We propose a neural network architecture that enables non-parametric calibration and generation of multivariate extreme value distributions (MEVs). MEVs arise from Extreme Value Theory (EVT) as the necessary class of models when extrapolating a distributional fit over large spatial and temporal scales based on data observed in intermediate scales. In turn, EVT dictates that $d$-max-decreasing, a stronger form of convexity, is an essential shape constraint in the characterization of MEVs. As far as we know, our proposed architecture provides the first class of non-parametric estimators for MEVs that preserve these essential shape constraints. We show that the architecture approximates the dependence structure encoded by MEVs at parametric rate. Moreover, we present a new method for sampling high-dimensional MEVs using a generative model. We demonstrate our methodology on a wide range of experimental settings, ranging from environmental sciences to financial mathematics and verify that the structural properties of MEVs are retained compared to existing methods.

It has been conjectured that the Fisher divergence is more robust to model uncertainty than the conventional Kullback-Leibler (KL) divergence. This motivates the design of a new class of robust generative auto-encoders (AE) referred to as Fisher auto-encoders. Our approach is to design Fisher AEs by minimizing the Fisher divergence between the intractable joint distribution of observed data and latent variables, with that of the postulated/modeled joint distribution. In contrast to KL-based variational AEs (VAEs), the Fisher AE can exactly quantify the distance between the true and the model-based posterior distributions. Qualitative and quantitative results are provided on both MNIST and celebA datasets demonstrating the competitive performance of Fisher AEs in terms of robustness compared to other AEs such as VAEs and Wasserstein AEs.

We propose a new generative modeling technique for learning multidimensional cumulative distribution functions (CDFs) in the form of copulas. Specifically, we consider certain classes of copulas known as Archimedean and hierarchical Archimedean copulas, popular for their parsimonious representation and ability to model different tail dependencies. We consider their representation as mixture models with Laplace transforms of latent random variables from generative neural networks. This alternative representation allows for computational efficiencies and easy sampling, especially in high dimensions. We describe multiple methods for optimizing the network parameters. Finally, we present empirical results that demonstrate the efficacy of our proposed method in learning multidimensional CDFs and its computational efficiency compared to existing methods.

This paper carries out a large dimensional analysis of a variation of kernel ridge regression that we call centered kernel ridge regression (CKRR), also known in the literature as kernel ridge regression with offset. This modified technique is obtained by accounting for the bias in the regression problem resulting in the old kernel ridge regression but with centered kernels. The analysis is carried out under the assumption that the data is drawn from a Gaussian distribution and heavily relies on tools from random matrix theory (RMT). Under the regime in which the data dimension and the training size grow infinitely large with fixed ratio and under some mild assumptions controlling the data statistics, we show that both the empirical and the prediction risks converge to a deterministic quantities that describe in closed form fashion the performance of CKRR in terms of the data statistics and dimensions. A key insight of the proposed analysis is the fact that asymptotically a large class of kernels achieve the same minimum prediction risk. This insight is validated with synthetic data.

We investigate random projections in the context of randomly projected linear discriminant analysis (LDA). We consider the case in which the data of dimension p is randomly projected onto a lower dimensional space before being fed to the classifier. Using fundamental results from random matrix theory and relying on some mild assumptions, we show that the asymptotic performance in terms of probability of misclassification approaches a deterministic quantity that only depends on the data statistics and the dimensions involved. Such results permits to reliably predict the performance of projected LDA as a function of the reduced dimension d <; p and thus helps to determine the minimum d to achieve a certain desired performance. Finally, we validate our results with finite-sample settings drawn from both synthetic data and the popular MNIST dataset.

This paper focuses on studying the performance of general regularized discriminant analysis (RDA) classifiers based on the Gaussian mixture model with different means and covariances. RDA offers a rich class of regularization options, covering as special cases the regularized linear discriminant analysis (RLDA) and the regularized quadratic discriminant analysis (RQDA) classifiers. Based on fundamental results from random matrix theory, we analyze RDA under the double asymptotic regime where the data dimension and the training size both increase in a proportional way. Under the double asymptotic regime and some mild assumptions, we show that the asymptotic classification error converges to a deterministic quantity that only depends on the data statistical parameters and dimensions. This result can be leveraged to select the optimal parameters that minimize the classification error, thus yielding the optimal classifier. Numerical results are provided to validate our theoretical findings on synthetic data showing high accuracy of our derivations.

This paper considers the problem of selecting a set of k measurements from n available sensor observations. The selected measurements should minimize a certain error function assessing the error in estimating a certain m dimensional parameter vector. The exhaustive search inspecting each of the (n) possible choices would require very high computational k complexity and as such is not practical for large n and k. Alternative methods with low complexity have recently been investigated but their main drawbacks are that they require perfect knowledge of the measurement matrix and they need to be applied at the pace of change of the measurement matrix. To overcome these issues, we consider the asymptotic regime in which k, n, and m grow large at the same pace. Tools from random matrix theory are then used to approximate in closed-form the most important error measures that are commonly used. The asymptotic approximations are then leveraged to properly select k measurements exhibiting low values for the asymptotic error measures. Two heuristic algorithms are proposed. The first one merely consists in applying the convex optimization artifice to the asymptotic error measure. The second algorithm is a low-complexity greedy algorithm that attempts to look for a sufficiently good solution for the original minimization problem. The greedy algorithm can be applied to both the exact and the asymptotic error measures and can be thus implemented in blind and channel-aware fashions. We present two potential applications where the proposed algorithms can be used, namely, antenna selection for uplink transmissions in large scale multiuser systems and sensor selection for wireless sensor networks. Numerical results are also presented and sustain the efficiency of the proposed blind methods in reaching the performances of channel-aware algorithms.

This letter focuses on the computation of the positive moments of one-side correlated random Gram matrices. Closed-form expressions for the moments can be obtained easily, but numerical evaluation thereof is prone to numerical stability, especially in high-dimensional settings. This letter provides a numerically stable method that efficiently computes the positive moments in closed form. The developed expressions are more accurate and can lead to higher accuracy levels when fed to moment-based approaches. As an application, we show how the obtained moments can be used to approximate the marginal distribution of the eigenvalues of random Gram matrices.

In this paper, we derive a closed-form expression for the inverse moments of one sided-correlated random Gram matrices. Such a question is mainly motivated by applications in signal processing and wireless communications for which evaluating this quantity is a question of major interest. This is for instance the case of the best linear unbiased estimator, in which the average estimation error corresponds to the first inverse moment of a random Gram matrix.

This paper presents a comprehensive performance analysis of full-duplex multiuser relay networks employing opportunistic scheduling with noisy and compressive feedback. Specifically, two feedback techniques based on compressive sensing (CS) theory are introduced and their effect on the system performance is analyzed. The problem of joint user identity and signal-to-noise ratio (SNR) estimation at the base-station is casted as a block sparse signal recovery problem in CS. Using existing CS block recovery algorithms, the identity of the strong users is obtained and their corresponding SNRs are estimated using the best linear unbiased estimator (BLUE). To minimize the effect of feedback noise on the estimated SNRs, a backoff strategy that optimally backsoff on the noisy estimated SNRs is introduced, and the error covariance matrix of the noise after CS recovery is derived. Finally, closed-form expressions for the end-to-end SNRs of the system are derived. Numerical results show that the proposed techniques drastically reduce the feedback air-time and achieve a rate close to that obtained by scheduling techniques that require dedicated error-free feedback from all network users. Key findings of this paper suggest that the choice of half-duplex or full-duplex SNR feedback is dependent on the channel coherence interval, and on low coherence intervals, full-duplex feedback is superior to the interference-free half-duplex feedback.

We propose a distributed user selection strategy in a network MIMO setting with M base stations serving K users. Each base station is equipped with L antennas, where LM ≪ K. The conventional selection strategy is based on a well known technique called semi-orthogonal user selection when the zero-forcing beamforming (ZFBF) is adopted. Such technique, however, requires perfect channel state information at the transmitter (CSIT), which might not be available or need large feedback overhead. This paper proposes an alternative distributed user selection technique where each user sets a timer that is inversely proportional to his channel quality indicator (CQI), as a means to reduce the feedback overhead. The proposed strategy allows only the user with the highest CQI to respond with a feedback. Such technique, however, remains collision free only if the transmission time is shorter than the difference between the strongest user timer and the second strongest user timer. To overcome the situation of longer transmission times, the paper proposes another feedback strategy that is based on the theory of compressive sensing, where collision is allowed and all users encode their feedback information and send it back to the base-stations simultaneously. The paper shows that the problem can be formulated as a block sparse recovery problem which is agnostic on the transmission time, which makes it a good alternative to the timer approach when collision is dominant.

Relay selection is a simple technique that achieves spatial diversity in cooperative relay networks. Nonetheless, relay selection algorithms generally require error-free channel state information (CSI) from all cooperating relays. Practically, CSI acquisition generates a great deal of feedback overhead that could result in significant transmission delays. In addition to this, the fed back channel information is usually corrupted by additive noise. This could lead to transmission outages if the central node selects the set of cooperating relays based on inaccurate feedback information. In this paper, we propose a relay selection algorithm that tackles the above challenges. Instead of allocating each relay a dedicated channel for feedback, all relays share a pool of feedback channels. Following that, each relay feeds back its identity only if its effective channel (source-relay-destination) exceeds a threshold. After deriving closed-form expressions for the feedback load and the achievable rate, we show that the proposed algorithm drastically reduces the feedback overhead and achieves a rate close to that obtained by selection algorithms with dedicated error-free feedback from all relays.

Relay selection is a simple technique that achieves spatial diversity in cooperative relay networks. Generally, relay selection algorithms require channel state information (CSI) feedback from all cooperating relays to make a selection decision. This requirement poses two important challenges, which are often neglected in the literature. Firstly, the fed back channel information is usually corrupted by additive noise. Secondly, CSI feedback generates a great deal of feedback overhead (air-time) that could result in significant performance hits. In this paper, we propose a compressive sensing (CS) based relay selection algorithm that reduces the feedback overhead of relay networks under the assumption of noisy feedback channels. The proposed algorithm exploits CS to first obtain the identity of a set of relays with favorable channel conditions. Following that, the CSI of the identified relays is estimated using least squares estimation without any additional feedback. Both single and multiple relay selection cases are considered. After deriving closed-form expressions for the asymptotic end-to-end SNR at the destination and the feedback load for different relaying protocols, we show that CS-based selection drastically reduces the feedback load and achieves a rate close to that obtained by selection algorithms with dedicated error-free feedback.

Relay selection is a simple technique that achieves spatial diversity in cooperative relay networks. However, for relay selection algorithms to make a selection decision, channel state information (CSI) from all cooperating relays is usually required at a central node. This requirement poses two important challenges. Firstly, CSI acquisition generates a great deal of feedback overhead (air-time) that could result in significant transmission delays. Secondly, the fed back channel information is usually corrupted by additive noise. This could lead to transmission outages if the central node selects the set of cooperating relays based on inaccurate feedback information. In this paper, we introduce a limited feedback relay selection algorithm for a multicast relay network. The proposed algorithm exploits the theory of compressive sensing to first obtain the identity of the "strong" relays with limited feedback. Following that, the CSI of the selected relays is estimated using linear minimum mean square error estimation. To minimize the effect of noise on the fed back CSI, we introduce a back-off strategy that optimally backs-off on the noisy estimated CSI. For a fixed group size, we provide closed form expressions for the scaling law of the maximum equivalent SNR for both Decode and Forward (DF) and Amplify and Forward (AF) cases. Numerical results show that the proposed algorithm drastically reduces the feedback air-time and achieves a rate close to that obtained by selection algorithms with dedicated error-free feedback channels.