Pith. sign in

REVIEW 3 major objections 5 minor 94 references

This paper shows that four lensed image positions carry enough information for a simple neural network to predict a lens's mass and ellipticity, cutting mass-model fitting from days to minutes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 04:22 UTC pith:BQA4OUQC

load-bearing objection A genuinely new ML initialisation scheme for quad lens modelling with an honest simulation test; the large speed-up claim over conventional fitting is not actually demonstrated. the 3 major comments →

arxiv 2607.22812 v1 pith:BQA4OUQC submitted 2026-07-24 astro-ph.GA astro-ph.COastro-ph.HE

Speeding up Gravitational Lens Mass Models with Machine Learning: Applications in X-ray Astronomy

Alex Ostridge , Rafael Mart\'inez-Galarza , J\'ulia M. Sisk-Reyn\'es , Daniel A. Schwartz , Anna Barnacka This is my paper
classification astro-ph.GA astro-ph.COastro-ph.HE
keywords gravitational lensingquadruply lensed quasarssingular isothermal ellipsoidneural networkmass modelcaustic methodX-ray astrometryparameter inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that the slowest step in strong-lens mass modelling — the search for good starting parameters — can be replaced by two plain fully connected neural networks. Given only the four observed positions of a quadruply lensed quasar, the networks predict the mass scale b' and ellipticity of a singular isothermal ellipsoid (SIE) lens, and those predictions feed a standard optimiser that converges to image positions at 0.5 milliarcsecond accuracy in minutes rather than days. The authors demonstrate this on 200 unseen simulated systems, 151 of which converge in a mean 126 seconds, and on seven real Gaia-observed quads, all of which converge within 22-319 seconds. A sympathetic reader would care because the caustic method for X-ray astrometry depends on fast mass models, and upcoming surveys are expected to add thousands of new quads to analyse.

Core claim

The central claim is that the geometry of four lensed image positions — coordinates measured from the intersection of opposing images plus ratios of image separations — is enough input for a neural network to predict the mass scale and ellipticity of an SIE lens. Two small feed-forward networks reach RMS errors of 0.148 arcsec on the mass parameter and 0.030 on ellipticity for held-out simulations without lens rotation; with random lens orientations the errors grow to 0.165 and 0.180, mostly from uncertainty in the analytic angle estimate. Seeded with these predictions, optimisation reaches chi-squared < 8 (0.5 mas per image coordinate) for 151 of 200 random simulated quads in a mean 126 sec

What carries the argument

The central objects are the SIE lens model — a power-law mass distribution with ellipticity and a mass scale b' that equals the Einstein radius only for circular lenses — and two independent fully connected networks, each with two hidden layers and ReLU activations, whose inputs are translation-invariant coordinates of the four images relative to the intersection of opposing image pairs, plus ratios of image separation distances; fluxes label only the brightest image as 'A' for consistent ordering. A third piece is the analytic ellipticity-angle estimate used in the paper's Eq. 9: the angle is one quarter of the sum of the four image position angles about the lens centre, plus or minus 45 de

Load-bearing premise

The whole method depends on the assumption that a network trained on noiseless SIE simulations with zero external shear and zero ellipticity rotation returns useful starting parameters for real quads, which contain external shear, astrometric noise, and possibly a perturbing galaxy.

What would settle it

Feed the trained networks the image positions of 200 simulated SIE quads that include external shear (say gamma 0.1-0.3) and 1 milliarcsecond astrometric noise, then run the published optimisation: if the convergence fraction to chi-squared < 8 falls well below the 151/200 (75 per cent) seen without shear, the zero-shear training assumption is the load-bearing weakness.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Mass modelling of a quad drops from roughly days to two minutes for the majority of systems, making it feasible to model large samples rather than a handful of hand-picked lenses.
  • Because only the image positions are needed, the method is instrument-agnostic: any facility resolving four point images can feed the same networks after centring and rotating the coordinates.
  • The fast models are sufficient for the caustic method's X-ray astrometry, which relies on the gradient of the lens potential (image positions) rather than the full potential shape needed for time-delay cosmography.
  • The failure analysis shows that non-convergence tracks the accuracy of the network's initial b' and ellipticity, so improving those predictions should directly raise the convergence fraction above 75 per cent.
  • The paper's simulation grid and code are released, so the training set can be extended to include external shear and random orientation, which the paper identifies as the immediate next step.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same architecture were retrained on SIE+shear simulations, the ellipticity-shear degeneracy could be absorbed by adding a third output for shear; the paper's observed sixfold jump in ellipticity error when random rotations are introduced suggests the angle estimate, not the network, is the dominant error source.
  • A purely geometric image-labelling scheme would remove the flux-ordering discontinuity that causes the sharp prediction-error lines in the caustic interior; the paper notes this but does not test it, so a simple testable extension would be to relabel images by distance pattern and measure whether the error contours smooth out.
  • The 0.5 mas 'convergence' threshold is a pragmatic engineering tolerance, not a calibrated fit quality: the paper itself acknowledges that the number of free parameters varies between optimisation steps, so chi-squared < 8 is not a uniform reduced chi-squared. Readers should compare models using the same stopping rule rather than interpreting the absolute numbers as goodness of fit.
  • A stress test following directly from the zero-shear training assumption would be to run the released networks on the shear-dominated quads the paper cites: if those systems require the optimiser to wander far from the network's start, the method's advertised speed-up may be confined to ellipticity-dominated systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a simulation-based machine-learning initialiser for SIE gravitational lens mass models. Two fully connected neural networks are trained on >6 million noiseless SIE quadruple-image configurations (zero external shear, ellipticities aligned with PA=0) to predict the mass parameter b' and ellipticity epsilon from the four image positions and distance ratios. The predictions are used as starting points for lensmodel optimisation of an SIE+shear model. On 200 unseen simulated systems with random ellipticity rotations, 151 converge to chi^2<8, with RMS errors of 0.165 in b' and 0.180 in epsilon; mean convergence time for successful cases is 126.1 s. The same pipeline is applied to seven real Gaia quads, which all converge in 22–319 s. The paper claims this accelerates mass modelling by several orders of magnitude and enables caustic-method X-ray astrometry on large forthcoming samples.

Significance. If the central speed-up claim were established, the method would be a practical contribution to strong-lensing mass modelling for the caustic method, and the public code and reproducibility are commendable. The simulation-based held-out evaluation and the explicit discussion of limitations are strengths. However, the paper's headline claims go beyond what the evidence supports: there is no baseline comparison to random or manual initialisation, and the transfer to real systems with external shear is not tested despite the authors' own citation that 15/39 observed quads are shear-dominated. The real-data demonstration also fits the same Gaia positions that are later described as 'predicted', so it cannot serve as an external validation. These issues are fixable within the paper's scope, but they are load-bearing for the claimed acceleration and for the method's applicability to a large fraction of real quads.

major comments (3)
  1. [§4.2 and Table 4] The central claim of 'several orders of magnitude' acceleration is not supported by any baseline comparison. The paper reports that NN-initialised simulated systems converge in a mean of 126.1 s and real systems in 22–319 s, but it gives no runtime for random initial guesses, literature-based initialisations, or a manual 'educated guess' pipeline on the same systems. Without such a baseline, the observed convergence times cannot be attributed to the NN predictions. Please add a controlled comparison on the 200 simulated systems and the seven real quads, e.g. random b', epsilon draws and/or the authors' previous manual procedure, with the same optimisation budget and stopping criterion.
  2. [§2.1, §5.1.1 and Eq. (9)] The training set contains zero external shear and the validation set (Sec. 4.2) also has zero shear, yet the method is intended for real quads, 15/39 of which are shear-dominated according to the paper's own citation of Luhtaru et al. (2021). The ellipticity-angle estimate of Eq. (9) assumes no shear, and the network has never seen shear-distorted image configurations. The 200-system validation therefore cannot probe the failure mode most relevant to a large fraction of real systems. Please test the pipeline on simulated SIE+shear systems with shear values representative of the observed population, and report convergence rates and runtimes separately for ellipticity- and shear-dominated configurations. Table 4 makes this concern concrete: WFI2033-4723 converges to epsilon=0.82, outside the training domain (0.05–0.60), and gamma=0.50, suggesting the NN initial estimate was far from the fi
  3. [§5.1 and Table 4] The real-data validation is partly circular as presented. The Gaia image positions are used both to construct the input features and as the optimisation target; the statement that the final models 'predict the observed lensed image positions in Gaia Data Release 3' is therefore a description of the fitting target, not an out-of-sample prediction. That all seven systems reach chi^2<8 with up to 7 free parameters is expected and does not by itself validate the initialiser. Please provide an out-of-sample check, such as withholding one image position during optimisation, using independent HST/radio positions, or testing whether the NN initial parameters are substantially closer to the converged solution than random draws on these same systems.
minor comments (5)
  1. [Abstract and §1.4] Please qualify the phrase 'several orders of magnitude'; no baseline runtime is reported for manual or random initialisation, so this specific magnitude claim is currently unsupported.
  2. [Table 2] The column 'Percentage Error' is undefined. State whether it is the mean absolute percentage error, the RMS percentage error, or another metric, and give the corresponding formula.
  3. [§3.2] The image labelling scheme uses model-predicted fluxes to select image A. In real data, observed fluxes are affected by microlensing and variability. The paper discusses this in Sec. 5.3.3, but it would help to state explicitly which observed flux catalogue (e.g. Gaia G-band) is used for labelling the seven real quads and whether any system is close to a labelling ambiguity.
  4. [§5.1.1] The sentence 'most resulting models are ellipticity dominated and cannot have large external shear' appears inconsistent with Table 4, where WFI2033-4723 has gamma=0.50. Please reconcile or clarify whether 'most' refers to the simulated population only.
  5. [§3.4] The fixed chi^2<8 threshold is acknowledged to be a pragmatic stopping criterion rather than a statistically optimal one. This is acceptable, but the statement in the abstract and conclusions that models reproduce positions to <0.005 arcsec should be phrased as 'to within the adopted stopping criterion', since the optimisation is halted at this threshold and could be refined further.

Circularity Check

1 steps flagged

The core ML training/evaluation is independent, but the real-data 'prediction' of Gaia image positions is a refit of the same positions, and the caustic-method motivation rests on self-citations.

specific steps
  1. fitted input called prediction [Abstract; Sec. 5.1 'Observed Lensed Sources'; Table 4 caption]
    "The final optimised mass models for each quasar predict the observed lensed image positions in Gaia Data Release 3. ... All seven systems converged to χ2 < 8 for their predicted image positions and reproduced the observed image positions to within ±0.5 milliarcseconds."

    The optimisation in Eqs. (10)–(11) minimises χ2 between 'Pred. Image Pos.' and 'Measured Image Pos.', and Table 4 reports the resulting models after stopping at χ2 < 8. Thus the claim that the optimised mass models 'predict' the observed Gaia positions is a restatement that the optimiser fit those same positions; no independent or held-out Gaia positions are used. This is fitted input called prediction. It does not invalidate the simulated holdout evaluation, but the real-data wording overstates the evidential content.

full rationale

The central machine-learning claim is not circular: the networks are trained on >6 million simulated SIE lens configurations and evaluated on held-out simulated systems, with training and test sets separated at random. The NN predictions of b' and ellipticity are checked against ground-truth simulation parameters, not against the values used to build the network, so this is genuine supervised learning rather than a self-fulfilling construction. The 200-system optimisation test also uses unseen simulated systems. The main circular element is the abstract/real-data language: the seven quads are fit to Gaia DR3 positions and then those same positions are called 'predicted.' That is a fit-quality statement, not a prediction. The self-citations (Barnacka 2017, 2018; Sisk-Reynés et al. 2026) motivate the caustic method and provide background, but they are not load-bearing for the NN speed-up result, which is independently demonstrated on simulations. The paper also candidly admits the training set excludes external shear and that this may limit real-data representativeness, which is a robustness concern rather than circularity. Overall, the derivation is largely self-contained, with one notable circular phrase and a modest self-citation footprint, hence a score of 4.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 0 invented entities

The central method rests on a physical model choice (SIE, alpha=1), a training-distribution choice (no shear, no noise, zero ellipticity angle), and several geometric approximations (lens centre via ellipse fit, ellipticity angle via Eq. 9). There are no new particles or forces. The fitted mass, ellipticity, shear, centre, and angle for each real system are the free parameters that make the demonstration work; the network hyperparameters are additional hand-chosen degrees of freedom.

free parameters (6)
  • Lens mass parameter b' (arcsec) for each real quad = 0.755–2.454 (Table 4)
    Predicted by the neural network and refined by chi-square optimisation; it is the central scale of the SIE model and is not fixed by theory.
  • Lens ellipticity epsilon = 0.0211–0.8218 (Table 4)
    Predicted by the neural network and refined by optimisation; degenerate with external shear, so values are not unique.
  • External shear gamma = 0.0721–0.5008 (Table 4)
    Not present in the training data; introduced during optimisation to fit real quads. Its inclusion is ad hoc relative to the trained model.
  • Lens centre (x_c, y_c) = Not reported; estimated from ellipse fit with RMS ~1.28 arcsec
    Estimated from the four image positions and then optimised; the uncertainty is large compared to the 0.5 mas target.
  • Ellipticity position angle theta = Two candidates: theta ± 45 deg from Eq. 9
    Computed from image position angles assuming a known lens centre; mean error ~2.73 deg on simulated tests; critical input to the network.
  • Network hyperparameters (layer sizes, dropout, learning rate) = 100-40/50 nodes, dropout 0.2, learning rate 0.001
    Chosen by hand/validation and affect prediction accuracy, but they are not physical parameters.
axioms (8)
  • domain assumption SIE with alpha=1 (isothermal) describes the mass distribution of galaxy-scale deflectors.
    Invoked in Sec. 2.1 and used for all training simulations; if real lenses deviate strongly from isothermal SIE, the network input distribution is mismatched.
  • domain assumption No external shear in the training data.
    Sec. 2.1: 'we assume no external shear as this parameter is degenerate with the ellipticity of the lensing galaxy'; real systems may contain shear, and the paper introduces shear only during optimisation.
  • domain assumption The ellipticity angle estimator of Eq. (9) is valid for SIE lenses producing quads and remains useful when shear is present.
    Sec. 2.2 uses Kassiola & Kovner (1995); for real systems with shear this is only an approximate initial estimate, and the paper relies on it before network input.
  • domain assumption The lens centre can be approximated by the centre of an ellipse fitted to the four image positions.
    Sec. 2.2.1 reports RMS error ~1.28 arcsec in each coordinate; this is much larger than the 0.5 mas target and could bias the feature preprocessing.
  • ad hoc to paper Training on noiseless simulated positions transfers to noisy Gaia astrometry.
    Sec. 5.4 acknowledges the simulated data have no measurement errors and that performance on real data may be affected; the paper does not test robustness to input noise.
  • domain assumption The chi-squared threshold <8 with sigma=0.5 mas is an adequate stopping criterion for caustic-method mass models.
    Sec. 3.4 notes this is not chi2_red=1 and is a pragmatic bound tied to Gaia astrometry; the number of active free parameters varies, so the statistical meaning is not uniform.
  • domain assumption A single SIE+shear model is sufficient to reproduce observed image positions for the seven real quads.
    Sec. 5.1 asserts this suffices for X-ray astrometry despite known degeneracies; the authors acknowledge the models are not general enough for time-delay cosmography.
  • standard math gravlens/lensmodel correctly computes image positions for given lens parameters.
    The paper relies entirely on this external software for simulations and optimisation; no independent verification is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 23811 in / 12826 out tokens · 129867 ms · 2026-08-01T04:22:02.458333+00:00 · methodology

0 comments
read the original abstract

Multi-wavelength observations of quadruply lensed quasars constitute a powerful probe of cosmology, dark matter substructure along the line of sight, and the structure of X-ray emitting regions in high-redshift quasars. These investigations are conditional on acquiring an accurate model for the surface mass density of matter lensing these quasars. We propose a simulation-based machine learning method to accelerate parameter inference in real quadruply lensed systems by several orders of magnitude. We simulate a grid of quadruply lensed sources with Singular Isothermal Ellipsoid (SIE) lenses and use the projected positions of the four lensed images to train two fully connected neural networks that predict the mass parameter and ellipticity. For a large fraction of simulated systems, the neural network-initialised mass models converge in time-scales of a few minutes and recover the source position at the <0.''005 level for a broad range of lens masses and ellipticities. We apply our neural networks to seven quadruply lensed quasars, lensed by isolated galaxies or a galaxy-perturber pair, which have archival Chandra observations. The final optimised mass models for each quasar predict the observed lensed image positions in Gaia Data Release 3. These mass models enable the caustic method, which locates the X-ray-to-optical emission regions to milliarcsecond precision in these otherwise unresolvable systems, improving the effective angular resolution of Chandra at high-z by up to two orders of magnitude. Our approach accelerates this mass modelling by supplying informed initial parameters, enabling application to the many new quadruply lensed systems expected from forthcoming surveys.

Figures

Figures reproduced from arXiv: 2607.22812 by Alex Ostridge, Anna Barnacka, Daniel A. Schwartz, J\'ulia M. Sisk-Reyn\'es, Rafael Mart\'inez-Galarza.

Figure 1
Figure 1. Figure 1: Schematic overview of the end-to-end methodology used in this work to infer and optimise gravitational lens mass models. Supervised training (minimise loss) Simulated image positions (RA, Dec) Relative image separations (distance ratios) Trained neural network (𝑏 ′ , 𝜖 regressor) Validation on unseen simulations [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of the neural network training workflow. hyper-parameters used. In Sec. 4 we describe the results of training the neural network on the simulated data, and present some metrics to evaluate the success of our method. Finally, in Sec. 5 we discuss the astrophysical implications of applying our method to real systems. MNRAS 000, 1–16 (0000) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The 10,008 source positions plotted in the source plane for a mass parameter of 2.4, ellipticity of 0.15, with a lens centre (0,0). The number of sources at each distance from the lens centre is decided by a truncated normal distribution. ellipse is aligned with the declination axis). This choice is made for simplicity, as it has no discernible effect on the underlying physics. Training a fully connected n… view at source ↗
Figure 5
Figure 5. Figure 5: A histogram with 30 bins, displaying the absolute error (in degrees) in the estimated angle of ellipticity of the simulated lenses, for a batch of 100 random lens configurations and source positions. 2.73 degrees. This ellipticity angle estimate is within an acceptable range for the majority of cases such that varying the ellipticity angle during optimisation should lead to an accurate final mass model. 2.… view at source ↗
Figure 6
Figure 6. Figure 6: The absolute error on the mass parameter predicted by the neural network for 1,000,000 unseen source positions generated from a uniform distribution across the inner caustic of an SIE lens configuration with a mass parameter of 2. ′′5 and ellipticity 0.25. Sources are placed up to within 0. ′′01 of the caustic to avoid producing 3 images due to lensmodel software limitations [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 7
Figure 7. Figure 7: The absolute error on the mass parameter predicted by the neural network for 1,000,000 unseen source positions generated from a uniform distribution across the inner caustic of an SIE lens configuration with a mass parameter of 5. ′′5 and ellipticity 0.55. Sources are placed up to within 0. ′′01 of the caustic to avoid producing 3 images due to lensmodel software limitations [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 8
Figure 8. Figure 8: Violin plots of the predicted mass parameter distribution for all testing data sources (1.2 million) across the 55 simulated mass parameters. A dashed identity line is plotted highlighting the true mass parameter values. The width of each violin is fixed and therefore cannot be compared. The red lines on the plots mark the 2.5 th and 97.5 th percentiles for each mass parameter [PITH_FULL_IMAGE:figures/ful… view at source ↗
Figure 9
Figure 9. Figure 9: Violin plots of the predicted ellipticity distribution for all testing data sources (1.2 million) across the 12 simulated ellipticities. A dashed identity line is plotted highlighting the true ellipticity values. The width of each violin is fixed and therefore cannot be compared. The red lines on the plot mark the 2.5 th and 97.5 th percentiles for each ellipticity. MNRAS 000, 1–16 (0000) [PITH_FULL_IMAGE… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

94 extracted references · 19 canonical work pages · 3 internal anchors

  1. [1]

    Mass-sheet degeneracy, power-law models and external convergence: Impact on the determination of the Hubble constant from gravitational lensing , volume =

    Schneider, Peter and Sluse, Dominique , year =. Mass-sheet degeneracy, power-law models and external convergence: Impact on the determination of the Hubble constant from gravitational lensing , volume =. Astronomy & Astrophysics , doi =

  2. [2]

    Monthly Notices of the Royal Astronomical Society , volume =

    Wong, Kenneth C and Suyu, Sherry H and Chen, Geoff C-F and Rusu, Cristian E and Millon, Martin and Sluse, Dominique and Bonvin, Vivien and Fassnacht, Christopher D and Taubenberger, Stefan and Auger, Matthew W and Birrer, Simon and Chan, James H H and Courbin, Frederic and Hilbert, Stefan and Tihhonova, Olga and Treu, Tommaso and Agnello, Adriano and Ding...

    2020

  3. [3]

    On model-dependent bounds on H(0) from gravitational images Application of Q0957 + 561A,B , volume =

    Falco, Emilio and Gorenstein, Marc and Shapiro, Irwin , year =. On model-dependent bounds on H(0) from gravitational images Application of Q0957 + 561A,B , volume =. The Astrophysical Journal , doi =

  4. [4]

    Weak lensing by line-of-sight haloes as the origin of flux-ratio anomalies in quadruply lensed QSOs: Weak lensing by line-of-sight haloes , volume=

    Inoue, Kaiki Taro and Takahashi, Ryuichi , year=. Weak lensing by line-of-sight haloes as the origin of flux-ratio anomalies in quadruply lensed QSOs: Weak lensing by line-of-sight haloes , volume=. Monthly Notices of the Royal Astronomical Society , publisher=. doi:10.1111/j.1365-2966.2012.21915.x , number=

    arXiv 2012

  5. [5]

    Chen, Jacqueline and Rozo, Eduardo and Dalal, Neal and Taylor, James E. , year=. Astrometric Perturbations in Substructure Lensing , volume=. The Astrophysical Journal , publisher=. doi:10.1086/512002 , number=

    doi:10.1086/512002

  6. [6]

    Monthly Notices of the Royal Astronomical Society , volume =

    Rusu, Cristian E and Wong, Kenneth C and Bonvin, Vivien and Sluse, Dominique and Suyu, Sherry H and Fassnacht, Christopher D and Chan, James H H and Hilbert, Stefan and Auger, Matthew W and Sonnenfeld, Alessandro and Birrer, Simon and Courbin, Frederic and Treu, Tommaso and Chen, Geoff C-F and Halkola, Aleksi and Koopmans, Léon V E and Marshall, Philip J ...

    doi:10.1093/mnras/stz3451 2020

  7. [7]

    and Schneider, P

    Bradač, M. and Schneider, P. and Steinmetz, M. and Lombardi, M. and King, L. J. and Porcas, R. , year=. B1422+231: The influence of mass substructure on strong lensing , volume=. Astronomy & Astrophysics , publisher=. doi:10.1051/0004-6361:20020559 , number=

    work page doi:10.1051/0004-6361:20020559

  8. [8]

    Nierenberg, A. M. and Treu, T. and Wright, S. A. and Fassnacht, C. D. and Auger, M. W. , year=. Detection of substructure with adaptive optics integral field spectroscopy of the gravitational lens B1422+231 , volume=. Monthly Notices of the Royal Astronomical Society , publisher=. doi:10.1093/mnras/stu862 , number=

    doi:10.1093/mnras/stu862

  9. [9]

    Nierenberg, A. M. and Treu, T. and Brammer, G. and Peter, A. H. G. and Fassnacht, C. D. and Keeton, C. R. and Kochanek, C. S. and Schmidt, K. B. and Sluse, D. and Wright, S. A. , year=. Probing dark matter substructure in the gravitational lens HE 0435−1223 with the WFC3 grism , volume=. Monthly Notices of the Royal Astronomical Society , publisher=. doi:...

    doi:10.1093/mnras/stx1400

  10. [10]

    , keywords =

    Determining the Spatial Origin of X-Ray and Optical Emission in the z = 3.1 Strongly Lensed Radio-quiet Quasar GraL J065904.1+162909 to Hundreds of Parsecs. , keywords =. doi:10.3847/1538-4357/ae3f2c , archivePrefix =. 2509.22797 , primaryClass =

    Pith/arXiv arXiv

  11. [11]

    and Storfer, C

    Baltasar, Saul and Ratier-Werbin, Nicolas and Huang, Xiaosheng and Sheu, W. and Storfer, C. J. and Hsu, Y.-M. and Xu, Sean and Schlegel, David J. , title =. arXiv e-prints , year =. 2601.18787 , archivePrefix =

    arXiv

  12. [12]

    , keywords =

    Recoiling black holes: prospects for detection and implications of spin alignment. , keywords =. doi:10.1093/mnras/stv2646 , archivePrefix =. 1508.01524 , primaryClass =

    Pith/arXiv arXiv

  13. [13]

    arXiv e-prints , keywords =

    POLISH'ing the Sky: Wide-Field and High-Dynamic Range Interferometric Image Reconstruction with Application to Strong Lens Discovery. arXiv e-prints , keywords =

  14. [14]

    , keywords =

    Strong gravitational lensing with upcoming wide-field radio surveys. , keywords =. doi:10.1093/mnras/staf1381 , archivePrefix =. 2412.01746 , primaryClass =

    Pith/arXiv arXiv

  15. [15]

    , keywords =

    Milliarcsecond X-Ray Positions and X-Ray Varstrometry for the Strongly Lensed Active Galactic Nucleus HE 0435-1223. , keywords =. doi:10.3847/1538-4357/add71a , archivePrefix =. 2505.08077 , primaryClass =

    Pith/arXiv arXiv

  16. [16]

    2021 , note =

    Astronomy and Astrophysics , author =. 2021 , note =. doi:10.1051/0004-6361/202141758 , abstract =

    doi:10.1051/0004-6361/202141758 2021

  17. [17]

    2021 , note =

    Astronomy and Astrophysics , author =. 2021 , note =. doi:10.1051/0004-6361/202039574 , abstract =

    doi:10.1051/0004-6361/202039574 2021

  18. [18]

    Astronomy and Astrophysics , author =

    Modeling lens potentials with continuous neural fields in galaxy-scale strong lenses , volume =. Astronomy and Astrophysics , author =. 2023 , note =. doi:10.1051/0004-6361/202245126 , abstract =

    work page doi:10.1051/0004-6361/202245126 2023

  19. [19]

    The Astrophysical Journal , author =

    Hierarchical. The Astrophysical Journal , author =. 2021 , note =. doi:10.3847/1538-4357/abdf59 , abstract =

    work page doi:10.3847/1538-4357/abdf59 2021

  20. [20]

    Huang, Kuan-Wei and Chih-Fan Chen, Geoff and Chang, Po-Wen and Lin, Sheng-Chieh and Hsu, Chia-Jung and Thengane, Vishal and Yao-Yu Lin, Joshua , month = oct, year =. Strong. doi:10.48550/arXiv.2210.04143 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.04143

  21. [21]

    and Schechter, Paul L

    Wynne, Raymond A. and Schechter, Paul L. , month = aug, year =. Robust modeling of quadruply lensed quasars (and random quartets) using. doi:10.48550/arXiv.1808.06151 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1808.06151

  22. [22]

    An overview of gradient descent optimization algorithms , url =

    Ruder, Sebastian , month = jun, year =. An overview of gradient descent optimization algorithms , url =. doi:10.48550/arXiv.1609.04747 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1609.04747

  23. [23]

    Proceedings of the International Astronomical Union , author =

    Lensed radio arcs at milli-arcsecond resolution:. Proceedings of the International Astronomical Union , author =. 2022 , keywords =. doi:10.1017/S1743921323004106 , abstract =

    work page doi:10.1017/s1743921323004106 2022

  24. [24]

    Monthly Notices of the Royal Astronomical Society , author =

    Testing strong lensing subhalo detection with a cosmological simulation , volume =. Monthly Notices of the Royal Astronomical Society , author =. 2022 , note =. doi:10.1093/mnras/stac2779 , abstract =

    doi:10.1093/mnras/stac2779 2022

  25. [25]

    2016 , note =

    The Astrophysical Journal , author =. 2016 , note =. doi:10.3847/0004-637X/819/2/154 , abstract =

    doi:10.3847/0004-637x/819/2/154 2016

  26. [26]

    The Astrophysical Journal , author =

    A. The Astrophysical Journal , author =. 1994 , note =. doi:10.1086/187520 , abstract =

    doi:10.1086/187520 1994

  27. [27]

    Dynamics and

    Bovy, Jo , month = feb, year =. Dynamics and

  28. [28]

    The Astrophysical Journal , author =

    Resolving. The Astrophysical Journal , author =. 2021 , note =. doi:10.3847/1538-4357/ac0909 , abstract =

    doi:10.3847/1538-4357/ac0909 2021

  29. [29]

    , keywords =

    Proper motion in lensed radio jets at redshift 3: A possible dual super-massive black hole system in the early Universe. , keywords =. doi:10.1051/0004-6361/201935427 , archivePrefix =. 1908.11756 , primaryClass =

    Pith/arXiv arXiv 1908

  30. [30]

    The Astrophysical Journal , author =

    Milliarcsecond. The Astrophysical Journal , author =. 2022 , note =. doi:10.3847/1538-4357/ac68eb , abstract =

    doi:10.3847/1538-4357/ac68eb 2022

  31. [31]

    , volume =

    Isothermal elliptical gravitational lens models. , volume =. Astronomy and Astrophysics , author =. 1994 , note =

    1994

  32. [32]

    Quasar Lensing

    Courbin, F. and Saha, P. and Schechter, P. L. , month = aug, year =. Quasar. doi:10.48550/arXiv.astro-ph/0208043 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.astro-ph/0208043

  33. [33]

    Annual Review of Astronomy and Astrophysics , author =

    Strong. Annual Review of Astronomy and Astrophysics , author =. 2010 , note =. doi:10.1146/annurev-astro-081309-130924 , abstract =

    doi:10.1146/annurev-astro-081309-130924 2010

  34. [34]

    The Astrophysical Journal , author =

    What. The Astrophysical Journal , author =. 2021 , note =. doi:10.3847/1538-4357/abfda1 , abstract =

    doi:10.3847/1538-4357/abfda1 2021

  35. [35]

    Monthly Notices of the Royal Astronomical Society , author =

    Gravitational lensing and galaxy shape , volume =. Monthly Notices of the Royal Astronomical Society , author =. 1985 , note =. doi:10.1093/mnras/213.4.871 , abstract =

    work page doi:10.1093/mnras/213.4.871 1985

  36. [36]

    2021 , month=Apr, pages=

    <i>Gaia</i>Early Data Release 3: Parallax bias versus magnitude, colour, and position , volume=. 2021 , month=Apr, pages=. doi:10.1051/0004-6361/202039653 , journal=

    doi:10.1051/0004-6361/202039653 2021

  37. [37]

    10.1051/0004-6361/202039709

    <i>Gaia</i> Early Data Release 3 - The astrometric solution , DOI= "10.1051/0004-6361/202039709", url= "https://doi.org/10.1051/0004-6361/202039709", journal =

    doi:10.1051/0004-6361/202039709

  38. [38]

    Monthly Notices of the Royal Astronomical Society , author =

    Gravitationally lensed quasars and supernovae in future wide-field optical imaging surveys , volume =. Monthly Notices of the Royal Astronomical Society , author =. 2010 , note =. doi:10.1111/j.1365-2966.2010.16639.x , abstract =

    arXiv 2010

  39. [39]

    , month = feb, year =

    Keeton, Charles R. , month = feb, year =. Computational

  40. [40]

    Software for

    Keeton, Chuck , month = jan, year =. Software for

  41. [41]

    , month = jun, year =

    Keeton, Charles R. , month = jun, year =. A. doi:10.48550/arXiv.astro-ph/0102341 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.astro-ph/0102341

  42. [42]

    The Astrophysical Journal , author =

    Fermat's. The Astrophysical Journal , author =. 1986 , note =. doi:10.1086/164709 , abstract =

    doi:10.1086/164709 1986

  43. [43]

    arXiv e-prints , author =

    Lectures on. arXiv e-prints , author =. 1996 , note =. doi:10.48550/arXiv.astro-ph/9606001 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.astro-ph/9606001 1996

  44. [44]

    Introduction to basic concepts of gravitational lensing , volume =

    Narayan, Ramesh and Wallington, Sylvanie , editor =. Introduction to basic concepts of gravitational lensing , volume =. Gravitational. 1992 , doi =

    1992

  45. [45]

    Saha, Prasenjit and Sluse, Dominique and Wagner, Jenny and Williams, Liliya L. R. , month = jan, year =. Essentials of strong gravitational lensing , url =. doi:10.48550/arXiv.2401.04165 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.04165

  46. [46]

    Reports on Progress in Physics , author =

    The dark matter of gravitational lensing , volume =. Reports on Progress in Physics , author =. 2010 , pages =. doi:10.1088/0034-4885/73/8/086901 , abstract =

    doi:10.1088/0034-4885/73/8/086901 2010

  47. [47]

    arXiv e-prints , keywords =

    Lens Modeling of STRIDES Strongly Lensed Quasars using Neural Posterior Estimation. arXiv e-prints , keywords =. doi:10.48550/arXiv.2410.10123 , archivePrefix =. 2410.10123 , primaryClass =

    work page doi:10.48550/arxiv.2410.10123

  48. [48]

    The SL2S Galaxy-scale Lens Sample. V. Dark Matter Halos and Stellar IMF of Massive Early-type Galaxies Out to Redshift 0.8. , keywords =. doi:10.1088/0004-637X/800/2/94 , archivePrefix =. 1410.1881 , primaryClass =

    Pith/arXiv arXiv

  49. [49]

    arXiv e-prints , keywords =

    Searching for strong lensing by late-type galaxies in UNIONS. arXiv e-prints , keywords =. doi:10.48550/arXiv.2503.10610 , archivePrefix =. 2503.10610 , primaryClass =

    work page doi:10.48550/arxiv.2503.10610

  50. [50]

    H _ 0 Lenses in COSMOGRAIL's Wellspring: program overview

    H0LiCOW - I. H _ 0 Lenses in COSMOGRAIL's Wellspring: program overview. , keywords =. doi:10.1093/mnras/stx483 , archivePrefix =. 1607.00017 , primaryClass =

    Pith/arXiv arXiv

  51. [51]

    , keywords =

    Microlensing of Strongly Lensed Quasars. , keywords =. doi:10.1007/s11214-024-01043-8 , archivePrefix =. 2312.00931 , primaryClass =

    Pith/arXiv arXiv

  52. [52]

    Gaia GraL: Gaia DR2 Gravitational Lens Systems. VI. Spectroscopic Confirmation and Modeling of Quadruply Imaged Lensed Quasars. , keywords =. doi:10.3847/1538-4357/ac0f04 , archivePrefix =. 2012.10051 , primaryClass =

    Pith/arXiv arXiv 2012

  53. [53]

    , keywords =

    STRIDES: automated uniform models for 30 quadruply imaged quasars. , keywords =. doi:10.1093/mnras/stac2235 , archivePrefix =. 2206.04696 , primaryClass =

    Pith/arXiv arXiv

  54. [54]

    Discovery of 24 lensed quasars

    Gravitationally lensed quasars in Gaia - II. Discovery of 24 lensed quasars. , keywords =. doi:10.1093/mnras/sty911 , archivePrefix =. 1803.07601 , primaryClass =

    Pith/arXiv arXiv

  55. [55]

    150 new lenses, quasar pairs, and projected quasars

    Gravitationally lensed quasars in Gaia - IV. 150 new lenses, quasar pairs, and projected quasars. , keywords =. doi:10.1093/mnras/stac3721 , archivePrefix =. 2206.07714 , primaryClass =

    Pith/arXiv arXiv

  56. [56]

    , keywords =

    Searching for Strong Gravitational Lenses. , keywords =. doi:10.1007/s11214-024-01042-9 , archivePrefix =. 2310.13466 , primaryClass =

    Pith/arXiv arXiv

  57. [57]

    , keywords =

    Observational Evidence of Active Galactic Nuclei Feedback. , keywords =. doi:10.1146/annurev-astro-081811-125521 , archivePrefix =. 1204.4114 , primaryClass =

    Pith/arXiv arXiv

  58. [58]

    , keywords =

    Coevolution (Or Not) of Supermassive Black Holes and Host Galaxies. , keywords =. doi:10.1146/annurev-astro-082708-101811 , archivePrefix =. 1304.7762 , primaryClass =

    Pith/arXiv arXiv

  59. [59]

    Handbook of X-ray and Gamma-ray Astrophysics , year = 2022, editor =

    AGN Feedback in Groups and Clusters of Galaxies. Handbook of X-ray and Gamma-ray Astrophysics , year = 2022, editor =. doi:10.1007/978-981-16-4544-0_122-1 , adsurl =

    doi:10.1007/978-981-16-4544-0_122-1 2022

  60. [60]

    , keywords =

    Properties and evolution of dual and offset AGN in the ASTRID simulation at z 2. , keywords =. doi:10.1093/mnras/stad834 , archivePrefix =. 2208.04970 , primaryClass =

    Pith/arXiv arXiv

  61. [61]

    , keywords =

    The Gaia mission. , keywords =. doi:10.1051/0004-6361/201629272 , archivePrefix =. 1609.04153 , primaryClass =

    Pith/arXiv arXiv

  62. [62]

    New COSMOGRAIL time delays of HE 0435-1223: H _ 0 to 3.8 per cent precision from strong lensing in a flat CDM model

    H0LiCOW - V. New COSMOGRAIL time delays of HE 0435-1223: H _ 0 to 3.8 per cent precision from strong lensing in a flat CDM model. , keywords =. doi:10.1093/mnras/stw3006 , archivePrefix =. 1607.01790 , primaryClass =

    Pith/arXiv arXiv

  63. [63]

    , keywords =

    PS J1721+8842: a gravitationally lensed dual AGN system at redshift 2.37 with two radio components. , keywords =. doi:10.1093/mnrasl/slab106 , archivePrefix =. 2109.03253 , primaryClass =

    Pith/arXiv arXiv

  64. [64]

    , keywords =

    Andromeda s Parachute: A Bright Quadruply Lensed Quasar at z = 2.377. , keywords =. doi:10.3847/1538-4357/aaaeb7 , archivePrefix =. 1707.05873 , primaryClass =

    Pith/arXiv arXiv

  65. [65]

    , keywords =

    J1721+8842: a gravitationally lensed binary quasar with a proximate damped Lyman- absorber. , keywords =. doi:10.1051/0004-6361/202142138 , archivePrefix =. 2109.01144 , primaryClass =

    Pith/arXiv arXiv

  66. [66]

    , keywords =

    J1721+8842: The first Einstein zigzag lens. , keywords =. doi:10.1051/0004-6361/202452970 , archivePrefix =. 2411.04177 , primaryClass =

    Pith/arXiv arXiv

  67. [67]

    Lens mass model of HE 0435-1223 and blind measurement of its time-delay distance for cosmology

    H0LiCOW - IV. Lens mass model of HE 0435-1223 and blind measurement of its time-delay distance for cosmology. , keywords =. doi:10.1093/mnras/stw3077 , archivePrefix =. 1607.01403 , primaryClass =

    Pith/arXiv arXiv

  68. [68]

    Summary of the astrometric, photometric, and survey properties

    Gaia Data Release 1. Summary of the astrometric, photometric, and survey properties. , keywords =. doi:10.1051/0004-6361/201629512 , archivePrefix =. 1609.04172 , primaryClass =

    Pith/arXiv arXiv

  69. [69]

    Summary of the contents and survey properties

    Gaia Early Data Release 3. Summary of the contents and survey properties. , keywords =. doi:10.1051/0004-6361/202039657 , archivePrefix =. 2012.01533 , primaryClass =

    Pith/arXiv arXiv 2012

  70. [70]

    Summary of the contents and survey properties

    Gaia Data Release 2. Summary of the contents and survey properties. , keywords =. doi:10.1051/0004-6361/201833051 , archivePrefix =. 1804.09365 , primaryClass =

    Pith/arXiv arXiv

  71. [71]

    , keywords =

    Dual AGN in the Horizon-AGN simulation and their link to galaxy and massive black hole mergers, with an excursus on multiple AGN. , keywords =. doi:10.1093/mnras/stac1217 , archivePrefix =. 2112.07193 , primaryClass =

    Pith/arXiv arXiv

  72. [72]

    , keywords =

    The cosmic evolution of massive black holes in the Horizon-AGN simulation. , keywords =. doi:10.1093/mnras/stw1123 , archivePrefix =. 1602.01941 , primaryClass =

    Pith/arXiv arXiv

  73. [73]

    2025 , eprint=

    Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine A -- System overview and lens catalogue. 2025 , eprint=

    2025

  74. [74]

    2025 , eprint=

    Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine B -- Early strong lens candidates from visual inspection of high velocity dispersion galaxies , author=. 2025 , eprint=

    2025

  75. [75]

    2025 , eprint=

    Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine C -- Finding lenses with machine learning , author=. 2025 , eprint=

    2025

  76. [76]

    2025 , eprint=

    Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine D -- Double-source-plane lens candidates , author=. 2025 , eprint=

    2025

  77. [77]

    2025 , eprint=

    Euclid Quick Data Release (Q1): The Strong Lensing Discovery Engine E -- Ensemble classification of strong gravitational lenses: lessons for Data Release 1 , author=. 2025 , eprint=

    2025

  78. [78]

    Astronomy and Astrophysics , author =

    Streamlined lensed quasar identification in multiband images via ensemble networks , volume =. Astronomy and Astrophysics , author =. 2023 , note =. doi:10.1051/0004-6361/202347332 , abstract =

    work page doi:10.1051/0004-6361/202347332 2023

  79. [79]

    2024 , note =

    The Astronomical Journal , author =. 2024 , note =. doi:10.3847/1538-3881/ad395e , abstract =

    work page doi:10.3847/1538-3881/ad395e 2024

  80. [80]

    Research Notes of the AAS , author =

    Predictions for. Research Notes of the AAS , author =. 2020 , note =. doi:10.3847/2515-5172/abc4ea , abstract =

    doi:10.3847/2515-5172/abc4ea 2020

Showing first 80 references.