Scribd
Upload
7 views2 pages

Estimation theory in signal processing

The document discusses current research directions in signal processing, focusing on estimation theory, linear prediction, spectral analysis, and time delay estimation. It highlights the relationship between linear prediction and system modeling, emphasizing the use of autoregressive processes and ladder filters for optimal predictors. Additionally, it touches on the challenges and advancements in spectral analysis and time delay estimation techniques, acknowledging the support of the National Science Foundation for ongoing research.

Uploaded by

Ayub Wazir
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
7 views2 pages

Estimation theory in signal processing

The document discusses current research directions in signal processing, focusing on estimation theory, linear prediction, spectral analysis, and time delay estimation. It highlights the relationship between linear prediction and system modeling, emphasizing the use of autoregressive processes and ladder filters for optimal predictors. Additionally, it touches on the challenges and advancements in spectral analysis and time delay estimation techniques, acknowledging the support of the National Science Foundation for ongoing research.

Uploaded by

Ayub Wazir
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

C opyright © IFAC Identifica tion a nd System Parameter

Estimation 1982 , W ashington D ,C .. USA 1982

ESTIMATION THEORY AND SIGNAL PROCESSING

B. W. Dickinson
Department of Electrical Engineering and Computer Science, Pn'nceton University,
Pnnceton, NJ 08544 , USA

Abstract. In keeping with the objective of the workshop, we will discuss


some current research directions in signal processing related to estimation
problems. Topics to be considered include linear prediction, spectral
analysis methods, and time delay estimation.

Keywords. Identification; parameter estimation; prediction; signal proces-


sing; spectral analysis; state estimation.

INTRODUCTION the output of a dynamical system excited by


a white noise input. When only second or-
If we adopt the broad view of signal proces- der statistical information is specified,
sing as a set of techniques for enhancing the system may be taken to be linear. This
and extracting the information content of system is essentially the inverse of the
observed signals, then estimation theory pro- Kalman one-step prediction filter for the
vides one framewor~ for modeling signal pro- process. Thus there is a close relation-
cessing problems , Such a probabilistic ap- ship between linear prediction of a process
proach provides a foundation upon which op- and the task of modeling it as the output
timum signal processing techniques may be of a linear system. The techniques of time
erected. It will not be appropriate in every series analysis and system identification
imaginable situation; deterministic models are the essential tools involved . The most
and approximation theory may be combined to convenient class of processes to model are
give optimal methods for certain problems. the autoregressive processes . This can be
On the other hand, some formulations lead seen in many ways, but from the point of
naturally to minimum norm problems in a Hil- view of estimation theory it is due to the
bert space setting, and solutions are ob- fact that their optimal predictors are time-
tained by the geometric procedure of orthog- varying only for a transient period equal to
onal projection. Only the interpretation of the system order. After that they are sim-
the solution relies on the nature of the prob- ply time-invariant FIR filters.
lem formulation -- for example as an esti-
mation problem. There is a remarkable filter structure that
may be associated with optimal predictors,
We will use this written supplement to our the (normalized) ladder filter. Derived
workshop presentation a s a general discus- first in an estimation theory context (i.e.
sion of the va rious sign a l processing topi cs known statistical structure), it has been
to be covered. It is not our goal to write extended to exact least squares prediction,
a survey on this vast area but only to in- in a deterministic setting, for an observed
dicate some areas of recent research activ- signal by introduc ing an appropriate
ity. Because of time limitations for our Hilbert space formulation. A main goal of
presentation, we have been selective in the our talk will be to explain the ladder
choice of topics without attempting to be structure and its uses in such linear pre-
s ystematic. We will consider only discrete diction problems. The exact least squares
time problems in this paper. These are the ladder algorithm can replace approximate
ones of most interest, at least in the case modeling techniques, such as those based on
of time signals, because of the total domi- gradient algorithms, without additional com-
nation of digital computers in physical putational requirements. Its uses include
realizations of the signal proc essing algor- such signal processing applications as lin-
itrns. We don't intend to imply that the ear predictive analysis for speech com-
problems selected are the only interesting pression and equalization of telephone
ones around : channels for high rate digital communica-
tions. Clever algorithmic methods have
LINEAR PREDICTION been designed to facilitate hardware im-
plementation on a VLSI chip.
The innovations representation for a random
process prov ides a model f or the process as

91
92 B. W. Dickinson

SPECTRAL ANALYSIS TIME DELAY ESTIMATION

One use of the models fit by linear predic- A good illustration of the multichannel sig-
tive methods is to estimate spectral infor- nal processing problems arising in many ap-
mation about the observed process. For nar- plications is the time delay estimation
rowband signals in high signal-to-noise ra- problem. Suppose that two noisy signals are
tio environments, fitting of autoregressive obtained at different spatial locations and
models and using the estimated narrow band- the information-bearing signal appears in
width poles to obtain frequency estimates each one, although not synchronously. Esti-
works reasonably well. More general pole- mation of time delay may be used to estimate
and-zero models may also be used with bet- directional information about the signal
ter results for more computational effort. source. There are a variety of methods em-
There is an expected loss of resolving ployed for solving this problem and its var-
power as the SNR decreases, including a iants, including spectral techniques and
threshold effect. Recently proposed modifi- techniques based on linear prediction. An
cations employ singular value decomposition extended Kalman filtering technique may also
in various ways to obtain threshold exten- be devised. One of our main reasons for in-
sion. For sinusoidal signals in additive cluding a brief mention of this kind of sig-
noise, it is possible to separate the nal processing problem is the availability
(linear) least squares estimation of the am- of an excellent source of recent papers
plitudes and the (nonlinear) least squares that interested readers may consult for fur-
estimation of the frequencies. The resul- ther information. There was a special
ting frequency estimates are optimal in the issue of the IEEE Transactions on Acoustics,
Cramer-Rao sense and give substantial Speech, and Signal Processing devoted to
threshold extension over linear predictive time delay estimation in June, 1981.
estimates.
ACKNOWLEDGMENT
The breadth of recent work in spectral anal-
ysis is great, but it is an area where the I would like to gratefully acknowledge the
tools of estimation theory need to be ap- support of the National Science Foundation
plied more systematically in the study of for my research in signal processing over
performance of various methods. Convergence the past few years through grant ENG77-28523.
theorems and performance bounds appear in
far fewer papers than simulation results.
We look forward to more progress on these
problems in the future.

You might also like