Search

Image classification based on over-parametrized convolutional neural networks with a global average-pooling layer is considered. The weights of the network are learned by gradient descent. A bound on the rate of convergence of the difference between the misclassification risk of the newly introduced convolutional neural network estimate and the minimal possible value is derived.

We study the probability distribution of the first return time to the initial state of a quantum many-body system subject to stroboscopic projective measurements. We show that this distribution can be interpreted as a continuation of the canonical partition function of a spin chain with non-interacting domains at equilibrium, which is entirely characterised by the Loschmidt amplitude of the quantum many-body system. This allows us to show that this probability may decay either algebraically or exponentially asymptotically in time, depending on whether the spin model displays a ferromagnetic or a paramagnetic phase. We illustrate this idea on the example of the return time of $N$ adjacent fermions in a tight-binding model, revealing a rich phase behaviour, which can be tuned by scaling the probing time with $N$. Our analytical predictions are corroborated by exact numerical computations., Comment: 6 pages, 4 figures, 12 pages supplemental material

We develop a framework for the stochastic thermodynamics of a probe coupled to a fluctuating medium with spatio-temporal correlations, described by a scalar field. For a Brownian particle dragged by a harmonic trap through a fluctuating Gaussian field, we show that near criticality (where the field displays long-range spatial correlations) the spatially-resolved average heat flux develops a dipolar structure, where heat is absorbed in front and dissipated behind the dragged particle. Moreover, a perturbative calculation reveals that the dissipated power displays three distinct dynamical regimes depending on the drag velocity., Comment: 7 pages, 3 figures (main) + 21 pages, 1 figure (SM)

We present a method to infer the arbitrary space-dependent drift and diffusion of a nonlinear stochastic model driven by multiplicative fractional Gaussian noise from a single trajectory. Our method, fractional Onsager-Machlup optimisation (fOMo), introduces a maximum likelihood estimator by minimising a field-theoretic action which we construct from the observed time series. We successfully test fOMo for a wide range of Hurst exponents using artificial data with strong nonlinearities, and apply it to a data set of daily mean temperatures. We further highlight the significant systematic estimation errors when ignoring non-Markovianity, underlining the need for nonlinear fractional inference methods when studying real-world long-range (anti-)correlated systems., Comment: 21 pages, 5 figures, appendices

Convolutional neural network image classifiers are defined and the rate of convergence of the misclassification risk of the estimates towards the optimal misclassification risk is analyzed. Here we consider images as random variables with values in some functional space, where we only observe discrete samples as function values on some finite grid. Under suitable structural and smoothness assumptions on the functional a posteriori probability, which includes some kind of symmetry against rotation of subparts of the input image, it is shown that least squares plug-in classifiers based on convolutional neural networks are able to circumvent the curse of dimensionality in binary image classification if we neglect a resolution-dependent error term. The finite sample size behavior of the classifier is analyzed by applying it to simulated and real data.

We provide a perturbative framework to calculate extreme events of non-Markovian processes, by mapping the stochastic process to a two-species reaction diffusion process in a Doi-Peliti field theory combined with the Martin-Siggia-Rose formalism. This field theory treats interactions and the effect of external, possibly self-correlated noise in a perturbation about a Markovian process, thereby providing a systematic, diagrammatic approach to extreme events. We apply the formalism to Brownian Motion and calculate its survival probability distribution subject to self-correlated noise., Comment: 24 pages, 4 figures; added figures, comments and references

Image classification is considered, and a hierarchical max-pooling model with additional local pooling is introduced. Here the additional local pooling enables the hierachical model to combine parts of the image which have a variable relative distance towards each other. Various convolutional neural network image classifiers are introduced and compared in view of their rate of convergence. The finite sample size performance of the estimates is analyzed by applying them to simulated and real data.

We introduce a perturbative method to calculate all moments of the first-passage time distribution in stochastic one-dimensional processes which are subject to both white and coloured noise. This class of non-Markovian processes is at the centre of the study of thermal active matter, that is self-propelled particles subject to diffusion. The perturbation theory about the Markov process considers the effect of self-propulsion to be small compared to that of thermal fluctuations. To illustrate our method, we apply it to the case of active thermal particles (i) in a harmonic trap (ii) on a ring. For both we calculate the first-order correction of the moment-generating function of first-passage times, and thus to all its moments. Our analytical results are compared to numerics., Comment: 25 pages, 10 figures

Branching processes are used to model diverse social and physical scenarios, from extinction of family names to nuclear fission. However, for a better description of natural phenomena, such as viral epidemics in cellular tissues, animal populations and social networks, a spatial embedding---the branching random walk (BRW)---is required. Despite its wide range of applications, the properties of the volume explored by the BRW so far remained elusive, with exact results limited to one dimension. Here we present analytical results, supported by numerical simulations, on the scaling of the volume explored by a BRW in the critical regime, the onset of epidemics, in general environments. Our results characterise the spreading dynamics on regular lattices and general graphs, such as fractals, random trees and scale-free networks, revealing the direct relation between the graphs' dimensionality and the rate of propagation of the viral process. Furthermore, we use the BRW to determine the spectral properties of real social and metabolic networks, where we observe that a lack of information of the network structure can lead to differences in the observed behaviour of the spreading process. Our results provide observables of broad interest for the characterisation of real world lattices, tissues, and networks., Comment: 11 pages, 6 figures

We present an algorithm to efficiently sample first-passage times for fractional Brownian motion. To increase the resolution, an initial coarse lattice is successively refined close to the target, by adding exactly sampled midpoints, where the probability that they reach the target is non-negligible. Compared to a path of $N$ equally spaced points, the algorithm achieves the same numerical accuracy $N_{\rm eff}$, while sampling only a small fraction of all points. Though this induces a statistical error, the latter is bounded for each bridge, allowing us to bound the total error rate by a number of our choice, say $P_{\rm error}^{\rm tot}=10^{-6}$. This leads to significant improvements in both memory and speed. For $H=0.33$ and $N_{\rm eff}=2^{32}$, we need $5\,000$ times less CPU time and $10\, 000$ times less memory than the classical Davies Harte algorithm. The gain grows for $H=0.25$ and $N_{\rm eff} = 2^{42}$ to $3\cdot 10^{5}$ for CPU and $10^6$ for memory. We estimate our algorithmic complexity as ${\cal C}^{\rm ABSec}(N_{\rm eff}) = {\cal O}\left(\left( \ln N_{\rm eff}\right)^{3}\right)$, to be compared to Davies Harte which has complexity ${\cal C}^{\rm DH}(N) = {\cal O}\left(N \ln N \right)$. Decreasing $P_{\rm error}^{\rm tot}$ results in a small increase in complexity, proportional to $\ln (1/P_{\rm error}^{\rm tot})$. Our current implementation is limited to the values of $N_{\rm eff}$ given above, due to a loss of floating-point precision. The algorithm can be adapted to other extreme events and arbitrary Gaussian processes. It enables one to numerically validate theoretical predictions that were hitherto inaccessible., Comment: 15 pages, 10 figues

We study the first-passage time, the distribution of the maximum, and the absorption probability of fractional Brownian motion of Hurst parameter $H$ with both a linear and a non-linear drift. The latter appears naturally when applying non-linear variable transformations. Via a perturbative expansion in $\epsilon = H-1/2$, we give the first-order corrections to the classical result for Brownian motion analytically. Using a recently introduced adaptive bisection algorithm, which is much more efficient than the standard Davies-Harte algorithm, we test our predictions for the first-passage time on grids of effective sizes up to $N_{\rm eff}=2^{28}\approx 2.7\times 10^{8}$ points. The agreement between theory and simulations is excellent, and by far exceeds in precision what can be obtained by scaling alone., Comment: 18 pages, 19 figures

Atomic force spectroscopy and microscopy (AFM) are invaluable tools to characterize nanostructures and biological systems. Most experiments, including state-of-the-art images of molecular bonds, are achieved by driving probes at their mechanical resonance. This resonance reaches the MHz for the fastest AFM micro-cantilevers, with typical motion amplitude of a few nanometres. Next-generation investigations of molecular scale dynamics, including faster force imaging and higher-resolution spectroscopy of dissipative interactions, require more bandwidth and vibration amplitudes below interatomic distance, for non-perturbative short-range tip-matter interactions. Probe frequency is a key parameter to improve bandwidth while reducing Brownian motion, allowing large signal-to-noise for exquisite resolution. Optomechanical resonators reach motion detection at 10^(-18) m.(Hz)^(-1/2), while coupling light to bulk vibration modes whose frequencies largely surpass those of cantilevers. Here we introduce an optically operated resonating optomechanical atomic force probe of frequency 2 decades above the fastest functional AFM cantilevers while Brownian motion is 4 orders below. Based on a Silicon-On-Insulator technology, the probe demonstrates high-speed sensing of contact and non-contact interactions with sub-picometre driven motion, breaking open current locks for faster and finer atomic force spectroscopy.