arXiv subjects

Konstantin Karchev

Publications and source records attributed to Konstantin Karchev.

15 recordsLinked to original sources

arXiv2026

Recovering Sharp Conductivity Features in the Finite-Data Calderón Problem with Physics-Informed Neural Networks

Physics-informed neural networks (PINNs) have recently emerged as a promising framework for addressing the Calderón inverse problem from limited boundary data. In this work, we revisit neural Calderón inversion by introducing multiscale boundary excitations based on randomized wavelet functions and investigating the role of Fourier-feature encoding (FFE) for representing sharp conductivity variations. We propose a physics-informed reconstruction framework that represents the unknown conductivity and the associated family of electric potentials with separate neural networks conditioned on the applied boundary excitations. The governing elliptic PDE is enforced through physics-informed residuals, while finite Dirichlet-to-Neumann (DtN) data are incorporated through boundary losses. Using synthetic data from a finite-difference forward solver, we evaluate the method on conductivity fields with inclusions, sharp interfaces, smooth profiles, and heterogeneous media. Results show that the framework recovers dominant conductivity structures from finite boundary measurements with relative errors between $3\%-12\%$ approximately. We show that FFE improves the reconstruction of localized sharp features, particularly for inclusions and interfaces, but are not universally optimal, with raw-coordinate networks performing competitively for smoother fields. These results highlight coordinate representations and boundary excitation design as key factors in neural Calderón inversion.

cs.LG↗

arXiv2026

The colour variability of low-z SNe Ia is entirely explained by dust

The relative importance of intrinsic colour variability of supernovae type Ia (SN Ia) versus dust-induced reddening remains an open question with important ramifications for understanding their environmental dependence, as well as for the validity of the traditionally employed Tripp linear correction for cosmological inference. We revisit this question in the light of two low-redshift, homogeneous datasets, the ZTF DR2 and Foundation DR1, which we analyse within the framework of the Bayesian hierarchical model Simple-BayeSN. We demonstrate both with simulation and on real data that traditional colour cuts, which remove highly reddened samples, induce a previously unrecognized selection effect, which may have biased previous conclusions on the origin of SN Ia colour variability. Once this is accounted for, we are able to explain the entirety of the colour--magnitude correlation as due to dust effects, with no need for an intrinsic colour correlation. This result is robust with respect to a host galaxy mass split and projected distance from the center of the host. Our findings imply that the traditional linear Tripp correction maintains an empirical validity, even though it should be ascribed to dust rather than intrinsic colour variation.

astro-ph.CO↗

arXiv2026

CIGaRS I: Combined simulation-based inference from type Ia supernovae and host photometry

Using type Ia supernovae as cosmological probes requires empirical corrections that are correlated with their host environment. Here we present a unified Bayesian hierarchical model designed to infer, from purely photometric observations, the intrinsic dependence of the brightness of type Ia supernovae on progenitor properties (metallicity and age), the delay-time distribution that governs their rate as a function of age, and cosmology, as well as the redshifts of all hosts. The model incorporates physics-based prescriptions for star formation and chemical evolution from Prospector-beta, dust extinction of both galaxy and supernova light, and observational selection effects. We show with simulations that intrinsic dependences on metallicity and age have distinct observational signatures, with metallicity mimicking the well-known step of magnitudes of type Ia supernovae across a host stellar mass of $\sim 10^{10}M_{\odot}$. We then demonstrate neural simulation-based inference of all model parameters from mock observations of ~16,000 type Ia supernovae and their hosts up to redshift 0.9. Our joint physics-based approach delivers robust and precise photometric redshifts (~0.01 median scatter) and improves cosmological constraints by a factor of ~4 over analyses of the small fraction of objects with spectroscopic follow-up. This approach unlocks the full power of photometric data and paves the way for an end-to-end simulation-based analysis pipeline in the LSST era.

astro-ph.CO↗

arXiv2025

STAR NRE: Solving supernova selection effects with set-based truncated auto-regressive neural ratio estimation

Accounting for selection effects in supernova type Ia (SN Ia) cosmology is crucial for unbiased cosmological parameter inference -- even more so for the next generation of large, mostly photometric-only surveys. The conventional "bias correction" procedure has a built-in systematic bias towards the fiducial model used to derive it and fails to account for the additional Eddington bias that arises in the presence of significant redshift uncertainty. On the other hand, likelihood-based analyses within a Bayesian hierarchical model, e.g. using MCMC, scale poorly with the data set size and require explicit assumptions for the selection function that may be inaccurate or contrived. To address these limitations, we introduce STAR NRE, a simulation-based approach that makes use of a conditioned deep set neural network and combines efficient high-dimensional global inference with subsampling-based truncation in order to scale to very large survey sizes while training on sets with varying cardinality. Applying it to a simplified SN Ia model consisting of standardised brightnesses and redshifts with Gaussian uncertainties and a selection procedure based on the expected LSST sensitivity, we demonstrate precise and unbiased inference of cosmological parameters and the redshift evolution of the volumetric SN Ia rate from ~100 000 mock SNae Ia. Our inference procedure can incorporate arbitrarily complex selection criteria, including transient classification, in the forward simulator and be applied to complex data like light curves. We outline these and other steps aimed at integrating STAR NRE into an end-to-end simulation-based pipeline for the analysis of future photometric-only SN Ia data.

astro-ph.CO↗

arXiv2025

Near-instantaneous Atmospheric Retrievals and Model Comparison with FASTER

In the era of the James Webb Space Telescope (JWST), the dramatic improvement in the spectra of exoplanetary atmospheres demands a corresponding leap forward in our ability to analyze them: atmospheric retrievals need to be performed on thousands of spectra, applying to each large ensembles of models (that explore atmospheric chemistry, thermal profiles and cloud models) to identify the best one(s). In this limit, traditional Bayesian inference methods such as nested sampling become prohibitively expensive. We introduce FASTER (Fast Amortized Simulation-based Transiting Exoplanet Retrieval), a neural-network based method for performing atmospheric retrieval and Bayesian model comparison at a fraction of the computational cost of classical techniques. We demonstrate that the marginal posterior distributions of all parameters within a model as well as the posterior probabilities of the models we consider match those computed using nested sampling both on mock spectra, and for the real NIRSpec PRISM spectrum of WASP-39b. The true power of the FASTER framework comes from its amortized nature, which allows the trained networks to perform practically instantaneous Bayesian inference and model comparison over ensembles of spectra -- real or simulated -- at minimal additional computational cost. This offers valuable insight into the expected results of model comparison (e.g., distinguishing cloudy from cloud-free and isothermal from non-isothermal models), as well as their dependence on the underlying parameters, which is computationally unfeasible with nested sampling. This approach will constitute as large a leap in spectral analysis as the original retrieval methods based on Markov Chain Monte Carlo have proven to be.

astro-ph.EP↗

arXiv2025

One never walks alone: the effect of the perturber population on subhalo measurements in strong gravitational lenses

Analyses of extended arcs in strong gravitational lensing images to date have constrained the properties of dark matter by measuring the parameters of one or two individual subhalos. However, since such analyses are reliant on likelihood-based methods like Markov-chain Monte Carlo or nested sampling, they require various compromises to the realism of lensing models for the sake of computational tractability, such as ignoring the numerous other subhalos and line-of-sight halos in the system, assuming a particular form for the source model and requiring the noise to have a known likelihood function. Here we show that a simulation-based inference method calledTruncated Marginal Neural Ratio Estimation (TMNRE) makes it possible to relax these requirements by training neural networks to directly compute marginal posteriors for subhalo parameters from lensing images. By performing a set of inference tasks on mock data, we verify the accuracy of TMNRE and show it can compute posteriors for subhalo parameters marginalized over populations of hundreds of substructures, as well as lens and source uncertainties. We also find the \gls*{mlp} Mixer network works far better for such tasks than the convolutional architectures explored in other lensing analyses. Furthermore, we show that since \gls*{tmnre} learns a posterior function it enables direct statistical checks that would be extremely expensive with likelihood-based methods. Our results show that TMNRE is well-suited for analyzing complex lensing data, and that the full subhalo and line-of-sight halo population must be included when measuring the properties of individual dark matter substructures with this technique.

astro-ph.CO↗

arXiv2024

Dark Matter Halo Parameters from Overheated Exoplanets via Bayesian Hierarchical Inference

Dark Matter (DM) can become captured, deposit annihilation energy, and hence increase the heat flow in exoplanets and brown dwarfs. Detecting such a DM-induced heating in a population of exoplanets in the inner kpc of the Milky Way thus provides potential sensitivity to the galactic DM halo parameters. We develop a Bayesian Hierarchical Model to investigate the feasibility of DM discovery with exoplanets and examine future prospects to recover the spatial distribution of DM in the Milky Way. We reconstruct from mock exoplanet datasets observable parameters such as exoplanet age, temperature, mass, and location, together with DM halo parameters, for representative choices of measurement uncertainty and the number of exoplanets detected. We find that detection of $\mathcal{O}(100)$ exoplanets in the inner Galaxy can yield quantitative information on the galactic DM density profile, under the assumption of 10% measurement uncertainty. Even as few as $\mathcal{O}(10)$ exoplanets can deliver meaningful sensitivities if the DM density and inner slope are sufficiently large.

astro-ph.IM↗

arXiv2024

SIDE-real: Supernova Ia Dust Extinction with truncated marginal neural ratio estimation applied to real data

We present the first fully simulation-based hierarchical analysis of the light curves of a population of low-redshift type Ia supernovae (SNae Ia). Our hardware-accelerated forward model, released in the Python package slicsim, includes stochastic variations of each SN's spectral flux distribution (based on the pre-trained BayeSN model), extinction from dust in the host and in the Milky Way, redshift, and realistic instrumental noise. By utilising truncated marginal neural ratio estimation (TMNRE), a neural network-enabled simulation-based inference technique, we implicitly marginalise over 4000 latent variables (for a set of $\approx 100$ SNae Ia) to efficiently infer SN Ia absolute magnitudes and host-galaxy dust properties at the population level while also constraining the parameters of individual objects. Amortisation of the inference procedure allows us to obtain coverage guarantees for our results through Bayesian validation and frequentist calibration. Furthermore, we show a detailed comparison to full likelihood-based inference, implemented through Hamiltonian Monte Carlo, on simulated data and then apply TMNRE to the light curves of 86 SNae Ia from the Carnegie Supernova Project, deriving marginal posteriors in excellent agreement with previous work. Given its ability to accommodate arbitrarily complex extensions to the forward model -- e.g. different populations based on host properties, redshift evolution, complicated photometric redshift estimates, selection effects, and non-Ia contamination -- without significant modifications to the inference procedure, TMNRE has the potential to become the tool of choice for cosmological parameter inference from future, large SN Ia samples.

astro-ph.CO↗

arXiv2023

SimSIMS: Simulation-based Supernova Ia Model Selection with thousands of latent variables

We present principled Bayesian model comparison through simulation-based neural classification applied to SN Ia analysis. We validate our approach on realistically simulated SN Ia light curve data, demonstrating its ability to recover posterior model probabilities while marginalizing over >4000 latent variables. The amortized nature of our technique allows us to explore the dependence of Bayes factors on the true parameters of simulated data, demonstrating Occam's razor for nested models. When applied to a sample of 86 low-redshift SNae Ia from the Carnegie Supernova Project, our method prefers a model with a single dust law and no magnitude step with host mass, disfavouring different dust laws for low- and high-mass hosts with odds in excess of 100:1.

astro-ph.CO↗

arXiv2023

Analytic auto-differentiable $Λ$CDM cosmography

I present general analytic expressions for distance calculations (comoving distance, time coordinate, and absorption distance) in the standard $Λ$CDM cosmology, allowing for the presence of radiation and for non-zero curvature. The solutions utilise the symmetric Carlson basis of elliptic integrals, which can be evaluated with fast numerical algorithms that allow trivial parallelisation on GPUs and automatic differentiation without the need for additional special functions. I introduce a PyTorch-based implementation in the phytorch.cosmology package and briefly examine its accuracy and speed in comparison with numerical integration and other known expressions (for special cases). Finally, I demonstrate an application to high-dimensional Bayesian analysis that utilises automatic differentiation through the distance calculations to efficiently derive posteriors for cosmological parameters from up to $10^6$ mock type Ia supernovae using variational inference.

astro-ph.IM↗

arXiv2023

SICRET: Supernova Ia Cosmology with truncated marginal neural Ratio EsTimation

Type Ia supernovae (SNae Ia), standardisable candles that allow tracing the expansion history of the Universe, are instrumental in constraining cosmological parameters, particularly dark energy. State-of-the-art likelihood-based analyses scale poorly to future large datasets, are limited to simplified probabilistic descriptions, and must explicitly sample a high-dimensional latent posterior to infer the few parameters of interest, which makes them inefficient. Marginal likelihood-free inference, on the other hand, is based on forward simulations of data, and thus can fully account for complicated redshift uncertainties, contamination from non-SN Ia sources, selection effects, and a realistic instrumental model. All latent parameters, including instrumental and survey-related ones, per-object and population-level properties, are implicitly marginalised, while the cosmological parameters of interest are inferred directly. As a proof of concept, we apply truncated marginal neural ratio estimation (TMNRE), a form of marginal likelihood-free inference, to BAHAMAS, a Bayesian hierarchical model for SALT parameters. We verify that TMNRE produces unbiased and precise posteriors for cosmological parameters from up to 100 000 SNae Ia. With minimal additional effort, we train a network to infer simultaneously the O(100 000) latent parameters of the supernovae (e.g. absolute brightnesses). In addition, we describe and apply a procedure that utilises local amortisation of the inference to convert the approximate Bayesian posteriors into frequentist confidence regions with exact coverage. Finally, we discuss the planned improvements to the model that are enabled by using a likelihood-free inference framework, like selection effects and non-Ia contamination.

astro-ph.CO↗

arXiv2022

Estimating the warm dark matter mass from strong lensing images with truncated marginal neural ratio estimation

Precision analysis of galaxy-galaxy strong gravitational lensing images provides a unique way of characterizing small-scale dark matter halos, and could allow us to uncover the fundamental properties of dark matter's constituents. Recently, gravitational imaging techniques made it possible to detect a few heavy subhalos. However, gravitational lenses contain numerous subhalos and line-of-sight halos, whose subtle imprint is extremely difficult to detect individually. Existing methods for marginalizing over this large population of sub-threshold perturbers to infer population-level parameters are typically computationally expensive, or require compressing observations into hand-crafted summary statistics, such as a power spectrum of residuals. Here, we present the first analysis pipeline to combine parametric lensing models and a recently-developed neural simulation-based inference technique called truncated marginal neural ratio estimation (TMNRE) to constrain the warm dark matter halo mass function cutoff scale directly from multiple lensing images. Through a proof-of-concept application to simulated data, we show that our approach enables empirically testable inference of the dark matter cutoff mass through marginalization over a large population of realistic perturbers that would be undetectable on their own, and over lens and source parameters uncertainties. To obtain our results, we combine the signal contained in a set of images with Hubble Space Telescope resolution. Our results suggest that TMNRE can be a powerful approach to put tight constraints on the mass of warm dark matter in the multi-keV regime, which will be relevant both for existing lensing data and in the large sample of lenses that will be delivered by near-future telescopes.

astro-ph.CO↗

arXiv2022

Strong-Lensing Source Reconstruction with Denoising Diffusion Restoration Models

Analysis of galaxy--galaxy strong lensing systems is strongly dependent on any prior assumptions made about the appearance of the source. Here we present a method of imposing a data-driven prior / regularisation for source galaxies based on denoising diffusion probabilistic models (DDPMs). We use a pre-trained model for galaxy images, AstroDDPM, and a chain of conditional reconstruction steps called denoising diffusion reconstruction model (DDRM) to obtain samples consistent both with the noisy observation and with the distribution of training data for AstroDDPM. We show that these samples have the qualitative properties associated with the posterior for the source model: in a low-to-medium noise scenario they closely resemble the observation, while reconstructions from uncertain data show greater variability, consistent with the distribution encoded in the generative model used as prior.

astro-ph.IM↗

arXiv2021

Strong-lensing source reconstruction with variationally optimised Gaussian processes

Strong-lensing images provide a wealth of information both about the magnified source and about the dark matter distribution in the lens. Precision analyses of these images can be used to constrain the nature of dark matter. However, this requires high-fidelity image reconstructions and careful treatment of the uncertainties of both lens mass distribution and source light, which are typically difficult to quantify. In anticipation of future high-resolution datasets, in this work we leverage a range of recent developments in machine learning to develop a new Bayesian strong-lensing image analysis pipeline. Its highlights are: (A) a fast, GPU-enabled, end-to-end differentiable strong-lensing image simulator; (B) a new, statistically principled source model based on a computationally highly efficient approximation to Gaussian processes that also takes into account pixellation; and (C) a scalable variational inference framework that enables simultaneously deriving posteriors for tens of thousands of lens and source parameters and optimising hyperparameters via stochastic gradient descent. Besides efficient and accurate parameter estimation and lens model uncertainty quantification, the main aim of the pipeline is the generation of training data for targeted simulation-based inference of dark matter substructure, which we will exploit in a companion paper.

astro-ph.IM↗

arXiv2020

Targeted Likelihood-Free Inference of Dark Matter Substructure in Strongly-Lensed Galaxies

The analysis of optical images of galaxy-galaxy strong gravitational lensing systems can provide important information about the distribution of dark matter at small scales. However, the modeling and statistical analysis of these images is extraordinarily complex, bringing together source image and main lens reconstruction, hyper-parameter optimization, and the marginalization over small-scale structure realizations. We present here a new analysis pipeline that tackles these diverse challenges by bringing together many recent machine learning developments in one coherent approach, including variational inference, Gaussian processes, differentiable probabilistic programming, and neural likelihood-to-evidence ratio estimation. Our pipeline enables: (a) fast reconstruction of the source image and lens mass distribution, (b) variational estimation of uncertainties, (c) efficient optimization of source regularization and other hyperparameters, and (d) marginalization over stochastic model components like the distribution of substructure. We present here preliminary results that demonstrate the validity of our approach.

astro-ph.CO↗