TY - JOUR
AB - We extend the proof of the local semicircle law for generalized Wigner matrices given in MR3068390 to the case when the matrix of variances has an eigenvalue -1. In particular, this result provides a short proof of the optimal local Marchenko-Pastur law at the hard edge (i.e. around zero) for sample covariance matrices X*X, where the variances of the entries of X may vary.
AU - Ajanki, Oskari H
AU - Erdös, László
AU - Krüger, Torben H
ID - 2179
JF - Electronic Communications in Probability
TI - Local semicircle law with imprimitive variance matrix
VL - 19
ER -
TY - JOUR
AB - Weighted majority votes allow one to combine the output of several classifiers or voters. MinCq is a recent algorithm for optimizing the weight of each voter based on the minimization of a theoretical bound over the risk of the vote with elegant PAC-Bayesian generalization guarantees. However, while it has demonstrated good performance when combining weak classifiers, MinCq cannot make use of the useful a priori knowledge that one may have when using a mixture of weak and strong voters. In this paper, we propose P-MinCq, an extension of MinCq that can incorporate such knowledge in the form of a constraint over the distribution of the weights, along with general proofs of convergence that stand in the sample compression setting for data-dependent voters. The approach is applied to a vote of k-NN classifiers with a specific modeling of the voters' performance. P-MinCq significantly outperforms the classic k-NN classifier, a symmetric NN and MinCq using the same voters. We show that it is also competitive with LMNN, a popular metric learning algorithm, and that combining both approaches further reduces the error.
AU - Bellet, Aurélien
AU - Habrard, Amaury
AU - Morvant, Emilie
AU - Sebban, Marc
ID - 2180
IS - 1-2
JF - Machine Learning
TI - Learning a priori constrained weighted majority votes
VL - 97
ER -
TY - JOUR
AB - We describe a simple adaptive network of coupled chaotic maps. The network reaches a stationary state (frozen topology) for all values of the coupling parameter, although the dynamics of the maps at the nodes of the network can be nontrivial. The structure of the network shows interesting hierarchical properties and in certain parameter regions the dynamics is polysynchronous: Nodes can be divided in differently synchronized classes but, contrary to cluster synchronization, nodes in the same class need not be connected to each other. These complicated synchrony patterns have been conjectured to play roles in systems biology and circuits. The adaptive system we study describes ways whereby this behavior can evolve from undifferentiated nodes.
AU - Botella Soler, Vicente
AU - Glendinning, Paul
ID - 2183
IS - 6
JF - Physical Review E Statistical Nonlinear and Soft Matter Physics
TI - Hierarchy and polysynchrony in an adaptive network
VL - 89
ER -
TY - JOUR
AB - Given topological spaces X,Y, a fundamental problem of algebraic topology is understanding the structure of all continuous maps X→ Y. We consider a computational version, where X,Y are given as finite simplicial complexes, and the goal is to compute [X,Y], that is, all homotopy classes of suchmaps.We solve this problem in the stable range, where for some d ≥ 2, we have dim X ≤ 2d-2 and Y is (d-1)-connected; in particular, Y can be the d-dimensional sphere Sd. The algorithm combines classical tools and ideas from homotopy theory (obstruction theory, Postnikov systems, and simplicial sets) with algorithmic tools from effective algebraic topology (locally effective simplicial sets and objects with effective homology). In contrast, [X,Y] is known to be uncomputable for general X,Y, since for X = S1 it includes a well known undecidable problem: testing triviality of the fundamental group of Y. In follow-up papers, the algorithm is shown to run in polynomial time for d fixed, and extended to other problems, such as the extension problem, where we are given a subspace A ⊂ X and a map A→ Y and ask whether it extends to a map X → Y, or computing the Z2-index-everything in the stable range. Outside the stable range, the extension problem is undecidable.
AU - Čadek, Martin
AU - Krcál, Marek
AU - Matoušek, Jiří
AU - Sergeraert, Francis
AU - Vokřínek, Lukáš
AU - Wagner, Uli
ID - 2184
IS - 3
JF - Journal of the ACM
TI - Computing all maps into a sphere
VL - 61
ER -
TY - CONF
AB - We revisit the classical problem of converting an imperfect source of randomness into a usable cryptographic key. Assume that we have some cryptographic application P that expects a uniformly random m-bit key R and ensures that the best attack (in some complexity class) against P(R) has success probability at most δ. Our goal is to design a key-derivation function (KDF) h that converts any random source X of min-entropy k into a sufficiently "good" key h(X), guaranteeing that P(h(X)) has comparable security δ′ which is 'close' to δ. Seeded randomness extractors provide a generic way to solve this problem for all applications P, with resulting security δ′ = O(δ), provided that we start with entropy k ≥ m + 2 log (1/δ) - O(1). By a result of Radhakrishnan and Ta-Shma, this bound on k (called the "RT-bound") is also known to be tight in general. Unfortunately, in many situations the loss of 2 log (1/δ) bits of entropy is unacceptable. This motivates the study KDFs with less entropy waste by placing some restrictions on the source X or the application P. In this work we obtain the following new positive and negative results in this regard: - Efficient samplability of the source X does not help beat the RT-bound for general applications. This resolves the SRT (samplable RT) conjecture of Dachman-Soled et al. [DGKM12] in the affirmative, and also shows that the existence of computationally-secure extractors beating the RT-bound implies the existence of one-way functions. - We continue in the line of work initiated by Barak et al. [BDK+11] and construct new information-theoretic KDFs which beat the RT-bound for large but restricted classes of applications. Specifically, we design efficient KDFs that work for all unpredictability applications P (e.g., signatures, MACs, one-way functions, etc.) and can either: (1) extract all of the entropy k = m with a very modest security loss δ′ = O(δ·log (1/δ)), or alternatively, (2) achieve essentially optimal security δ′ = O(δ) with a very modest entropy loss k ≥ m + loglog (1/δ). In comparison, the best prior results from [BDK+11] for this class of applications would only guarantee δ′ = O(√δ) when k = m, and would need k ≥ m + log (1/δ) to get δ′ = O(δ). - The weaker bounds of [BDK+11] hold for a larger class of so-called "square- friendly" applications (which includes all unpredictability, but also some important indistinguishability, applications). Unfortunately, we show that these weaker bounds are tight for the larger class of applications. - We abstract out a clean, information-theoretic notion of (k,δ,δ′)- unpredictability extractors, which guarantee "induced" security δ′ for any δ-secure unpredictability application P, and characterize the parameters achievable for such unpredictability extractors. Of independent interest, we also relate this notion to the previously-known notion of (min-entropy) condensers, and improve the state-of-the-art parameters for such condensers.
AU - Dodis, Yevgeniy
AU - Pietrzak, Krzysztof Z
AU - Wichs, Daniel
ED - Nguyen, Phong
ED - Oswald, Elisabeth
ID - 2185
TI - Key derivation without entropy waste
VL - 8441
ER -