arXiv subjects
Publications and source records attributed to Ben Krause.
Let $a_n$ be the random increasing sequence of natural numbers which takes each value independently with probability $n^{-a}$, $0 < a < 1/2$, and let $p(n) = n^{1+ε}$, $0 < ε< 1$. We prove that, almost surely, for every measure-preserving system $(X,T)$ and every $f \in L^1(X)$ the modulated, random averages \[ \frac{1}{N} \sum_{n = 1}^N e(p(n)) T^{a_n(ω)} f\] converge to $0$ pointwise almost everywhere.
We study $L^p(μ) \to L^q(ν)$ mapping properties of the convolution operator $ T_λf(x)=λ*(fμ)(x)$ and of the corresponding maximal operator $ {\mathcal T}_λf(x)=\sup_{t>0} |λ_t*(fμ)(x)|$, where $λ$ is a tempered distribution, and $μ$ and $ν$ are compactly supported measures satisfying the polynomial growth bounds $μ(B(x,r)) \leq Cr^{s_μ}$ and $ν(B(x,r)) \leq Cr^{s_ν}$. As a result, we prove variants of the classical $L^p$-improving (Littman; Strichartz) and maximal (Stein) inequalities in a setting where the Plancherel formula is not available. Connections with the David-Semmes conjecture are also discussed.
Recurrent Neural Networks (RNNs) have long been recognized for their potential to model complex time series. However, it remains to be determined what optimization techniques and recurrent architectures can be used to best realize this potential. The experiments presented take a deep look into Hessian free optimization, a powerful second order optimization method that has shown promising results, but still does not enjoy widespread use. This algorithm was used to train to a number of RNN architectures including standard RNNs, long short-term memory, multiplicative RNNs, and stacked RNNs on the task of character prediction. The insights from these experiments led to the creation of a new multiplicative LSTM hybrid architecture that outperformed both LSTM and multiplicative RNNs. When tested on a larger scale, multiplicative LSTM achieved character level modelling results competitive with the state of the art for RNNs using very different methodology.
In this note we prove the following good-$λ$ inequality, for $r>2$, all $λ> 0$, $δ\in \big(0, \frac{1}{2} \big)$ \[ ν\big\{ V_r(f) > 3 λ; \mathcal{M}(f) \leq δλ\big\} \leq 4 ν\{s(f) > δλ\} + {δ^2 \left(1+\frac{16}{r-2}\right)^2} \cdot ν\big\{ V_r(f) > λ\big\}, \] where $\mathcal{M}(f)$ is the martingale maximal function, $s(f)$ is the conditional martingale square function. This immediately proves that $V_r(f)$ is bounded on $L^p$, $1 < p <\infty$ and moreover is integrable when the maximal function is.
The aim of this paper is to exhibit a wide class of sparse deterministic sets, $\mathbf B \subseteq \mathbb{N}$, so that \[ \limsup_{N \to \infty} N^{-1}|\mathbf B \cap [1,N]|= 0, \] for which the Hardy--Littlewood majorant property holds: \[ \sup_{|a_n|\le 1} \Big\| \sum_{n\in\mathbf B\cap[1, N]} a_n e^{2 πi n ξ}\Big \|_{L^p(\mathbb{T}, {\mathrm d} ξ)} \leq \mathbf{C}_p \Big\| \sum_{n\in\mathbf B\cap[1, N]} e^{2 πi n ξ} \Big\|_{L^p(\mathbb{T}, {\mathrm d} ξ)}, \] where $p \geq p_{\mathbf{B}}$ is sufficiently large, the implicit constant $\mathbf{C}_p$ is independent of $N$, and the supremum is taken over all complex sequences $ (a_n : n \in \mathbb{N})$ such that $|a_n| \leq 1$.
Our aim is to establish the first two-parameter version of Bourgain's maximal logarithmic inequality on $L^2(\mathbb R^2)$ for the rational frequencies. We achieve this by introducing a variant of a two-parameter Rademacher--Menschov inequality. The method allows us to control an oscillation seminorm as well.
We extend the results of Jones, Rosenblatt, and Wierdl concerning higher-dimensional oscillation in ergodic theory in a variety of ways. We do so by transference to the integer lattice, where we employ technique from (discrete) harmonic analysis.
We prove weighted and vector-valued variational estimates for ergodic averages on $\mathbb{R}^d$. The weighted square function estimate relating ergodic averages to the dyadic martingale is obtained using an $\ell^r$ version of a reverse Hölder inequality for variation seminorms.
Let $T$ be an ergodic measure-preserving transformation on a non-atomic probability space $(X,Σ,μ)$. We prove uniform extensions of the Wiener-Wintner theorem in two settings: For averages involving weights coming from Hardy field functions, $p$: \[ \{\frac{1}{N} \sum_{n\leq N} e(p(n)) T^{n}f(x) \} \] and for "twisted" polynomial ergodic averages: \[ \{\frac{1}{N} \sum_{n\leq N} e(n θ) T^{P(n)}f(x) \} \] for certain classes of badly approximable $θ\in [0,1]$. We also give an elementary proof that the above twisted polynomial averages converge pointwise $μ$-a.e. for $f \in L^p(X), \ p >1,$ and arbitrary $θ\in [0,1]$.
For $m \geq 2$, let $(\mathbb{Z}_{m+1}^N, |\cdot|)$ denote the group equipped with the so-called $l^0$ metric, \[ |y| = \left| \big( y(1), \dots, y(N) \big) \right| := | \{1 \leq i \leq N : y(i) \neq 0 \} |,\] and define the $L^1$-normalized indicator of the $r$-sphere, \[ σ_r := \frac{1}{|\{|x| = r\}|} 1_{\{|x| =r\}}.\] We study the $L^p \to L^p$ mapping properties of the maximal operator \[ M^{N} f (x) := \sup_{r \leq N} | σ_r*f| \] acting on functions defined on $\mathbb{Z}_{m+1}^N$. Specifically, we prove that for all $p>1$, there exist absolute constants $C_{m,p}$ so that \[ \| M^{N} f \|_{L^p(\mathbb{Z}_{m+1}^N)} \leq C_{m,p} \| f \|_{L^p(\mathbb{Z}_{m+1}^N)} \] for all $N$.
Let $L^2(X,Σ,μ,τ)$ be a measure-preserving system, with $τ$ a $\mathbb{Z}$-action. In this note, we prove that the ergodic averages along integer-valued polynomials, $P(n)$, \[ M_N(f):= \frac{1}{N}\sum_{n \leq N} τ^{P(n)} f \] converge pointwise for $f \in L^2(X)$. We do so by proving that, for $r>2$, the $r$-variation, $\mathcal{V}^r(M_N(f))$, extends to a bounded operator on $L^2$. We also prove that our result is sharp, in that $\mathcal{V}^2(M_N(f))$ is an unbounded operator on $L^2$.
We use Oberlin, Nazarov, and Thiele's Multi-Frequency Calderón-Zygmund decomposition to lower estimates on maximal multipliers in $L^p$. We also improve on classical multiplier results of Coifman, Rubio de Francia, and Semmes.