arXiv subjects

Rahul Ramesh

Publications and source records attributed to Rahul Ramesh.

At least 19 recordsLinked to original sources

arXiv2026

Magnetogenesis by galactic processes: impact on circumgalactic and intergalactic fields

We investigate the origin and evolution of cosmic magnetic fields using a suite of large-volume cosmological magnetohydrodynamic simulations (L$_\mathrm{box}=25$ Mpc/h) run with the moving-mesh code AREPO. Atop the IllustrisTNG galaxy formation model, we implement additional recipes for magnetogenesis in which magnetic energy is injected during supernovae (SNe) and supermassive black hole (SMBH) feedback events, and compare these to simulations initialized with uniform primordial seed fields. Halo magnetic field strengths at $z=0$ are largely similar across seeding models and are primarily amplified and sustained by small-scale and halo-scale dynamo action. Nevertheless, we find differences in magnetic field topology, with SMBH-driven models exhibiting systematically smaller coherence lengths than primordial-only and SNe-only runs. We find that feedback-driven injection accelerates the onset of dynamo growth, leading to more rapid convergence of magnetic field strengths with numerical resolution, particularly in low-mass halos. In the intergalactic medium (IGM), SNe-only injection underproduces magnetic fields relative to inferred lower limits from $\gamma$-ray cascade constraints at both $z=0$ and $z \sim 3$, whereas our specific SMBH-based injection prescription satisfies present-day constraints but remains in mild tension at high redshifts. Reconciling these specific high-$z$ constraints therefore likely requires either modified feedback prescriptions or an additional primordial seeding component.

astro-ph.GA↗

arXiv2026

Zooming in on the circumgalactic medium with GIBLE: Cloud-scale simulations with cosmological initial conditions

We conduct simulations of $\sim$kpc-scale cool clouds in the circumgalactic medium (CGM), using initial conditions sampled from a highly resolved cosmological magneto-hydrodynamical zoom-in of a Milky Way-like galaxy. We select ten distinct cold clouds with masses of $m_{\rm{cloud}}$ $\sim$ $10^{4.5-5}$ M$_\odot$, originally resolved at a mass resolution of $m_{\rm{gas}}$ $\sim$ $200$ M$_\odot$. To further resolve small-scale features and physics, we implement a targeted refinement scheme within spherical regions co-moving with each cloud, thereby boosting the local mass resolution by a factor of 1000, reaching $m_{\rm{gas}}$ $\sim$ $0.2$ M$_\odot$ (spatial resolution, $r_{\rm{gas,cloud}}$ $\sim$ $O(\rm{pc})$). The selected clouds have diverse properties, across a broad parameter space, resulting in heterogeneous evolution. For the clouds we study, radiative cooling is the dominant physical process enabling cloud survival, while magnetic fields play a comparatively smaller role. The motion of these clouds is governed not only by drag forces that decelerate them, but also by acceleration from momentum exchange with the complex background velocity field, which can cause them to move faster than ballistic projectiles set by their initial velocities. Our results suggest that the non-trivial details of realistic cosmological initial conditions -- specifically the complex density, temperature, and velocity fields -- may play an important role in subsequent cloud evolution, and that sampling the output of an existing large-scale simulation provides a self-consistent approach to capture these effects without ad hoc assumptions.

astro-ph.GA↗

arXiv2025

The Synthetic Absorption Line Spectral Almanac (SALSA)

We create the first large-scale mock spectroscopic survey of gas absorption sightlines traversing the interstellar medium (ISM), circumgalactic medium (CGM), and intergalactic medium (IGM) surrounding galaxies of virtual Universes. That is, we create mock, or synthetic, absorption spectra by drawing lines-of-sight through cosmological hydrodynamical simulations, using a new mesh-free Voronoi ray-tracing algorithm. The result is the Synthetic Absorption Line Spectral Almanac (SALSA), which is publicly released on a feature-rich online science platform (www.tng-project.org/spectra). It spans a range of ions, transitions, instruments, observational characteristics, assumptions, redshifts, and simulations. These include, but are not limited to: (ions) HI, OI, CI, MgI, MgII, FeII, SiII, CaII, ZnII, SiIII, SiIV, NV, CII, CIV, OVI; (instruments) SDSS-BOSS, KECK-HIRES, UVES, COS, DESI, 4MOST, WEAVE, XSHOOTER; (model choices) with/without dust depletion, noise, quasar continua, foregrounds; (redshift) from z=0 to z~6; (ancillary data) integrated equivalent widths, column densities, distances and properties of nearby galaxies; (simulations) IllustrisTNG including TNG50, TNG-Cluster, EAGLE, and SIMBA. This scope is not fixed, and will grow and evolve with community interest and requests over time -- suggestions are welcome. The resulting dataset is generic and broadly applicable, enabling diverse science goals such as: (i) studies of the underlying physical gas structures giving rise to particular absorption signatures, (ii) galaxy-absorber and halo-absorber correlations, (iii) virtual surveys and survey strategy optimization, (iv) stacking experiments and the identification of faint absorption features, (v) assessment of data reduction methods and completeness calculations, (vi) inference of physical properties from observables, and (vii) apples-to-apples comparisons between simulations and data.

astro-ph.GA↗

arXiv2025

Numerical effects on the stripping of dark matter and stars in IllustrisTNG galaxy groups and clusters

The stellar haloes and intra-cluster light around galaxies are crucial test beds for dark matter (DM) physics and galaxy formation models. We consider the role that the numerical resolution plays in the modelling of these systems by studying the stripping of satellites in the IllustrisTNG cosmological simulations. We focus on host haloes of total halo mass $M_{\mathrm 200c}=10^{12-15}M_{\odot}$ and satellites of stellar mass $>10^{7}$$M_{\odot}$, and compare stellar halo / satellite properties across 9 IllustrisTNG runs with baryonic particle mass resolution between $8.5\times10^4M_{\odot}$ and $7\times10^8$$M_{\odot}$, using a Lagrangian-region technique to identify counterpart satellites across different resolution simulations of the same volume. We publish the corresponding catalogues alongside this paper. We demonstrate that the stripping of DM from satellites that orbit in group- and cluster-mass hosts is largely independent of resolution at least until 90 per cent of their initial mass at infall has been stripped. We do not find evidence for spurious disruption of galaxies due to insufficient resolution for the satellite masses we consider. By contrast, the stripping of stellar mass is strongly resolution-dependent: each factor of 8 improvement in particle stellar mass typically adds 2Gyr to the stripping time. Improved numerical resolution within the IllustrisTNG model generally results in more compact satellites with larger stellar masses, which in turn generate more centrally concentrated stellar haloes and intra-cluster mass profiles. However, the concomitant increase in stellar mass with increased resolution of both satellites and hosts may still be the cause for the overprediction of the stellar halo mass at large host radii relative to observations seen in some previous studies.

astro-ph.GA↗

arXiv2025

From Linearity to Non-Linearity: How Masked Autoencoders Capture Spatial Correlations

Masked Autoencoders (MAEs) have emerged as a powerful pretraining technique for vision foundation models. Despite their effectiveness, they require extensive hyperparameter tuning (masking ratio, patch size, encoder/decoder layers) when applied to novel datasets. While prior theoretical works have analyzed MAEs in terms of their attention patterns and hierarchical latent variable models, the connection between MAE hyperparameters and performance on downstream tasks is relatively unexplored. This work investigates how MAEs learn spatial correlations in the input image. We analytically derive the features learned by a linear MAE and show that masking ratio and patch size can be used to select for features that capture short- and long-range spatial correlations. We extend this analysis to non-linear MAEs to show that MAE representations adapt to spatial correlations in the dataset, beyond second-order statistics. Finally, we discuss some insights on how to select MAE hyper-parameters in practice.

cs.CV↗

arXiv2025

Tracing the cosmological origin of gas that fuels in situ star formation in TNG50 galaxies

Based on their cosmological origin, the stars of a galaxy can be divided into two categories: those that enter through merger events (ex situ) and those born in the main progenitor (in situ). We used the TNG50 cosmological magnetohydrodynamical simulation and its Lagrangian tracer particles to explore and quantify the origin of gas that ultimately forms the in situ stars of galaxies. We tracked back the baryonic mass contributing to the $z=0$ in situ stellar populations of galaxies, studying trends in mass from dwarfs to group-scale halos. We find that more massive halos acquire this matter earlier than lower-mass halos, reflecting an overall earlier assembly of their in situ stellar mass. Defining the Lagrangian half-mass radius R$_{\rm L, 1/2}$ of a galaxy as the distance containing half of the mass that will form its in situ stars by $z=0$, we find that R$_{\rm L, 1/2}$ is larger for more massive halos at early times, reflecting larger "in situ Lagrangian regions." However, the dependence of this radius on halo mass becomes flat at $z \simeq 3$ and then inverts toward $z=0$. In addition, R$_{\rm L, 1/2}$ increases rapidly with redshift, surpassing the virial radii of halos at $z \sim 2$. This marks the cosmic epoch at which most of the gas that eventually forms the in situ stars of galaxies leaves the intergalactic medium (IGM) and enters halos, a transition that occurs earlier for more massive halos. The formation redshift of the in situ stellar component increases with halo mass, while the formation redshift of the dark matter halo decreases, indicative of a differential assembly history between these two components. Finally, we decomposed the $z=0$ in situ stellar mass into its distinct modes of accretion. Smooth accretion from the IGM is the most important for low-mass galaxies, while mergers and satellite-stripped gas become relevant and even dominant only for high-mass galaxies.

astro-ph.GA↗

arXiv2024

Prospective Learning: Learning for a Dynamic Future

In real-world applications, the distribution of the data, and our goals, evolve over time. The prevailing theoretical framework for studying machine learning, namely probably approximately correct (PAC) learning, largely ignores time. As a consequence, existing strategies to address the dynamic nature of data and goals exhibit poor real-world performance. This paper develops a theoretical framework called "Prospective Learning" that is tailored for situations when the optimal hypothesis changes over time. In PAC learning, empirical risk minimization (ERM) is known to be consistent. We develop a learner called Prospective ERM, which returns a sequence of predictors that make predictions on future data. We prove that the risk of prospective ERM converges to the Bayes risk under certain assumptions on the stochastic process generating the data. Prospective ERM, roughly speaking, incorporates time as an input in addition to the data. We show that standard ERM as done in PAC learning, without incorporating time, can result in failure to learn when distributions are dynamic. Numerical experiments illustrate that prospective ERM can learn synthetic and visual recognition problems constructed from MNIST and CIFAR-10. Code at https://github.com/neurodata/prolearn.

stat.ML↗

arXiv2024

Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing

Knowledge Editing (KE) algorithms alter models' weights to perform targeted updates to incorrect, outdated, or otherwise unwanted factual associations. However, recent work has shown that applying KE can adversely affect models' broader factual recall accuracy and diminish their reasoning abilities. Although these studies give insights into the potential harms of KE algorithms, e.g., performance evaluations on benchmarks, little is understood about why such destructive failures occur. Motivated by this, we define a novel synthetic task in which a Transformer is trained from scratch to internalize a "structured" knowledge graph. The structure enforces relationships between entities of the graph, such that editing a factual association has "trickling effects" on other entities (e.g., altering X's parent is Y to Z affects who X's siblings' parent is). Through evaluations of edited models on this task, we show that KE inadvertently affects representations of entities beyond the targeted one, distorting relevant structures that allow a model to infer unseen knowledge about an entity. We call this phenomenon representation shattering and demonstrate that it degrades models' factual recall and reasoning performance. We further corroborate our findings in naturalistic settings with pre-trained Llama and Mamba models as well. Overall, our work yields a precise mechanistic hypothesis to explain why KE has adverse effects on model abilities.

cs.LG↗

arXiv2024

IllustrisTNG + Cosmic Rays with a Simple Transport Model: From Dwarfs to L$^\star$ Galaxies

We use a simple model for cosmic ray (CR) production and transport to assess the impact of CRs on $z$\,$=$\,$0$ galaxy, circumgalactic medium (CGM), and halo properties. To do so, we run the first suite of large-volume cosmological magnetohydrodynamical simulations (25\,Mpc\,h$^{-1}$ boxes) with the IllustrisTNG galaxy formation model including CR physics. We select CR transport parameters that yield a reasonable trade off between realistic large-scale integrated properties, and galactic CR pressure profiles predicted by more complex models. The resulting simulations show that, at fixed halo mass, including CRs does not strongly impact the temperature, density, or (total) pressure structure of the CGM with respect to the fiducial TNG model. However, cosmic rays add significant non-thermal pressure support to the halo. This suppresses the star formation activity and thus stellar masses of galaxies, from dwarf to L$^\star$ halos. The cosmic star formation rate density, stellar mass function, and stellar mass to halo mass relation are all reshaped by CRs. Galaxy sizes and halo-scale gas fractions are more mildly affected, while lower gas densities in the interstellar medium inhibit supermassive black hole growth. Halo gas is also less magnetized, and less metal enriched. These differences are primarily driven by suppressed gas accretion onto halos and galaxies, as well as weaker galactic outflows in the presence of CRs. Our findings are in qualitative agreement with previous studies of the impact of CRs on galactic outflows, and motivate the inclusion of CR physics in future large-scale cosmological simulations.

astro-ph.GA↗

arXiv2024

Many Perception Tasks are Highly Redundant Functions of their Input Data

We show that many perception tasks, from visual recognition, semantic segmentation, optical flow, depth estimation to vocalization discrimination, are highly redundant functions of their input data. Images or spectrograms, projected into different subspaces, formed by orthogonal bases in pixel, Fourier or wavelet domains, can be used to solve these tasks remarkably well regardless of whether it is the top subspace where data varies the most, some intermediate subspace with moderate variability--or the bottom subspace where data varies the least. This phenomenon occurs because different subspaces have a large degree of redundant information relevant to the task.

cs.CV↗

arXiv2024

Zooming in on the Circumgalactic Medium with GIBLE: Tracing the Origin and Evolution of Cold Clouds

We use the GIBLE suite of cosmological zoom-in simulations of Milky Way-like galaxies with additional super-Lagrangian refinement in the circumgalactic medium (CGM) to quantify the origin and evolution of CGM cold gas clouds. The origin of $z$\,$=$\,$0$ clouds can be traced back to recent ($\lesssim$\,$2$\,Gyr) outflows from the central galaxy ($\sim$\,45\,$\%$), condensation out of the hot phase of the CGM in the same time frame ($\sim$\,45\,$\%$), and to a lesser degree to satellite galaxies ($\lesssim$\,5\,$\%$). We find that in-situ condensation results from rapid cooling around local over-densities primarily seeded by the dissolution of the previous generation of clouds into the hot halo. About $\lesssim$\,10\,$\%$ of the cloud population is long lived, with their progenitors having already assembled $\sim$\,$2$\,Gyr ago. Collective cloud-cloud dynamics are crucial to their evolution, with coalescence and fragmentation events occurring frequently ($\gtrsim$\,20\,Gyr$^{-1}$). These interactions are modulated by non-vanishing pressure imbalances between clouds and their interface layers. The gas content of clouds is in a constant state of flux, with clouds and their surroundings exchanging mass at a rate of \mbox{$\gtrsim$\,$10^3$\,M$_\odot$\,Myr$^{-1}$}, depending on cloud relative velocity and interface vorticity. Furthermore, we find that a net magnetic tension force acting against the density gradient is capable of inhibiting cloud-background mixing. Our results show that capturing the distinct origins of cool CGM clouds, together with their physical evolution, requires high-resolution, cosmological galaxy formation simulations with both stellar and supermassive black hole feedback-driven outflows.

astro-ph.GA↗

arXiv2024

Zooming in on the Circumgalactic Medium with GIBLE: the Topology and Draping of Magnetic Fields around Cold Clouds

We use a cosmological zoom-in simulation of a Milky Way-like galaxy to study and quantify the topology of magnetic field lines around cold gas clouds in the circumgalactic medium (CGM). This simulation is a new addition to Project GIBLE, a suite of cosmological magnetohydrodynamical simulations of galaxy formation with preferential super-Lagrangian refinement in the CGM, reaching an unprecedented (CGM) gas mass resolution of $\sim$ $225$ M$_\odot$. To maximize statistics and resolution, we focus on a sample of $\sim$ $200$ clouds with masses of $\sim$ $10^6$ M$_\odot$. The topology of magnetic field lines around clouds is diverse, from threading to draping, and there is large variation in the magnetic curvature ($\kappa$) within cloud-background interfaces. We typically find little variation of $\kappa$ between upstream and downstream cloud faces, implying that strongly draped configurations are rare. In addition, $\kappa$ correlates strongly with multiple properties of the interface and the ambient background, including cloud overdensity and relative velocity, suggesting that cloud properties impact the topology of interface magnetic fields.

astro-ph.GA↗

arXiv2024

Towards an Understanding of Stepwise Inference in Transformers: A Synthetic Graph Navigation Model

Stepwise inference protocols, such as scratchpads and chain-of-thought, help language models solve complex problems by decomposing them into a sequence of simpler subproblems. Despite the significant gain in performance achieved via these protocols, the underlying mechanisms of stepwise inference have remained elusive. To address this, we propose to study autoregressive Transformer models on a synthetic task that embodies the multi-step nature of problems where stepwise inference is generally most useful. Specifically, we define a graph navigation problem wherein a model is tasked with traversing a path from a start to a goal node on the graph. Despite is simplicity, we find we can empirically reproduce and analyze several phenomena observed at scale: (i) the stepwise inference reasoning gap, the cause of which we find in the structure of the training data; (ii) a diversity-accuracy tradeoff in model generations as sampling temperature varies; (iii) a simplicity bias in the model's output; and (iv) compositional generalization and a primacy bias with in-context exemplars. Overall, our work introduces a grounded, synthetic framework for studying stepwise inference and offers mechanistic hypotheses that can lay the foundation for a deeper understanding of this phenomenon.

cs.LG↗

arXiv2023

Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks

Transformers trained on huge text corpora exhibit a remarkable set of capabilities, e.g., performing basic arithmetic. Given the inherent compositional nature of language, one can expect the model to learn to compose these capabilities, potentially yielding a combinatorial explosion of what operations it can perform on an input. Motivated by the above, we train autoregressive Transformer models on a synthetic data-generating process that involves compositions of a set of well-defined monolithic capabilities. Through a series of extensive and systematic experiments on this data-generating process, we show that: (1) autoregressive Transformers can learn compositional structures from small amounts of training data and generalize to exponentially or even combinatorially many functions; (2) generating intermediate outputs when composing functions is more effective for generalizing to new, unseen compositions than not generating any intermediate outputs (3) biases in the order of the compositions in the training data result in Transformers that fail to compose some combinations of functions; and (4) the attention layers select which capability to apply while the feed-forward layers execute the selected capability.

cs.LG↗

arXiv2023

The physical origins of gas in the circumgalactic medium using observationally-motivated TNG50 mocks

Absorbers in the spectrum of background objects probe the circumgalactic medium (CGM) surrounding galaxies, but its physical properties remain unconstrained. We use the cosmological hydrodynamical simulation TNG50 to statistically trace the origins of HI Ly-$\alpha$ absorbers around galaxies at $z = 0.5$ with stellar masses ranging from 10$^8$ to 10$^{11}$ M$_\odot$. We emulate observational CGM studies by considering all gas within a line of sight velocity range of $\pm 500$ km s$^{-1}$ from the central, to quantitatively assess the impact of other galaxy haloes and overdense gas in the IGM that intersect sightlines. The impact of satellites to the total absorber fraction is most significant at impact parameters $0.5 R_{\rm vir} < b < R_{\rm vir}$ and satellites with masses below typical detection limits ($M_* < 10^8$ M$_\odot$) account for 10 (40) per cent of absorbers that intersect any satellite bound to $10^{10}$ and $10^{11}$ $(10^9)$ M$_\odot$ centrals. After confirming outflows are more dominant along the minor axis, we additionally show that at least 20 per cent of absorbers exhibit no significant radial movement, indicating that absorbers can also trace quasi-static gas. The metallicity of absorbers also depends on the azimuthal angle, but this signal is largely driven by enriched inflowing and quasi-static gas. Our work shows that determining the stellar mass of galaxies at $z_{\rm abs}$ is essential to constrain the physical origin of the gas traced in absorption, which in turn is key to characterising the kinematics and distribution of gas and metals in the CGM.

astro-ph.GA↗

arXiv2023

Zooming in on the circumgalactic medium: resolving small-scale gas structure with the GIBLE cosmological simulations

We introduce Project GIBLE (Gas Is Better resoLved around galaxiEs), a suite of cosmological zoom-in simulations where gas in the circumgalactic medium (CGM) is preferentially simulated at ultra-high numerical resolution. Our initial sample consists of eight galaxies, all selected as Milky Way-like galaxies at $z=0$ from the TNG50 simulation. Using the same galaxy formation model as IllustrisTNG, and the moving-mesh code AREPO, we re-simulate each of these eight galaxies maintaining a resolution equivalent to TNG50-2 ($m_{\rm{gas}}$ $\sim$ $8 \times 10^5 {\rm M}_{\odot}$). However, we use our super-Lagrangian refinement scheme to more finely resolve gas in the CGM around these galaxies. Our highest resolution runs achieve 512 times better mass resolution ($\sim$ $10^3 {\rm M}_{\odot}$). This corresponds to a median spatial resolution of $\sim$ $75$ pc at $0.15~R_{\rm{200,c}}$, which coarsens with increasing distance to $\sim$ $700$ pc at the virial radius. We make predictions for the covering fractions of several observational tracers of multi-phase CGM gas: HI, MgII, CIV and OVII. We then study the impact of improved resolution on small scale structure. While the abundance of the smallest cold, dense gas clouds continues to increase with improving resolution, the number of massive clouds is well converged. We conclude by quantifying small scale structure with the velocity structure function and the auto-correlation function of the density field, assessing their resolution dependence. The GIBLE cosmological hydrodynamical simulations enable us to improve resolution in a computationally efficient manner, thereby achieving numerical convergence of a subset of key CGM gas properties and observables.

astro-ph.GA↗

arXiv2023

Azimuthal Anisotropy of Magnetic Fields in the Circumgalactic Medium Driven by Galactic Feedback Processes

We use the TNG50 cosmological magnetohydrodynamical simulation of the IllustrisTNG project to show that magnetic fields in the circumgalactic medium (CGM) have significant angular structure. This azimuthal anisotropy at fixed distance is driven by galactic feedback processes that launch strong outflows into the halo, preferentially along the minor axes of galaxies. These feedback-driven outflows entrain strong magnetic fields from the interstellar medium, dragging fields originally amplified by small-scale dynamos into the CGM. At the virial radius, $z=0$ galaxies with M$_\star \sim 10^{10}\,\rm{M_\odot}$ show the strongest anisotropy ($\sim 0.35$ dex). This signal weakens with decreasing impact parameter, and is also present but weaker for lower mass as well as higher mass galaxies. Creating mock Faraday rotation measure (RM) sightlines through the simulated volume, we find that the angular RM trend is qualitatively consistent with recent observational measurements. We show that rich structure is present in the circumgalactic magnetic fields of galaxies. However, TNG50 predicts small RM amplitudes in the CGM that make detection difficult as a result of other contributions along the line of sight.

astro-ph.GA↗

arXiv2023

The Training Process of Many Deep Networks Explores the Same Low-Dimensional Manifold

We develop information-geometric techniques to analyze the trajectories of the predictions of deep networks during training. By examining the underlying high-dimensional probabilistic models, we reveal that the training process explores an effectively low-dimensional manifold. Networks with a wide range of architectures, sizes, trained using different optimization methods, regularization techniques, data augmentation techniques, and weight initializations lie on the same manifold in the prediction space. We study the details of this manifold to find that networks with different architectures follow distinguishable trajectories but other factors have a minimal influence; larger networks train along a similar manifold as that of smaller networks, just faster; and networks initialized at very different parts of the prediction space converge to the solution along a similar manifold.

cs.LG↗