arXiv subjects

Xinyu Dai

Publications and source records attributed to Xinyu Dai.

At least 127 records · Page 7Linked to original sources

arXiv2017

Neural Machine Translation with Word Predictions

In the encoder-decoder architecture for neural machine translation (NMT), the hidden states of the recurrent structures in the encoder and decoder carry the crucial information about the sentence.These vectors are generated by parameters which are updated by back-propagation of translation errors through time. We argue that propagating errors through the end-to-end recurrent structures are not a direct way of control the hidden vectors. In this paper, we propose to use word predictions as a mechanism for direct supervision. More specifically, we require these vectors to be able to predict the vocabulary in target sentence. Our simple mechanism ensures better representations in the encoder and decoder without using any extra data or annotation. It is also helpful in reducing the target side vocabulary and improving the decoding efficiency. Experiments on Chinese-English and German-English machine translation tasks show BLEU improvements by 4.53 and 1.3, respectively

cs.CL↗

arXiv2017

Top-Rank Enhanced Listwise Optimization for Statistical Machine Translation

Pairwise ranking methods are the basis of many widely used discriminative training approaches for structure prediction problems in natural language processing(NLP). Decomposing the problem of ranking hypotheses into pairwise comparisons enables simple and efficient solutions. However, neglecting the global ordering of the hypothesis list may hinder learning. We propose a listwise learning framework for structure prediction problems such as machine translation. Our framework directly models the entire translation list's ordering to learn parameters which may better fit the given listwise samples. Furthermore, we propose top-rank enhanced loss functions, which are more sensitive to ranking errors at higher positions. Experiments on a large-scale Chinese-English translation task show that both our listwise learning framework and top-rank enhanced listwise losses lead to significant improvements in translation quality.

cs.CL↗

arXiv2017

Accelerating gravitational microlensing simulations using the Xeon Phi coprocessor

Recently Graphics Processing Units (GPUs) have been used to speed up very CPU-intensive gravitational microlensing simulations. In this work, we use the Xeon Phi coprocessor to accelerate such simulations and compare its performance on a microlensing code with that of NVIDIA's GPUs. For the selected set of parameters evaluated in our experiment, we find that the speedup by Intel's Knights Corner coprocessor is comparable to that by NVIDIA's Fermi family of GPUs with compute capability 2.0, but less significant than GPUs with higher compute capabilities such as the Kepler. However, the very recently released second generation Xeon Phi, Knights Landing, is about 5.8 times faster than the Knights Corner, and about 2.9 times faster than the Kepler GPU used in our simulations. We conclude that the Xeon Phi is a very promising alternative to GPUs for modern high performance microlensing simulations.

astro-ph.IM↗

arXiv2017

Extended X-ray Monitoring of Gravitational Lenses with Chandra and Joint Constraints on X-ray Emission Regions

We present an X-ray photometric analysis of six gravitationally lensed quasars, with observation campaigns spanning from 5 to 14 years, measuring the total (0.83 - 21.8 keV restframe), soft (0.83 - 3.6 keV), and hard (3.6 - 21.8 keV) band image flux ratios or each epoch. Using the ratios of the model-predicted macro-magnifications as baselines, we build differential microlensing light curves and obtain joint likelihood functions for the average X-ray emission region sizes. Our analysis yields a Probability Distribution Function for the average half-light radius of the X-Ray emission region in the sample that peaks slightly above 1 gravitational radius and with nearly indistinguishable 68% confidence (one-sided) upper limits of 17.8 and 18.9 gravitational radii for the soft and hard X-ray emitting regions, assuming a mean stellar mass of 0.3 solar masses. We see hints of energy dependent microlensing between the soft and hard bands in two of the objects. In a separate analysis on the root-mean-square of the microlensing variability, we find significant differences between the soft and hard bands but the sign of the difference is not consistent across the sample. This suggests the existence of some kind of spatial structure to the X-ray emission in an otherwise extremely compact source. We also discover a correlation between the RMS microlensing variability and the average microlensing amplitude.

astro-ph.HE↗

arXiv2016

Constraining Gamma-Ray Emission from Luminous Infrared Galaxies with Fermi-LAT; Tentative Detection of Arp 220

Star-forming galaxies produce gamma-rays primarily via pion production, resulting from inelastic collisions between cosmic ray protons and the interstellar medium (ISM). The dense ISM and high star formation rates of luminous and ultra-luminous infrared galaxies (LIRGs and ULIRGs) imply that they should be strong gamma-ray emitters, but so far only two LIRGs have been detected. Theoretical models for their emission depend on the unknown fraction of cosmic ray protons that escape these galaxies before interacting. We analyze Fermi-LAT data for 82 of the brightest IRAS LIRGs and ULIRGs. We examine each system individually and carry out a stacking analysis to constrain their gamma-ray fluxes. We report the detection of the nearest ULIRG Arp 220 (~4.6sigma). We observe a gamma-ray flux (0.8--100 GeV) of 2.4e-10 phot cm^-2 s^-1 with photon index of 2.23 (8.2e10^41 ergs s^-1 at 77 Mpc) We also derive upper limits for the stacked LIRGs and ULIRGs. The gamma-ray luminosity of Arp~220 and the stacked upper limits agree with calorimetric predictions for dense star-forming galaxies. With the detection of Arp 220, we extend the gamma-ray--IR luminosity correlation to the high luminosity regime with log(L_[0.1-100 GeV]) = 1.25log(L_[8-1000]) + 26.7 as well as the gamma-ray--radio continuum luminosity correlation with log(L_[0.1-100 GeV]) = 1.22log(L_[1.4 GHz]) + 13.3. The current survey of Fermi-LAT is on the verge of detecting more LIRGs/ULIRGs in the local universe, and we expect even more detections with deeper Fermi-LAT observations or the next generation of gamma-ray detectors.

astro-ph.HE↗

arXiv2016

The Swift AGN and Cluster Survey. II. Cluster Confirmation with SDSS Data

We study 203 (of 442) Swift AGN and Cluster Survey extended X-ray sources located in the SDSS DR8 footprint to search for galaxy over-densities in three dimensional space using SDSS galaxy photometric redshifts and positions near the Swift cluster candidates. We find 104 Swift clusters with a >3sigma galaxy over-density. The remaining targets are potentially located at higher redshifts and require deeper optical follow-up observations for confirmation as galaxy clusters. We present a series of cluster properties including the redshift, BCG magnitude, BCG-to-X-ray center offset, optical richness, and X-ray luminosity. We also detect red sequences in ~85% of the 104 confirmed clusters. The X-ray luminosity and optical richness for the SDSS confirmed Swift clusters are correlated and follow previously established relations. The distribution of the separations between the X-ray centroids and the most likely BCG is also consistent with expectation. We compare the observed redshift distribution of the sample with a theoretical model, and find that our sample is complete for z <~ 0.3 and is still 80% complete up to z ~= 0.4, consistent with the SDSS survey depth. These analysis results suggest that our Swift cluster selection algorithm has yielded a statistically well-defined cluster sample for further studying cluster evolution and cosmology. We also match our SDSS confirmed Swift clusters to existing cluster catalogs, and find 42, 23 and 1 matches in optical, X-ray and SZ catalogs, respectively, so the majority of these clusters are new detections.

astro-ph.CO↗

arXiv2016

The optical-UV emissivity of quasars: dependence on black hole mass and radio loudness

We analyzed a large sample of radio-loud and radio-quiet quasar spectra at redshift 1.0 < z < 1.2 to compare the inferred underlying quasar continuum slopes (after removal of the host galaxy contribution) with accretion disk models. The latter predict redder (decreasing) alpha_3000 continuum slopes (L_ν~ν^alpha at 3000Ang) with increasing black hole mass, bluer alpha_3000 with increasing luminosity at 3000Ang, and bluer alpha_3000 with increasing spin of the black hole, when all other parameters are held fixed. We find no clear evidence for any of these predictions in the data. In particular we find that: (i) alpha_3000 shows no significant dependence on black hole mass or luminosity. Dedicated Monte Carlo tests suggest that the substantial observational uncertainties in the black hole virial masses can effectively erase any intrinsic dependence of alpha_3000 on black hole mass, in line with some previous studies. (ii) The mean slope alpha_3000 of radio-loud sources, thought to be produced by rapidly spinning black holes, is comparable to, or even redder than, that of radio-quiet quasars. Indeed, although quasars appear to become more radio loud with decreasing luminosity, we still do not detect any significant dependence of alpha_3000 on radio loudness. The predicted mean alpha_3000 slopes tend to be bluer than in the data. Disk models with high inclinations and dust extinction tend to produce redder slopes closer to empirical estimates. Our mean alpha_3000 values are close to the ones independently inferred at z<0.5 suggesting weak evolution with redshift, at least for moderately luminous quasars.

astro-ph.GA↗

arXiv2015

A probable Milli-Parsec Supermassive Binary Black Hole in the Nearest Quasar Mrk 231

Supermassive binary black holes (BBHs) are unavoidable products of galaxy mergers and are expected to exist in the cores of many quasars. Great effort has been made during the past several decades to search for BBHs among quasars; however, observational evidence for BBHs remains elusive and ambiguous, which is difficult to reconcile with theoretical expectations. In this paper, we show that the distinct optical-to-UV spectrum of Mrk 231 can be well interpreted as emission from accretion flows onto a BBH, with a semimajor axis of ~590AU and an orbital period of ~1.2 year. The flat optical and UV continua are mainly emitted from a circumbinary disk and a mini-disk around the secondary black hole (BH), respectively; and the observed sharp drop off and flux deficit at wavelength lambda ~ 4000-2500 Angstrom is due to a gap (or hole) opened by the secondary BH migrating within the circumbinary disk. If confirmed by future observations, this BBH will provide a unique laboratory to study the interplay between BBHs and accretion flows onto them. Our result also demonstrates a new method to find sub-parsec scale BBHs by searching for deficits in the optical-to-UV continuum among the spectra of quasars.

astro-ph.HE↗

arXiv2015

Algorithms And Programs For Strong Gravitational Lensing In Kerr Space-time Including Polarization

Active galactic nuclei (AGNs) and quasars are important astrophysical objects to understand. Recently, microlensing observations have constrained the size of the quasar X-ray emission region to be of the order of 10 gravitational radii of the central supermassive black hole. For distances within a few gravitational radii, light paths are strongly bent by the strong gravity field of the central black hole. If the central black hole has nonzero angular momentum (spin), a photon's polarization plane will be rotated by the gravitational Faraday effect. The observed X-ray flux and polarization will then be influenced significantly by the strong gravity field near the source. Consequently, linear gravitational lensing theory is inadequate for such extreme circumstances. We present simple algorithms computing strong lensing effects of Kerr black holes, including effects on polarization. Our algorithms are realized in a program "KERTAP" in two versions: MATLAB and Python. The key ingredients of KERTAP are: a graphic user interface, a {\it backward} ray-tracing algorithm, a polarization propagator dealing with gravitational Faraday rotation, and algorithms computing observables such as flux magnification and polarization angles. Our algorithms can be easily realized in other programming languages such as FORTRAN, C, and C++. The MATLAB version of KERTAP is parallelized using the MATLAB Parallel Computing Toolbox and the Distributed Computing Server. The Python code was sped up using Cython and supports full implementation of MPI using 'mpi4py' package. As an example, we investigate the inclination angle dependence of the observed polarization and the strong lensing magnification of AGN X-ray emission. We conclude that it is possible to perform complex numerical-relativity-related computations using interpreted languages such as MATLAB and Python.

astro-ph.HE↗

arXiv2015

An integrated Sachs-Wolfe effect vs redshift test for the cosmological parameters

We describe a method using the integrated Sachs-Wolfe (ISW) effect caused by individual inhomogeneities to determine the cosmological parameters, $H_0$, $Ω_{\rm m}$, and $Ω_Λ$, etc. This ISW-redshift test requires detailed knowledge of the internal kinematics of a set of individual density perturbations, e.g., galaxy clusters and/or cosmic voids, in particular their density and velocity profiles, and their mass accretion rates. It assumes the density perturbations are isolated and imbedded (equivalently compensated) and makes use of the newly found relation between the ISW temperature perturbation of the CMB and the Fermat potential of the lens. Given measurements of the amplitudes of the temperature variations in the CMB caused by such clusters or voids at various redshifts and estimates of their angular sizes or masses, one can constrain the cosmological parameters. More realistically, the converse is more likely, i.e., if the background cosmology is sufficiently constrained, measurement of ISW profiles of clusters and voids (e.g., hot and cold spots and rings) can constrain dynamical properties of the dark matter, including accretion, associated with such lenses and thus constrain the evolution of these objects with redshift.

astro-ph.CO↗

arXiv2015

Embedded Lensing Time Delays, the Fermat Potential, and the Integrated Sachs-Wolfe Effect

We derive the Fermat potential for a spherically symmetric lens embedded in an FLRW cosmology and use it to investigate the late-time integrated Sachs-Wolfe (ISW) effect, i.e., secondary temperature fluctuations in the cosmic microwave background (CMB) caused by individual large scale clusters and voids. We present a simple analytical expression for the temperature fluctuation in the CMB across such a lens as a derivative of the lens' Fermat potential. This formalism is applicable to both linear and nonlinear density evolution scenarios, to arbitrarily large density contrasts, and to all open and closed background cosmologies. It is much simpler to use and makes the same predictions as conventional approaches. In this approach the total temperature fluctuation can be split into a time-delay part and an evolutionary part. Both parts must be included for cosmic structures that evolve and both can be equally important. We present very simple ISW models for cosmic voids and galaxy clusters to illustrate the ease of use of our formalism. We use the Fermat potentials of simple cosmic void models to compare predicted ISW effects with those recently extracted from WMAP and \emph{Planck} data by stacking large cosmic voids using the aperture photometry method. If voids in the local universe with large density contrasts are no longer evolving we find that the time delay contribution alone predicts values consistent with the measurements. However, we find that for voids still evolving linearly, the evolutionary contribution cancels a significant part of the time delay contribution and results in predicted signals that are much smaller than recently observed.

astro-ph.CO↗

arXiv2015

A Simple Gravitational Lens Model For Cosmic Voids

We present a simple gravitational lens model to illustrate the ease of using the embedded lensing theory when studying cosmic voids. It confirms the previously used repulsive lensing models for deep voids. We start by estimating magnitude fluctuations and weak lensing shears of background sources lensed by large voids. We find that sources behind large ($\sim$$90\,\rm Mpc$) and deep voids (density contrast about $-0.9$) can be magnified or demagnified with magnitude fluctuations of up to $\sim$$0.05\,\rm mag$ and that the weak-lensing shear can be up to the $\sim$$10^{-2}$ level in the outer regions of large voids. Smaller or shallower voids produce proportionally smaller effects. We investigate the "wiggling" of the primary cosmic microwave background (CMB) temperature anisotropies caused by intervening cosmic voids. The void-wiggling of primary CMB temperature gradients is of the opposite sign to that caused by galaxy clusters. Only extremely large and deep voids can produce wiggling amplitudes similar to galaxy clusters, $\sim$$\rm 15\,μK$ by a large void of radius $\sim$$4^\circ$ and central density contrast $-0.9$ at redshift 0.5 assuming a CMB background gradient of $\sim$$\rm10\,μK\, arcmin^{-1}$. The dipole signal is spread over the entire void area, and not concentrated at the lens' center as it is for clusters. Finally we use our model to simulate CMB sky maps lensed by large cosmic voids. Our embedded theory can easily be applied to more complicated void models and used to study gravitational lensing of the CMB, to probe dark-matter profiles, to reduce the lensing-induced systematics in supernova Hubble diagrams, as well as study the integrated Sachs-Wolfe effect.

astro-ph.CO↗

arXiv2015

The SWIFT AGN and Cluster Survey I: Number Counts of AGN and Galaxy Clusters

The Swift AGN and Cluster Survey (SACS) uses 125 deg^2 of Swift XRT serendipitous fields with variable depths surrounding gamma-ray bursts to provide a medium depth (4e-15 erg/s/cm^2) and area survey filling the gap between deep, narrow Chandra/XMM-Newton surveys and wide, shallow ROSAT surveys. Here we present a catalog of 22,563 point sources and 442 extended sources and examine the number counts of the AGN and galaxy cluster populations. SACS provides excellent constraints on the AGN number counts at the bright end with negligible uncertainties due to cosmic variance, and these constraints are consistent with previous measurements. We use Wise mid-infrared (MIR) colors to classify the sources. For AGN we can roughly separate the point sources into MIR-red and MIR-blue AGN, finding roughly equal numbers of each type in the soft X-ray band (0.5-2 keV), but fewer MIR-blue sources in the hard X-ray band (2-8 keV). The cluster number counts, with 5% uncertainties from cosmic variance, are also consistent with previous surveys but span a much larger continuous flux range. Deep optical or IR follow-up observations of this cluster sample will significantly increase the number of higher redshift (z > 0.5) X-ray-selected clusters.

astro-ph.CO↗

arXiv2015

Non-linear Learning for Statistical Machine Translation

Modern statistical machine translation (SMT) systems usually use a linear combination of features to model the quality of each translation hypothesis. The linear combination assumes that all the features are in a linear relationship and constrains that each feature interacts with the rest features in an linear manner, which might limit the expressive power of the model and lead to a under-fit model on the current data. In this paper, we propose a non-linear modeling for the quality of translation hypotheses based on neural networks, which allows more complex interaction between features. A learning framework is presented for training the non-linear models. We also discuss possible heuristics in designing the network structure which may improve the non-linear learning performance. Experimental results show that with the basic features of a hierarchical phrase-based machine translation system, our method produce translations that are better than a linear model.

cs.CL↗

arXiv2015

Unifying X-ray Scaling Relations from Galaxies to Clusters

We examine a sample of $\sim 250 000$ "locally brightest galaxies" selected from the Sloan Digital Sky Survey to be central galaxies within their dark matter halos. These were previously stacked by the Planck Collaboration to measure the Sunyaev-Zel'dovich signal as a function of central galaxy stellar mass. Here, we stack the X-ray emission from these halos using data from the ROSAT All-Sky Survey. We detect emission across almost our entire sample, including emission which we attribute to hot gas around galaxies spanning a range of 1.2 dex in stellar mass (corresponding to nearly two orders of magnitude in halo mass) down to $M* = 10^{10.8} M_{\odot}$ ($M_{500} \approx 10^{12.6} M_{\odot}$). Over this range, the X-ray luminosity can be fit by a power-law, either of stellar mass or of halo mass. A single unified scaling relation between mass and $L_X$ applies for galaxies, groups, and clusters. This relation has a steeper slope than expected for self-similarity, in contrast to the $Y_{SZ}$-$M_{500}$ relation, showing the importance of non-gravitational heating. If this heating is predominantly due to AGN feedback, the lack of a break in our relation suggests that AGN feedback is tightly self-regulated and fairly gentle, in agreement with recent high-resolution simulations. Our results are consistent with earlier measurements of the $L_X$-$L_K$ relation for elliptical galaxies and of the $L_X$-$M_{500}$ relation for optically-selected galaxy clusters. However, our $L_X$-$M_{500}$ relation lies more than a factor of two below most previous relations based on X-ray-selected cluster samples. We argue that optical selection offers a less biased view of the $L_X$-$M_{500}$ relation.

astro-ph.CO↗

arXiv2014

New Limits On Gamma-Ray Emission From Galaxy Clusters

Galaxy clusters are predicted to produce gamma-rays through cosmic ray interactions and/or dark matter annihilation, potentially detectable by the Fermi Large Area Telescope (Fermi-LAT). We present a new, independent stacking analysis of Fermi-LAT photon count maps using the 78 richest nearby clusters (z<0.12) from the Two Micron All-Sky Survey (2MASS) cluster catalog. We obtain the lowest limit on the photon flux to date, 2.3e-11 ph/s/cm^2 (95% confidence) per cluster in the 0.8-100 GeV band, which corresponds to a luminosity limit of 3.5e44 ph/s. We also constrain the emission limits in a range of narrower energy bands. Scaling to recent cosmic ray acceleration and gamma-ray emission models, we find that cosmic rays represent a negligible contribution to the intra-cluster energy density and gas pressure.

astro-ph.HE↗

arXiv2014

The Optical, Ultraviolet, and X-ray Structure of the Quasar HE 0435-1223

Microlensing has proven an effective probe of the structure of the innermost regions of quasars, and an important test of accretion disk models. We present light curves of the lensed quasar HE 0435-1223 in the R band and in the ultraviolet, and consider them together with X-ray light curves in two energy bands that are presented in a companion paper. Using a Bayesian Monte Carlo method, we constrain the size of the accretion disk in the rest-frame near- and far-UV, and constrain for the first time the size of the X-ray emission regions in two X-ray energy bands. The R-band scale size of the accretion disk is about 10^15.23 cm (~23 r_g), slightly smaller than previous estimates, but larger than would be predicted from the quasar flux. In the UV, the source size is weakly constrained, with a strong prior dependence. The UV to R-band size ratio is consistent with the thin disk model prediction, with large error bars. In soft and hard X-rays, the source size is smaller than ~10^14.8 cm (~10 r_g) at 95% confidence. We do not find evidence of structure in the X-ray emission region, as the most likely value for the ratio of the hard X-ray size to the soft X-ray size is unity. Finally, we find that the most likely value for the mean mass of stars in the lens galaxy is ~0.3 M_sun, consistent with other studies.

astro-ph.CO↗

arXiv2014

Unveiling the Intrinsic X-ray Properties of Broad Absorption Line Quasars with a Relatively Unbiased Sample

There is growing evidence of a higher intrinsic fraction of broad absorption line quasars (BALQSOs) than that obtained in optical surveys, on which most previous X-ray studies of BALQSOs have focused. Here we present Chandra observations of 18 BALQSOs at $z\sim2$, selected from a near-infrared (2MASS) sample, where the BALQSO fraction is likely to be close to the intrinsic fraction. We measure photon indices using the stacked spectra of the optically-faint ($i-K_s\geq 2.3$ mag) and optically-bright ($i-K_s < 2.3$ mag) samples to be $Γ\simeq 1.5$--$2.1$. We constrain their intrinsic column density by modelling the X-ray fractional hardness ratio, finding a mean column density of $3.5\times10^{22}$ \cmsq\ assuming neutral absorption. We incorporate SDSS optical measurements (rest frame UV) to study the broadband spectral index between the X-ray and UV bands, and compare this to a large sample of normal quasars. We estimate that the optically-faint BALQSOs are X-ray weaker than the optically-bright ones, and the entire sample of BALQSOs are intrinsically X-ray weak when compared to normal AGN. Correcting for magnification of X-ray emission via gravitational lensing by the central black hole viewed at large inclination angles makes these BALQSOs even more intrinsically X-ray weak. Finally, we estimate AGN kinetic feedback efficiencies of a few percent for an X-ray wind of $0.3c$ in high-ionization BALQSOs. Combined with energy carried by low-ionization BALQSOs and UV winds, the total kinetic energy in BALQSOs can be sufficient to provide AGN kinetic feedback required to explain the co-evolution between black holes and host galaxies.

astro-ph.HE↗