arXiv subjects

Yashar Hezaveh

Publications and source records attributed to Yashar Hezaveh.

At least 19 recordsLinked to original sources

arXiv2026

POSTELLAR: Posterior Stellar Spectrum Sampling - An Alternative to Approximate Stellar Spectra for Exoplanetary Analysis

We present a novel approach to perform posterior sampling of the underlying stellar spectrum in high-resolution spectroscopic observations. Our method, postellar, coherently combines information from empirical observations and physics-based models, enabling more accurate spectral recovery while providing uncertainties that can be propagated into downstream analyses. This is accomplished by treating the intrinsic stellar spectrum as a latent variable and performing posterior sampling under a Gaussian likelihood with an informative prior constructed using a score-based diffusion model trained on PHOENIX stellar models. We validate the framework on synthetic SPIRou radial velocity (RV) observations generated from the empirical spectra of Barnard's Star and Proxima Centauri. Spectra inferred with postellar more accurately recover the ground truth than standard empirical templates. For medium signal-to-noise observations, postellar improves RV accuracy by up to a factor of three, while RVs derived from empirical templates tend to be biased. This method performs particularly well in low signal-to-noise and low-cadence regimes. The postellar framework is broadly applicable to other high-resolution spectroscopy science cases including stellar abundance analyses and exoplanet atmospheric characterization.

astro-ph.EP↗

arXiv2026

Do Tabular Foundation Models Know Physics? Contamination, Units, and the Deterministic Limit

Tabular foundation models (TFMs) learn to fill in tables the way language models fill in text, and tables are arguably the format in which most physical measurement arrives. Did they learn any physics in the process? They are Bayesian by construction, so the question is what their prior contains. We probe it directly, evaluating four of them (TabPFN-3, TabICLv2, TabDPT and Real-TabPFN-2.5) against six baselines on datasets sampled from 316 physical equations, in and out of domain. TFMs dominate, out of the box and after tuning. But we show that their prior can represent neither a noiseless mechanism nor physical units, which is why they interpolate physics without yet being able to act as physical models.

cs.LG↗

arXiv2026

PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation

We propose a likelihood-free method for comparing two distributions given samples from each, with the goal of assessing the quality of generative models. The proposed approach, PQMass, provides a statistically rigorous method for assessing the performance of a single generative model or the comparison of multiple competing models. PQMass divides the sample space into non-overlapping regions and applies chi-squared tests to the number of data samples that fall within each region, giving a p-value that measures the probability that the bin counts derived from two sets of samples are drawn from the same multinomial distribution. PQMass does not depend on assumptions regarding the density of the true distribution, nor does it rely on training or fitting any auxiliary models. We evaluate PQMass on data of various modalities and dimensions, demonstrating its effectiveness in assessing the quality, novelty, and diversity of generated samples. We further show that PQMass scales well to moderately high-dimensional data and thus obviates the need for feature extraction in practical applications.

stat.ML↗

arXiv2026

LSST Strong Lensing Systems Dark Matter Sensitivity Analysis with Neural Ratio Estimators

Strong gravitational lensing offers a unique probe of dark matter (DM) on sub-galactic scales, where the abundance and distribution of low-mass halos are highly sensitive to the underlying properties of DM particles. In this work, we forecast LSST's sensitivity to DM substructure in galaxy-galaxy strong lenses using simulated samples and neural ratio estimators (NREs). Our simulations include both subhalos within the main deflector and line-of-sight (LOS) halos, with halo masses down to $\sim 10^7 M_\odot$ under the expected LSST ten-year survey imaging quality. We show that the constraining power on halo mass function (HMF) parameters improves significantly with sample size. Analyses based on a few hundred lenses yield broad posteriors comparable with other probes like the Ly-$α$ forest. By contrast, when combining 2500 lenses, $\approx 74\%$ and $\approx 36\%$ of the prior volume considered can be excluded at the $3σ$ and $5σ$ levels respectively, enabling statistically significant exclusions of non-$Λ$CDM scenarios. We further demonstrate that the sensitivity arises not only from the high-mass end of the HMF but also from low-mass halos: masking halos below $\log (m_{\rm halo}/M_\odot) \leq 7.5$ induces a measurable shift in the inferred posteriors. Finally, we find that LOS halos contribute significantly to the constraining power, with increasing importance of LOS halos at higher redshifts. While this analysis assumes perfect knowledge of the data-generating process and cannot be directly applied to data analysis, it quantifies constraints achievable with LSST alone and motivates the development of robust inference methods for real survey data.

astro-ph.CO↗

arXiv2026

Microlensing Detection and Inference via Learned Bayes Factors

We present a unified framework for gravitational microlensing event detection and parameter inference. Traditional pipelines use deterministic hard cuts on photometric statistics, systematically missing low-magnification events in the finite-source regime. We instead frame detection as Bayesian model comparison using Evidence Networks, which learn calibrated Bayes factors from binary-labeled simulations, and combine this with Neural Posterior Estimation (NPE) for amortized parameter inference. Both share a transformer encoder that handles irregularly-sampled time series without imputation. On simulated Roman Space Telescope data, our Evidence Network achieves $99.9\%$ detection efficiency with a false-positive rate below $6\times10^{-4}$ on simulated data with augmentation and noise, but no astrophysical confounders. Gains are most dramatic in the extreme finite-source regime ($ρ\gtrsim 5$), where detection rates reach ${\sim}95\%$ versus ${\sim}65\%$ for hard cuts, precisely the short-duration free-floating planet events most constraining for formation scenarios. Our NPE provides calibrated posteriors, working towards real-time analysis at survey scale.

astro-ph.IM↗

arXiv2026

Strong Gravitational Lensing Posterior Sampling in Pixel-Space Using Diffusion Models and Recurrent Inference Machines

Modeling galaxy-galaxy strong gravitational lenses to infer the brightness of the source galaxy and the mass distribution of the foreground galaxy is computationally challenging, particularly for high-resolution, high signal-to-noise ratio observations. In this regime, high-dimensional representations of both the source and the foreground mass distribution are necessary to model the data down to the noise level. This inference problem has been challenging for both traditional and machine learning-based methods because of its high dimensionality and its non-linearity in the foreground mass distribution. We present a method to generate joint posterior samples of the source galaxy and foreground mass distribution as pixelated images conditioned on observations. The method combines diffusion-based generative modeling and recurrent inference machines. It can model realistic gravitational lensing simulations with background and foreground galaxies drawn from cosmological hydrodynamical simulations down to the noise level.

astro-ph.IM↗

arXiv2026

MIRA: A Score for Conditional Distribution Accuracy and Model Comparison

We introduce Mira, a sample-based score for assessing the accuracy of a candidate conditional distribution using only joint samples from the true data-generating process. Relying on the principle that distributions coincide if they assign equal probability mass to all regions, we derive an analytic expression for the Mira statistic, whose average defines the Mira score. This formulation further allows us to compute theoretical reference values and uncertainty estimates when the candidate distribution matches the true one. This framework enables model comparison by quantifying the alignment between the conditional distribution of a candidate model and the true data generating process. Consequently, Mira enables Bayesian model comparison through direct posterior validation, bypassing the challenging evidence computation. We demonstrate its effectiveness across several toy problems and Bayesian inference tasks.

stat.ML↗

arXiv2026

The Interpolation Constraint in the RV Analysis of M-Dwarfs Using Empirical Templates

Precise radial velocity (pRV) measurements of M dwarfs in the near-infrared (NIR) rely on empirical templates due to the lack of accurate stellar spectral models in this regime. Templates are assumed to approximate the true spectrum when constructed from many observations or in the high signal-to-noise limit. We develop a numerical simulation that generates SPIRou-like pRV observations from PHOENIX spectra, constructs empirical templates, and estimates radial velocities. This simulation solely considers photon noise and evaluates when empirical templates remain reliable for pRV analysis. Our results reveal a previously unrecognized noise source in templates created from stacking registered observations, establishing a noise floor for such template-based pRV measurements. We find that these templates inherently include distortions in stellar line shapes due to imperfect interpolation at the detector's sampling resolution. The magnitude of this interpolation error depends on sampling resolution and RV content. Consequently, for stars with higher RV content, such as cooler M-dwarfs, interpolation noise has a larger relative impact, making their performance comparable to hotter M-dwarfs when using detectors with low sampling. For a typical M4V star, SPIRou's spectral and sampling resolution imposes an RV uncertainty floor of 0.5-0.8 m/s, independent of the star's magnitude or the telescope's aperture. These findings reveal a limitation of template-based pRV methods, underscoring the need for improved spectral modeling and better-than-Nyquist detector sampling to reach the next level of RV precision.

astro-ph.EP↗

arXiv2025

Transformer Embeddings for Fast Microlensing Inference

The search for free-floating planets (FFPs) is a key science driver for upcoming microlensing surveys like the Nancy Grace Roman Galactic Exoplanet Survey. These rogue worlds are typically detected via short-duration microlensing events, the characterization of which often requires analyzing noisy, irregularly-sampled observations. We present a pipeline for this task using simulation-based inference. We use a Transformer encoder to learn a compressed summary representation of the raw time-series data, which in turn conditions a neural posterior estimator. We demonstrate that our method produces accurate and well-calibrated posteriors over three orders of magnitude faster than traditional methods. We also demonstrate its performance on KMT-BLG-2019-2073, a short-duration FFP candidate event.

astro-ph.IM↗

arXiv2025

Neural Deprojection of Galaxy Stellar Mass Profiles

We introduce a neural approach to dynamical modeling of galaxies that replaces traditional imaging-based deprojections with a differentiable mapping. Specifically, we train a neural network to translate Nuker profile parameters into analytically deprojectable Multi Gaussian Expansion components, enabling physically realistic stellar mass models without requiring optical observations. We integrate this model into SuperMAGE, a differentiable dynamical modelling pipeline for Bayesian inference of supermassive black hole masses. Applied to ALMA data, our approach finds results consistent with state-of-the-art models while extending applicability to dust-obscured and active galaxies where optical data analysis is challenging.

astro-ph.GA↗

arXiv2025

Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images

We introduce a novel framework for upsampled Point Spread Function (PSF) modeling using pixel-level Bayesian inference. Accurate PSF characterization is critical for precision measurements in many fields including: weak lensing, astrometry, and photometry. Our method defines the posterior distribution of the pixelized PSF model through the combination of an analytic Gaussian likelihood and a highly expressive generative diffusion model prior, trained on a library of HST ePSF templates. Compared to traditional methods (parametric Moffat, ePSF template-based, and regularized likelihood), we demonstrate that our PSF models achieve orders of magnitude higher likelihood and residuals consistent with noise, all while remaining visually realistic. Further, the method applies even for faint and heavily masked point sources, merely producing a broader posterior. By recovering a realistic, pixel-level posterior distribution, our technique enables the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis. An implementation of our posterior sampling procedure is available on GitHub.

astro-ph.IM↗

arXiv2025

Mind the Information Gap: Unveiling Detailed Morphologies of z 0.5-1.0 Galaxies with SLACS Strong Lenses and Data-Driven Analysis

We present new state-of-the-art lens models for strong gravitational lensing systems from the Sloan Lens ACS (SLACS) survey, developed within a Bayesian framework that employs high-dimensional (pixellated), data-driven priors for the background source, foreground lens light, and point-spread function (PSF). Unlike conventional methods, our approach delivers high-resolution reconstructions of all major physical components of the lensing system and substantially reduces model-data residuals compared to previous work. For the majority of 30 lensing systems analyzed, we also provide posterior samples capturing the full uncertainty of each physical model parameter. The reconstructions of the background sources reveal high significance morphological structures as small as 200 parsecs in galaxies at redshifts of z 0.5-1.0, demonstrating the power of strong lensing and the analysis method to be used as a cosmic telescope to study the high redshift universe. This study marks the first application of data-driven generative priors to modeling real strong-lensing data and establishes a new benchmark for strong lensing precision modeling in the era of large-scale imaging surveys.

astro-ph.GA↗

arXiv2025

Robustness of Neural Ratio and Posterior Estimators to Distributional Shifts for Population-Level Dark Matter Analysis in Strong Gravitational Lensing

We investigate the robustness of Neural Ratio Estimators (NREs) and Neural Posterior Estimators (NPEs) to distributional shifts in the context of measuring the abundance of dark matter subhalos using strong gravitational lensing data. While these data-driven inference frameworks can be accurate on test data from the same distribution as the training sets, in real applications, it is expected that simulated training data and true observational data will differ in their distributions. We explore the behavior of a trained NRE and trained sequential NPEs to estimate the population-level parameters of dark matter subhalos from a large sample of images of strongly lensed galaxies with test data presenting distributional shifts within and beyond the bounds of the training distribution in the nuisance parameters (e.g., the background source morphology). While our results show that NREs and NPEs perform well when tested perfectly in distribution, they exhibit significant biases when confronted with slight deviations from the examples seen in the training distribution. This indicates the necessity for caution when applying NREs and NPEs to real astrophysical data, where high-dimensional underlying distributions are not perfectly known.

astro-ph.CO↗

arXiv2025

Blind Strong Gravitational Lensing Inversion: Joint Inference of Source and Lens Mass with Score-Based Models

Score-based models can serve as expressive, data-driven priors for scientific inverse problems. In strong gravitational lensing, they enable posterior inference of a background galaxy from its distorted, multiply-imaged observation. Previous work, however, assumes that the lens mass distribution (and thus the forward operator) is known. We relax this assumption by jointly inferring the source and a parametric lens-mass profile, using a sampler based on GibbsDDRM but operating in continuous time. The resulting reconstructions yield residuals consistent with the observational noise, and the marginal posteriors of the lens parameters recover true values without systematic bias. To our knowledge, this is the first successful demonstration of joint source-and-lens inference with a score-based prior.

astro-ph.IM↗

arXiv2025

Bridging Simulators with Conditional Optimal Transport

We propose a new field-level emulator that bridges two simulators using unpaired simulation datasets. Our method leverages a flow-based approach to learn the likelihood transport from one simulator to the other. Since multiple transport maps exist, we employ Conditional Optimal Transport Flow Matching (COT-FM) to ensure that the transformation minimally distorts the underlying structure of the data. We demonstrate the effectiveness of this approach by bridging weak lensing simulators: a Lagrangian Perturbation Theory (LPT) to a N-body Particle-Mesh (PM). We demonstrate that our emulator captures the full correction between the simulators by showing that it enables full-field inference to accurately recover the true posterior, validating its accuracy beyond traditional summary statistics.

astro-ph.CO↗

arXiv2025

Predicting the Subhalo Mass Functions in Simulations from Galaxy Images

Strong gravitational lensing provides a powerful tool to directly infer the dark matter (DM) subhalo mass function (SHMF) in lens galaxies. However, comparing observationally inferred SHMFs to theoretical predictions remains challenging, as the predicted SHMF can vary significantly between galaxies - even within the same cosmological model - due to differences in the properties and environment of individual galaxies. We present a machine learning framework to infer the galaxy-specific predicted SHMF from galaxy images, conditioned on the assumed inverse warm DM particle mass $M^{-1}_{\rm DM}$. To train the model, we use 1024 high-resolution hydrodynamical zoom-in simulations from the DREAMS suite. Mock observations are generated using Synthesizer, excluding gas particle contributions, and SHMFs are computed with the Rockstar halo finder. Our neural network takes as input both the galaxy images and the inverse DM mass. This method enables scalable, image-based predictions for the theoretical DM SHMFs of individual galaxies, facilitating direct comparisons with observational measurements.

astro-ph.CO↗

arXiv2025

On Diffusion Modeling for Anomaly Detection

Known for their impressive performance in generative modeling, diffusion models are attractive candidates for density-based anomaly detection. This paper investigates different variations of diffusion modeling for unsupervised and semi-supervised anomaly detection. In particular, we find that Denoising Diffusion Probability Models (DDPM) are performant on anomaly detection benchmarks yet computationally expensive. By simplifying DDPM in application to anomaly detection, we are naturally led to an alternative approach called Diffusion Time Estimation (DTE). DTE estimates the distribution over diffusion time for a given input and uses the mode or mean of this distribution as the anomaly score. We derive an analytical form for this density and leverage a deep neural network to improve inference efficiency. Through empirical evaluations on the ADBench benchmark, we demonstrate that all diffusion-based anomaly detection methods perform competitively for both semi-supervised and unsupervised settings. Notably, DTE achieves orders of magnitude faster inference time than DDPM, while outperforming it on this benchmark. These results establish diffusion-based anomaly detection as a scalable alternative to traditional methods and recent deep-learning techniques for standard unsupervised and semi-supervised anomaly detection settings.

cs.LG↗

arXiv2025

Solving Bayesian inverse problems with diffusion priors and off-policy RL

This paper presents a practical application of Relative Trajectory Balance (RTB), a recently introduced off-policy reinforcement learning (RL) objective that can asymptotically solve Bayesian inverse problems optimally. We extend the original work by using RTB to train conditional diffusion model posteriors from pretrained unconditional priors for challenging linear and non-linear inverse problems in vision, and science. We use the objective alongside techniques such as off-policy backtracking exploration to improve training. Importantly, our results show that existing training-free diffusion posterior methods struggle to perform effective posterior inference in latent space due to inherent biases.

cs.LG↗