arXiv subjects
Publications and source records attributed to Ruixiang Zhang.
We prove some weighted Fourier restriction estimates using polynomial partitioning and refined Strichartz estimates. As application we obtain improved spherical average decay rates of the Fourier transform of fractal measures, and therefore improve the results for the Falconer distance set conjecture in three and higher dimensions.
Recurrent neural networks (RNNs) have been successfully applied to various natural language processing (NLP) tasks and achieved better results than conventional methods. However, the lack of understanding of the mechanisms behind their effectiveness limits further improvements on their architectures. In this paper, we present a visual analytics method for understanding and comparing RNN models for NLP tasks. We propose a technique to explain the function of individual hidden state units based on their expected response to input texts. We then co-cluster hidden state units and words based on the expected response and visualize co-clustering results as memory chips and word clouds to provide more structured knowledge on RNNs' hidden states. We also propose a glyph-based sequence visualization based on aggregate information to analyze the behavior of an RNN's hidden state at the sentence-level. The usability and effectiveness of our method are demonstrated through case studies and reviews from domain experts.
We present a new proof of the Joints Theorem without taking derivatives. Then we generalize the proof to prove the Multijoints Conjecture and Carbery's generalization. All results are in any dimension over an arbitrary field.
Despite the successes in capturing continuous distributions, the application of generative adversarial networks (GANs) to discrete settings, like natural language tasks, is rather restricted. The fundamental reason is the difficulty of back-propagation through discrete random variables combined with the inherent instability of the GAN training objective. To address these problems, we propose Maximum-Likelihood Augmented Discrete Generative Adversarial Networks. Instead of directly optimizing the GAN objective, we derive a novel and low-variance objective using the discriminator's output that follows corresponds to the log-likelihood. Compared with the original, the new objective is proved to be consistent in theory and beneficial in practice. The experimental results on various discrete datasets demonstrate the effectiveness of the proposed approach.
In [Sharir and Solomon 2015], Sharir and Solomon showed that the number of incidences between $m$ distinct points and $n$ distinct lines in $\mathbb R^4$ is $$O^*\left(m^{2/5}n^{4/5}+ m^{1/2}n^{1/2}q^{1/4} + m^{2/3}n^{1/3}s^{1/3} + m + n\right),$$ provided that no 2-flat contains more than $s$ lines, and no hyperplane or quadric contains more than $q$ lines, where the $O^*$ hides a multiplicative factor of $2^{c\sqrt {\log m}}$ for some absolute constant $c$. In this paper we prove that, for integers $m,n,$ satisfying $n^{9/8}<m<n^{3/2}$, there exist $m$ points and $n$ lines on the quadratic hypersurface in $\mathbb{R}^4$ $$ \{(x_1,x_2,x_3,x_4)\in \mathbb R^4 \mid x_1 = x_2^2 + x_3^2 - x_4^2\}, $$ such that (i) at most $s=O(1)$ lines lie on any 2-flat, (ii) at most $q=O(n/m^{1/3})$ lines lie on any hyperplane, and (iii) the number of incidences between the points and the lines is $Θ(m^{2/3}n^{1/2})$, which is asymptotically larger than the upper bound by Sharir and Solomon. This shows that the assumption that no quadric contains more than $q$ lines (in the above mentioned theorem of Sharir and Solomon) is necessary in this regime of $m$ and $n$. By a suitable projection from this quadratic hypersurface onto $\mathbb{R}^3$, we obtain $m$ points and $n$ lines in $\mathbb{R}^3$, with at most $s=O(1)$ lines on a common plane, such that the number of incidences between the $m$ points and the $n$ lines is $Θ(m^{2/3}n^{1/2})$. It remains an interesting question to determine if this bound is also tight in general.
In this article we estimate the number of integers up to $X$ which can be represented by a positive-definite, binary integral quadratic form of discriminant which is small relative to $X$. This follows from understanding the vector of signs when computing the Legendre symbol of small integers $n$ at multiple primes.
We prove the folklore endpoint multilinear $k_j$-plane conjecture originated from the paper \cite{bennett2006multilinear} of Bennett, Carbery and Tao. Along the way we prove a more general result, namely the endpoint multilinear $k_j$-variety theorem. Finally, we generalize our results to the endpoint perturbed Beascamp-Lieb inequality using techniques in this paper.
We prove new bounds on the number of incidences between points and higher degree algebraic curves. The key ingredient is an improved initial bound, which is valid for all fields. Then we apply the polynomial method to obtain global bounds on $\mathbb{R}$ and $\mathbb{C}$.
We prove the discrete analogue of Kakeya conjecture over $\mathbb{R}^n$. This result suggests that a (hypothetically) low dimensional Kakeya set cannot be constructed directly from discrete configurations. We also prove a generalization which completely solves the discrete analogue of the Furstenberg set problem in all dimensions. The difference between our theorems and the (true) problems is only the (still difficult) issue of continuity since no transversality-at-incidences assumptions are imposed. The main tool of the proof is a theorem of Wongkew \cite{wongkew2003volumes} which states that a low degree polynomial cannot have its zero set being too dense inside the unit cube, coupled with Dvir-type polynomial arguments \cite{dvir2009size}. From the viewpoint of the proofs, we also state a conjecture that is stronger than and almost equivalent to the (lower) Minkowski version of the Kakeya conjecture and prove some results towards it. We also present our own version of the proof of the theorem in \cite{wongkew2003volumes}. Our proof shows that this theorem follows from a combination of properties of zero sets of polynomials and a general proposition about hypersurfaces which might be of independent interest. Finally, we discuss how to generalize Bourgain's conjecture to high dimensions, which is closely related to the theme here.
In this paper, we consider the so-called "Furstenberg set problem" in high dimensions. First, following Wolff's work on the two dimensional real case, we provide "reasonable" upper bounds for the problem for $\mathbb{R}$ or $\mathbb{F}_p$. Next we study the "critical" case and improve the "trivial" exponent by $Ω(\frac{1}{n^2})$ for $\mathbb{F}_p^n$. Our key tool to obtain this lower bound is a theorem about how things behave when the Loomis-Whitney inequality is nearly sharp, as it helps us to reduce the problem down to dimension two.
We show that for a real polynomial of degree $n$ in two variables $x$ and $y$, any local "sharp turn" must have its "size" $\gtrsim e^{-Cn^{2}}$. We also show that there is indeed an example that has a sharp turn of size $\lesssim e^{-Cn}$. This gives a quite satisfactory answer to a problem raised by Guth. The problem was inspired by applications of the polynomial method in the study of Kakeya conjecture.