Publications

While overfitting and, more generally, double descent are ubiquitous in machine learning, increasing the number of parameters of the most widely used tensor network, the matrix product state (MPS), has generally lead to monotonic improvement of test performance in previous studies. To better understand the generalization properties of architectures parameterized by MPS, we construct artificial data which can be exactly modeled by an MPS and train the models with different number of parameters. We observe model overfitting for one-dimensional data, but also find that for more complex data overfitting is less significant, while with MNIST image data we do not find any signatures of overfitting. We speculate that generalization properties of MPS depend on the properties of data: with one-dimensional data (for which the MPS ansatz is the most suitable) MPS is prone to overfitting, while with more complex data which cannot be fit by MPS exactly, overfitting may be much less significant.

We update the capabilities of the open-knowledge software instrument Modules for Experiments in Stellar Astrophysics (MESA). The new auto_diff module implements automatic differentiation in MESA, an enabling capability that alleviates the need for hard-coded analytic expressions or finite difference approximations. We significantly enhance the treatment of the growth and decay of convection in MESA with a new model for time-dependent convection, which is particularly important during late-stage nuclear burning in massive stars and electron degenerate ignition events. We strengthen MESA's implementation of the equation of state, and we quantify continued improvements to energy accounting and solver accuracy through a discussion of different energy equation features and enhancements. To improve the modeling of stars in MESA we describe key updates to the treatment of stellar atmospheres, molecular opacities, Compton opacities, conductive opacities, element diffusion coefficients, and nuclear reaction rates. We introduce treatments of starspots, an important consideration for low-mass stars, and modifications for superadiabatic convection in radiation-dominated regions. We describe new approaches for increasing the efficiency of calculating monochromatic opacities and radiative levitation, and for increasing the efficiency of evolving the late stages of massive stars with a new operator split nuclear burning mode. We close by discussing major updates to MESA's software infrastructure that enhance source code development and community engagement.

Small cell clusters exhibit numerous phenomena typically associated with complex systems, such as division of labour and programmed cell death. A conserved class of such clusters occurs during oogenesis in the form of germline cysts that give rise to oocytes. Germline cysts form through cell divisions with incomplete cytokinesis, leaving cells intimately connected through intercellular bridges that facilitate cyst generation, cell fate determination and collective growth dynamics. Using the well-characterized Drosophila melanogaster female germline cyst as a foundation, we present mathematical models rooted in the dynamics of cell cycle proteins and their interactions to explain the generation of germline cell lineage trees (CLTs) and highlight the diversity of observed CLT sizes and topologies across species. We analyse competing models of symmetry breaking in CLTs to rationalize the observed dynamics and robustness of oocyte fate specification, and highlight remaining gaps in knowledge. We also explore how CLT topology affects cell cycle dynamics and synchronization and highlight mechanisms of intercellular coupling that underlie the observed collective growth patterns during oogenesis. Throughout, we point to similarities across organisms that warrant further investigation and comment on the extent to which experimental and theoretical findings made in model systems extend to other species.

We present the first detailed study comparing the populations of stellar streams in cosmological simulations to observed Milky Way dwarf galaxy streams. In particular, we compare streams identified around Milky Way analogs in the FIRE-2 simulations to stellar streams observed by the Southern Stellar Stream Spectroscopic Survey (S5). For an accurate comparison between the stream populations, we produce mock Dark Energy Survey (DES) observations of the FIRE streams and estimate the detectability of their tidal tails and progenitors. The number and stellar mass distributions of detectable stellar streams is consistent between observations and simulations. However, there are discrepancies in the distributions of pericenters and apocenters, with the detectable FIRE streams, on average, forming at larger pericenters (out to > 110 kpc) and surviving only at larger apocenters (> 40 kpc) than those observed in the Milky Way. We find that the population of high-stellar mass dwarf galaxy streams in the Milky Way is incomplete. Interestingly, a large fraction of the FIRE streams would only be detected as satellites in DES-like observations, since their tidal tails are too low-surface brightness to be detectable. We thus predict a population of yet-undetected tidal tails around Milky Way satellites, as well as a population of fully undetected low surface brightness stellar streams, and estimate their detectability with the Rubin Observatory. Finally, we discuss the causes and implications of the discrepancies between the stream populations in FIRE and the Milky Way, and explore future avenues for tests of satellite disruption in cosmological simulations.

In the era of large-scale spectroscopic surveys in the Local Group, we can explore using chemical abundances of halo stars to study the star formation and chemical enrichment histories of the dwarf galaxy progenitors of the Milky Way (MW) and M31 stellar halos. In this paper, we investigate using the chemical abundance ratio distributions (CARDs) of seven stellar halos from the Latte suite of FIRE-2 simulations. We attempt to infer galaxies' assembly histories by modeling the CARDs of the stellar halos of the Latte galaxies as a linear combination of template CARDs from disrupted dwarfs, with different stellar masses M⋆ and quenching times t100. We present a method for constructing these templates using present-day dwarf galaxies. For four of the seven Latte halos studied in this work, we recover the mass spectrum of accreted dwarfs to a precision of <10 percent. For the fraction of mass accreted as a function of t100, we find the residuals of 20–30 percent for five of the seven simulations. We discuss the failure modes of this method, which arise from the diversity of star formation and chemical enrichment histories that dwarf galaxies can take. These failure cases can be robustly identified by the high model residuals. Although the CARDs modeling method does not successfully infer the assembly histories in these cases, the CARDs of these disrupted dwarfs contain signatures of their unusual formation histories. Our results are promising for using CARDs to learn more about the histories of the progenitors of the MW and M31 stellar halos.

We predict that twisted bilayers of 1T-ZrS

We study the ground state and thermodynamic properties of the spin-half XXZ model, with an Ising interaction J

In recent years, strong-field physics in condensed-matter was pioneered as a novel approach for controlling material properties through laser-dressing, as well as for ultrafast spectroscopy via nonlinear light-matter interactions (e.g. harmonic generation). A potential controversy arising from these advancements is that it is sometimes vague which band-picture should be used to interpret strong-field experiments: the field-free bands, the adiabatic (instantaneous) field-dressed bands, Floquet bands, or some other intermediate picture. We here try to resolve this issue by performing 'theoretical experiments' of time- and angle-resolved photoelectron spectroscopy (Tr-ARPES) for a strong-field laser-pumped solid, which should give access to the actual observable bands of the irradiated material. To our surprise, we find that the adiabatic band-picture survives quite well, up to high field intensities ( 10

We compute the frequency dependent conductivity of the two dimensional square lattice Hubbard model at zero temperature as a function of density to second order in the interaction strength, and compare the results to the predictions of single-site dynamical mean field theory computed at the same order. We find that despite the neglect of vertex corrections, the single site dynamical mean field approximation produces semiquantitatively accurate results for most carrier concentrations, but fails qualitatively for the nearly empty or nearly filled band cases where the theory exhibits an emergent Galilean invariance. The theory also becomes qualitatively inaccurate very near half filling if nesting is important.

We studied the low-energy electronic response of the prototypical correlated metal SrVO