arXiv subjects

Carl Nuzman

Publications and source records attributed to Carl Nuzman.

5 recordsLinked to original sources

arXiv2025

Demonstrating Interoperable Channel State Feedback Compression with Machine Learning

Neural network-based compression and decompression of channel state feedback has been one of the most widely studied applications of machine learning (ML) in wireless networks. Various simulation-based studies have shown that ML-based feedback compression can result in reduced overhead and more accurate channel information. However, to the best of our knowledge, there are no real-life proofs of concepts demonstrating the benefits of ML-based channel feedback compression in a practical setting, where the user equipment (UE) and base station have no access to each others' ML models. In this paper, we present a novel approach for training interoperable compression and decompression ML models in a confidential manner, and demonstrate the accuracy of the ensuing models using prototype UEs and base stations. The performance of the ML-based channel feedback is measured both in terms of the accuracy of the reconstructed channel information and achieved downlink throughput gains when using the channel information for beamforming. The reported measurement results demonstrate that it is possible to develop an accurate ML-based channel feedback link without having to share ML models between device and network vendors. These results pave the way for a practical implementation of ML-based channel feedback in commercial 6G networks.

eess.SP↗

arXiv2025

Energy-Efficient Flat Precoding for MIMO Systems

This paper addresses the suboptimal energy efficiency of conventional digital precoding schemes in multiple-input multiple-output (MIMO) systems. Through an analysis of the power amplifier (PA) output power distribution associated with conventional precoders, it is observed that these power distributions can be quite uneven, resulting in large PA backoff (thus low efficiency) and high power consumption. To tackle this issue, we propose a novel approach called flat precoding, which aims to control the flatness of the power distribution within a desired interval. In addition to reducing PA power consumption, flat precoding offers the advantage of requiring smaller saturation levels for PAs, which reduces the size of PAs and lowers the cost. To incorporate the concept of flat power distribution into precoding design, we introduce a new lower-bound per-antenna power constraint alongside the conventional sum power constraint and the upper-bound per-antenna power constraint. By adjusting the lower-bound and upper-bound values, we can effectively control the level of flatness in the power distribution. We then seek to find a flat precoder that satisfies these three sets of constraints while maximizing the weighted sum rate (WSR). In particular, we develop efficient algorithms to design weighted minimum mean squared error (WMMSE) and zero-forcing (ZF)-type precoders with controllable flatness features that maximize WSR. Numerical results demonstrate that complete flat precoding approaches, where the power distribution is a straight line, achieve the best trade-off between spectral efficiency and energy efficiency for existing PA technologies. We also show that the proposed ZF and WMMSE precoding methods can approach the performance of their conventional counterparts with only the sum power constraint, while significantly reducing PA size and power consumption.

cs.IT↗

arXiv2023

On convergence and optimality of maximum-likelihood APA

Affine projection algorithm (APA) is a well-known algorithm in adaptive filtering applications such as audio echo cancellation. APA relies on three parameters: $P$ (projection order), $μ$ (step size) and $δ$ (regularization parameter). It is known that running APA for a fixed set of parameters leads to a tradeoff between convergence speed and accuracy. Therefore, various methods for adaptively setting the parameters have been proposed in the literature. Inspired by maximum likelihood (ML) estimation, we derive a new ML-based approach for adaptively setting the parameters of APA, which we refer to as ML-APA. For memoryless Gaussian inputs, we fully characterize the expected misalignment error of ML-APA as a function of iteration number and show that it converges to zero as $O({1\over t})$. We further prove that the achieved error is asymptotically optimal. ML-APA updates its estimate once every $P$ samples. We also propose incremental ML-APA (IML-APA), which updates the coefficients at every time step and outperforms ML-APA in our simulations results. Our simulation results verify the analytical conclusions for memoryless inputs and show that the new algorithms also perform well for strongly correlated input signals.

cs.IT↗

arXiv2019

Efficient Deep Learning of GMMs

We show that a collection of Gaussian mixture models (GMMs) in $R^{n}$ can be optimally classified using $O(n)$ neurons in a neural network with two hidden layers (deep neural network), whereas in contrast, a neural network with a single hidden layer (shallow neural network) would require at least $O(\exp(n))$ neurons or possibly exponentially large coefficients. Given the universality of the Gaussian distribution in the feature spaces of data, e.g., in speech, image and text, our result sheds light on the observed efficiency of deep neural networks in practical classification problems.

cs.LG↗

arXiv2004

Large deviation asymptotics for occupancy problems

In the standard formulation of the occupancy problem one considers the distribution of r balls in n cells, with each ball assigned independently to a given cell with probability 1/n. Although closed form expressions can be given for the distribution of various interesting quantities (such as the fraction of cells that contain a given number of balls), these expressions are often of limited practical use. Approximations provide an attractive alternative, and in the present paper we consider a large deviation approximation as r and n tend to infinity. In order to analyze the problem we first consider a dynamical model, where the balls are placed in the cells sequentially and ``time'' corresponds to the number of balls that have already been thrown. A complete large deviation analysis of this ``process level'' problem is carried out, and the rate function for the original problem is then obtained via the contraction principle. The variational problem that characterizes this rate function is analyzed, and a fairly complete and explicit solution is obtained. The minimizing trajectories and minimal cost are identified up to two constants, and the constants are characterized as the unique solution to an elementary fixed point problem. These results are then used to solve a number of interesting problems, including an overflow problem and the partial coupon collector's problem.

math.PR↗