REVIEW
ALPACA I: Controlling source and PSF systematics in JWST time-delay cosmography with differentiable lens modeling
Instrumentation and Methods for Astrophysics · Submitted 3 Sep 2026 · 89 references
Pith's verdict: A well-validated methods paper that makes a real practical point: joint PSF and pixelated-source reconstruction is affordable and removes most of the PSF systematic on D_Delta_t in realistic mocks, with the main caveat being that the mock PSF errors are spatially uniform.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
A fixed but imperfect JWST point spread function biases inferred time-delay distances by as much as 14 percent, and reconstructing the PSF jointly with the lens and source brings that bias to the percent level.
The reading
What carries the argument
The load-bearing object is the parametric PSF model: an analytic elliptical Moffat core plus a pixelated background grid, anchored to an initial star-based kernel estimate by a Gaussian prior scaled by a PSF error map, and regularized by a starlet-transform $\ell^1$ penalty. Because the entire forward model is differentiable, this kernel can be optimized — and later sampled — simultaneously with an elliptical power-law mass, S\'ersic lens light, a pixelated correlated-field source (a Gaussian-process prior on source pixels with an inferred power spectrum), quasar image positions, and $D_{\Delta t}$ in one likelihood. Multi-start gradient descent locates the joint maximum, and Hamiltonian Monte Carlo with geometric variational inference maps the posterior, including the inner PSF-core pixels. This combination is what reduces a full per-lens analysis from weeks to under 15 hours on a single GPU, making a flexible source and a free PSF affordable.
Extended reading notes
Core claim
The central claim is that a fixed, imperfect PSF is the dominant image-level systematic in JWST time-delay cosmography, and that fitting the PSF jointly with the lens and source in a single differentiable likelihood removes it. In a controlled ladder of ten mock JWST/NIRCam lenses, supplying distorted kernels with 0.4%, 2.0%, and 12.4% root-mean-square error from the truth produces median biases in the time-delay distance of $-$2.1%, $-$4.5%, and $-$14.1%, because the point-source residuals left by the wrong kernel are absorbed into the mass parameters. The joint reconstruction reverses this trend: median biases become $+0.7\%$, $+1.2\%$, and $+1.9\%$, comparable to the 0.84% combined statistical precision of the exact-PSF quad sample, while the recovered kernel itself reaches about 1.4% RMSE from truth in quads. The inner $(6\times\mathrm{FWHM})^2$ pixels of that kernel can be promoted to sampled parameters, and covariance maps show the remaining coupling to the mass slope and $D_{\Delta t}$ is confined to the bright core. Propagating the ten-lens posteriors through hierarchical inference recovers $H_0 = 68.95 \pm 0.83$ km s$^{-1}$ Mpc$^{-1}$ against the input 70, an offset the authors judge within expected sample variance.
Load-bearing premise
The validation assumes that a realistic PSF error can be represented as the smooth, low-order affine field distortion of Equation (23); if real JWST PSF errors are dominated by chromatic, detector, or time-varying components outside that family, the measured reduction of the $D_{\Delta t}$ bias may not fully transfer to real observations.
Editorial extensions
If this is right
- A fixed, imperfect PSF biases $D_{\Delta t}$ monotonically with mismatch size — median $-$2.1%, $-$4.5%, and $-$14.1% — so any JWST analysis that uses only a star-derived kernel carries a systematic that can exceed its statistical error bar.
- Joint PSF reconstruction brings the median bias to $+0.7\%$, $+1.2\%$, and $+1.9\%$ across the three distortion levels, placing the residual bias in the same percent regime as the 0.84% statistical precision of the exact-PSF quad sample.
- Quadruple-image systems constrain the PSF roughly twice as well as doubles — recovered kernel RMSE about 1.4% versus 2.5% — since four quasar images provide more independent constraints on the kernel.
- Sampling the inner $(6\times\mathrm{FWHM})^2$ PSF pixels together with the lens parameters is computationally feasible and shows the residual PSF–lens coupling is confined to the bright core.
- One fixed pipeline configuration, applied without per-system tuning, fits all ten lenses and recovers the input cosmology through hierarchical inference, with $H_0 = 68.95 \pm 0.83$ km s$^{-1}$ Mpc$^{-1}$ against an input of 70.
Reading between the lines
- The affine-field distortions of the PSF experiment are smooth and low-order; real JWST PSF errors also contain chromatic, temporal, and detector components, so the measured bias reduction is best read as a floor on the difficulty, with joint sampling as the mechanism that would expose un-modelled components in real data.
- The near-resolution-independent gradient cost on a GPU points toward the roughly thousand-lens LSST time-delay sample, where per-lens runtimes of hours rather than weeks are what make uniform modeling of the whole sample feasible.
- Because the mass model is fixed to an elliptical power law plus shear throughout, the mass-sheet degeneracy is excluded by construction; pairing the same joint likelihood with explicit mass-sheet exploration would extend this systematic control to the mass-profile axis.
- The PSF–parameter covariance maps suggest a transferable diagnostic: in real data, a wrong fixed PSF should leave a detectable, coherent correlation between quasar-image residuals and the recovered mass slope and $D_{\Delta t}$, giving a data-driven check on the adequacy of the PSF prior.
Original abstract
Time-delay cosmography provides an independent probe of the Hubble constant ($H_0$), but exploiting the high-resolution imaging of the James Webb Space Telescope (JWST) requires strong-lens models that capture complex source-galaxy morphologies and the point spread function (PSF) with high fidelity. Flexible source and joint PSF reconstruction are computationally expensive, motivating more rigid parameterizations that can bias the inferred time-delay distance $D_{\Delta t}$. We present ALPACA, a semi-automatic, GPU-accelerated, differentiable strong-lens modeling pipeline built on Herculens, Starred, and JAX. ALPACA jointly reconstructs a pixelated correlated-field source, the lens mass, and the PSF within a single likelihood, reducing per-system runtimes from weeks to less than 15 hours. We validate the pipeline on realistic mock JWST/NIRCam observations built from morphologically complex COSMOS galaxy sources. Holding an imperfect PSF (RMSE = 12% from truth) fixed biases $D_{\Delta t}$ by up to 14%, whereas joint PSF reconstruction reduces this bias to the percent level. We further show that the reconstructed PSF core can be marginalized over jointly with the lens parameters during posterior sampling. Propagating the per-lens posteriors of a ten-lens mock sample through hierarchical inference, we recover the input cosmology without significant bias. ALPACA shows that rigorous control of source and PSF systematics is computationally affordable, supporting reliable percent-level precision in the JWST era.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
Time-delay cosmography provides an independent probe of the Hubble constant ($H_0$), but exploiting the high-resolution imaging of the James Webb Space Telescope (JWST) requires strong-lens models that capture complex source-galaxy morphologies and the point spread function (PSF) with high fidelity. Flexible source and joint PSF reconstruction are computationally expensive, motivating more rigid parameterizations that can bias the inferred time-delay distance $D_{\Delta t}$. We present ALPACA, a semi-automatic, GPU-accelerated, differentiable strong-lens modeling pipeline built on Herculens, Starred, and JAX. ALPACA jointly reconstructs a pixelated correlated-field source, the lens mass, and the PSF within a single likelihood, reducing per-system runtimes from weeks to less than 15 hours. We validate the pipeline on realistic mock JWST/NIRCam observations built from morphologically complex COSMOS galaxy sources. Holding an imperfect PSF (RMSE = 12% from truth) fixed biases $D_{\Delta t}$ by up to 14%, whereas joint PSF reconstruction reduces this bias to the percent level. We further show that the reconstructed PSF core can be marginalized over jointly with the lens parameters during posterior sampling. Propagating the per-lens posteriors of a ten-lens mock sample through hierarchical inference, we recover the input cosmology without significant bias. ALPACA shows that rigorous control of source and PSF systematics is computationally affordable, supporting reliable percent-level precision in the JWST era.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Time-delay cosmography provides an independent probe of the Hubble constant ($H_0$), but exploiting the high-resolution imaging of the James Webb Space Telescope (JWST) requires strong-lens models that capture complex source-galaxy morphologies and the point spread function (PSF) with high fidelity. | background 0.99 | 0.06 | 0.04 | asserted 0.99 |
| s1 | Flexible source and joint PSF reconstruction are computationally expensive, motivating more rigid parameterizations that can bias the inferred time-delay distance $D_{\Delta t}$. | background 0.63 | 0.09 | 0.03 | asserted 0.88 |
| s2 | We present ALPACA, a semi-automatic, GPU-accelerated, differentiable strong-lens modeling pipeline built on Herculens, Starred, and JAX. | contribution 0.72 | 0.91 | 0.04 | asserted 0.99 |
| s3 | ALPACA jointly reconstructs a pixelated correlated-field source, the lens mass, and the PSF within a single likelihood, reducing per-system runtimes from weeks to less than 15 hours. | method 0.70 | 0.95 | 0.14 | asserted 1.00 |
| s4 | We validate the pipeline on realistic mock JWST/NIRCam observations built from morphologically complex COSMOS galaxy sources. | method 0.87 | 0.96 | 0.03 | asserted 0.99 |
| s5 | Holding an imperfect PSF (RMSE = 12% from truth) fixed biases $D_{\Delta t}$ by up to 14%, whereas joint PSF reconstruction reduces this bias to the percent level. | result 1.00 | 0.93 | 0.17 | asserted 0.99 |
| s6 | We further show that the reconstructed PSF core can be marginalized over jointly with the lens parameters during posterior sampling. | method 0.64 | 0.94 | 0.05 | asserted 0.97 |
| s7 | Propagating the per-lens posteriors of a ten-lens mock sample through hierarchical inference, we recover the input cosmology without significant bias. | result 1.00 | 0.96 | 0.13 | asserted 1.00 |
| s8 | ALPACA shows that rigorous control of source and PSF systematics is computationally affordable, supporting reliable percent-level precision in the JWST era. | interpretation 0.99 | 0.90 | 0.89 | asserted 0.97 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2609.04312 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.