Searching for Lorentz invariance violation with artificial neural networks
Karlo Mrakovčić · 2025
Read Searching for Lorentz invariance violation with artificial neural networks by Karlo Mrakovčić (2025) 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-2509-14818.txt.
Introduction
Tomislav Terzić and Karlo Mrakovčić search for violations of Lorentz invariance with neural networks in a 2025 astrophysics paper. Lorentz invariance is a core symmetry of relativity, but some theories predict tiny breaks at very high energies. Such breaks could change how gamma rays are emitted, travel, and get detected, through several effects at once.
Earlier experiments usually tested each possible effect on its own. Terzić and Mrakovčić instead train artificial neural networks to look for traces of several effects together in a single analysis. Their paper presents the method and its first results, and places those results alongside earlier experimental searches. A combined search matters because real violations, if they exist, would likely show several signatures at the same time.
Catalog record
| CREATOR | Karlo Mrakovčić |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-14818
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Searching for Lorentz invariance violation with artificial neural networks Karlo Mrakovčić Abstract Lorentz invariance violation (LIV) can have multiple consequences on very-high energy gamma rays’ emission, propagation, and detection, such as energy-dependent photon group velocity, photon instability, vacuum birefringence, and modified electromagnetic interaction. Depending on the underlying theoretical model, several of these effects can coexist. Nevertheless, in experimental tests of LIV, each effect is tested separately and independently. Here, we are performing a search for traces of several coexisting effects in a single analysis. We present our analysis method based on artificial neural networks and put our very first results in the context of experimental searches for LIV. [a]Tomislav Terzić 1 Introduction The search for the theory of quantum gravity (QG) is hampered by the lack of experimental evidence. The Planck energy scale ( E_{\mathrm{Pl}}\approx 1.2\times 10^{19} GeV), as the expected domain of QG is far beyond the reach of accelerator experiments. At the end of 20^{\text{th}} century, an idea appeared that highly energetic particles from astrophysical sources might be able to probe spacetime fluctuations, expected to occur at the Planck scale distances [ 1 ] . Gamma rays are particularly good probes because they are detected with relatively high statistics, reach energies of up to a PeV scale [ 2 ] , and propagate on straight lines, which facilitates source identification. The interaction of gamma rays with spacetime is expected to manifest itself as energy-dependent photon group velocity. This is a possible consequence of Lorentz invariance violation (LIV). Other possible effects of LIV include modified reaction thresholds, as well as modified interaction dynamics, vacuum birefringence, etc. A comprehensive review of different QG models and tests of LIV using cosmic messengers (cosmic rays, gamma rays, gravitational waves, neutrinos) can be found in [ 3 , 4 ] , while [ 5 ] offers a critical discussion and comparison of tests and analysis methods performed using gamma rays. Depending on the underlying theoretical model, several LIV effects can manifest simultaneously. However, testing multiple LIV effects simultaneously remains a major challenge. Most of the existing analyses focus on individual phenomena—such as energy-dependent time delays or modified reaction—in isolation. Here, we explore a joint approach using two distinct machine learning architectures: a dense feedforward artificial neural network (ANN) based on parameterized fits of observed data, and a sequence-to-regression transformer trained on full-photon time-energy sequences. 2 Theoretical framework LIV is typically parametrized through a modified photon dispersion relation: E^{2}\approx p^{2}c^{2}\left[1+\sum_{n=1}^{\infty}S_{n}\left(\frac{E}{E_{\mathrm{QG},n}}\right)^{n}\right], (1) where E and p represent the photon energy and momentum, respectively, E_{\mathrm{QG},n} denotes the quantum gravity energy scale, and S_{n}=\pm 1 encodes the sign of the violation. S_{n}=+1 would result in a photon group velocity greater than c (the so-called superluminal behaviour), and the opposite is valid for S_{n}=-1 ( subluminal behaviour). Different orders in the series are treated independently. The sensitivity of current measurements allows us to probe only the first two orders. The first-order modification ( n=1 ) follows from theoretical models with CPT-odd operators. These allow vacuum birefringence, which was strongly constrained independently (see, e.g., [ 6 , 7 , 8 ] ). Vacuum birefringence does not appear in the n=2 class of models. Moreover, superluminal behaviour allows photons to split to three gamma rays or to decay to electron-positron pairs, which was also strongly constrained [ 9 , 10 ] . In this case, we restricted ourselves to second-order correction in the subluminal scenario ( n=2 , S_{2}=-1 ). This choice is motivated by several reasons: second-order ( n=2 ) operators dominate in effective field theories, preserving CPT and gauge invariance. Furthermore, they yield both time-of-flight effects and modifications to reaction thresholds and cross sections. Additionally, the second-order modification is observationally less constrained than the first-order one. The specific effects that we have considered were the time-of-flight and modified absorption of gamma rays on extragalactic background light (EBL). The LIV-induced delay in the arrival of gamma rays from a source at redshift z_{s} , compared to the special relativistic case is given by \Delta t_{2}=-\frac{3}{2}S_{2}\frac{E^{2}}{E_{\mathrm{QG},2}^{2}}\kappa_{2}(z_{s}), (2) where \kappa_{2}(z_{s}) represents the distance contribution. We used the standard measure, first introduced in [ 11 ] : \kappa_{2}(z_{s})=\frac{1}{H_{0}}\int_{0}^{z_{s}}\frac{(1+z)^{2}}{\sqrt{\Omega_{\Lambda}+\Omega_{m}(1+z)^{3}}}\,dz. (3) For the standard cosmological parameters of the \Lambda CDM cosmology model, we use H_{0}=70\,\mathrm{km\,Mpc}^{-1}\,\mathrm{s}^{-1} , \Omega_{\Lambda}=0.7 and \Omega_{m}=0.3 . The probability of gamma rays avoiding scattering on EBL and reaching Earth is given by P(E,z_{s})=\exp[-\tau(E,z_{s})], (4) where the opacity of the universe to gamma rays is given as \tau(E,z_{s})=\int_{0}^{z_{\mathrm{s}}}dz\,\frac{dl}{dz}\int_{-1}^{1}d\cos\theta\left(\frac{1-\cos\theta}{2}\right)\int_{\omega_{\text{th}}(E,\theta)}^{\infty}d\omega\;n(\omega,z)\,\sigma(E(1+z),\omega,\theta). (5) Here, n is the EBL spectral density of the background radiation, and \omega_{\text{th}} the reaction threshold of the soft photon. LIV affects both \omega_{\text{th}} and the cross section \sigma(E(1+z),\omega,\theta) . We stress that the expressions for cross-section usually used in the literature are approximations, and not particularly good ones. Therefore, we followed the recommendation from [ 12 ] , which offers the most complete calculation of the cross section for pair production in any LIV scenario to date. 3 Data set and analysis The analysis method development and testing was performed on simulated data sets. These were based on the 2014 flare from Mrk 501 observed by the High Energy Stereoscopic System (H.E.S.S.) Collaboration [ 13 ] . Based on this data set, LIV studies of both time-of-flight and modified universe transparency were performed. Although these two LIV effects were tested independently, using this dataset as a sandbox will allow us to compare our results to the results of independent LIV tests. Moreover, both the spectrum and the light curve have rather simple functional forms, which reduces the extent of possible systematic effects. According to [ 13 ] , the data set consisted of \sim 2000 gamma rays between 1.3 and 20 TeV. The intrinsic spectrum was well fitted with a power law with a slope of \alpha=2.03\pm 0.04_{\mathrm{stat}}\pm 0.2_{\mathrm{sys}} . The light curve at energies below 3.25 TeV was fitted with two Gaussians with parameters (A_{1},\mu_{1}(\mathrm{s}),\sigma_{1}(\mathrm{s}))=(80.5\pm 6,2361\pm 185,2153\pm 301) , (A_{2},\mu_{2}(\mathrm{s}),\sigma_{2}(\mathrm{s}))=(60.5\pm 11,6564\pm 220,676\pm 283) , where A_{i} , \mu_{i} , \sigma_{i} are the prefactor, mean, and width of each Gaussian, respectively. 3.1 Artificial neural networks architecture We employed two different models of neural networks. A basic ANN consisting of the input layer with eight nodes, three hidden layers of 1000 nodes each, and the output layer of nine nodes. Each node in the input layer is reserved for one of the parameters that describe the spectrum or the light curve as detected on Earth. Of the nine output parameters, eight are reserved for parameters that describe the spectrum or the light curve at emission. The additional parameter represents E_{\mathrm{QG,2}} . All input and output parameters are mapped to a [0,1] interval. We use a simple mean square error to construct a cost function \mathcal{L}=\frac{1}{2mn}\sum_{i,j=1}^{m,n}\left(\hat{y}_{i}^{(j)}-y_{i}^{(j)}\right)^{2}, (6) where m is the numeber of parameters in the output layer, n is the total number of training examples, and \hat{y}_{i}^{(j)} and y_{i}^{(j)} are the reconstructed and true (targeted) values of parameters, respectively. As an alternative, a transformer, originally introduced by [ 14 ] , was chosen for its intrinsic ability to model non-local dependencies across long sequences. In this case, the input layer consists of 4000 nodes, representing the arrival time and energy of each of 2000 gamma rays in the data set. The output layer