Our current research focuses on data science, machine learning, and statistical signal processing with applications ranging from
mobile communications to biomedicine and medicine. Some of our
current and past projects include the following.
Improper Gaussian signaling schemes for interferencelimited communications

This project aims at developing novel signaling schemes for
interferencelimited mobile communication networks. Already today,
multiuser interference presents the major limiting factor of the endtoend
performance in wireless communications. In pursuit of efficient
interference management schemes, current research focuses on
interference coordination by means of a joint precoder design that
exploits the channel state information. The vast majority of these
works assumes that the transmitted signals are distributed as proper
Gaussian random signals, which are known to be optimal when the
system is not interferencelimited. Recently, alternative signaling
schemes that deviate from these standard assumptions have been
proposed in the literature. It has been shown that the transmission of
improper Gaussian signals, whose real and imaginary parts are
correlated and/or have unequal power, outperform proper Gaussian
signals in various interferencelimited networks. This project studies
improper signaling schemes for different interferencelimited
networks. It has the following objectives:

To provide insights on when improper signaling enables better performance than its proper
counterpart. To this end, we will focus on singleantenna and partially
connected interference channels, where the analysis is tractable and
permits insights that can be extended to more general scenarios.

To design algorithms that optimize the transmission parameters for
general interference channels. An improper signaling scheme
contains not only the parameters associated with a proper scheme,
but also additional ones that describe the impropriety of the signal,
which makes the optimization more challenging.

To apply improper signaling to underlay cognitive radio scenarios, which can be
regarded as a paradigm for lowcooperative multiuser networks. In
these scenarios, the interference is managed by setting interference
constraints, so that its effect can be upperbounded and thereby
require less cooperation. We will explore new interference constraints
taking impropriety into account, as well as the optimization of the
transmission parameters subject to these constraints.
Interference management in wireless communications with realistic assumptions on the
knowledge of channel state information at the transmitters

This project aims at developing efficient transmission and reception
schemes for interferencelimited multiuser communication networks
with realistic assumptions on the availability of channel state
information (CSI) at the transmitter side. Extensive research on the
capacity limits of interferencelimited networks in the last several
years has proved that knowledge of CSI at the transmitters (CSIT)
can provide considerable performance improvements. Indeed, for
many networks it has been shown that at high SNR the optimal
performance can be achieved by exploiting the CSIT using a method
called interference alignment (IA). However, the global CSI of the
network must be available at each transmitter to perform IA in a multiuser
network. In this proposal, we consider more realistic IA scenarios
where the transmitters only have access to a limited amount of CSI.
The goal is to develop schemes that optimally exploit whatever CSI is
available to improve spectral efficiency.
In our first scenario, we
assume that the transmitters have access to delayed CSI, which
means that the available information is completely outdated with
respect to the current channel. It has been previously shown that
using only delayed CSI it is still possible to achieve multiplexing gains
over the case where no CSI is available. However, many of the
solutions in the delayedCSI problem are based on heuristic methods.
We are looking for unified and compact solutions that can be applied
to general settings.
The second scenario is the case where some
transmitters do not create interference for some receivers. This
situation is common in the real world where some devices are at
distant locations. In partially connected networks, it has been
previously shown that IA can be effective even without CSI at the
transmitter. Our objective here is to improve the IAbased schemes by
exploiting network properties. For example, multiantenna nodes at
the transceivers can receive or transmit in reduceddimensional
subspaces while creating a virtual partially connected network.
Finally, we investigate several realistic scenarios and adapt our solutions to
account for practical impairments. We study joint application of
different IA methods as multiple conditions coexist in reality.
A unifying framework for detecting cyclostationarity with applications to cognitive radio

Cyclostationary (CS) signals can model periodic phenomena
occurring in a wide range of areas in science and technology. The detection of CS signals is a
particularly important problem. For instance, detection of CS signals is a key
ingredient in the dynamic spectrum management of cognitive radio
(CR), where cognitive users are allowed to access unused licensed
bands. This requires testing for the presence of licensed users, which
transmit CS signals. Because detection of cyclostationarity is such an
important problem, many detectors have been proposed for it.
However, a close analysis of these detectors reveals that most of the
proposed techniques are not based on sound statistical theory. While
they may be sensible adhoc detectors, they do not offer any kind of
optimality. Moreover, most of these detectors make unrealistic
assumptions such as known cycle period or not accounting for the
fact that sampled CS signals are generally only almost CS. Thus, the main objective of this project is the
development of detectors for cyclostationarity
based on solid statistical arguments, without the need for unrealistic
simplifying assumptions. These detectors will be based on wellestablished
statistical techniques, such as the generalized likelihood ratio test, the
locally most powerful invariant test, and  if it exists  the uniformly
most powerful invariant test. We also consider the case of unknown
cycle period. Since the sampling of a
cyclostationary signal generally results in an almost cyclostationary
signal this will involve the development of tests for almost cyclostationarity.
Signal processing for identifying coupled effects in highdimensional data

This project is concerned with identifying coupled effects in highdimensional data. The goal is to extract only
a few modes that explain much of the joint variability between two or more sets of data. This is a common
objective, with numerous applications in many areas of the natural and social sciences and engineering. In this
project, we deal with highdimensional problems with extremely low sample support, where the coupling must be
identified from very few measurements. In such a small sample scenario, only very few, dominant, modes are
trustworthy, as the remaining modes are due to the spurious effects of noise or sample variability. This
requires the right tradeoff between bias and variance. Too simple a model is a poor representation of the data,
causing large bias. Too complicated a model overfits the data, causing large variance. We need to strike the
right balance between underfitting and overfitting. Getting this balance right is the problem of modelorder
selection. Modelorder selection for single data sets, where a single model order needs to be determined and
sample support is sufficient, is a wellstudied problem. There are also techniques that work for small sample
support and techniques that work for two or more datasets. However, the combination  two or more datasets with
small sample support  is still a very challenging open problem.
Medical image processing of Xray images

We assisted one of the world's leading medical technology companies in their development of a revolutionary
product in computer assisted surgery (CAS). CAS has been successfully used in brain and spine surgeries for
many years. However, classical CAS systems have some disadvantages (changed operating room workflow and
increased complexity)
so that they are rarely used in trauma surgery, where efficiency and simplicity are critical. The new product is
completely
adaptive and requires little to no interaction between surgeon and system. In order to achieve this, the system
automatically
processes Xray images using sophisticated image processing algorithms, some of which we contributed.
Nonparametric techniques for analyzing directional
structure in spacetime random fields

The analysis of directional structure in images and spacetime data is crucial to many applications
since onedimensional patterns often correspond to important features such as object contours or
trajectories. Estimating the orientation of such patterns is particularly important. For example, Synthetic
Aperture Radar (SAR) images are frequently analyzed for structures such as oceanic waves, hurricane rain bands,
tsunamis, etc., which all exhibit locally unidirectional structure. Other examples include the analysis of
texture
and optical flow by estimating multidimensional orientation, the efficient coding of local
differential structures in images, and the analysis of superimposed directional patterns, which
may occur in Xray projection imaging. Much of the existing work on orientation estimation treats deterministic
data, which means that it may fail for very noisy data. On the other hand, much of the work on random fields has
focused on the isotropic case. In this project, we develop statistical techniques for analyzing directional
(i.e., highly nonisotropic) structure in random images and spacetime random fields.
Iterative subspace expansions for spacetime
adaptive wireless communications, radar and sonar (completed)

This project addresses the fundamental challenge of high
computational complexity in receivers for bandwidthefficient, high
datarate wireless communications, radar and sonar. We study
conjugate gradient Wiener filters that enable effective low
complexity approximation if the signal correlation matrix has a small
number of distinct or clustered eigenvalues. In this case, these
filters converge with warp speed. Hence, by properly designing signal
transmissions in multiaccess communications, we can build receivers
that rapidly adapt. A critical question is the sensitivity of
repeated/clustered eigenvalues to perturbations, e.g., in
timevarying communication channels. We analyse this question from
the point of view of feedback control.