Anand D. Sarwate

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2006-3

January 23, 2006

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-3.pdf

The term "sensor network" encompasses a wide range of engineering systems with dramatically different characteristics. We consider a specific class of sensor networks whose objective is to reconstruct a source at a central terminal. Our objective in this thesis is to quantify the asymptotic error in reconstructing the source as the number of data sources, sensors, and model complexity increases. We consider three types of estimation systems -- unconstrained estimators for vector Gaussian sources that are allowed direct access to the sensor observations, estimators for discrete sources that receive information via rate constrained links from the sensors, and estimators for scalar Gaussians whose input is the output of a multiple-access channel.

We first establish bounds on the optimal estimator performance of these networks using a centralized estimator with access to all of the sensor observations. We assume the observations are noisy linear functions of the source and are thus specified by a matrix. Because the asymptotic error depends only on the spectral properties of this matrix, we can use tools from matrix analysis to give bounds on the spectrum and error in terms of the entries of the matrix for a number of different scenarios. Finally, we look at the case where the matrix is partially unknown. In some cases we can estimate the matrix directly from the data and in others we must minimize the worst mismatch distortion.

These problems can also be looked at in a more information-theoretic framework. We look at a lossless distributed source coding problem in which the joint distribution of the sources is partially unknown. Although for any finite number of sensors standard multi-terminal source codes can easily be adapted to handle the model uncertainty across time, we show a rate penalty is incurred if the number of sensors and blocklength go to $\infty$ simultaneously. This represents one kind of tradeoff between delay and complexity for the scaling behavior of these systems.

Finally, we look at the case where the sensors must communicate their observations across an additive white Gaussian noise multiple-access channel. With a known correlation structure, the optimal error converges to $0$ as $1/M$, where $M$ is the number of sensors. However, a simple feedback scheme using $K$ bits broadcast to all sensors can provide a distortion that scales to $0$ as $M^{-K/(K+2)}$. We conjecture that providing similar feedback to an optimal source code will not improve the performance beyond that of our protocol.

Advisors: Michael Gastpar

---

BibTeX citation:

@mastersthesis{Sarwate:EECS-2006-3,
    Author= {Sarwate, Anand D.},
    Title= {Observation Uncertainty in Gaussian Sensor Networks},
    School= {EECS Department, University of California, Berkeley},
    Year= {2006},
    Month= {Jan},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-3.html},
    Number= {UCB/EECS-2006-3},
    Abstract= {The term "sensor network" encompasses a wide range of engineering
systems with dramatically different characteristics.  We consider a
specific class of sensor networks whose objective is to reconstruct a
source at a central terminal.  Our objective in this thesis is to
quantify the asymptotic error in reconstructing the source as the
number of data sources, sensors, and model complexity increases.  We
consider three types of estimation systems -- unconstrained estimators
for vector Gaussian sources that are allowed direct access to the
sensor observations, estimators for discrete sources that receive
information via rate constrained links from the sensors, and
estimators for scalar Gaussians whose input is the output of a
multiple-access channel.

We first establish bounds on the optimal estimator performance of
these networks using a centralized estimator with access to all of the
sensor observations.  We assume the observations are noisy linear
functions of the source and are thus specified by a matrix.  Because
the asymptotic error depends only on the spectral properties of this
matrix, we can use tools from matrix analysis to give bounds on the
spectrum and error in terms of the entries of the matrix for a
number of different scenarios.  Finally, we look at the case where the
matrix is partially unknown.  In some cases we can estimate the matrix
directly from the data and in others we must minimize the worst
mismatch distortion.

These problems can also be looked at in a more information-theoretic
framework.  We look at a lossless distributed source coding problem in
which the joint distribution of the sources is partially unknown.
Although for any finite number of sensors standard multi-terminal
source codes can easily be adapted to handle the model uncertainty
across time, we show a rate penalty is incurred if the number of
sensors and blocklength go to $\infty$ simultaneously.  This
represents one kind of tradeoff between delay and complexity for the
scaling behavior of these systems.

Finally, we look at the case where the sensors must communicate their
observations across an additive white Gaussian noise multiple-access
channel.  With a known correlation structure, the optimal error
converges to $0$ as $1/M$, where $M$ is the number of sensors.
However, a simple feedback scheme using $K$ bits broadcast to all
sensors can provide a distortion that scales to $0$ as $M^{-K/(K+2)}$.
We conjecture that providing similar feedback to an optimal source
code will not improve the performance beyond that of our protocol.},
}

EndNote citation:

%0 Thesis
%A Sarwate, Anand D. 
%T Observation Uncertainty in Gaussian Sensor Networks
%I EECS Department, University of California, Berkeley
%D 2006
%8 January 23
%@ UCB/EECS-2006-3
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-3.html
%F Sarwate:EECS-2006-3