arXiv subjects

Lisa Hartung

Publications and source records attributed to Lisa Hartung.

At least 19 recordsLinked to original sources

arXiv2025

Fisher-Hartwig asymptotics for non-Hermitian random matrices

We prove the two-dimensional analogue of the asymptotics for Toeplitz determinants with Fisher-Hartwig singularities, for general real symbols. This formula has applications to random normal matrices with complex spectra: (i) the characteristic polynomial converges to a Gaussian multiplicative chaos random measure on the limiting droplet, in the subcritical phase; (ii) the electric potential converges pointwise to a logarithmically correlated field; (iii) the measure of its level sets (i.e. thick points) is identified; (iv) the associated free energy undergoes a freezing transition. This establishes emergence of the Liouville quantum gravity measure from free fermions in 2d, and universality with respect to the external potential.

math.PR↗

arXiv2025

Extremes of the zero-average Gaussian Free Field on random regular graphs

We study the extreme value statistics of the zero-average Gaussian free field (GFF) on random $r$-regular graphs and the Gaussian free field on $r$-regular trees. For random $r$-regular graphs of diverging size, for every fixed $r\ge3$, we show that the rescaled extremal point process of the field is asymptotically distributed, in the annealed sense, as a Poisson point process on the line with intensity $e^{-x}\,\mathrm{d}x$. The same limit behaviour is obeyed by the restriction of the GFF on $r$-regular trees to finite subsets of vertices. Our approach relies on a direct Gaussian comparison argument and precise Green function estimates.

math.PR↗

arXiv2025

Lift What You Can: Green Online Learning with Heterogeneous Ensembles

Ensemble methods for stream mining necessitate managing multiple models and updating them as data distributions evolve. Considering the calls for more sustainability, established methods are however not sufficiently considerate of ensemble members' computational expenses and instead overly focus on predictive capabilities. To address these challenges and enable green online learning, we propose heterogeneous online ensembles (HEROS). For every training step, HEROS chooses a subset of models from a pool of models initialized with diverse hyperparameter choices under resource constraints to train. We introduce a Markov decision process to theoretically capture the trade-offs between predictive performance and sustainability constraints. Based on this framework, we present different policies for choosing which models to train on incoming data. Most notably, we propose the novel $\zeta$-policy, which focuses on training near-optimal models at reduced costs. Using a stochastic model, we theoretically prove that our $\zeta$-policy achieves near optimal performance while using fewer resources compared to the best performing policy. In our experiments across 11 benchmark datasets, we find empiric evidence that our $\zeta$-policy is a strong contribution to the state-of-the-art, demonstrating highly accurate performance, in some cases even outperforming competitors, and simultaneously being much more resource-friendly.

cs.LG↗

arXiv2025

A Probabilistic Proof for Stable Fluctuations in the Extremal Process of Branching Brownian Motion

We give a probabilistic proof for the emergence of the Stable-$1$ Law for the random fluctuations of the mass of the extremal process of branching Brownian Motion away from its tip. This result was already shown by Mytnik et al. albeit using PDE techniques. As a consequence, we demystify the origin of these fluctuations and the meaning of the deterministic centering function required.

math.PR↗

arXiv2024

Soft Hoeffding Tree: A Transparent and Differentiable Model on Data Streams

We propose soft Hoeffding trees (SoHoT) as a new differentiable and transparent model for possibly infinite and changing data streams. Stream mining algorithms such as Hoeffding trees grow based on the incoming data stream, but they currently lack the adaptability of end-to-end deep learning systems. End-to-end learning can be desirable if a feature representation is learned by a neural network and used in a tree, or if the outputs of trees are further processed in a deep learning model or workflow. Different from Hoeffding trees, soft trees can be integrated into such systems due to their differentiability, but are neither transparent nor explainable. Our novel model combines the extensibility and transparency of Hoeffding trees with the differentiability of soft trees. We introduce a new gating function to regulate the balance between univariate and multivariate splits in the tree. Experiments are performed on 20 data streams, comparing SoHoT to standard Hoeffding trees, Hoeffding trees with limited complexity, and soft trees applying a sparse activation function for sample routing. The results show that soft Hoeffding trees outperform Hoeffding trees in estimating class probabilities and, at the same time, maintain transparency compared to soft trees, with relatively small losses in terms of AUROC and cross-entropy. We also demonstrate how to trade off transparency against performance using a hyperparameter, obtaining univariate splits at one end of the spectrum and multivariate splits at the other.

cs.LG↗

arXiv2024

Gaussian free field on the tree subject to a hard wall II: Asymptotics

This is the second in a series of two works which study the discrete Gaussian free field on the binary tree when all leaves are conditioned to be positive. In the first work ("Gaussian free field on the tree subject to a hard wall I: Bounds") we identified the repulsion profile followed by the field in order to fulfill this "hard-wall constraint" event. In this work, we use these findings to obtain a comprehensive, sharp asymptotic description of the law of the field under this conditioning. We provide asymptotics for both local statistics, namely the (conditional) law of the field in a neighborhood of a vertex, as well as global statistics, including the (conditional) law of the minimum, maximum, empirical population mean and all subcritical exponential martingales. We conclude that the laws of the conditional and unconditional fields are asymptotically mutually singular with respect to each other.

math.PR↗

arXiv2024

Gaussian free field on the tree subject to a hard wall I: Bounds

This is the first in a series of two works which study the discrete Gaussian free field on the binary tree when all leaves are conditioned to be positive. In this work, we obtain sharp asymptotics for the probability of this "hard-wall constraint" event, and identify the repulsion profile followed by the field in order to achieve it. We also provide estimates for the mean, fluctuations and covariances of the field under the conditioning, which show that in the first log-many generations the field is localized around its mean. These results are used in the sequel work ("Gaussian free field on the tree subject to a hard wall II: Asymptotics") to obtain a comprehensive asymptotic description of the law of the field under the conditioning.

math.PR↗

arXiv2024

Self-repellent branching random walk

We consider a system of particles performing a discrete-time binary branching random walk with independent standard normal increments subject to a penalty $\b$ for every pair of particles that get within distance $\e$ of each other at every time. We study the optimal configurations that minimise the sum of the spread out cost and the repulsion cost up to a given time horizon $N$. We show that at time $N$ particles are spread out over a distance $\asymp (\b\e)^{1/3} 2^{2N/3}$. We also show that the total cost of the optimal configurations up to time $N$ is $\asymp (\b\e)^{2/3} 2^{4N/3}$.

math.PR↗

arXiv2024

On the Growth of the Extremal and Cluster Level Sets in Branching Brownian Motion

We study the limiting extremal and cluster point processes of branching Brownian motion. The former records the heights of all extreme values of the process, while the latter records the relative heights of extreme values in a genealogical neighborhood of order unity around a local maximum thereof. For the extremal point process, we show that the mass of upper level sets $[-v, \infty)$ grows as $C_\star Z v e^{\sqrt{2} v}(1+o(1))$ as $v \to \infty$, almost surely, where $Z$ is the limit of the associated derivative martingale and $C_\star \in (0, \infty)$ is a universal constant. For the cluster point process, we show that the logarithm of the mass of $[-v, \infty)$ grow as $\sqrt{2}v$ minus random fluctuations of order $v^{2/3}$, which are governed by an explicit law in the limit. The first result improves upon the works of Cortines et al. (arXiv:1703.06529) and Mytnik et al. (arXiv:2009.02042) in which asymptotics are shown in probability, while the second makes rigorous the derivation in the physics literature by Mueller et al. (arXiv:1910.06382) and Le et al. (arXiv:2207.07672) and resolves a conjecture thereof.

math.PR↗

arXiv2024

The mean field stubborn voter model

We analyse the effect of agent-dependent heavy-tailed waiting times in the voter model on the complete graph with $N$ vertices. We derive a novel scaling limit and show the existence of a limiting infinite voter model on the slowest updating sites. We further derive the consensus probabilities in the limit model explicitly. In the mean-field setting, the limit is determined by the extreme-value landscape of the waiting times and depends only on the tail index. To obtain these results, we study the coalescing system of random walks that is dual to the limit voter model and prove, among other auxiliary results, that it comes down from infinity.

math.PR↗

arXiv2024

A Random Walk Approach to Broadcasting on Random Recursive Trees

In the broadcasting problem on trees, a $\{-1,1\}$-message originating in an unknown node is passed along the tree with a certain error probability $q$. The goal is to estimate the original message without knowing the order in which the nodes were informed. We show a connection to random walks with memory effects and use this to develop a novel approach to analyse the majority estimator on random recursive trees. With this powerful approach, we study the entire group of very simple increasing trees as well as shape exchangeable trees together. This also extends Addario-Berry et al. (2022) who investigated this estimator for uniform and linear preferential attachment random recursive trees.

math.PR↗

arXiv2024

From 1 to infinity: The log-correction for the maximum of variable-speed branching Brownian motion

We study the extremes of variable speed branching Brownian motion (BBM) where the time-dependent "speed functions", which describe the time-inhomogeneous variance, converge to the identity function. We consider general speed functions lying strictly below their concave hull and piecewise linear, concave speed functions. In the first case, the log-correction for the order of the maximum depends only on the rate of convergence of the speed function near 0 and 1 and exhibits a smooth interpolation between the correction in the i.i.d. case, $\frac{1}{2\sqrt{2}} \ln t$, and that of standard BBM, $\frac{3}{2\sqrt{2}} \ln t$. In the second case, we describe the order of the maximum in dependence of the form of speed function and show that any log-correction larger than $\frac{3}{2\sqrt{2}} \ln t$ can be obtained. In both cases, we prove that the limiting law of the maximum and the extremal process essentially coincide with those of standard BBM, using a first and second moment method which relies on the localisation of extremal particles. This extends the results of Bovier and Hartung for two-speed BBM.

math.PR↗

arXiv2023

The phase diagram of the complex continuous random energy model: The weak correlation regime

We identify the fluctuations of the partition function of the continuous random energy model on a Galton-Watson tree in the so-called weak correlation regime. Namely, when the ``speed functions'', that describe the time-inhomogeneous variance, lie strictly below their concave hull and satisfy a certain weak regularity condition. We prove that the phase diagram coincides with the one of the random energy model. However, the fluctuations are different and depend on the slope of the covariance function at $0$ and the final time $t$.

math.PR↗

arXiv2022

Sampling Rare Event Energy Landscapes via Birth-Death Augmented Dynamics

A common problem that affects simulations of complex systems within the computational physics and chemistry communities is the so-called sampling problem or rare event problem where proper sampling of energy landscapes is impeded by the presences of high kinetic barriers that hinder transitions between metastable states on typical simulation time scales. Many enhanced sampling methods have been developed to address this sampling problem and more efficiently sample rare event systems. An interesting idea, coming from the field of statistics, was introduced in a recent work (Y. Lu, J. Lu, and J. Nolen, arXiv:1905.09863, 2019) in the form of a novel sampling algorithm that augments overdamped Langevin dynamics with a birth-death process. In this work, we expand on this idea and show that this birth-death sampling scheme can efficiently sample prototypical rare event energy landscapes, and that the speed of equilibration is independent of the barrier height. We amend a crucial shortcoming of the original algorithm that leads to incorrect sampling of barrier regions by introducing a new approximation of the birth-death term. We establish important theoretical properties of the modified algorithm and prove mathematically that the relevant convergence results still hold. We investigate via numerical simulations the effect of various parameters, and we investigate ways to reduce the computational effort of the sampling scheme. We show that the birth-death mechanism can be used to accelerate sampling in the more general case of underdamped Langevin dynamics that is more commonly used in simulating physical systems. Our results show that this birth-death scheme is a promising method for sampling rare event energy landscapes.

physics.comp-ph↗

arXiv2022

Maximum of Branching Brownian Motion among mild obstacles

We study the height of the maximal particle at time $t$ of a one dimensional branching Brownian motion with a space-dependent branching rate. The branching rate is set to zero in finitely many intervals (obstacles) of order $t$. We obtain almost sure asymptotics of the first order of the maximum, describe the path of a particle reaching this height and describe its dependence on the size and location of the obstacles.

math.PR↗

arXiv2022

The speed of invasion in an advancing population

We derive rigorous estimates on the speed of invasion of an advantageous trait in a spatially advancing population in the context of a system of one-dimensional F-KPP equations. The model was introduced and studied heuristically and numerically in a paper by Venegas-Ortiz et al. In that paper, it was noted that the speed of invasion by the mutant trait is faster faster when the resident population ist expanding in space compared to the speed when the resident population is already present everywhere. We use the Feynman-Kac representation to provide rigorous estimates that confirm these predictions.

math.PR↗

arXiv2021

Maxima of a Random Model of the Riemann Zeta Function over Intervals of Varying Length

We consider a model of the Riemann zeta function on the critical axis and study its maximum over intervals of length $(\log T)^{\theta}$, where $\theta$ is either fixed or tends to zero at a suitable rate. It is shown that the deterministic level of the maximum interpolates smoothly between the ones of log-correlated variables and of i.i.d. random variables, exhibiting a smooth transition 'from $\frac34$ to $\frac14$' in the second order. This provides a natural context where extreme value statistics of log-correlated variables with time-dependent variance and rate occur. A key ingredient of the proof is a precise upper tail tightness estimate for the maximum of the model on intervals of size one, that includes a Gaussian correction. This correction is expected to be present for the Riemann zeta function and pertains to the question of the correct order of the maximum of the zeta function in large intervals.

math.PR↗

arXiv2021

Branching Brownian motion with self repulsion

We consider a model of branching Brownian motion with self repulsion. Self-repulsion is introduced via change of measure that penalises particles spending time in an $\e$-neighbourhood of each other. We derive a simplified version of the model where only branching events are penalised. This model is almost exactly solvable and we derive a precise description of the particle numbers and branching times. In the limit of weak penalty, an interesting universal time-inhomogeneous branching process emerges. The position of the maximum is governed by a F-KPP type reaction-diffusion equation with a time dependent reaction term.

math.PR↗