I mainly focus on statistical and adaptive signal processing techniques and their application to next generation communication systems. For my PhD thesis, I worked on the application of adaptive filtering techniques to cyclostationarity based spectrum sensing in cognitive radios. Right now, my research is focused on the following three broad themes.
The use of large MIMO systems promises increased spectral efficiency and data throughput. However, all the benefits offered by large MIMO systems depend on the availability of accurate CSI at both the transmitter and the receiver. This may not always hold true due to various physical impairments such as pilot contamination, channel estimation errors, channel aging, and reciprocity imperfections. Channel aging refers to the mismatch in the true and the acquired channel states due to the passage of time between training and actual data transmission. To understand the performance trade-offs in large dimensional MIMO systems due to channel aging we first characterized the effects of channel aging on point to point large MIMO systems, and show that aging limits the usable system dimensions. We have then extended this analysis to multiuser (massive) MIMO systems, and the effect of channel aging on the uplink and downlink performance of an FDD massive MIMO system. Using deterministic equivalent analysis, we have derived lower bounds on the achievable uplink and downlink spectral efficiencies, and shown that the effect of channel aging can be somewhat limited by optimally choosing the data transmission frame duration. Following this, we study the effect of channel aging on the achievable rate of time TDD massive MIMO systems using non orthogonal pilots and successive interference cancellation (SIC) at the receivers. The main contribution of this work is to show that the impact of channel aging depends heavily on the underlying protocol, and is often difficult predict using intuition . We also looked at the investigate the channel aging problem in the context of users moving at different velocities and conclude that not all users need to be retrained during each frame, and the slow moving users can be trained less often as compared to the fast moving ones. We also developed optimal and adaptive data aided channel tracking algorithms to mitigate the effect of aging in massive MIMO systems. Most of this work was done in a collaboration with Prof Chandra R Murthy and Prof Kumar Appaiah.
The canonical cell free massive MIMO (CF-mMIMO) setup employs several access points (APs), each equipped with a small antenna array or mostly a single antenna, distributed over a large area and simultaneously serving a large number of users over the same time frequency resource. All the APs are connected to a central processing unit (CPU) via a backbone link. Another interesting aspect of CF-mMIMO systems is the proximity between the users and the serving APs, increasing the probability of having a line of sight (LoS) link between the two. Most of the existing literature on CF-mMIMO systems assumes either Rayleigh fading channels with an LoS link existing with probability 0 or Rician fading channels with an LoS link existing with probability 1. However, practically the channel to a UE could be a “mixture” of LoS and NLoS channels, with the probability of encountering an LoS channel being dependent on the distance between the AP and the UE, as well as the on the local terrain, an aspect that has not been studied well in the literature. In this direction, my initial work was with my collegue, Dr. Sudarshan Mukherjee. Following up on this, Rishi, my BTP student from the 2021 batch analyzed the CSI requirements of such a system. Ashish, my first PhD student, chose this as the topic for his PhD thesis and has published follow up works related to power control, and feedback bit allocation . Another related work discussing the performance of canonical massive MIMO systems enabled by repeater swarms under the mixed LoS/NLoS channel model is currently under preparation with Dr. Anubhab Chowdhury and Prof Erik G Larsson. In the near future I plan to use insights gained from my past work on cellular and cell free massive MIMO systems to better characterize the time evolution of the mixed LoS/ NLoS channel, evaluating the effect of channel aging on CF-mMIMO systems and Repeater Swarm Enabled massive MIMO under mixed LoS/ NLoS channels.
Recently, the design of joint radar and communica tions (JRC) systems, especially that of jointly designed com munication and sensing (JCAS) systems has become an active area of research. This is due to the improved spectral efficiencies and hardware costs offered by these systems. These systems are seen as key enablers for the paradigm of intelligent transportation systems (ITS), popularly known as smart vehicles. In general, there exist three approaches towards the design of JRC systems, viz. coexistence, cooperation and co-design. It is apparent that while the co-design based approach optimizes the performances of both the underlying subsystems it is unsuited for most legacy hardware and necessitates a hard reboot the system architecture.
On the other hand, the coexistence based approach is the least disruptive approach for legacy hardware, therefore, my research focus lies in coexistance based design of a JRC system. Aparna, my first PhD student who has recently defended her thesis has worked on the performance analysis of coexistence between a massive MIMO system and a MIMO radar; and a CF-mMIMO communication system and a tracking radar. This has led us to the further analyze the problem coexistence among multiple radars, that is, the problem of multiple interfering radars trying to detect/ track a set of targets, unaware of each other’s presence. Sairaj, a former BTP student has derived some interesting initial results in this direction. I am activly working with Prof Bhavani Shankar, on this problem.
Recently, a significant amount of research activity in the broad domain of wireless communications has been directed towards the physical layer aspects of Federated learning (FL). This is due to the ubiquitous nature of the federated learning problems. The classical federated learning structure consists of a central server that aims to learn features of data, and user devices which contain data implicitly highlighting its distributed nature. This distributed nature of the training data results in two issues from a communications perspective, viz. the communications overhead required to transmit all this data to the central server, and concerns about the preservation user data privacy. The FL paradigm addresses these concerns via transmit ting the specific features of the data that suffice to train the central model. In such cases, each user device maintains its own local model and locally computes the gradient of the loss function with respect to the variable parameters of the model (for example, the weights and biases in a neural network) and sends this gradient vector to the parameter server. The central server collects these gradients from the user devices to update to the central model. This implies that means that the central model represents the data characteristics exhibited by all user devices combined. Finally, the central server shares its model parameters with the user devices at regular intervals at which point the user devices adopt the central model and start applying updates to the new model. This process is repeated until the central model converges.
However, this involves the cooperation of users in the training phase, thereby requiring the utilisation of local resources. In addition, over the air computation harnesses the superposition property of wireless fading MACs, and consequently, it becomes important to look at this problem from a physical layer perspective. Some recent works have started addressing this problem. However, much more work is required before concrete performance guarantees can be developed for such systems. My postdoctoral work with Prof Chandra R Murthy and Dr. Ramesh Annavajjala on physical layer data fusion for wireless sensor networks closely ties with this body of work . My previous UG studnets, Amit, Diya and Arya have done some preliminary analysis in this context and have come up with some exciting results.