Publications

We present data structures and algorithms for native implementations of discrete convolution operators over Adaptive Particle Representations (APR) of images on parallel computer architectures. The APR is a content-adaptive image representation that locally adapts the sampling resolution to the image signal. It has been developed as an alternative to pixel representations for large, sparse images as they typically occur in fluorescence microscopy. It has been shown to reduce the memory and runtime costs of storing, visualizing, and processing such images. This, however, requires that image processing natively operates on APRs, without intermediately reverting to pixels. Designing efficient and scalable APR-native image processing primitives, however, is complicated by the APR’s irregular memory structure. Here, we provide the algorithmic building blocks required to efficiently and natively process APR images using a wide range of algorithms that can be formulated in terms of discrete convolutions. We show that APR convolution naturally leads to scale-adaptive algorithms that efficiently parallelize on multi-core CPU and GPU architectures. We quantify the speedups in comparison to pixel-based algorithms and convolutions on evenly sampled data. We achieve pixel-equivalent throughputs of up to 1TB/s on a single Nvidia GeForce RTX 2080 gaming GPU, requiring up to two orders of magnitude less memory than a pixel-based implementation.

We present an efficient basis for imaginary time Green's functions based on a low rank decomposition of the spectral Lehmann representation. The basis functions are simply a set of well-chosen exponentials, so the corresponding expansion may be thought of as a discrete form of the Lehmann representation using an effective spectral density which is a sum of δ functions. The basis is determined only by an upper bound on the product βωmax, with β the inverse temperature and ωmax an energy cutoff, and a user-defined error tolerance ϵ. The number r of basis functions scales as (log(βωmax)log(1/ϵ)). The discrete Lehmann representation of a particular imaginary time Green's function can be recovered by interpolation at a set of r imaginary time nodes. Both the basis functions and the interpolation nodes can be obtained rapidly using standard numerical linear algebra routines. Due to the simple form of the basis, the discrete Lehmann representation of a Green's function can be explicitly transformed to the Matsubara frequency domain, or obtained directly by interpolation on a Matsubara frequency grid. We benchmark the efficiency of the representation on simple cases, and with a high precision solution of the Sachdev-Ye-Kitaev equation at low temperature. We compare our approach with the related intermediate representation method, and introduce an improved algorithm to build the intermediate representation basis and a corresponding sampling grid.

This case study is a reimplementation of the algorithm described in Barmherzig and Sun (2022) [1] as a Stan model. This requires the new features available in Stan 2.30

We build a field level emulator for cosmic structure formation that is accurate in the nonlinear regime. Our emulator consists of two convolutional neural networks trained to output the nonlinear displacements and velocities of N-body simulation particles based on their linear inputs. Cosmology dependence is encoded in the form of style parameters at each layer of the neural network, enabling the emulator to effectively interpolate the outcomes of structure formation between different flat ΛCDM cosmologies over a wide range of background matter densities. The neural network architecture makes the model differentiable by construction, providing a powerful tool for fast field level inference. We test the accuracy of our method by considering several summary statistics, including the density power spectrum with and without redshift space distortions, the displacement power spectrum, the momentum power spectrum, the density bispectrum, halo abundances, and halo profiles with and without redshift space distortions. We compare these statistics from our emulator with the full N-body results, the COLA method, and a fiducial neural network with no cosmological dependence. We find our emulator gives accurate results down to scales of $$k ∼ 1 Mpc −1 h,$$ representing a considerable improvement over both COLA and the fiducial neural network. We also demonstrate that our emulator generalizes well to initial conditions containing primordial non-Gaussianity, without the need for any additional style parameters or retraining.

Kidney Precision Medicine Project (KPMP) is building a spatially specified human kidney tissue atlas in health and disease with single-cell resolution. Here, we describe the construction of an integrated reference map of cells, pathways, and genes using unaffected regions of nephrectomy tissues and undiseased human biopsies from 56 adult subjects. We use single-cell/nucleus transcriptomics, subsegmental laser microdissection transcriptomics and proteomics, near-single-cell proteomics, 3D and CODEX imaging, and spatial metabolomics to hierarchically identify genes, pathways, and cells. Integrated data from these different technologies coherently identify cell types/subtypes within different nephron segments and the interstitium. These profiles describe cell-level functional organization of the kidney following its physiological functions and link cell subtypes to genes, proteins, metabolites, and pathways. They further show that messenger RNA levels along the nephron are congruent with the subsegmental physiological activity. This reference atlas provides a framework for the classification of kidney disease when multiple molecular mechanisms underlie convergent clinical phenotypes.

In this paper, we introduce a randomized iterative method for signal detection in uplink large-scale multiple-input multiple-output (MIMO) systems, which not only achieves a low computational complexity but also enjoys a global and exponentially fast convergence. First of all, by adopting the random sampling into the iterations, the randomized iterative detection algorithm (RIDA) is proposed for large-scale MIMO systems. We show that RIDA converges exponentially fast in terms of mean squared error (MSE). Furthermore, this global convergence always holds, and does not depend on the standard requirements such as N≫K , where N and K denote the numbers of antennas at the sides of base station and users. This broadly extends the applications of low-complexity detection in uplink large-scale MIMO systems. Then, based on a new conditional sampling, optimization and enhancements are given to further improve both the convergence and efficiency of RIDA, resulting in the modified randomized iterative detection algorithm (MRIDA). Meanwhile, with respect to MRIDA, further complexity reduction by exploiting the matrix structure is given while its implementation by deep neural networks (DNN) is also presented for a better detection performance.

Chemical abundances of stars in the Milky Way disk are empirical tracers of its enrichment history. However, they capture joint-information that is valuable to disentangle. In this work, we seek to quantify how individual abundances evolve across the present-day radius of the disk, at fixed supernovae contribution ([Fe/H], [Mg/Fe]). We use 18,135 APOGEE DR17 red clump stars and 7,943 GALAH DR3 main sequence stars to compare the abundance distributions conditioned on ([Fe/H], [Mg/Fe]) across 3−13 kpc and 6.5−9.5 kpc, respectively. In total we examine 15 elements: C, N, Al, K (light), O, Si, S, Ca, (α), Mn, Ni, Cr, Cu, (iron-peak) Ce, Ba (s-process) and Eu (r-process). We find that the conditional neutron capture and light elements most significantly trace variations in the disk's enrichment history, with absolute conditional radial gradients ≤0.03 dex/kpc. The other elements studied have absolute conditional gradients ≲0.01 dex/kpc. We uncover structured conditional abundance variations as a function of [Fe/H] for the low-α, but not the high-α sequence. The average scatter between the mean conditional abundances at different radii is σintrinsic≈ 0.02 dex (with Ce, Eu, Ba σintrinsic> 0.05 dex). These results serve as a measure of the magnitude via which different elements trace Galactic radial enrichment history once fiducial supernovae correlations are accounted for. Furthermore, we uncover subtle systematic variations in all moments of the conditional abundance distributions that will presumably constrain chemical evolution models of the Galaxy.

Extended, old, and round stellar halos appear to be ubiquitous around high-mass dwarf galaxies (108.5 < M⋆/M⊙ < 109.6) in the observed universe. However, it is unlikely that these dwarfs have undergone a sufficient number of minor mergers to form stellar halos that are composed of predominantly accreted stars. Here, we demonstrate that FIRE-2 (Feedback in Realistic Environments) cosmological zoom-in simulations are capable of producing dwarf galaxies with realistic structures, including both a thick disk and round stellar halo. Crucially, these stellar halos are formed in situ, largely via the outward migration of disk stars. However, there also exists a large population of "nondisky" dwarfs in FIRE-2 that lack a well-defined disk/halo and do not resemble the observed dwarf population. These nondisky dwarfs tend to be either more gas-poor or to have burstier recent star formation histories than the disky dwarfs, suggesting that star formation feedback may be preventing disk formation. Both classes of dwarfs underscore the power of a galaxy's intrinsic shape—which is a direct quantification of the distribution of the galaxy's stellar content—to interrogate the feedback implementation in simulated galaxies.

We derive efficient, closed-form, differentiable, and numerically stable solutions for the flux measured from a spherical planet or moon seen in reflected light, either in or out of occultation. Our expressions apply to the computation of scattered light phase curves of exoplanets, secondary eclipse light) curves in the optical, or future measurements of planet–moon and planet–planet occultations, as well as to photometry of solar system bodies. We derive our solutions for Lambertian bodies illuminated by a point source, but extend them to model illumination sources of finite angular size and rough surfaces with phase-dependent scattering. Our algorithm is implemented in Python within the open-source starry mapping framework and is designed with efficient gradient-based inference in mind. The algorithm is ∼4–5 orders of magnitude faster than direct numerical evaluation methods and ∼10 orders of magnitude more precise. We show how the techniques developed here may one day lead to the construction of two-dimensional maps of terrestrial planet surfaces, potentially enabling the detection of continents and oceans on exoplanets in the habitable zone.

We present the first results from an all-sky all-frequency (ASAF) search for an anisotropic stochastic gravitational-wave background using the data from the first three observing runs of the Advanced LIGO and Advanced Virgo detectors. Upper limit maps on broadband anisotropies of a persistent stochastic background were published for all observing runs of the LIGO-Virgo detectors. However, a broadband analysis is likely to miss narrowband signals as the signal-to-noise ratio of a narrowband signal can be significantly reduced when combined with detector output from other frequencies. Data folding and the computationally efficient analysis pipeline, {\tt PyStoch}, enable us to perform the radiometer map-making at every frequency bin. We perform the search at 3072 {\tt{HEALPix}} equal area pixels uniformly tiling the sky and in every frequency bin of width 1/32~Hz in the range 20−1726~Hz, except for bins that are likely to contain instrumental artefacts and hence are notched. We do not find any statistically significant evidence for the existence of narrowband gravitational-wave signals in the analyzed frequency bins. Therefore, we place 95% confidence upper limits on the gravitational-wave strain for each pixel-frequency pair, the limits are in the range (0.030−9.6)×10−24. In addition, we outline a method to identify candidate pixel-frequency pairs that could be followed up by a more sensitive (and potentially computationally expensive) search, e.g., a matched-filtering-based analysis, to look for fainter nearly monochromatic coherent signals. The ASAF analysis is inherently independent of models describing any spectral or spatial distribution of power. We demonstrate that the ASAF results can be appropriately combined over frequencies and sky directions to successfully recover the broadband directional and isotropic results.