Normalizing flows for all-orders QED corrections in lattice field theory

Nils Hermansson-Truedsson, Gurtej Kanwar · 2026

Read Normalizing flows for all-orders QED corrections in lattice field theory by Nils Hermansson-Truedsson, Gurtej Kanwar (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-2605-22444.txt.

Introduction

Hermansson-Truedsson and Kanwar develop normalizing-flow transformations for all-orders quantum electrodynamics corrections in lattice field theory. The method yields all-order corrections without extra Monte Carlo sampling, bypassing complex Wick-contraction diagrams at fixed order. Applied to lattice scalar QED in two, three and four spacetime dimensions, both analytical and machine-learned flows give estimates with far lower variance than standard methods. Flows trained on small lattices still work well on much larger lattices. A fermion generalisation is envisaged, pointing toward lattice quantum chromodynamics applications.

Catalog record

CREATORNils Hermansson-Truedsson, Gurtej Kanwar
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40510
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22444

Browse the full library · About the Tower of Babel

Full text

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

Normalizing flows for all-orders QED corrections in lattice field theory Nils Hermansson-Truedsson, Gurtej Kanwar Abstract This work develops a framework to apply normalizing-flow transformations of field configurations for all-orders Quantum Electrodynamics (QED) corrections in lattice field theory. This opens a new possibility to determine all-order corrections without the need for additional Monte Carlo sampling, generally bypassing the complexity in Wick-contraction diagrams needed at fixed order. The new method is applied to lattice scalar QED in two, three, and four spacetime dimensions, using both analytical and machine-learned flows, with both approaches yielding estimates with significantly reduced variance with respect to standard methods. It is further shown that flows can be trained using small lattice geometries and subsequently evaluated on much larger lattice geometries while maintaining good efficiency. A generalization to theories with fermions is envisaged, suggesting a path to applications in challenging field theories including lattice Quantum Chromodynamics. Introduction — Electromagnetic corrections and strong isospin-breaking effects from light-quark mass differences are needed for a growing number of low-energy precision tests of the Standard Model Aoki et al. 2026 . To date, these effects have been evaluated for a variety of observables from first-principles lattice quantum chromodynamics (QCD) with explicit quantum electrodynamics (QED) corrections Borsanyi et al. 2015 ; de Divitiis et al. 2013 ; Clark et al. 2022 ; Frezzotti et al. 2022 ; Bushnaq et al. 2023 ; Altherr et al. 2026a ; Carrasco et al. 2015 ; Giusti et al. 2018 ; Di Carlo et al. 2019 ; Boyle et al. 2023 ; Christ et al. 2023 ; Giusti et al. 2017a ; Blum et al. 2018 ; Giusti et al. 2017b ; Giusti et al. 2019 ; Aoyama et al. 2020 ; Borsanyi et al. 2021 ; Biloshytskyi et al. 2023 ; Chao et al. 2024 ; Boccaletti et al. 2026 ; Djukanovic et al. 2025 ; Parrino et al. 2025 ; Altherr et al. 2025 ; Aliberti et al. 2025 ; Bruno et al. 2026 . In this work we develop a new technique to determine electromagnetic corrections in lattice field theory with reduced variance, which has the potential to enable determining currently unknown quantities in QCD+QED of phenomenological interest, such as the neutron axial charge Tomalak and Yang 2026 ; Cirigliano et al. 2025 ; Gorchtein and Seng 2023 . We propose to construct lattice normalizing flows Tabak and Vanden-Eijnden 2010 ; Tabak and Turner 2013 ; Rezende and Mohamed 2015 ; Albergo et al. 2019 to deterministically transform field configurations from an originally uncoupled theory (e=0) to the fully interacting theory, including QED dynamics at all orders. Similar techniques have been recently applied to reduce variance of inserted operators or correlation functions in lattice QCD Abbott et al. 2024 ; Abbott et al. 2026 . Here, this method is extended and specialized for the context of QED corrections. A key innovation is the application of flows in way that exactly cancels the O(e) contributions to the variance of measured observables. This is explored first by constructing an analytical normalizing flow with this property, then training and applying machine-learned flows in a setup that enforces this cancellation and further improves variance over the analytical flow. Notably, this approach does not require new Monte Carlo sampling to be performed, making it directly applicable to existing ensembles which are often generated at great cost. This stands in contrast to several comparable methods, such as all-orders QCD+QED using C^{*} boundary conditions Lucini et al. 2016 ; Bushnaq et al. 2023 ; Altherr et al. 2025 or stochastic automatic differentiation for isospin-breaking beyond leading order Albandea et al. 2026 . This also allows QED effects to be precisely isolated by jointly analyzing correlated samples from the uncoupled and fully interacting theory. Finally, the flow approach evaluated at all orders has the computational benefit of avoiding demanding Wick contractions Borsanyi et al. 2015 ; Endres et al. 2016 ; Boyle et al. 2017 ; Altherr et al. 2025 ; Altherr et al. 2026b . To test the method, a numerical evaluation is presented for observables in scalar QED in two, three, and four dimensions. Our key numerical results are operator expectation values calculated in scalar QED to all orders in the electromagnetic coupling. Representative fits to extract the electromagnetic mass shift are also compared to results from traditional lattice methods Davoudi et al. 2019 ; Bijnens et al. 2019 , and in all cases the flow method provides controlled estimates at fixed statistics for much larger volumes than is accessible by direct evaluation. As there is no unique way to regularize QED in a finite volume Kronfeld and Wiese 1991 ; Wiese 1992 ; Kronfeld and Wiese 1993 ; Polley 1993 ; Duncan et al. 1997 ; Duncan et al. 1996 ; Hayakawa and Uno 2008 ; Asmussen et al. 2016 ; Endres et al. 2016 ; Feng and Jin 2019 ; Davoudi et al. 2019 ; Clark et al. 2022 ; Biloshytskyi et al. 2023 ; Di Carlo et al. 2025 , all numerical demonstrations here are performed for a typical choice of scheme, QED {}_{\text{L}} Hayakawa and Uno 2008 ; Davoudi et al. 2019 . However, the normalizing flow approach is generally applicable also for other finite-volume QED regularizations, and transferring flow models to these other choices represents interesting future work. These results open the door to analogous gains via machine-learned normalizing flows for lattice QCD+QED. Flows for all-orders QED effects — Starting from a theory with lattice action S_{0}(\phi) and field content generically written as \phi , QED effects can be included by introducing the photon field A_{\mu}(x) , the free photon action S_{A}(A) , and the QED interaction term S_{\rm int}(\phi,A;e) with coupling e . The free photon action typically introduces a distribution that is easy to sample, e.g., a Gaussian action which can be sampled exactly. As such, we will take as our starting point field configurations sampled according to the uncoupled theory described by the distribution r(\phi,A)=\frac{1}{Z_{0}}e^{-S_{0}(\phi)-S_{A}(A)} (1) normalized by Z_{0}\equiv\int d[\phi]\,d[A]\,e^{-S_{0}(\phi)-S_{A}(A)} . The factorized distribution allows configurations of A to be generated independently of \phi , for example being produced on-demand after an ensemble of fields \phi has already been sampled. Here we propose to use a normalizing flow to transform the uncorrelated prior distribution ( 1 ) to the all-orders interacting theory, described by the distribution p(\phi,A)=\frac{1}{Z_{e}}e^{-S_{0}(\phi)-S_{A}(A)-S_{\rm int}(\phi,A;e)}, (2) where Z_{e}\equiv\int d[\phi]\,d[A]\,e^{-S_{0}(\phi)-S_{A}(A)-S_{\rm int}(\phi,A;e)} . A normalizing flow transformation f is any a diffeomorphism with tractable Jacobian determinant Rezende and Mohamed 2015 , which here maps samples drawn from r(\phi,A) to new field configurations, f:(\phi,A)\mapsto(\phi^{\prime},A^{\prime}). (3) The transformed fields (\phi^{\prime},A^{\prime}) follow a new distribution described by the change of measure formula q(\phi^{\prime},A^{\prime})=r(\phi,A)|\det J_{f}|^{-1}, (4) where J_{f} is the Jacobian of f . By optimizing the flow f such that q(\phi^{\prime},A^{\prime})\approx p(\phi^{\prime},A^{\prime}) , this transformation provides a way to approximately map from an uncoupled ensemble to fields distributed according to the all-orders interacting theory. Even for an imperfect flow, exact reweighting factors can be computed as \hat{w}(\phi^{\prime},A^{\prime})\equiv\frac{e^{-S_{0}(\phi^{\prime})-S_{A}(A^{\prime})-S_{\rm int}(\phi^{\prime},A^{\prime};e)}}{e^{-S_{0}(\phi)-S_{A}(A)}}|\det J_{f}|. (5) These unnormalized reweighting factors provide an unbiased estimator of the ratio of partition functions, \left<\hat{w}\right>_{0}=Z_{e}/Z_{0} , leading to the normalized weights w(\phi^{\prime},A^{\prime})\equiv\hat{w}(\phi^{\prime},A^{\prime})/\left<\hat{w}\right>_{0}. (6) Here expectation values \left<\cdot\right>_{0} are taken with respect to (\phi,A) drawn from

Read the complete text