Sensitivities of Black Hole Images from GRMHD Simulations

Pedro Naethe Motta, Mário Raia Neto, Cora Prather, Alejandro Cárdenas-Avendaño · 2026

Read Sensitivities of Black Hole Images from GRMHD Simulations by Pedro Naethe Motta, Mário Raia Neto, Cora Prather, Alejandro Cárdenas-Avendaño (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-11869.txt.

Introduction

Motta, Neto, Prather, and Cárdenas-Avendaño use Jipole, a radiative transfer code, to compute how black hole images from magnetohydrodynamic simulations respond to model settings. Those pixel-wise sensitivities form a Jacobian mapping settings to image space. In mock tests, the fitting landscape proves structured, with anisotropies and local minima, yet gradient guidance keeps exploration tractable. Gradients still guide recovery well under idealised, blurred, and noisy conditions.

Catalog record

CREATORPedro Naethe Motta, Mário Raia Neto, Cora Prather, Alejandro Cárdenas-Avendaño
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58130
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.11869

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sensitivities of Black Hole Images from GRMHD Simulations Pedro Naethe Motta, Mário Raia Neto, Cora Prather, Alejandro Cárdenas-Avendaño Abstract The advent of high-fidelity imaging of supermassive black holes calls for efficient and robust data-analysis methods. In this work, we use Jipole , a differentiable, ipole -based radiative transfer code, to enable gradient-based analyses of images generated from state-of-the-art general relativistic magnetohydrodynamic (GRMHD) simulations. We compute image sensitivities, i.e., pixel-wise derivatives of the intensity with respect to model parameters, which form the Jacobian of the forward model and define a local map from parameter space to image space. Using these sensitivities in a mock data analysis, we find that GRMHD-based images generate a structured error landscape for parameter fitting, with anisotropies and local minima, making parameter exploration nontrivial but still tractable when guided by gradient information. We characterize this landscape through the Jacobian and assess the feasibility of gradient-based recovery under idealized, blurred, and noisy conditions. Our results show that automatic differentiation-computed image gradients can guide parameter exploration effectively even in the presence of noise. These findings establish a basis for efficient, high-precision model–data comparisons in black hole imaging and motivate the integration of these sensitivities into advanced inference frameworks. show][email protected] ][email protected] ][email protected] ][email protected] I Introduction General-relativistic magnetohydrodynamic (GRMHD) simulations form complete forward models useful for studying near-horizon accretion flows ( 28 ; 4 ) . Comparisons between simulation parameter studies and very long baseline interferometry (VLBI) observations have already provided constraints on accretion-flow properties and have also served as tests of the accuracy of current simulation methods ( 6 ) . As complete models, GRMHD simulations can be compared to any observation of a source, including resolved and spectral imaging ( 29 ) , jet power ( 25 ; 14 ) , and time variability ( 10 ) . To generate predictions of the horizon-scale radio synchrotron emission observable by the Event Horizon Telescope (EHT), GRMHD simulations are post-processed with a general-relativistic radiative transfer (GRRT) code to generate a predicted intensity distribution on the sky. This imaging step is carried out with ray-tracing codes such as GRay ( 3 ) , Odyssey ( 22 ) , ipole ( 15 ; 16 ) , RAPTOR ( 1 ) , Mahakala ( 24 ) , and Jipole ( 19 ) . The computational cost of this post-processing stage becomes significant when one seeks to compare each simulation over a broad space of possible imaging parameters, such as possible observer angles and electron thermodynamics models. In 19 , we introduced a differentiable imaging framework with Jipole , capable of computing both a simulated image and its derivatives with respect to selected input parameters. That is, for each ray-traced pixel, Jipole computes the derivatives of the specific intensity I_{\nu} with respect to the post-processing parameters P_{i} , denoted dI_{\nu}/dP_{i} . As discussed in 19 , these sensitivities can enable dramatic efficiency and capability improvements for both forward and backward modeling pipelines. However, differentiable imaging of GRMHD models is not necessarily guaranteed to be either useful or feasible beforehand. Its success depends on how smoothly the image varies as the input parameters are changed. GRMHD-based images are disrupted by sharp emissivity transitions, emission region cuts, and discrete structure in the underlying simulation data ( 2 ) . These features can lead to irregular changes in the image under small parameter variations and may therefore complicate both local optimization and gradient-informed inference. For this reason, before sensitivities can be used in data analysis, it is necessary to establish both that they are computed correctly and that they remain informative when imaging GRMHD simulations. In this work, for the first time, we compute these sensitivities for black hole images obtained from GRMHD snapshots, and use them to examine the questions raised above. We focus on derivatives with respect to two post-processing parameters: the observer inclination angle \theta_{\rm o} and the electron-heating parameter R_{\rm high} . By post-processing parameters we mean image-model parameters that do not alter the underlying GRMHD simulation itself. The two parameters chosen in this paper provide examples of different types of dependence: \theta_{\rm o} determines the observer’s inclination, whereas R_{\rm high} modifies the estimated plasma thermodynamics through an electron-heating prescription ( 17 ) . The same framework can be extended to other post-processing parameters, including the accretion rate, black hole mass, or a generalized electron energy distribution. We then use these sensitivities to conduct gradient-based fitting experiments as illustrative applications, rather than as a final inference strategy for black hole imaging. Existing Bayesian analyses of EHT data are typically carried out within stochastic sampling frameworks, for example with Comrade.jl ( 26 ) , in order to account for systematic and stochastic observational uncertainties. Our goal here is narrower: we use simple conjugate-gradient experiments to test whether the derivatives dI/d\theta_{\rm o} and dI/dR_{\rm high} provide useful local information for parameter recovery and to assess how they reflect the structure of the imaging error surface. These experiments should therefore be interpreted as a controlled demonstration of the information content of the sensitivities, and as a step toward their incorporation into more complete Bayesian inference pipelines. The rest of this paper is organized as follows. In Sec. II we present the governing equations and numerical methods used to compute both the images and their sensitivities. In Sec. III we validate Jipole against ipole , describe the numerical changes adopted in the present implementation, and compare sensitivities computed with finite differences and automatic differentiation. In Sec. IV we examine the normalized mean squared error (NMSE) landscape associated with the two parameters considered here, \theta_{\rm o} and R_{\rm high} . In Sec. V we present a simple mock-data analysis using these sensitivities, first for an injected image without blurring or noise, and then after adding blurring and Gaussian noise. Finally, in Sec. VI we discuss the implications of these results and outline directions for future work. II Governing Equations and Numerical Methods In this section, we briefly summarize the main ingredients of Jipole and refer the reader to 19 for more details. To compute the radiation intensity from an accretion disk in a given spacetime as seen by a distant observer, one simulates the ray paths by solving the following system of two first-order differential equations \displaystyle\frac{dx^{\mu}}{d\lambda}=k^{\mu}, (1) \displaystyle\frac{dk^{\mu}}{d\lambda}=\Gamma^{\mu}_{\alpha\beta}k^{\alpha}k^{\beta}. (2) In this formulation, x^{\mu} denotes the four-position and k^{\mu} the four-momentum of the photon. The geodesic is parameterized by the affine variable \lambda , while \Gamma^{\mu}_{\alpha\beta} represents the connection coefficients, and the Einstein summation convention is employed. The path of a photon is calculated by integrating this set of coupled equations backward from the camera’s location. Once the path is calculated, the intensity is integrated forward along the ray for each pixel, following the covariant form of the radiative transfer equation \frac{d}{d\lambda}\left(\frac{I_{\nu}}{\nu^{3}}\right)=\left(\frac{j_{\nu}}{\nu^{2}}\right)-(\alpha_{\nu}\nu)\left(\frac{I_{\nu}}{\nu^{3}}\right), (3) where \nu represents the photon frequency, I_{\nu} denotes the specific intensity, j_{\nu} is the specific emissivity, and \alpha_{\nu} represents the absorption coefficient. The terms grouped in parentheses within Eq. 3 represent quantities that are relativistically invariant. For the simulations performed in this work, we focus on the Kerr geometry and j_{\nu} is modeled as synchrotron emission from a thermal distribution of electrons ( 13 ) , \displaystyle j_{\nu}(\nu,\theta) \displaystyle=\frac{\sqrt{2}\pi e^{2}n_{e}\nu_{s}}{3cK_{2}(1/\Theta_{\rm e})}(X^{1/2}+2^{11/12}X^{1/6})^{2}\times \displaystyle\exp(-X^{1/3}), (4) where X\equiv\nu/\nu_{s} ,

Read the complete text