Published on 16 May 2024
Mapping cosmological models from standard to modified gravity using bias relations
This study investigates using a non-parametric bias model to generate mock halo catalogs for modified gravity cosmologies. Either the modified gravity or standard LambdaCDM matter distribution is relied on. The method aims to effectively capture distinct impacts of modified gravity with high accuracy in 2- and 3-point statistics. Experiments are designed to map modified gravity effects either using the consistent modified gravity density fields, or using the LambdaCDM density field, which could save computational costs. Using 7 high-resoluti...
Published on 16 May 2024
Phonon magnetization in Dirac materials
This paper develops a quantum theory to calculate phonon magnetic moments in Dirac semimetals. The key finding is that phonon magnetic moments can be enhanced by increasing electrical conductivity. First-principles calculations show moments on the order of a Bohr magneton in graphene and Cd3As2. This provides guidance to dynamically generate large magnetization in quantum materials.
Published on 16 May 2024
Visual transformation through image analogy
This paper proposes an inference-based method called Analogist that performs a wide range of visual tasks through image analogy, using a single example pair. It works by prompting a pretrained image diffusion model with carefully designed visual and textual cues during inference. Key benefits are flexibility, efficiency, and superior performance, without needing extra training or optimization.
Published on 16 May 2024
Creating 3D Scenes from Images
CAT3D is a two-step method that uses a multi-view diffusion model to generate consistent novel views of a scene from input images. These views are fed into a robust 3D reconstruction pipeline to create a 3D representation that can be rendered from any viewpoint.
Published on 15 May 2024
Simulation tools for customizable computer vision data
This paper introduces a simulation toolkit that allows generating customizable synthetic data to systematically evaluate computer vision models. It supports adjusting various parameters like scene layout, lighting, object models and poses, camera settings to create controlled experiments. Example uses include assessing model robustness and capabilities on the same images, and facilitating simulation-to-real transfer.
Published on 15 May 2024
Simplified spectral complexity of deep neural networks
This paper proposes using the angular power spectrum of deep neural networks to characterize their complexity. It defines sequences of random variables based on the spectrum, and shows how their asymptotic distribution as depth increases can classify networks as low-disorder, sparse, or high-disorder. For example, ReLU networks exhibit sparsity. Numerical simulations validate the theoretical results.
Published on 15 May 2024
Singular elliptic operators in the half-space: Dirichlet and oblique derivative boundary conditions
This paper studies elliptic and parabolic differential equations involving singular elliptic operators in the upper half-space under Dirichlet or oblique derivative boundary conditions. A special case is when the operator takes the form of a weighted trace of a second derivative plus lower order terms. The authors prove well-posedness results, characterize the domain using weighted Sobolev spaces, and establish maximal regularity.
Published on 15 May 2024
Multi-modality diffusion model diagnoses lymph node metastasis
This paper introduces a new multi-modality diffusion model to diagnose lymph node metastasis in esophageal cancer, using CT scans, clinical data, and radiomics. It constructs a heterogeneous graph to explore relationships between data types, then uses a conditional diffusion process to reduce redundancy and uncover prognostic correlations between tumor and lymph node images.
Published on 15 May 2024
Phonon-induced magnetization in strontium titanate
This paper presents a theoretical model explaining the experimentally observed magnetization in strontium titanate after excitation by circularly polarized terahertz light. The model is based on coupling between electrons and the material's infrared-active phonon modes, enabling transfer of angular momentum from the phonons to the electronic system, thus inducing a magnetization even without an external magnetic field present.
Published on 15 May 2024
Realizing supersolidity in ultracold quantum gases
This paper reviews the realization of supersolidity, an exotic state combining superfluidity and crystallinity, in ultracold quantum gases. Key achievements include observing density modulations indicating spontaneous translational symmetry breaking and signatures of persistence of coherence. Outstanding questions remain regarding observing quantum vortices in these novel crystals, elucidating dimensionality effects, further probing the crystalline nature via sound propagation studies, and elucidating the nature of phase transitions to the s...
Published on 15 May 2024
Boosting for probabilistic prediction via Wasserstein gradient
This paper proposes Wasserstein Gradient Boosting, a new framework that harnesses gradient boosting to produce a distributional estimate that approximates the posterior distribution over a model's output parameter. It works by fitting base learners to target an approximate Wasserstein gradient. The method is shown to enable superior probabilistic prediction compared to existing methods.
Published on 15 May 2024
Forward and Far-Forward Heavy Hadron Production: A High-Energy Viewpoint
This paper reviews studies on the production of light and heavy hadrons at forward and far-forward rapidities at the LHC. It utilizes a resummation technique called JETHAD to analyze rapidity rates and angular multiplicities. It explores reach of current LHC detectors and potential future Forward Physics Facilities, using a precise timing coincidence to detect a far-forward light/heavy hadron with a central particle at the LHC. The goal is assessing the impact of high-energy resummation beyond fixed-order calculations.
Published on 15 May 2024
Classical shadows via rank-dependent measurements
This paper presents a new measurement protocol for the classical shadows problem that requires fewer copies of an unknown quantum state to estimate observables. The sample complexity scales as the square root of the rank of the state times the observable bound, achieving nearly quadratic savings over previous methods in low-rank cases. Key innovations include extending a 'nice' Schur basis to higher dimensions, introducing a population classical shadows formulation, and using local symmetries to simplify calculations.
Published on 15 May 2024
Plasma water purification eliminates bacteria
This paper discusses a reactor that uses plasma glow discharge in water to create oxidants that eliminate bacteria colonies, as an alternative to traditional water purification methods involving chemicals. Materials, design, and testing are covered.
Published on 15 May 2024
Semi-supervised estimation of discrete distributions
This paper studies the problem of estimating a joint distribution over two dependent random variables, using both complete samples containing both variables, and incomplete samples missing one variable. It shows that a simple composition of univariate minimax estimators is optimal, achieving the minimax risk for both lp losses with 1 ≤ p ≤ 2, and for f-divergences like KL, χ2, Squared Hellinger and Le Cam.
Published on 15 May 2024
Learning to resolve garment intersections with neural networks
This paper presents a new method called ContourCraft that enables neural networks to simulate complex, multilayered clothing. It can reliably handle and resolve garment intersections that occur due to collisions or errors in outfit design. This is achieved through a novel 'intersection contour loss' that penalizes and encourages the network to resolve penetrations.
Published on 15 May 2024
Declarative neural-symbolic reasoning
The paper proposes a framework to make neural-symbolic systems fully declarative, allowing them to answer arbitrary queries without retraining. This is achieved by designing neural predicates to learn prototypes instead of mappings, separating instance generation from inference, and introducing a relational interpretation of the encoding-decoding scheme.
Published on 15 May 2024
Fluctuations in subcritical stochastic heat equations
This paper studies 2D stochastic heat equations with subcritical coupling, showing solutions exhibit Edwards-Wilkinson fluctuations asymptotically. Part of fluctuation comes from the original noise, the rest independent.
Published on 15 May 2024
Cost-benefit analysis for condition monitoring maintenance strategies for unmanned systems
This paper proposes a method to conduct cost-benefit analysis to determine the return on investment of potential condition-based maintenance strategies for unmanned systems. It uses modular dynamic fault tree analysis and Monte Carlo simulations to model an unmanned surface vessel's maintenance requirements under different strategies. Comparing maintenance cost differences and investment costs provides the basis to evaluate and select effective condition monitoring approaches for these complex, multi-component systems.
Published on 15 May 2024
Nonperturbative pion form factor at high energies
This paper constructs a model of the pion electromagnetic form factor at high energies that accounts for the pion wave function, gluonic exchanges, and quark Reggeization effects. It finds these nonperturbative dynamics can be probed by studying violations of asymptotic scaling predictions.
Published on 15 May 2024
Generalization guarantees for causal machine learning models
This paper proposes a theory using generalization bounds to provide quality guarantees for causal machine learning models, which currently lack supporting theory. A key innovation is introducing a change-of-measure inequality to tightly bound model loss based on treatment propensity deviation. Bounds hold even with hidden confounding and lack of positivity. Demonstrated remarkable tightness and utility of bounds on semi-synthetic and real datasets.
Published on 15 May 2024
Antiferromagnetic correlations enable pseudogap and alter charge density wave in strongly correlated electrons
This paper explores how short-range antiferromagnetic correlations impact the charge density wave phase and pseudogap phenomenon in strongly correlated electron systems. Using a Hubbard model and approximations, they find increased repulsive Coulomb interactions enhance these antiferromagnetic fluctuations, flattening a band near an anti-nodal point. This manifests as a pseudogap, prompting analysis across interaction and occupation values. Results show correlations significantly reshape the charge density wave state, reconstructing the Ferm...
Published on 15 May 2024
Task-oriented communication with distribution shift robustness
This paper proposes a novel approach to make task-oriented communication systems robust to unpredictable distribution shifts between training and deployment data. It extracts compact features with high capability for domain-shift generalization and semantic-shift detection. An invariant feature encoding method based on information bottleneck and invariant risk minimization achieves effective generalization. A label-dependent feature encoding approach further enables accurate semantic-shift detection.
Published on 15 May 2024
Gravitational waves from neutron star mergers using supervised learning
This paper develops a time-domain model to predict gravitational waves from equal-mass neutron star merger remnants using numerical relativity simulations and K-nearest neighbor regression. A large dataset of 157 simulations with the APR4 equation of state is produced. Models are built using training sets of 20 to 100 simulations and evaluated based on accuracy metrics when tested on an additional 30 simulations. Models with 40+ training simulations achieve high faithfulness scores around 0.95. Injection studies also demonstrate accurate rec...
Published on 15 May 2024
Voltage-induced transition from Mott insulator to metal
This paper examines how applying a voltage across a Mott insulator can cause it to transition into a metal. The model involves a chain of interacting spinless fermions connected to metal leads. At low to medium interaction strength, increasing voltage induces a conducting charge density wave state which transitions into a disordered metal. At high interaction strength, the system transitions into an insulating charge-separated state before becoming metallic. Features like hysteresis, sharp current onset, and negative differential conductivit...
Published on 15 May 2024
Bagging for stable statistical learning
This paper studies using bagging, a model averaging technique, to improve stability of statistical algorithms. Bagging trains models on resampled subsets of the data and averages predictions. The authors prove bagging provides finite-sample guarantees on stability when algorithm outputs are vectors, functions, or distributions, not just real numbers. Experiments confirm bagging stabilizes causal inference, Bayesian inference, regression, and classification methods.
Published on 15 May 2024
Simplified title focusing on key contributions
This paper explores instrumental variable models where the instrument, treatment, and outcome variables are all categorical, making important theoretical contributions. The authors provide a full characterization of constraints on the joint distribution of potential outcomes, study how potential outcome margins vary dependently, and assess compatibility of observed data with IV assumptions.
Published on 15 May 2024
Simplified Title Focusing on Double Robustness of Local Projections
This paper shows that conventional local projection (LP) confidence intervals for impulse responses are surprisingly robust to model misspecification, even when that misspecification is large. This robustness stems from a 'double robustness' property, analogous to that of some modern regression estimators. In contrast, standard vector autoregression (VAR) intervals can severely undercover even for small misspecification. Restoring coverage requires VAR intervals as wide as LPs.