arXiv subjects

Oliver Hahn

Publications and source records attributed to Oliver Hahn.

At least 37 records · Page 2Linked to original sources

arXiv2025

Scene-Centric Unsupervised Panoptic Segmentation

Unsupervised panoptic segmentation aims to partition an image into semantically meaningful regions and distinct object instances without training on manually annotated data. In contrast to prior work on unsupervised panoptic scene understanding, we eliminate the need for object-centric training data, enabling the unsupervised understanding of complex scenes. To that end, we present the first unsupervised panoptic method that directly trains on scene-centric imagery. In particular, we propose an approach to obtain high-resolution panoptic pseudo labels on complex scene-centric data, combining visual representations, depth, and motion cues. Utilizing both pseudo-label training and a panoptic self-training strategy yields a novel approach that accurately predicts panoptic segmentation of complex scenes without requiring any human annotations. Our approach significantly improves panoptic quality, e.g., surpassing the recent state of the art in unsupervised panoptic segmentation on Cityscapes by 9.4% points in PQ.

cs.CV↗

arXiv2025

cosmICweb: Cosmological Initial Conditions for Zoom-in Simulations in the Cloud

We present the online service cosmICweb (COSMological Initial Conditions on the WEB) - the first database and web interface to store, analyze, and disseminate initial conditions for zoom simulations of objects forming in cosmological simulations: from galaxy clusters to galaxies and more. Specifically, we store compressed information about the Lagrangian proto-halo patches for all objects in a typical simulation merger tree along with properties of the halo/galaxy across cosmic time. This enables a convenient web-based selection of the desired zoom region for an object fitting user-specified selection criteria. The information about the region can then be used with the MUSIC code to generate the zoom ICs for the simulation. In addition to some other simulations, we currently support all objects in the EAGLE simulation database, so that for example the Auriga simulations are easily reproduced, which we demonstrate explicitly. The framework is extensible to include other simulations through an API that can be added to an existing database structure and with which cosmICweb can then be interfaced. We make the web portal and database publicly available to the community.

astro-ph.CO↗

arXiv2025

Bridging magic and non-Gaussian resources via Gottesman-Kitaev-Preskill encoding

Although the similarity between non-stabilizer states -- also known as magic states -- in discrete-variable systems and non-Gaussian states in continuous-variable systems has widely been recognized, the precise connections between these two notions have still been unclear. We establish a fundamental link between these two quantum resources via the Gottesman-Kitaev-Preskill (GKP) encoding. We show that the negativity of the continuous-variable Wigner function for an encoded GKP state coincides with a magic measure we introduce, which matches the negativity of the discrete Wigner function for odd dimensions. We also provide a continuous-variable representation of the stabilizer Rényi entropy -- a recent proposal for a magic measure for multi-qubit states. With this in hand, we give a classical simulation algorithm with runtime scaling with the resource contents, quantified by our magic measures. We also employ our results to prove that implementing a multi-qubit logical non-Clifford operation in the GKP code subspace requires a non-Gaussian operation even at the limit of perfect encoding, despite the fact that the ideal GKP states already come with a large amount of non-Gaussianity.

quant-ph↗

arXiv2025

BullFrog: Multi-step perturbation theory as a time integrator for cosmological simulations

Modelling the cosmic large-scale structure can be done through numerical N-body simulations or by using perturbation theory. Here, we present an N-body approach that effectively implements a multi-step forward model based on Lagrangian Perturbation Theory (LPT) in a $Λ$CDM Universe. This is achieved by introducing the second-order accurate BullFrog integrator, which automatically performs 2LPT time steps to second order without requiring the explicit computation of 2LPT displacements. Importantly, we show that BullFrog trajectories rapidly converge to the exact solution as the number of time steps increases, at any moment in time, even though 2LPT becomes invalid after shell-crossing. As a validation test, we compare BullFrog against other N-body integrators and high-order LPT, both for a realistic $Λ$CDM cosmology and for simulations with a sharp UV cutoff in the initial conditions. The latter scenario enables controlled experiments against LPT and, in practice, is particularly relevant for modelling coarse-grained fluids arising in the context of effective field theory. We demonstrate that BullFrog significantly improves upon other LPT-inspired integrators, such as FastPM and COLA, without incurring any computational overhead compared to standard N-body integrators. Implementing BullFrog in any existing N-body code is straightforward, particularly if FastPM is already integrated.

astro-ph.CO↗

arXiv2024

Stochastic Super-resolution of Cosmological Simulations with Denoising Diffusion Models

In recent years, deep learning models have been successfully employed for augmenting low-resolution cosmological simulations with small-scale information, a task known as "super-resolution". So far, these cosmological super-resolution models have relied on generative adversarial networks (GANs), which can achieve highly realistic results, but suffer from various shortcomings (e.g. low sample diversity). We introduce denoising diffusion models as a powerful generative model for super-resolving cosmic large-scale structure predictions (as a first proof-of-concept in two dimensions). To obtain accurate results down to small scales, we develop a new "filter-boosted" training approach that redistributes the importance of different scales in the pixel-wise training objective. We demonstrate that our model not only produces convincing super-resolution images and power spectra consistent at the percent level, but is also able to reproduce the diversity of small-scale features consistent with a given low-resolution simulation. This enables uncertainty quantification for the generated small-scale features, which is critical for the usefulness of such super-resolution models as a viable surrogate model for cosmic structure formation.

astro-ph.CO↗

arXiv2024

Simulating the Universe from the cosmological horizon to halo scales

Ultra-large scales close to the cosmological horizon will be probed by the upcoming observational campaigns. They hold the promise to constrain single-field inflation as well as general relativity, but in order to include them in the forthcoming analyses, their modelling has to be robust. In particular, general relativistic effects may be mistaken for primordial signals, and no consensus has emerged either from analytical modelling nor from the numerical route, obstructed by the large dynamical range to be simulated. In this work, we present a numerical technique to overcome the latter limitation: we compute the general relativistic displacement field with the N-body relativistic code gevolution and combine it with the accurate Newtonian simulation Gadget-4. This combination leads to an effective simulation reproducing the desired behaviour at the level of the matter power spectrum and bispectrum. We then measure, for the first time in a simulation, the relativistic scale-dependent bias in Poisson gauge; at redshift $z=0$, we find $b_1^{\mathrm{GR}}=-5.7 \pm 1.7$. Our results at the field level are only valid in the Poisson gauge and need to be complemented with a relativistic ray tracing algorithm to compute the number count observable.

astro-ph.CO↗

arXiv2024

Boosting Unsupervised Semantic Segmentation with Principal Mask Proposals

Unsupervised semantic segmentation aims to automatically partition images into semantically meaningful regions by identifying global semantic categories within an image corpus without any form of annotation. Building upon recent advances in self-supervised representation learning, we focus on how to leverage these large pre-trained models for the downstream task of unsupervised segmentation. We present PriMaPs - Principal Mask Proposals - decomposing images into semantically meaningful masks based on their feature representation. This allows us to realize unsupervised semantic segmentation by fitting class prototypes to PriMaPs with a stochastic expectation-maximization algorithm, PriMaPs-EM. Despite its conceptual simplicity, PriMaPs-EM leads to competitive results across various pre-trained backbone models, including DINO and DINOv2, and across different datasets, such as Cityscapes, COCO-Stuff, and Potsdam-3. Importantly, PriMaPs-EM is able to boost results when applied orthogonally to current state-of-the-art unsupervised semantic segmentation pipelines. Code is available at https://github.com/visinf/primaps.

cs.CV↗

arXiv2024

Testing the assumptions of the Effective Field Theory of Large-Scale Structure

The Effective Field Theory of Large-Scale Structure (EFTofLSS) attempts to amend some of the shortcomings of the traditional perturbative methods used in cosmology. It models the evolution of long-wavelength perturbations above a cutoff scale without the need for a detailed description of the short-wavelength ones. Short-scale physics is encoded in the coefficients of a series of operators composed of the long-wavelength fields, and ordered in a systematic expansion. As applied in the literature, the EFTofLSS corrects a summary statistic (such as the power spectrum) calculated from standard perturbation theory by matching it to $N$-body simulations or observations. This `bottom-up' construction is remarkably successful in extending the range of validity of perturbation theory. In this work, we compare this framework to a `top-down' approach, which estimates the EFT coefficients from the stress tensor of an $N$-body simulation, and propagates the corrections to the summary statistic. We consider simple initial conditions, viz. two sinusoidal, plane-parallel density perturbations with substantially different frequencies and amplitudes. We find that the leading EFT correction to the power spectrum in the top-down model is in excellent agreement with that inferred from the bottom-up approach which, by construction, provides an exact match to the numerical data. This result is robust to changes in the wavelength separation between the two linear perturbations. However, in our setup, the leading EFT coefficient does not always grow linearly with the cosmic expansion factor as assumed in the literature based on perturbative considerations. Instead, it decreases after orbit crossing takes place.

astro-ph.CO↗

arXiv2024

Scale-dependent local primordial non-Gaussianity as a solution to the $S_8$ tension

For the last decade, several probes have pointed to a cosmological tension between the amplitude of density fluctuations extrapolated from the cosmic microwave background within the standard cosmological model and the one encapsulated by the $S_8$ parameter from large scale structure. The origin of this $S_8$ tension has not yet been elucidated and may hint at systematics in the data, unaccounted effects from baryonic physics, or new physics beyond the standard model of cosmology. Baryonic physics may in principle provide a nonlinear solution to the tension by suppressing the matter power spectrum more strongly on nonlinear scales than is traditionally assumed. Such a solution would not worsen the Hubble tension, contrary to many other proposed solutions to the $S_8$ tension. However, no realistic baryonic feedback in hydrodynamical simulations provides the needed suppression as a function of redshift. Here, we point out that a scale-dependence of local-type primordial non-Gaussianities (PNG), with significant PNG at scales of a few Mpc, can provide the needed suppression, since such PNG can suppress the power spectrum at slightly larger scales than baryons do. We demonstrate this by devising collisionless numerical simulations of structure formation in boxes of 0.5 Gpc/$h$ with scale-dependent local-type PNG. Our simple models show that, as a proof of principle, scale-dependent PNG, with a Gaussian random field for primordial density fluctuations on large scales and $f_{\rm NL} \simeq -300$ at $\lesssim 10$ Mpc scales, together with state-of-the-art baryonification of the matter power spectrum, can in principle solve the $S_8$ tension. The $S_8$ tension would then be a smoking-gun of non-trivial inflationary physics.

astro-ph.CO↗

arXiv2024

DISCO-DJ I: a differentiable Einstein-Boltzmann solver for cosmology

We present the Einstein-Boltzmann module of the DISCO-DJ (DIfferentiable Simulations for COsmology - Done with JAX) software package. This module implements a fully differentiable solver for the linearised cosmological Einstein-Boltzmann equations in the JAX framework, and allows computing Jacobian matrices of all solver output with respect to all input parameters using automatic differentiation. This implies that along with the solution for a given set of parameters, the tangent hyperplane in parameter space is known as well, which is a key ingredient for cosmological inference and forecasting problems as well as for many other applications. We discuss our implementation and demonstrate that our solver agrees at the per-mille level with the existing non-differentiable solvers CAMB and CLASS, including massive neutrinos and a dark energy fluid with parameterised equation of state. We illustrate the dependence of various summary statistics in large-scale structure cosmology on model parameters using the differentiable solver, and finally demonstrate how it can be easily used for Fisher forecasting, with a forecast for Euclid as an example. Since the implementation is significantly shorter and more modular than existing solvers, we believe it will be more straightforward to extend our solver to include additional physics, such as additional dark energy and dark matter models, modified gravity, or other non-standard physics in the future.

astro-ph.CO↗

arXiv2024

Perturbation-theory informed integrators for cosmological simulations

Large-scale cosmological simulations are an indispensable tool for modern cosmology. To enable model-space exploration, fast and accurate predictions are critical. In this paper, we show that the performance of such simulations can be further improved with time-stepping schemes that use input from cosmological perturbation theory. Specifically, we introduce a class of time-stepping schemes derived by matching the particle trajectories in a single leapfrog/Verlet drift-kick-drift step to those predicted by Lagrangian perturbation theory (LPT). As a corollary, these schemes exactly yield the analytic Zel'dovich solution in 1D in the pre-shell-crossing regime (i.e. before particle trajectories cross). One representative of this class is the popular FastPM scheme by Feng et al. 2016, which we take as our baseline. We then construct more powerful LPT-inspired integrators and show that they outperform FastPM and standard integrators in fast simulations in two and three dimensions with $\mathcal{O}(1 - 100)$ timesteps, requiring less steps to accurately reproduce the power spectrum and bispectrum of the density field. Furthermore, we demonstrate analytically and numerically that, for any integrator, convergence is limited in the post-shell-crossing regime (to order 3/2 for planar wave collapse), owing to the lacking regularity of the acceleration field, which makes the use of high-order integrators in this regime futile. Also, we study the impact of the timestep spacing and of a decaying mode present in the initial conditions. Importantly, we find that symplecticity of the integrator plays a minor role for fast approximate simulations with a small number of timesteps.

astro-ph.CO↗

arXiv2024

A Perspective on Deep Vision Performance with Standard Image and Video Codecs

Resource-constrained hardware, such as edge devices or cell phones, often rely on cloud servers to provide the required computational resources for inference in deep vision models. However, transferring image and video data from an edge or mobile device to a cloud server requires coding to deal with network constraints. The use of standardized codecs, such as JPEG or H.264, is prevalent and required to ensure interoperability. This paper aims to examine the implications of employing standardized codecs within deep vision pipelines. We find that using JPEG and H.264 coding significantly deteriorates the accuracy across a broad range of vision tasks and models. For instance, strong compression rates reduce semantic segmentation accuracy by more than 80% in mIoU. In contrast to previous findings, our analysis extends beyond image and action classification to localization and dense prediction tasks, thus providing a more comprehensive perspective.

cs.CV↗

arXiv2024

Starting Cosmological Simulations from the Big Bang

The cosmic large-scale structure (LSS) provides a unique testing ground for connecting fundamental physics to astronomical observations. Modeling the LSS requires numerical $N$-body simulations or perturbative techniques that both come with distinct shortcomings. Here we present the first unified numerical approach, enabled by new time integration and discreteness reduction schemes, and demonstrate its convergence at the field level. In particular, we show that our simulations (1) can be initialized directly at time zero, and (2) can be made to agree with high-order Lagrangian perturbation theory in the fluid limit. This enables fast, self-consistent, and UV-complete forward modeling of LSS observables.

astro-ph.CO↗

arXiv2024

A deep learning framework for jointly extracting spectra and source-count distributions in astronomy

Astronomical observations typically provide three-dimensional maps, encoding the distribution of the observed flux in (1) the two angles of the celestial sphere and (2) energy/frequency. An important task regarding such maps is to statistically characterize populations of point sources too dim to be individually detected. As the properties of a single dim source will be poorly constrained, instead one commonly studies the population as a whole, inferring a source-count distribution (SCD) that describes the number density of sources as a function of their brightness. Statistical and machine learning methods for recovering SCDs exist; however, they typically entirely neglect spectral information associated with the energy distribution of the flux. We present a deep learning framework able to jointly reconstruct the spectra of different emission components and the SCD of point-source populations. In a proof-of-concept example, we show that our method accurately extracts even complex-shaped spectra and SCDs from simulated maps.

astro-ph.IM↗

arXiv2023

Semantic Self-adaptation: Enhancing Generalization with a Single Sample

The lack of out-of-domain generalization is a critical weakness of deep networks for semantic segmentation. Previous studies relied on the assumption of a static model, i. e., once the training process is complete, model parameters remain fixed at test time. In this work, we challenge this premise with a self-adaptive approach for semantic segmentation that adjusts the inference process to each input sample. Self-adaptation operates on two levels. First, it fine-tunes the parameters of convolutional layers to the input image using consistency regularization. Second, in Batch Normalization layers, self-adaptation interpolates between the training and the reference distribution derived from a single test sample. Despite both techniques being well known in the literature, their combination sets new state-of-the-art accuracy on synthetic-to-real generalization benchmarks. Our empirical study suggests that self-adaptation may complement the established practice of model regularization at training time for improving deep network generalization to out-of-domain data. Our code and pre-trained models are available at https://github.com/visinf/self-adaptive.

cs.CV↗

arXiv2023

The FLAMINGO project: cosmological hydrodynamical simulations for large-scale structure and galaxy cluster surveys

We introduce the Virgo Consortium's FLAMINGO suite of hydrodynamical simulations for cosmology and galaxy cluster physics. To ensure the simulations are sufficiently realistic for studies of large-scale structure, the subgrid prescriptions for stellar and AGN feedback are calibrated to the observed low-redshift galaxy stellar mass function and cluster gas fractions. The calibration is performed using machine learning, separately for three resolutions. This approach enables specification of the model by the observables to which they are calibrated. The calibration accounts for a number of potential observational biases and for random errors in the observed stellar masses. The two most demanding simulations have box sizes of 1.0 and 2.8 Gpc and baryonic particle masses of $1\times10^8$ and $1\times10^9 \text{M}_\odot$, respectively. For the latter resolution the suite includes 12 model variations in a 1 Gpc box. There are 8 variations at fixed cosmology, including shifts in the stellar mass function and/or the cluster gas fractions to which we calibrate, and two alternative implementations of AGN feedback (thermal or jets). The remaining 4 variations use the unmodified calibration data but different cosmologies, including different neutrino masses. The 2.8 Gpc simulation follows $3\times10^{11}$ particles, making it the largest ever hydrodynamical simulation run to $z=0$. Lightcone output is produced on-the-fly for up to 8 different observers. We investigate numerical convergence, show that the simulations reproduce the calibration data, and compare with a number of galaxy, cluster, and large-scale structure observations, finding very good agreement with the data for converged predictions. Finally, by comparing hydrodynamical and `dark-matter-only' simulations, we confirm that baryonic effects can suppress the halo mass function and the matter power spectrum by up to $\approx20$ per cent.

astro-ph.CO↗

arXiv2023

Phase-space simulations of prompt cusps: simulating the formation of the first haloes without artificial fragmentation

The first generation of haloes forms from the collapse of the smallest peaks in the initial density field. $N$-body simulations of this process suggest a prompt formation of a steep power-law cusp, but these calculations are plagued by numerical artifacts which casts some doubt on this result. Here, we develop new simulation methods based on the dark matter phase-space sheet approach and present results which are entirely free of artificial clumps. We find that a cusp with density $ρ\propto r^{-1.5}$ is indeed formed promptly, subsequently accreting a more extended halo and participating in the hierarchical growth of later halo generations. However, our simulations also suggest that the presence of artificial clumps just before peak collapse can significantly shallow the inner profiles of the cusps. We use $N$-body simulations with controlled amounts of small-scale power to place a conservative upper limit on the scales affected by artificial clumps. Finally, we used these results to simulate the collapse of the first generation of peaks of various types and in different cosmologies, finding prompt cusps to form in all cases. We conclude that prompt cusps are a generic feature of the collapse of peaks on the free-streaming scale of the initial density field, and their structure can safely be studied using $N$-body simulations provided care is taken to excise the region potentially affected by artificial clumps.

astro-ph.GA↗

arXiv2023

Hydrodynamical simulations of galaxy formation with non-Gaussian initial conditions

Collisionless simulations of structure formation with significant local primordial non-Gaussianities at Mpc scales have shown that a non-Gaussian tail favouring underdensities, with a negative $f_{\rm NL}$ parameter, can significantly change the merging history of galaxy-sized dark matter halos, which then typically assemble later than in vanilla $Λ$CDM. Moreover, such a small-scale negative $f_{\rm NL}$ could have interesting consequences for the cosmological $S_8$ tension. Here, we complement our previous work on collisionless simulations with new hydrodynamical simulations of galaxy formation in boxes of 30 Mpc/$h$, using the {\sc RAMSES} code. In particular, we show that all feedback prescriptions being otherwise identical, simulations with a negative $f_{\rm NL} \sim -1000$ on small scales, hence forming galaxies a bit later than in vanilla $Λ$CDM, allow to form simulated galaxies with more disky kinematics than in the vanilla case. Therefore, such small-scale primordial non-Gaussianities could potentially help alleviate, simultaneously, tensions in cosmology and galaxy formation. These hydrodynamical simulations on small scales will need to be complemented with larger box simulations with scale-dependent non-Gaussianities, to statistically confirm these trends and explore their observational consequences in further detail.

astro-ph.CO↗