
consists only of floor noise (~!2,d/ added %y 3DC. The actual result depicted in Figure 0a
shows periodicity in the spectrum$ which can not %e avoided for uniformly sampling se"uences.
#n contrast$ the (DFT applied to the nonuniform test se"uence gives the correct power spectrum$
although this re"uires more calculations ! *1 iterations are performed to o%tain an estimate in
Figure 0%. The relative fre"uency resolution of nonuniform (DFT and DFT are compared in
Figure 0c. The relative resolution of the nonuniform DFT is calculated as *>(2f
u
!
s
@,.1 and it is
half the normal resolution %ecause of analysis is performed in two Dy"uist 'ones. Devertheless$
the s"uares under %lue and %lac& plots in Figure 0c are e"ual to oneFs depicted in Figure -d. The
maximum increase in the fre"uency resolution -,,,>27~31 times is achieved on a complex
Figure 0. The estimates o%tained in the extended fre"uency range:
(a .ower spectrum of uniform se"uence ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
(% .ower spectrum of nonuniform se"uence ! True (red$ DFT (%lue and (DFT (*1
th
iteration$
(c ;elative fre"uency resolution of nonuniform se"uence! DFT (%lue and (DFT (*1
th
iteration.
Extended summary of Dr.Sc.Comp. thesis 16
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
exponent at fre"uency ,.01 :'. The (DFT should also increase the resolution in half to process a
%and!limited noise component (R!,.1...!,.-1Q :' and a pulse (R,...,.-1Q :' with the normal
fre"uency resolution e"ual to *$ as it is indicated %y the red doted lines in Figure 0c. :ence the
conclusion that (DFT can handle nonuniformly sampled signals in multiple Dy"uist 'ones$ %ut
the spectrum of the signal components if its sum$ still must not exceed a one Dy"uist 'one. 9etFs
chec& fulfilling of this condition on a test se"uence. Since spectrum of the uniform se"uence (see
the red color lines in Figure * covers more than half of Dy"uist 'one$ (DFT should %e a%le to
handle it with mean sampling period !
s
S-!. The increase of !
s
will %e achieved %y s&ipping of
samples from the original se"uence randomly. The result can %e considered as nonuniformly
sampled se"uences as the distance %etween ad)acent readings %ecomes une"ual [5]. The power
spectrum depicted in Figure 7 shows the impact of sample s&ipping on the performance of DFT
and (DFT. The inputs are modeled %y removing *2$ -7 and 0- samples randomly from the
uniform 27!point test se"uence$ and leads in increase of mean sampling period !
s
$ 27>76!=*$00s$
27>7,!=*$2s and 27>0-!=-s$ correspondingly. The length of DFT is &ept 3@*,,, and
fre"uencies are uniformly spread in the range R!,.1...,.1R :'. The simulations showing that DFT
is not a%le to handle se"uences with missed samples$ while (DFT is still applica%le if one
Dy"uist 'one condition for the signal is satisfied (Figure 7a and 7%$ otherwise the results
Figure 7. The power spectrum ! True (red$ DFT (%lue and (DFT (*,
th
iteration$
of test se"uence with randomly s&ipped *2 (a$ -7 (% and 0- (c samples.
Extended summary of Dr.Sc.Comp. thesis 17
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
%ecomes worse (Figure 7c. 3lso processing of time locali'ed pulse re"uires the denser sampling.
ThatFs why a rectangular pulse in range R,...,.-1Q :' not recovered %y the (DFT in Figure 7%.
The third test se"uence used in computer simulations is well!&nown <arpleTUay data set
ta&en form [1]. #t is 27!point real sample se"uence from a process consisting of two unit power
harmonics with fre"uencies of ,.- and ,.-* :'$ a third harmonic with a power of ,.* (-, d/
down at ,.* :' and a colored noise in fre"uency range R,.-A,.1Q :' (see red color lines in
Figure 1. The signal upper fre"uency is f
u
@,.1 :' and the length of DFT is selected 3@*,,,.
Cnly 1,, positive fre"uencies are shown$ %ecause of the <arpleTUay se"uence is real!valued
and negative fre"uencies$ if depicted$ gives a symmetrical pattern to 'ero fre"uency. The Figure 1
shows the power spectrum of the DFT$ (DFT and :;DFT approaches in a common view$ while
separately these plots have %een presented in [3] and [7]. The performance of other well!&nown
spectral analysis methods for <arpleTUay data set can %e found in [1]$ including <inimum
Gariance approach$ named in the Section 1.* as traditional Capon filter (05.
The simulation results in the Figure 1a$% demonstrate$ that the classical DFT and (DFT a%le to
evaluate not only the spectrum of sinusoids$ %ut also the shape of continuous spectrum of other
signal components$ whereas :;DFT on Figure 1c is suita%le mostly for the estimation of line
spectrum. The plot in Figure 1a showing that due to limited fre"uency resolution the classical
Figure 1. The power spectrum o%tained for <arpleTUay data set %y
(a DFT$ (% (DFT (*,
th
iteration$ (c :;DFT (*,
th
iteration.
Extended summary of Dr.Sc.Comp. thesis 18
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
DFT cannot resolve sinusoids at fre"uencies ,.- and ,.-*. 3lthough the first (DFT iteration
coincides with the DFT$ in su%se"uent iterations the (DFT is a%le to increase the fre"uency
resolution around the powerful signal components and all three sinusoids are clearly
distinguished after *, iterations in Figure 1%.
3ll the three DFT have one common feature ! the a%ility to get %ac& 27 samples of <arpleTUay
data set %y applying #DFT to the output of each of these methods. Since the length of DFT is
chosen e"ual to *,,,$ the inverse transform returns *,,,!27 additional samples$ which are
plotted in Figure 2 (%lac&. The samples 21$ 22$ 25$... are considered as a forward extrapolation$
%ut samples *,,,$ +++$ ++6$... as a %ac&ward extrapolation of &nown 27!sample se"uence (%lue.
Cf course$ <arpleTUay se"uence outside of given data set is un&nown$ and plots on Figure 2 are
)ust a three possi%le versions of its extrapolation. The classical DFT (Fig.2a suggests that
<arpleTUay se"uence outside of given 27 samples will %e 'eros$ :;DFT (Fig.2c shows that
the extrapolated data even will increase in power$ while (DFT (Fig.2% expects that the se"uence
%eyond will have approximately the same power$ which only gradually decreases in time.
3ny approach$ that claims that it is a high fre"uency resolution method$ in accordance with the
=ncertainty .rinciple must ma&e certain assumptions a%out the data outside of the o%servation
period$ even if %y itself it is not a%le to recover the signal. The advantage of the proposed method
Figure 2. True <arpleTUay real!value se"uence (%lue and extrapolated data (%lac& %y
(a DFT$ (% (DFT (*,
th
iteration$ (c :;DFT (*,
th
iteration.
Extended summary of Dr.Sc.Comp. thesis 19
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
over similar ones is that (DFT %ased on a solution that satisfies the minimum least s"uares
criteria (2$ ma&ing it an accurate$ relia%le and sta%le.
;un <3T93/ program (DFTVF#4.m availa%le on file exchange (see lin& %elow to recreate the
computer simulation results presented in this section.
/ E!FT algorit%m in 0AT)AB code
The (DFT pac&age consisting of programs written in a simple <3T93/ code and are created
to demonstrate the (xtended DFT capa%ilities descri%ed in the previous sections. (ach
program contains commented (W help text section where its syntax$ algorithm$ usage and
features are descri%ed.
The programs D(DFT.m and the inverse transform #D(DFT.m can %e applied for uniform or
nonuniform input>output data and fre"uency sets.
function RF$S$StopitQ@nedft(X$t&$fn$#$8
W D(DFT ! Donuniform (xtended Discrete Fourier Transform.
W
W SYDT3X
W a. <andatory inputs>outputs
W F@nedft(X$t&$fn
W Function D(DFT returns discrete Fourier transform F of input se"uence X sampled at ar%itrary
W selected time moments t&: X(t& NNN F(fn$ where fre"uencies fn$ in general$ also may selected
W ar%itrary. #f fn is less than X$ input se"uences X and t& will %e truncated.
W %. <andatory and optional inputs>outputs
W RF$S$StopitQ@nedft(X$t&$fn$#$8
W # Cptional input parameter # can %e used for limiting maximum num%er of iterations. #f # is not
W specified in input arguments$ default value for # is set %y parameter F<iterationF$ that is$
W nedft(X$t&$fn@nedft(X$t&$fn$<iteration. To complete iteration process faster$ the value for
W F<iterationF should %e decreased.
W 8 #nput weight vector 8$ if specified$ override the default values 8@ones(si'e(fn. 8 must have
W at least length(X non'ero elements.
W S The second output argument S represents the 3mplitude spectrum. .ea& values of a%s(S can %e
W used for estimate amplitudes of sinusoids in the input se"uence X.
W Stopit is an informative output parameter. The first row of Stopit showing the num%er of performed iteration$
W the second row indicate %rea&ing of iteration reason and may have the following values:
W ,! <aximum num%er of iteration performed.
W *! Sum of outputs division sum(F.>S is not e"ual to UHD within ;elative deviation F;deviatF.
W the calculations is interrupted %ecause of results could %e inaccurate. #f this occur in the first
W D(DFT iteration$ then outputs F and S are 'eros.
W -! ;elative threshold F;threshF reached. To complete iteration process faster$ the value for
W F;threshF should %e increased.
W 394C;#T:<
W #nput:
W X! input se"uence
W (! complex exponents matrix (Fourier transform %asis ! (@exp(!iH-HpiHt&.FHfnI
W #! (optional num%er of maximum iteration.
W 8! (optional weight vector 8. #f not specified$ 8 @ ones(*$si'e(fn used for the first iteration.
W Cutput F and S for each D(DFT iteration are calculated %y following formulas:
W *. ;@(Hdiag(8>DH(FI
W -. F@8.H(XHinv(;H(I
W S@(XHinv(;H(.>diag((FHinv(;H(.FI
W 0. 8@S.Hcon)(SI ! the weight vector 8 for the next iteration.
W 3 special case: if length(X is e"ual to length(fn$ the D(DFT output do not depend on selected weight
W vector 8 and is calculated in non!iterative way.
W Tips for selection of mandatory D(DFT inputs X(t& and fn:
W *. #nput se"uence X(t& for D(DFT can %e sampled uniformly or nonuniformly. =niform sampling
W can %e considered as a special case of nonuniform sampling$ where t&@R,$*$...$U!*QHT and T is
W sampling period. Donuniform sampling can %e reali'ed in many different ways$ li&e as:
W ! uniform sampling with randomly missed samples (&nown as sparse dataI
W ! uniform sampling with missed data segments (&nown as gapped dataI
W ! uniform sampling with )itter: t&@(R,$*$...$U!*Q B )itterHrand(*$UHTs$ where value for )itter is selected
W in range R,...*R and Ts is the mean sampling periodI
W ! additive nonuniform sampling: t&@t&!* B (*B)itterH(rand!,.1HTs$ &@*$...U!*$ t,@,I
W ! signal dependent sampling$ e.g$ level!crossing sampling$ etc... .
W -. Fre"uencies for fn can %e selected ar%itrary. This mean$ that user can choose not only the length
W of D(DFT (num%er of fre"uencies in fn$ %ut also the way how to distri%ute fre"uencies along the
Extended summary of Dr.Sc.Comp. thesis 20
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
W fre"uency axis. Cn other hand$ to get ade"uate se"uence X representation$ fre"uencies fn should
W %e selected to cover overall range$ where the input se"uence X spectrum is supposed to %e found$
W otherwise$ in result of D(DFT$ all components having spectra outside fn will %e incorporated.
W Dote that fn should contain negative fre"uencies too$ and for a real value X(t& analysis each positive
W fre"uency in fn should have corresponding negative one.
W 0. Fre"uencies for vector fn can %e added in any order. Therefore it is possi%le to com%ine different
W fre"uency sets in one or )ust add individual fre"uencies of interest to fn$ e.g$ fn@Rfn* fn- f* f-Q$ where
W fn* and fn- are different fre"uency sets$ f*$f- ! specific fre"uencies. D(DFT outputs will %e calculated
W accordingly! F(fn@RF(fn* F(fn- F(f* F(f-Q$ S(Fn@RS(Fn* S(fn- S(f* S(f-Q.
W F(3T=;(S
W *. D(DFT output F(fn is the discrete Fourier transform of se"uence X(t&.
W The .ower Spectral Density function of nonuniform se"uence X(t& can %e estimated %y the following
W formula: a%s(F.Z->(DHTs$ Ts ! mean sampling period.
W -. #n general$ the function Y@inedft(F$fn$tn (see attached program is used to calculate the reconstructed
W se"uence Y(tn. #f fre"uencies fn are selected on the same grid as used %y FFT algorithm$ then ifft(F
W can %e applied to get uniformly re!sampled and extrapolated to length(fn version of input se"uence X(t&.
W 0. D(DFT output S(fn estimate amplitudes and phases of sinusoidal components in se"uence X(t&.
W 7. D(DFT can increase fre"uency resolution length(fn>length(X times. Division of outputs *>(TsH(F.>S
W demonstrate the fre"uency resolution of D(DFT. he following is true for any D(DFT iteration:
W ,SF.>SS@length(fn$
W sum(F.>S@length(fnHlength(X.
W 1. #f input arguments are matrixes$ the D(DFT operation is applied to each column.
W
W See also FFT$ #FFT$ FFTS:#FT$ (DFT$ #D(DFT.
W@@@@@@@@@@@@@@@@@@@@@@@ Set default parameters for D(DFT @@@@@@@@@@@@@@@@@@@@@@@@@@@@
<iteration@0,I W 9imit for maximum num%er of iteration (Stopit ,.
;deviat@,.,,,1I W Galue for relative deviation (Stopit *.
;thresh@,.,,,*I W Galue for relative threshold (Stopit -.
W@@@@@@@@@@@@@@@@@@@@@@@ Chec& D(DFT input arguments @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if narginS0$error(FDot enough input arguments. See help nedft.F$end
if sum(any(isinf(Xsum(any(isnan(X$ error(F#nput argument X contain #nf or DaD. See help nedft.F$ end
if si'e(X$*@@*$ W Chec& si'e of input se"uence X.
trf@,I
else
X@X.FI t&@t&.FI fn@fn.FI trf@*I
end
R9 UQ@si'e(XI W U ! length of input se"uence X.
if si'e(t&$*O@9 M si'e(t&$-O@U$ error(FSi'e of input arguments X and t& must %e e"ual. See help nedft.F$ end
if si'e(fn$*O@9$ error(F#ncorrect si'e of input argument fn. See :elp nedft.F$ end
D@si'e(fn$-I W D ! length of DFT.
if DSU$ W Truncate se"uence X if DSU.
X@X(:$*:DI t&@t&(:$*:DI U@DI
end
if narginS7$ W Set value for maximum num%er of iterations.
#@<iterationI W Default value for #.
else
if isempty(#$#@<iterationIend$ #@floor(#(*I W Chec& input argument #.
end
if narginN7$ W Chec& of input argument 8.
if trf@@*$8@8.FIend
if (si'e(8$-O@DM(si'e(8$*O@9$error(F#ncorrect si'e of input argument 8. See help nedft.F$ end
8@8.Hcon)(8I
if any(find(sum(8N,SU$ error(FToo many 'eros in input argument 8. See help edft.F$ end
else
8@ones(9$DI W Default values for 8.
end
W@@@@@@@@@@@@@@@@@@@@@@@ Chec& for a special cases @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if U@@D$ #@*I 8@ones(9$UI end W #f U@D$ perform )ust one D(DFT iteration.
W@@@@@@@@@@@@@@@@@@@@@@@ Set default values for D(DFT output arguments @@@@@@@@@@@@@@@
F@'eros(9$DI S@'eros(9$DI W Fill 'eros in output matrixes F and S.
Stopit@R#Hones(*$9I 'eros(*$9QI W Stopit ,: Set values for default Stopit.
W@@@@@@@@@@@@@@@@@@@@@@@ Calculate D(DFT for each X column l @@@@@@@@@@@@@@@@@@@@@@
for l@*:9$
(@exp(!iH-HpiHt&(l$:.FHfn(l$:I W Calculate the complex exponents matrix (.
for it@*:#$ W Start iterations...
W Calculate the correlation matrix ; %y using a loop structure.
for n@*:U$
for &@n:U$
;(&$n@sum(8(l$:.Hcon)(((n$:.H((&$:>DI
if nO@&$
;(n$&@con)(;(&$nI
else
Extended summary of Dr.Sc.Comp. thesis 21
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
;(n$n@real(;(n$nI
end
end
end
W Calculate the correlation matrix ; %y using vectori'ed form and ;(@;[( (an alternative approach.
W ;@(Hdiag(8(l$:>DH(FI
W ;(@;[(I
W Calculate ;(@inv(;H( and (;(@diag((FHinv(;H(.F@sum(con)((.H;(.
;(@inv(;H(I
(;(@sum(con)((.H;(I
W Stopit *: /rea& iterations if sum(F.>S is not e"ual to DHU.
if a%s((;(H8(l$:.F>D>U!*N;deviat$ Stopit(:$l@Rit!*I *QI %rea&$ end
W Calculate outputs for iteration (it: D!point D(DFT (F and 3mplitude Spectrum (S.
F(l$:@X(l$:H;(I
S(l$:@F(l$:.>(;(I
F(l$:@F(l$:.H8(l$:I
W Calculate weight (8 for the next iteration.
8(l$:@S(l$:.Hcon)(S(l$:I
W Stopit -: /rea& iterations if relative threshold reached.
S8(it@sum(8(l$:I
if itN*$
thit@a%s(S8(it!*!S8(it>S8(*I
if thitS@;thresh$ Stopit(:$l@RitI -QI %rea&$ end
end
end W ... end iterations.
end
W@@@@@@@@@@@@@@@@@@@@@@@ 3d)ust si'e of D(DFT output @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if trf@@*$F@F.FIS@S.FIend W 3d)ust si'e of D(DFT outputs.
function Y@inedft(F$fn$tn
W#D(DFT #nverse Donuniform (xtended Discrete Fourier Transform.
W
W Y@inedft(F$fn$tn is the inverse discrete Fourier transform of vector
W F estimated %y D(DFT function at ar%itrary fre"uency set fn:
W F(fn !N Y(tn$
W where time moments tn for reconstructed se"uence Y can %e uniformly or
W nonuniformly spaced in time. #n the special case of uniform vectors fn and
W tn$ the #D(DFT function can %e replaced %y well &nown <3T93/ function #FFT.
W
W #f input arguments are matrixes$ the #D(DFT operation is applied to each column.
W
W See also #FFT$ (DFT$ D(DFT.
W@@@@@@@@@@@@@@@@@@@@@@@ Chec& #D(DFT input arguments @@@@@@@@@@@@@@@@@@@@@@@@@@@
if narginS0$error(FDot enough input arguments. See help inedft.F$end
W Chec&ing si'e of input arguments.
if si'e(F$*@@*$
trf@*IF@F.FI tn@tn.FI
else
trf@,Ifn@fn.FI
end
RD 9Q@si'e(FI
if si'e(fn$-O@D$ error(FSi'es of input arguments F and fn must %e e"ual. See help inedft.F$ end
if si'e(tn$-O@9$ error(F#ncorrect si'e of input argument tn. See help inedft.F$ end
W@@@@@@@@@@@@@@@@@@@@@@@ Calculate #D(DFT for each X column l @@@@@@@@@@@@@@@@@@@@@@
for l@*:9
(@exp(iH-HpiHtn(:$lHfn(l$:I
Y(:$l@(HF(:$l>DI
end
W@@@@@@@@@@@@@@@@@@@@@@@ 3d)ust si'e of #D(DFT output @@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if trf@@*$Y@Y.FIend
From the calculations complexity viewpoint$ it is reasona%le to select the fre"uencies on the
same grid as used %y the Fast Fourier Transform algorithm (FFT.m in <3T93/ li%rary. The
program (DFT.m is designed as a faster reali'ation of the proposed method$ where the
algorithm descri%ed in [5] is implemented. This program is applica%le for uniformly sampled
signals and for signals with missing samples or data segments (gaps inside of the input
se"uence. The inverse transform to (DFT.m is <3T93/ li%rary program #FFT.m.
Extended summary of Dr.Sc.Comp. thesis 22
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
function RF$S$StopitQ@edft(X$D$#$8
W (DFT ! (xtended Discrete Fourier Transform.
W
W Function (DFT produce discrete D!point Fourier transform F and amplitude spectrum S of the
W data vector X. Data X may contain DaD (Dot!a!Dum%er.
W
W SYDT3X
W RF$S$StopitQ@edft(X$D for DNlength(X calculate F and S iteratively (see an 394C;#T:< %elow.
W #f data X do not contain DaD and DS@length(X or D is not specified$ (DFT return the
W same results as fast Fourier transform: F@fft(X$D or F@fft(X and S@F>D.
W RF$S$StopitQ@edft(X$D$# performs edft(X$D with limit # for maximum num%er of iterations.
W Default value for # is set %y parameter F<iterationF$ that is$ edft(X$D@edft(X$D$<iteration.
W To complete iteration process faster$ the value for F<iterationF should %e decreased.
W RF$S$StopitQ@edft(X$D$#$8 execute edft(X$D$# with initial conditions defined %y weight vector 8.
W Default values for 8 are ones(si'e(F. 8 must have at least length(X non'ero elements.
W Stopit is an informative (optional output parameter. The first row of Stopit showing the num%er of
W performed iteration$ the second row indicate %rea&ing of iteration reason and may have
W the following values:
W , ! <aximum num%er of iteration performed. #f length(XS@D$ only one (DFT iteration is
W performed (#@*.
W * ! Sum of outputs division sum(F.>S is not e"ual to UHD within ;elative deviation
W F;deviatF. The calculations were interrupted %ecause of results could %e inaccurate.
W #f this occur in the first (DFT iteration$ then outputs F and S are 'eros.
W - ! ;elative threshold F;threshF reached. To complete iteration process faster$ the value
W for F;threshF should %e increased.
W 394C;#T:<
W #nput:
W X ! input data.
W D ! length of discrete Fourier transform.
W # ! (optional num%er of maximum iteration. #f not specified$ #@0,.
W 8 ! (optional weight vector 8. #f not specified$ 8 @ ones(*$DI used for the first iteration.
W ( ! Fourier transform %asis matrix: (@exp(!iH-HpiH(,:length(X!*FH(,:D!*>DI
W #f part of un&nown data in X are replaced %y DaD then the time vector (,:length(X!* is
W changed to exclude time moments where DaD inserted.
W Cutput F and S for each (DFT iteration are calculated %y following formulas:
W *. ;@(Hdiag(8>DH(FI
W (DFT using function ifft to calculate ; faster.
W -. F@8.H(XHinv(;H(I
W S@(XHinv(;H(.>diag((FHinv(;H(.FI
W 9evinson!Dur%in recursion used for inverse of toeplit' ;.
W Function fft applied to speed up matrix multiplications.
W 0. 8@S.Hcon)(SI 8 used as input to the next (DFT iteration.
W 3 special case: if length(X is e"ual to D$ the (DFT output do not depend on selected weight
W vector 8 and is calculated in non!iterative way.
W F(3T=;(S
W *. (DFT output F is the D!point Fourier transform of data X.
W The .ower Spectral Density (.SD function can %e calculated %y the following formula:
W a%s(F.Z->(DHT$ T ! sampling period.
W -. (DFT can extrapolate input data X to length D. That is$ if apply (DFT for DNlength(X$
W get the results: F@edft(X$D@edft(Y@fft(YI Y@ifft(F$ where Y is input X plus non!'ero
W forward and %ac&ward extrapolation of X to length D.
W 0. (DFT output S estimate amplitudes and phases of sinusoidal components in input data X.
W 7. (DFT can increase fre"uency resolution D>length(X times. Division of outputs *>(THF.>S
W demonstrate the fre"uency resolution of (DFT. The following is true for any (DFT iteration:
W ,SF.>SS@D$
W sum(F.>S@DHlength(X.
W 1. (DFT input data X may contain DaD. DaD indicates unavaila%le data or missing samples
W or data segments in X. (DFT Cutputs F and S are calculated %y applying slower algorithm
W then in case of X without DaD.
W 2. #f X is a matrix$ the (DFT operation is applied to each column.
W
W See also FFT$ #FFT$ FFTS:#FT.
W@@@@@@@@@@@@@@@@@@@@@@@ Set default parameters for (DFT @@@@@@@@@@@@@@@@@@@@@@@@@@@@
<iteration@0,I W 9imit for maximum num%er of iteration (Stopit ,.
;deviat@,.,,,1I W Galue for relative deviation (Stopit *.
;thresh@,.,,,*I W Galue for relative threshold (Stopit -.
W@@@@@@@@@@@@@@@@@@@@@@@ Chec& (DFT input arguments @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if nargin@@,$ error(FDot enough input arguments. See help edft.F$ end W Chec& input argument X.
if sum(any(isinf(X$ error(F#nput argument X contain #nf. See help edft.F$ end
if si'e(X$*@@*$
Extended summary of Dr.Sc.Comp. thesis 23
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
X@X.FItrf@*I W X is row vector
else
trf@,I W X is - dim array
end
RU 9Q@si'e(XI W U ! length of input data X
if narginN*$ W Chec&ing input argument D.
if isempty(D$D@UIend
D@floor(D(*I
if DSU$ X@X(*:D$:IU@DI end W Truncate X if has more than D points
else
D@UI
end W Chec&ing X on DaDs:
Xnan@Oisnan(XI W Xnan ! indicate samples as F*F $ DaD as F,F
if D@@*$
UU@XnanI
else
UU@sum(XnanI W UU ! length of input data X without DaD
end
if narginS0$ W Chec&ing input argument #.
#@<iterationI W Set default value for #.
else
if isempty(#$#@<iterationIend
#@floor(#(*I
end
if narginS7$ W Chec&ing of input argument 8.
8@ones(D$9I W Set default values for 8
else
if trf@@*$8@8.FIend
if (si'e(8$*O@DM(si'e(8$-O@9$ error(F#ncorrect si'e of input argument 8. See help edft.F$ end
8@8.Hcon)(8I
if any(find(sum(8N,SUU$ error(FToo many 'eros in input argument 8. See help edft.F$ end
end
W@@@@@@@@@@@@@@@@@@@@@@@ Set default values for (DFT output arguments @@@@@@@@@@@@@@@
F@'eros(D$9IS@'eros(D$9I W Fill with 'eros output matrixes F$S.
Stopit@R#Hones(*$9I'eros(*$9QI W Set default value for Stopit.
W@@@@@@@@@@@@@@@@@@@@@@@ Calculate (DFT for each X column l @@@@@@@@@@@@@@@@@@@@@@
for l@*:9$
W@@@@@@@@@@@@@@@@@@@@@@@ Chec& for a special cases @@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if UU(l@@DMUU(l@@,$ W #f length(X@D or X(:$l has all DaDs then
F(:$l@fft(X(:$l$DI W (DFT output (F$S e"uals to FFT.
S(:$l@F(:$l>DI
Stopit(:$l@R*I ,QI
elseif U@@*TDO@*$ W Special case$ the length(X@*$
F(:$l@fft(X(:$l$D.FI W (DFT output (F$S e"uals to FFT.
S(:$l@F(:$l>DI
Stopit(:$l@R*I ,QI
elseif find(X(:$l@@RQTUU(lN,$ W #f input X(:$l has all 'eros or 'erosTDaD
Stopit(:$l@R*I ,QI W then (DFT output (F$S is 'eros.
W@@@@@@@@@@@@@@@@@@@@@@@ /asic (DFT algorithm started @@@@@@@@@@@@@@@@@@@@@@@@@@@
elseif UU(l@@U$ W #nput X(:$l does not contain DaD
W@@@@@@@@@@@@@@@@@@@@@@@ 3pply F3ST(; algorithm @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
for it@*:#$ W Start iterations...
r@ifft(8(:$lI W Calculate correlation vector (r.
W .erform inverse of correlation matrix: 9evinson!Dur%in recursion.
a@!r(->r(*I
G@r(*!r(-Hcon)(r(->r(*I
for n@*:U!-$
alfa@R* a.FQHr(nB-:!*:-I
rho@!alfa>GI
G@GBrhoHcon)(alfaI
a@RaBrhoHcon)(flipud(aIrhoQI
end
a@R*IaQI
W #nverse %y <atla% %ac&slash operator (an alternative approach.
W a@R*I toeplit'(con)(r(*:U!*[(!r(-:UQI
W G@a.FHcon)(r(*:UI
W Calculate (;(@diag((FHinv(;H( and X;@XHinv(;.
X;@'eros(U$*I;(@'eros(U$*Irc@aI
for &@*:U>-$
&,@U!&B*I
&*@-:U!-H&B*I
&-@&B*:U!&I
&0@&:U!&B*I
Extended summary of Dr.Sc.Comp. thesis 24
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
;((*@;((*B-Hrc(&I
;((&,!&B*@;((&,!&B*B-Hrc(&,I
;((&*@;((&*B7Hrc(&-I
X;(&@X;(&Brc(&0FHX(&0$lI
X;(&,@X;(&,B(flipud(rc(&0.FHX(&0$lI
X;(&-@X;(&-Brc(&-HX(&$lBflipud(con)(rc(&-HX(&,$lI
rc(&-@rc(&-!*Bcon)(a(&B*Ha(&-!a(&,Hflipud(con)(a(&-B*I
end
if round(U>-NU>-$
;((*@;((*Brc(&B*I
X;(&B*@X;(&B*BX(&B*$lHrc(&B*I
end
(;(@real(fft(;($DI
8(:$l@8(:$l>real(GI
W Stopit *: /rea& iterations if sum(F.>S is not e"ual to DHU.
if a%s((;(.FH8(:$l>D>U!*N;deviat$ Stopit(:$l@Rit!*I *QI %rea&$ end
W Calculate outputs for iteration (it: D!point (DFT (F and 3mplitude Spectrum (S.
F(:$l@fft(X;$DI
S(:$l@F(:$l.>(;(I
F(:$l@F(:$l.H8(:$lI
W Calculate weight (8 for the next iteration.
8(:$l@S(:$l.Hcon)(S(:$lI
W Stopit -: /rea& iterations if relative threshold reached.
S8(it@sum(8(:$lI
if itN*$ thit@a%s(S8(it!*!S8(it>S8(*I
if thitS@;thresh$ Stopit(:$l@RitI -QI %rea&$ end
end
end W ... end iterations.
W@@@@@@@@@@@@@@@@@@@@@@@ (nd of F3ST(; algorithm @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
else W #nput X(:$l contains DaD
W@@@@@@@@@@@@@@@@@@@@@@@ 3pply S9C8(; algorithm @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
#DG;@'eros(UI(;@'eros(U$*I
X(find(OXnan(:$l$l@'eros(U!UU(l$*I W ;eplace DaD %y , in X
t@find(Xnan(:$lI W Sample num%ers vector (t
for it@*:#$ W Start iterations...
W Calculate correlation matrix (; %y applying ifft and inverse of ;.
;T@ifft(8(:$lI
;@toeplit'(;T(*:UI
#DG;(t$t@inv(;(t$tI
(;(*@trace(#DG;I
for &@*:U!*
(;(&B*$*@sum(diag(#DG;$&Bcon)(diag(#DG;$!&I
end
W Calculate (;(@diag((FHinv(;H(.F %y applying fft..
(;(@real(fft((;$DI
W Stopit *: /rea& iterations if sum(F.>S is not e"ual to DHUU.
if a%s((;(.FH8(:$l>D>UU(l!*N;deviat$ Stopit(:$l@Rit!*I *QI %rea&$ end
W Calculate outputs for iteration (it: D!point (DFT (F and 3mplitude Spectrum (S.
F(:$l@fft(con)(#DG;HX(:$l$DI
S(:$l@F(:$l.>(;(I
F(:$l@F(:$l.H8(:$lI
W Calculate weight (8 for the next iteration.
8(:$l@S(:$l.Hcon)(S(:$lI
W Stopit -: /rea& iterations if relative threshold reached.
S8(it@sum(8(:$lI
if itN*$ thit@a%s(S8(it!*!S8(it>S8(*I
if thitS@;thresh$Stopit(:$l@RitI -QI%rea&$end
end
end W ... end iterations.
W@@@@@@@@@@@@@@@@@@@@@@@ (nd of S9C8(; algorithm @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
end
end
W@@@@@@@@@@@@@@@@@@@@@@@ 3d)ust si'e of (DFT output @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
if trf@@*$F@F.FIS@S.FIend
The next program demonstrates applica%ility of the (xtended DFT in -!dimensional signal
processing. The program (DFT-.m is created on %asis of standard <3T93/ li%rary program
FFT-.m where FFT.m calls are simply replaced %y (DFT.m. The inverse transform to
(DFT-.m is the <3T93/ li%rary program #FFT-.m.
Extended summary of Dr.Sc.Comp. thesis 25
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
function f @ (DFT-(x$ mrows$ ncols
W (DFT- Two!dimensional (xtended Discrete Fourier Transform.
W (DFT-(X returns the two!dimensional Fourier transform of matrix X.
W /efore run (DFT- un&nown data (if any inside of X should %e replaced
W %y DaD (Dot!a!Dum%er.
W #f X is a vector$ the result will have the same orientation.
W (DFT-(X$<;C8S$DCC9S performing si'e <;C8S!%y!DCC9S Fourier transform
W without padding of matrix X with 'eros.
W
W See also (DFT$ #FFT-.
W
W (DFT- is created on %asis of <3T93/ program FFT- (\.D. 9ittle *->*6>*+61
W Do input.
if nargin@@,$ error(FDot enough input arguments. See help edft-.F$ end
Rm$ nQ @ si'e(xI
W /asic algorithm.
if (nargin @@ * T (m N * T (n N *
W f @ fft(fft(x.F.FI
f @ edft(edft(x.F.FI
returnI
end
W .adding for vector input.
if nargin S 0$ ncols @ nI end
if nargin S -$ mrows @ mI end
mpad @ mrowsI npad @ ncolsI
if m @@ * T mpad N m$ x(-$ * @ ,I m @ -I end
if n @@ * T npad N n$ x(*$ - @ ,I n @ -I end
if m @@ *$ mpad @ npadI npad @ *I end W For row vector.
W Transform.
Wf @ fft(x$ mpadI
Wif m N * T n N *$ f @ fft(f.F$ npad.FI end
f @ edft(x$ mpadI
if m N * T n N *$ f @ edft(f.F$ npad.FI end
The first version of (DFT (file 4DFT.m was su%mitted on *,>5>*++5 as <3T93/ 7.* code.
The renewed code version uploaded on 6>1>-,,2 and availa%le online
http:>>www.mathwor&s.com>matla%central>fileexchange>**,-,!extended!dft.
.lease note that programs have not %een tested on the latest <3T93/ versions and therefore
have opportunities to performance improvements (see for example [23$ 26].
1 -eferences
[1] S.<. Uay$ S.9. <arple. Spectrum analysis ! a modern perspective. .roc. #((($ Gol.2+$ Do.**$
*+6*.
[2] G. 9iepinPsh. 3 method for spectral analysis of %and!limited signals$ 3utomatic Control
and Computer Sciences. Gol.-5$ Do.1$ pp. 12!20$ *++0.
[3] G. 9iepinPsh. 3 method for spectrum evaluation applica%le to analysis of periodically and non!
regularly digiti'ed signals. 3utomatic Control and Computer Sciences$ Gol.-5$ Do.2$ pp. 15!27$
*++0.
[4] G. 9iepinPsh. 3 spectral estimation method of nonuniformly sampled %and!limited signals.
3utomatic Control and Computer Sciences$ Gol.-6 $ Do.-$ pp. 22!50$ *++7.
[5] G. 9iepinPsh. 3n algorithm for evaluation a discrete Fourier transform for incomplete data.
3utomatic Control and Computer Sciences$ Gol.0,$ Do.0$ pp.-5!7,$ *++2.
[6] Gilnis 9iepins. :igh!resolution spectral analysis %y using %asis function adaptation
approach. Doctoral Thesis for Scientific Degree of Dr. Sc. Comp. >in Lat;ian>$ =niversity of
9atvia$ *++5. 3%stract availa%le on http:>>hdl.handle.net>*,,26>00,6*2 .
[7] <.D. Sacchi$ T.\. =lrych$ C. 8al&er. #nterpolation and extrapolation using a high!
resolution discrete Fourier transform. #((( Trans. on Signal .rocessing$ Gol.72$ Do.*$ pp. 0*!
06$ *++6.
Extended summary of Dr.Sc.Comp. thesis 26
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
[8] <odris 4reitans. <ulti%and signal processing %y using nonuniform sampling and iterative
updating of autocorrelation matrix. .roceedings of the -,,* #nternational Conference on
Sampling Theory and 3pplication$ <ay *0!*5$ -,,*$ Crlando$ Florida$ =S3$ pp. 61!6+.
[9] <odris 4reitans. Spectral analysis %ased on signal dependent transformation. The -,,1
#nternational 8or&shop on Spectral <ethods and <ultirate Signal .rocessing $ (S<<S.
-,,1$ \une -,!--$ -,,1$ ;iga$ 9atvia.
[10] \ayme 4arcia 3rnal /ar%edo$ 3mauri 9opes$ .atric& \. 8olfe. :igh Time!;esolution
(stimation of <ultiple Fundamental Fre"uencies. .roceedings of the 6th #nternational
Conference on <usic #nformation ;etrieval$ #S<#; -,,5$ Sept. -0!-5$ Gienna$ 3ustria$ pp.
0++!7,-.
[11] David Dolenc$ /ar%ara ;omanowic'$ .aul <c4ill$ 8illiam 8ilcoc&. C%servations of
infragravity waves at the ocean!%ottom %road%and seismic stations (ndeavour (U(// and
(xplorer (UX//. 4eochemistry$ 4eophysics$ 4eosystems$ Gol.+$ #ssue 1$ <ay -,,6.
[12] ]uanming ^hang$ :ui)in 9iu$ :ong&un Chen$ ]ionglin 9i$ ^henhuan ^hang. 3 .recise
and 3daptive 3lgorithm for #nterharmonics <easurement /ased on #terative DFT. #((( Trans
on .ower Delivery$ Gol.-0$ #ssue 7$ pp. *5-6!*501$ Cct. -,,6.
[13] .etre Stoica$ \ian 9i$ :ao :e. Spectral 3nalysis of nonuniformly Sampled Data: 3 Dew
3pproach Gersus the .eriodogram. #((( Trans on Signal .rocessing$ Gol.15$ #ssue 0$ pp. 670!
616$ <arch -,,+.
[14] (ric 4reenwood$ Fredric :. Schmit'. Separation of <ain and Tail ;otor Doise Sources
from 4round!/ased 3coustic <easurements =sing Time!Domain De!Doppleri'ation. 01th
(uropean ;otorcraft Forum -,,+$ Sept. --!-1$ :am%urg$ 4ermany.
[15] \ayme 4arcia 3rnal /ar%edo$ 3mauri 9opes$ .atric& \. 8olfe. (mpirical <ethods to
Determine the Dum%er of Sources in Single!Channel <usical Signals. #((( Transactions on
3udio$ Speech and 9anguage .rocessing$ Gol.*5$ #ssue 5$ pp. *701!*777$ Sept. -,,+.
[16] Tari& Yardi%i$ \ean 9i$ .etre Stoica$ <ing Xue$ 3rthur /. /aggeroer. Source locali'ation
and sensing: 3 nonparametric iterative adaptive approach %ased on weighted least s"uares.
#((( Transactions on 3erospace and (lectronic Systems$ Gol.72$ pp. 7-1!770$ \an. -,*,.
[17] <. Caciotta$ S. 4iarnetti$ F. 9eccese$ ^. 9eonowic'. Comparison %etween DFT$ adaptive
window DFT and (DFT for power "uality fre"uency spectrum analysis. <odern (lectric
.ower Systems (<(.S$ -,*, .roceedings of the #nternational Symposium$ Sept. -,!--$
-,*,$ 8roclaw$ pp. *!1.
[18] 9i 9i$ Yan ^heng$ 8ang Xing!'hi. #nter!harmonic 3nalysis =sing #44 and (xtended
Fourier. .roceedings of the Chinese Society of =niversities for (lectric .ower System and its
3utomation$ --(0$ -,*,.
[19] (ri& 4udmundson$ 3ndreas \a&o%sson$ \_rgen \ensen$ .eter Stoica. 3n #terative
3daptive 3pproach for /lood Gelocity (stimation =sing =ltrasound. *6th (uropean Signal
.rocessing Conference ((=S#.CC$ 3al%org$ Denmar&$ 3ug -0!-5$ -,*,$ pp. 076!01-.
[20] <odris 4reitans$ ;olands Shavelis. ;econstruction of se"uences of ar%itrary shaped
pulses from its low pass or %and pass approximations using spectrum extrapolation.
(=S#.CC -,*,$ 3ug. -0!-5$ 3al%org$ Denmar&$ pp. *2,5!*2**.
[21] \uggrapong Treetrong. Fault Detection of (lectric <otors /ased on Fre"uency and Time!
Fre"uency 3nalysis using (xtended DFT. #nternational \ournal of Control and 3utomation$
Gol.7 Do.*$ <arch -,**.
[22] \esper ;indom \ensen$ <ads 4r`s%all Christensen$ Saren :oldt \ensen. 3 Single
Snapshot Cptimal Filtering <ethod for Fundamental Fre"uency (stimation. 02th #nternational
Conference on 3coustics$ Speech and Signal .rocessing (#C3SS.$ .rague$ C'ech ;epu%lic$
<ay --!-5$ -,**$ pp. 7-5-!7-51.
Extended summary of Dr.Sc.Comp. thesis 27
Dr.Sc.Comp. Vilnis Liepi Email: vilnislp@gmail.com
[23] <ing Xue$ 9u'hou Xu$ \ian 9i. #33 spectral estimation: fast implementation using the
4oh%erg!Semencul factori'ation. #C3SS.-,**$ <ay --!-5$ .rague$ C'ech ;epu%lic$ pp.
0-1*!0-2*.
[24] /onifatius 8ilhelmus Tilma. Supervisor: <.U. SmitI Co!promotor: (.3.\.<. /ente.
#ntegrated tuna%le "uantum!dot laser for optical coherence tomography in the *.5bm
wavelength region. (indhoven$ Technische =niversiteit (indhoven$ Diss.$ \un. -,**.
[25] (lmar <air$ <ichael Fleps$ <ichael Suppa$ and Darius /ursch&a. Spatio!temporal
initiali'ation for #<= to camera registration. #n .roceedings of the #((( #nternational
Conference on ;o%otics and /iomimetics (;C/#C$ Dec. -,**$ pp. 115!127.
[26] 4eorge!Cthan 4lentis$ 3ndreas \a&o%sson. Superfast 3pproximative #mplementation of
the #33 Spectral (stimate. #((( Transactions on Signal .rocessing$ Gol.2,$ #ssue *$ \an.
-,*-$ pp. 75-!756.
[27] /. 8. Tilma$ Yu"ing \iao$ \. Uotani$ /. Smal%rugge$ :. .. <. <. 3m%rosius$ .. \. Thi)s$
X. \. <. 9ei)tens$ ;. Dot'el$ <. U. Smit$ (. 3. \. <. /ente. #ntegrated Tuna%le ]uantum!Dot
9aser for Cptical Coherence Tomography in the *.5 m 8avelength ;egion. #((( \ournal of
]uantum (lectronics$ Gol.76$ Do.-$ Fe%. -,*-$ pp. 65!+6.
[28] T. Cdstrcil$ <. Cdstrcil$ C. 4rover*$ G. Svo%oda*$ #. curan*$ and \. <lynde. 9ow cost
alternative of high speed visi%le light camera for to&ama& experiments. ;eview of Scientific
#nstruments. Cct. -,*-$ Gol.60$ #ssue *,.
[29] (lmar <air. Co!promotor: 4regory Donald :ager. (fficient and ;o%ust .ose (stimation
fghij lm nmiopqgr gmj sqhtgr uimhqmvw xymz{im| }iz{mqhz{i ~mqiohqpp xymz{im| qhhw|
-,*-.
[30] 3&ash U Singh. ]uantum!Dot 9aser CCT. #nternational \ournal of (ngineering ;esearch
and 3pplications (#\(;3$ Gol.-$ #ssue 2$ Dov.! Dec. -,*-$ pp.07,!05*.
[31] (lliot /riggs$ CFD< .hysical 9ayer 3rchitecture and ;eal!Time <ulti!.ath Fading
Channel (mulation for the 04.. 9ong Term (volution Downlin&. .hD Thesis. Texas Tech
=niversity$ Dec. -,*-.
[32] Uwadwo S. 3gyepong$ Fang!:an :su$ (dward ;. Dougherty and (rchin Serpedin.
Spectral 3nalysis on Time!Course (xpression Data: Detecting .eriodic 4enes =sing a ;eal!
Galued #terative 3daptive 3pproach. 3dvances in /ioinformatics$ Gol. -,*0.
Extended summary of Dr.Sc.Comp. thesis 28

Você também pode gostar