numerical reconstruction of signals - ima · definition 1.1 ([9, 10]) let denote the unique complex...
TRANSCRIPT
September 14, 2011 NUMERICAL RECONSTRUCTION OF SIGNALS
1
NUMERICAL RECONSRUCTION OF SIGNALS AND ORTHOGONAL POLYNOMIALS
Nicholas J. Daras
Department of Mathematics,
University of Military Education, Hellenic Military Academy,
166 73, Vari Attikis, Greece
E-mail: [email protected]
Abstract. Given only a few of initial Fourier coefficients for a continuous-time periodic signal, we construct efficient rational
approximants to the whole signal everywhere on his domain of definition. The convergence of these approximants depends
on the orthonormality of the chosen generating polynomial system into . The form of
each is characterized by recurrence relations dues to the connection between Schur and Szegö theories.
MSC2010: 41A20, 41A21, 42A16, 42C05, 65B99, 65D05, 65T40
Key words: approximation by rational functions, Padé-type and/or Padé approximants, Fourier coefficients, orthogonal
polynomials, acceleration of convergence, interpolation, trigonometric approximation and interpolation
INTRODUCTION Given a signal Fourier analysis easily calculates the frequencies and the amplitudes of those frequencies
which make up the signal. However, Fourier methods are not always a good tool to recapture the signal,
particularly if it is highly non-smooth. Especially, there is no analytical way to reconstruct with exactitude a
signal if only a few of its initial Fourier coefficients are known.
In this paper, we will consider a numerical version of the Carathéodory-Fejér interpolation problem ([4,
5], [13, 14]) in the trigonometric context. In particular, we will investigate a numerical method for
constructing efficient approximants to any continuous-time periodic signal by using only a few of its
initial Fourier coefficients. These approximants are real parts of rational functions with numerators
determined by the condition that the Fourier series expansion of the approximants matches the Fourier
series of the signal as far as possible. Motivated by this crucial property, the obtained approximants will be
called Padé-type approximants. The convergence of these approximants depends strongly on the
orthonormality of the chosen generating polynomial system into . The
form of is characterized by recurrence relations dues to the connection between Schur and Szegö
theories.
The detailed definition of a Padé-type approximant to a continuous-time periodic signal and their
principal properties are presented in Section 1. Section 2 is devoted to a study of the convergence of a
sequence of such approximants, with emphasis on the assumptions under which, for every sequence of
functions converging to a periodic continuous signal there is a Padé-type approximation sequence
converging point wise to that signal faster than the first sequence. Finally, in Section 3 numerical examples
are given making use of Padé-type approximants to signals.
1. CONSTRUCTION OF RATIONAL APPROXIMANTS TO A PERIODIC CONTINUOUS TIME SIGNAL Consider a periodic continuous time signal defined over an interval .
Suppose and has a finite number of extrema and discontinuities in any given
interval. Then, has a Fourier series expansion defined by
. The
problem we will investigate is the following: if only a few Fourier coefficients of the signal are given,
construct efficient rational approximants to the whole signal (almost) everywhere on .
Putting , the signal converts to a periodic continuous time signal defined over
the standard interval , with Fourier series expansion . It is clear that can be
NUMERICAL RECONSTRUCTION OF SIGNALS September 14, 2011
2
identified with a real-valued function in of the unit circle by setting
. From the solution of the Dirichlet problem in the unit disk , it follows that the extended
real-valued function is harmonic in the open unit disk and such that
and the Fourier series expansion of the restriction of
to any circle of radius is given by .
We can consider Padé-type approximants to the harmonic function . Their fundamental property is that the Fourier series expansion of their restriction to any circle of radius coincides with the Fourier series expansion of as far as possible. Indeed, is the real part of an analytic function in . So, where denotes complex conjugate of . If
is the Taylor power series
expansion of around , then and
( ). Define the linear functional , where is the
vector space of all complex analytic polynomials. An application of Cauchy’s integral formula shows that
whenever . By density, there is a
continuous extension of into the space of all functions which are analytic in an open neighborhood
of . In particular, for every fixed point , the number is well defined and
equals . Hence, it holds for any . If the function
is replaced by a polynomial , then is approximated by . This is
an approximate quadrature formula and leads to a Padé-type approximant to the harmonic real-valued function .
DEFINITION 1.1 ([9, 10]) Let denote the unique complex polynomial of degree at most in , which interpolates the Cauchy kernel at points , i.e., for any
and . The real-valued function
is said to be a Padé-type approximant to , with
generating polynomial ( ). ■
The approximant in Definition 1.1 can be interpreted as real part of a rational function. Indeed, we put
and
.
THEOREM 1.2 ([8, 9]) is the real part of a rational complex function of type : + = + .The error of the respective approximation equals
.
The fundamental property of these approximants is the following. The Fourier series
expansion
of the restriction of to the circle of
radius fulfills
for any . ■
Since the limit is uniform on – , the function is the
Poisson integral of a continuous function on the unit circle.
DEFINITION 1.3 ([8, 9])The radial limit
is said to be a Padé-type approximant to , with generating polynomial
( ). ■
By Theorem 1.2, we get the following properties of these approximants.
THEOREM 1.4 The Padé-type approximant is the real part of a rational complex
function of type : .
The error of the respective approximation is given by the following theoretical formula
where the limit is taken in the –norm. The Fourier series expansion
of
fulfills
for any .
September 14, 2011 NUMERICAL RECONSTRUCTION OF SIGNALS
3
Proof We will use the standard identification of the periodic signal ( ) with the function ( , ). Let be any sequence such
that . Since, by Theorem 1.2, ,
the uniform convergence of the sequence
to the radial limit function
implies that , and the first assertion
is proved. To prove the second assertion, recall that
. Further, by the
uniform convergence of the sequence
to the radial limit function
we also get
. Letting ,
from Theorem 1.2, it follows that there exists a such that
for any . Now, set
. An application of the continuity property for the linear
functional shows that
.
Hence
,
for any , which proves the second assertion. It remains to prove the third assertion. Since every harmonic function in the unit disk, with continuous boundary values, is the Poisson integral of its continuous restriction to the unit circle, we have
.
Henceforth, the Fourier series expansion of is
. From Theorem 1.2, it
follows that
for any , and the proof is complete. ■
From Theorem 1.4, it follows immediately that if only a few Fourier coefficients of the signal are given, then one can approximate by an approximant in the Padé-type:
.
2. CONVERGENCE ACCELERATION AND ORTHOGONAL POLYNOMIALS In this Section, we shall study the convergence of a sequence of Padé-type approximants to a signal.
Let
be an infinite triangular interpolation matrix with complex entries
. For any fixed , let denote the unique
polynomial of degree at most in , which interpolates the Cauchy kernel in the nodes
of the th row of , i.e.,
for any .
THEOREM 2.1 Suppose the family of generating polynomials is an
orthonormal bounded system in and
, for a . Then for
any real-valued periodic continuous-time signal , the associated
sequence
of Padé-type approximants to , with
generating polynomials , converges to almost everywhere on .
Especially, if , the sequence
converges to everywhere on . Proof Let be a real-valued periodic continuous-time signal. Let also and be a strictly increasing sequence of positive numbers such that . Fix any . By
Theorem 1.4, there is a subsequence
of such that
for almost all . Denote by the set of all points with this property. For every , one can find a such that
( ) and
NUMERICAL RECONSTRUCTION OF SIGNALS September 14, 2011
4
the function is in , so, by Mercer’s Theorem, the
Fourier coefficients of this function with respect to the orthonormal family
, tend to zero. This means that there exists a = , such that
(for any and ).
Combination of these inequalities shows that for any . This implies that
for almost all . If , the convergence holds for
any . ■
COROLLARY 2.2 Let the generating polynomials of a Padé-type approximation
( ) be such that
( ) and
( ).
If there are two constants and fulfilling
( ) and ( and ).
Then for any real-valued periodic continuous-time signal , the associated
sequence
of Padé-type approximants
to , with generating polynomials , converges to almost everywhere
on . Especially, if , the sequence
= , converges to everywhere on , . ■
In [10], we gave a stronger sufficient condition in terms of the entries only. If the
interpolation points are chosen so that and
, then,
for any real-valued periodic continuous-time signal , the associated sequence
of Padé-type approximants to , with generating polynomials
, converges to everywhere on .
One can also obtain interesting results about the form of , dues to the connection between
Schur and Szegö theories. This connection is often attributed to Akhiezer ([1]), but it appears earlier and in
greater detail in the papers of Geronimo ([11, 12]). It based on important recurrence relations which were
first given in previous Szegö’s work ([15]). Denoting by the polynomial , the
recurrence relations written in terms of the monic polynomials are of
the general form for certain parameters ( ).
In current terminology the numbers are called Szegö parameters. Let us see what the
Schur parameters are. To do so, we may remind the following Schur’s construction.
THEOREM 2.3 ([13, 14])There is a one-to-one correspondence between the class of analytic
functions which are bounded by one on the unit disk and the set of all sequences of
complex numbers which are bounded by and such that if some term has unit modulus, then all
subsequent terms are zero.
Proof Given any , define a sequence
.
If for some , then is constant and we taken for any . This occurs if and
only if is finite Blaschke product of factors:
,
where and …, are points in The numbers
( ) are
defined to be the Schur parameters for . ■
Since
has positive real part in the open unit disk and
value at the origin, the function belongs to . In 1943,
Geronimo showed that the Shur parameters
for coincide with the numbers :
whenever ([15]).
September 14, 2011 NUMERICAL RECONSTRUCTION OF SIGNALS
5
We shall now study assumptions under which, for every sequence of signals converging to a continuous periodic signal, there is always a Padé-type approximation sequence converging pointwise to the signal faster than the sequence. This property, due to the free choice of the interpolation points permits us to construct better and better approximations to continuous functions.
Using techniques similar to those proposed by Bromwich and Clark ([3, 6]), we can prove the
PROPOSITION 2.4 Let be the operator of differences. Let also and be two
convergent sequences of real numbers. Suppose is strictly monotone and .
If , the sequence converges faster than the
sequence . ■
Combination of Proposition 2.4 with Theorem 2.1 gives the following result. THEOREM 2.5 Let be any strictly monotone converging sequence. Suppose the family is an orthonormal bounded system in and
, for a . Then, for any real-valued periodic continuous-time
signal , the associated sequence
of
Padé-type approximants converges to faster than .everywhere on
.■
3. NUMERICAL EXAMPLES In this section, we give examples making use of Padé-type approximants to periodic continuous-
time signals . EXAMPLE 4.1 Let be the signal ( ). If and are the roots of the fifth Legendre
polynomial in , then and therefore
( ). Indicatively, we have and
. If and , then and therefore
( ). Indicatively, we have ,
, , and
.
EXAMPLE 4.2 Let be the signal ( ). If and
, then and therefore
( ).
In particular, it holds
,
,
,
,
and
.
If and and , then and therefore
NUMERICAL RECONSTRUCTION OF SIGNALS September 14, 2011
6
( ). Observe that is well defined everywhere on , with the exception of
. This is a consequence of our choice , which in particular implies . However,
,
,
and
.
EXAMPLE 4.3 Let finally be the signal
If and , , then and
( ).
In particular, we have , ,
, and
.
REFERENCES [1] N. I. Akhiezer: The Classical Moment Problem, Hafner, 1965.
[2] C. Brezinski: The Asymptotic Behavior of Sequences and New Series Transformations Based on the Cauchy
Product, Rocky Mountain I. Math. 21(1) (1991), 71-84.
[3] T. J. Bromwich: An Introduction to the Theory of Infinite Series, 2nd edition, Macmillan, London, 1949.
[4] C. Carathéodory: Über den Variabilitätsbereich der Koeffizienten von Potenzreihen, die gegebene Werte
nicht annehmen, Math.Ann. 64(1907), 95-115.
[5] C. Carathéodory: Über den Variabilitätsbereich der Fourierschen Konstanten von positive harmonischen
Functionen, Rend. di Palermo 32(1911), 193-217.
[6] W. D. Clark: Infinite Series Transformations and their Applications, Thesis, University of Texas, 1967.
[7] T. Constantinescu: Schur parameters, Factorization and Dilation Problems, Operator Theory: Advances
and Applications, 2, Birkhäuser, 1996.
[8] N. J. Daras: Rational approximation to harmonic functions, Numer. Algorithms 20 (1999), 285-301.
[9] N. J. Daras: Padé and Padé-type approximation to periodic functions, Acta Applicandae
Mathematicae 65 (2000), 245-343.
[10] N. J. Daras: Interpolation methods for the evaluation of a periodic finite Baire measure,
Approximation Theory and its Applications 17(2) (2001), 1-27.
[11] J. Geronimo: On polynomials orthogonal on the circle, on trigonometric moment problem, and on allied
Carathéodory and Schur functions, C.R. (Doklady) Acad.Sci.USSR (N.S.) 39 (1943), 291-295.
[12] J. Geronimo: On polynomials orthogonal on the circle, on trigonometric moment problem, and on allied
Carathéodory and Schur functions, Mat. Sb. 105 (147) (1978), English transl. in Math. USSR Sb. 34(2)
(1978), 131-145.
[13] I. Schur: On power series which are bounded in the interior of the unit circle (I. Gohberg, editor),
Methods in Operator Theory and Signal Processing, Operator Theory: Advances and Applications, 18
(1986), pp. 31-59 (Original (in German): J. Reine Angew. Math. 147 (1917), 205-232).
[14] I. Schur: On power series which are bounded in the interior of the unit circle II (I. Gohberg, editor),
Methods in Operator Theory and Signal Processing, Operator Theory: Advances and Applications, 18
(1986), pp. 68-88 (Original (in German): J. Reine Angew. Math. 184 (1918), 122-145).
[15] G. Szegö: Orthogonal Polynomials, Amer. Math. Soc. Colloq. Publ. 23 (4th edition with revisions, 1975;
1st edition 1939), Amer. Math. Soc., Providence, RI.