Kernel Methods for Interferometric Imaging

Dimitrios Psaltis; Feryal Ozel; Yassine Ben Zineb

REVIEW 3 major objections 4 minor 20 references

Kernel Methods for Interferometric Imaging

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read KRISP, a kernel-regression imaging algorithm, reconstructs the complete Fourier map of an interferometric observation up to its longest baseline using only the data, with no user tuning or training on prior images.

desk verdict KRISP is a genuinely new, training-free interpolation scheme for sparse u-v coverage, and the synthetic tests are credible, but the 'no tuning' claim is overstated and the method needs baseline comparisons and image-domain metrics before I'd use it on real data. read the letter →

arxiv 2412.01908 v1 pith:P77UGO7R submitted 2024-12-02 astro-ph.IM

Dimitrios Psaltis , Feryal Ozel , Yassine Ben Zineb This is my paper · ORCID

classification astro-ph.IM

keywords interferometricimagingkernelregressionGaussianprocesssparseu-vcoverageEventHorizonTelescopeFourier-domainreconstructionblackholeMatern

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

KRISP is a new image-reconstruction algorithm for sparse interferometric arrays. It claims that the complete Fourier map of a source, up to the longest observed baseline, can be recovered from the measured visibilities alone, without user-tuned regularizers or training on prior images. The key quantitative claim is that reconstruction fidelity is controlled by a single ratio: the typical gap in baseline coverage divided by the correlation length of the Fourier map, which is set by the size of the source. When this ratio is at or below one, KRISP produces high-fidelity images, with reported rms errors below about 1 percent for uniform and two-satellite sampling and below about 7 percent for EHT-like 2017 Sgr A* coverage. The paper demonstrates this on analytic rings and crescents and on GRMHD simulation snapshots, including complex brightness asymmetries.

What carries the argument

The central object is Gaussian process regression with a Matern kernel, applied separately to the real and imaginary parts of the visibilities. The Matern kernel (with smoothness parameter $\nu=1.5$ and length scale $\rho$ learned from the data) defines an infinite-dimensional function space and yields a closed-form posterior mean and variance; the variance acts as an uncertainty map for the reconstructed Fourier plane. The second essential mechanism is the placement of randomly located ghost points with zero visibility in the annulus between $b_{\rm max}$ and $2b_{\rm max}$, which suppresses unmeasured high spatial frequencies while preserving conjugate symmetry. The argument that carries the method is the correlation-length condition: for a compact image the Fourier map decorrelates on a scale inversely proportional to the image size, so gaps smaller than that scale are interpolable.

What would settle it

Construct a synthetic image whose Fourier map has a sharp feature or significant power beyond the maximum baseline, sample it with EHT-like gaps at or slightly above one correlation length, run KRISP, and compare the reconstructed Fourier map and image to the ground truth; if the rms error substantially exceeds the reported few-percent level, the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that kernel regression on the real and imaginary parts of the complex visibilities, using a Matern kernel with fixed smoothness parameter $\nu=1.5$ and a correlation length learned from the data, reconstructs the full Fourier map across the observed $u$-$v$ domain even when that domain is sparsely sampled. Interpolation is possible because any compact source has a Fourier correlation length inversely proportional to image size; as long as the typical gap between measured baselines is smaller than this correlation length, the Gaussian-process posterior anchors the function between data points. The algorithm also inserts random ghost points with zero visibility between the maximum baseline and twice that baseline, forcing a gradual roll-off and enforcing conjugate symmetry so that the reconstructed map corresponds to a real image. The reported errors are typically below one percent for uniform and two-satellite coverage and below about seven percent for EHT-like coverage whenever the gap-to-correlation-length ratio is within one; the method also produces formal uncertainty maps from the Gaussian-process posterior, which naturally identify regions of poor coverage.

Load-bearing premise

The visibility field is a zero-mean stationary Gaussian process with a Matern covariance, so unmeasured baselines can be interpolated from neighboring measurements within one learned correlation length, and all Fourier power beyond the longest baseline can be safely set to zero.

Editorial extensions

If this is right

KRISP reconstructs the full Fourier plane up to the maximum observed baseline from sparse interferometric data with no user-tuned regularizer and no training images.
Reconstruction fidelity is governed by the ratio of the typical $u$-$v$ gap to the Fourier correlation length, with errors near 1 percent for uniform and two-satellite sampling and below about 7 percent for EHT-like 2017 Sgr A* coverage when the ratio is at or below one.
Ghost points between $b_{\rm max}$ and $2b_{\rm max}$ enforce a gradual roll-off to zero and conjugate symmetry, suppressing artifacts from unobserved high spatial frequencies.
Separate regression of the real and imaginary parts handles the different correlation lengths of the two components and yields formal uncertainty bands for the reconstructed Fourier map.
The method reproduces the size, width, brightness asymmetry, and substructure of complex GRMHD snapshot images at the resolution of the array.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The Gaussian-process posterior could be used to design array configurations, choosing baselines that minimize the worst-case gap-to-correlation-length ratio for a target image size.
Because the method is agnostic to the image, it should transfer to other sparse interferometric regimes, such as optical or infrared long-baseline arrays, whenever the visibility field is smooth at the gap scale; the paper itself focuses on horizon-scale black hole imaging.
A natural testable extension is to add random short-baseline points to EHT-like coverage and verify that the rms error drops toward the 1 percent level as the typical gap shrinks below one correlation length.
The same machinery could be adapted to heteroscedastic or correlated measurement errors, which the paper identifies as future work, by replacing the scalar variance with a full noise covariance in the Gaussian-process update.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Here's my take on arXiv:2412.01908. The paper does something new: instead of training dictionaries (PRIMO) or enforcing image-domain priors (CLEAN, RML), it does Gaussian process regression directly on the real and imaginary visibilities, using a Matern kernel and zero-valued 'ghost points' beyond the maximum baseline to kill the extrapolation. That separation of real and imaginary parts is smart, because their correlation lengths genuinely differ. The central scaling result - reconstruction error stays below a few percent when the typical baseline gap is within one correlation length of the Fourier map - is clean and reproducible from the description; the synthetic tests on rings, crescents, and GRMHD snapshots are appropriate for the claim.

Now the soft spots, in order of size.

First, the abstract's 'based entirely on the data without tuning by a user' is not true. The Matern smoothness nu is fixed at 1.5, and the ghost points are a user-specified band-limiting prior. The stress-test concern is real: for a compact component whose visibility decays slowly, the zero-valued ghost points in [bmax, 2bmax] will pull the GP prediction down inside bmax, and Eq. (6) gives error bars that ignore that model misspecification. That's a legitimate caveat and the paper should state it openly.

Second, there are no comparisons against existing imaging algorithms. The paper claims high fidelity, but without imaging CLEAN or RML on the same synthetic data, it's hard to know whether KRISP is actually better or just different. There are also no quantitative image-domain metrics (e.g., normalized cross-correlation or structural similarity) for the final images; the reported rms errors are all in the Fourier domain.

Third, the method assumes directly measurable phases. That excludes ground-based EHT-like observations without closure phases. The paper acknowledges this and defers to future work, which is honest, but it means the headline application is not yet testable on real EHT data. No code or data are released, which makes independent verification harder.

These are not fatal flaws. The idea is novel, the experiments support the central claim for the tested configurations, and the limitations are acknowledged at least in part. But the framing needs to be toned down, and the evidence needs to be widened before this is a tool I'd trust on real sparse-array data.

Who's this for? Anyone working on sparse-array imaging, especially EHT/ngEHT and space interferometry. I'd send it to peer review - it deserves serious referee time, likely major revision. I'd probably cite it if I were working in that space.

Net: worth engaging, not ready as is.

Referee Report

3 major / 4 minor

Summary. The paper introduces KRISP, a Gaussian-process / kernel-regression method for reconstructing the complex visibility function (Fourier map) of a compact source from sparsely sampled interferometric data. The method fits the real and imaginary parts of the visibilities separately using a Matérn kernel with fixed smoothness ν=1.5 and data-optimized length scale, and it appends randomly located zero-valued "ghost points" in the annulus [b_max, 2 b_max] to suppress extrapolation beyond the maximum baseline. The authors demonstrate the method on synthetic thin-ring, crescent, and GRMHD images, reporting Fourier-domain rms errors of about 1% for uniform and two-satellite sampling and up to about 7% for EHT-like sampling when the typical uv-gap is within one correlation length. They also provide formal GP uncertainty bands for the reconstructed Fourier maps. The central claims are that KRISP reconstructs the complete Fourier map up to b_max "based entirely on the data," is agnostic to the underlying image, requires no user tuning or training set, and yields high-fidelity images for gap sizes comparable to the correlation length.

Significance. If the claims hold, KRISP would be a simple, training-free alternative to PRIMO and a statistically well-motivated interpolation tool for sparse arrays such as EHT and proposed space interferometers. The paper's strengths are its clear presentation of the GP framework and the synthetic tests across several array geometries and image complexities; the demonstrated dependence of the error on the ratio of baseline spacing to Fourier correlation length is a concrete, falsifiable prediction. However, the central novelty claim of "data-only, no-prior" reconstruction is undermined by the ghost-point band-limiting prior, and the formal uncertainty formula appears incorrect as written. The absence of quantitative image-domain error metrics and of comparisons to existing algorithms (CLEAN, RML, PRIMO) further limits the strength of the high-fidelity claims. These issues are fixable, but they require additional analysis and revision before the paper can be accepted.

major comments (3)

[Abstract; §3] The claim that KRISP reconstructs the complete Fourier map "based entirely on the data without tuning by a user" is not supported by the algorithm as described. The ghost points of §3 are user-specified pseudo-data set to zero at random locations in the annulus [b_max, 2 b_max], with the number of ghost points and the zero-value range chosen by the user; in addition, the Matérn smoothness ν=1.5 is fixed. These choices impose a band-limited GP prior on the reconstruction. If the true visibility function has significant power near b_max (for example, because the image contains a compact hot spot or photon-ring substructure), the regression will trade off the measured visibilities against the zero ghost values and bias the reconstruction inside b_max. A concrete test would be to add a compact Gaussian component to the crescent model and measure the bias as a function of its flux and angular size. The paper should either qualify the "based entirely on the data" claim or demonstrate with such a test that the ghost-point prior does not materially bias the reconstruction in the regimes of interest.
[§2, Eqs. (5)–(6)] The posterior variance formula in Eq. (6) is incorrect for Gaussian-process regression with measurement noise. For independent Gaussian noise of variance σ², the predictive mean and variance should be <V(b)> = k^T (K + σ² I)^{-1} V and σ_b² = k(b,b) − k^T (K + σ² I)^{-1} k. Equation (6), as written, uses K^{-1} without the noise term and subtracts from σ²; it is not the GP posterior variance and can become negative. Since §6 uses Eq. (6) to draw the uncertainty bands in Figures 8, 10, and 12, the quoted error bars and the "statistically robust" claim need to be re-derived and recomputed with the correct formula.
[§5–§6, Figures 3–12] The quantitative evaluation is limited to Fourier-domain rms error measured over the domain up to 0.9 b_max. There are no image-domain error metrics (e.g., normalized root-mean-square error, structural similarity index, or image fidelity) and no quantitative comparisons to existing imaging algorithms such as CLEAN, RML, or PRIMO. The claims of "high fidelity" for the crescent and GRMHD reconstructions in Figures 7, 9, and 11 are supported only by visual inspection. Adding quantitative image-domain metrics and at least one baseline comparison to an established method is necessary to substantiate the central claim that KRISP reproduces images with high fidelity.

minor comments (4)

[§3] The number of ghost points and the random seed used in the reconstructions are not stated; specifying them would improve reproducibility.
[§5] The definition of "typical rms error" and the exact domain over which it is computed should be stated precisely; the phrase "up to a fraction (typically 0.9) of the maximum baseline length" is ambiguous about the mask and whether ghost-point regions are excluded.
[§4.1] The two-satellite coverage is modeled by the analytic curve in Eq. (9); the text should clarify that this is a toy model and not the actual uv-coverage from orbital simulations.
[§2] The symbol k is used both for the kernel function in Eq. (3) and for the magnitude of the spatial-frequency vector in Eq. (10); renaming one of these would remove ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GP interpolation, benchmark comparisons, and ghost-point prior are explicit and self-contained, with only a non-load-bearing self-citation.

full rationale

KRISP is a standard Gaussian-process regression: Eq. (4)-(6) give the posterior mean and variance of an interpolated visibility field given measured visibilities, a Matérn kernel (Eq. 8), and a learned lengthscale ρ. The reconstruction at unmeasured baselines is not a refit of the same data points presented as a prediction; it is an interpolation at new locations under a data-driven covariance. Optimizing ρ from the same data is standard empirical Bayes, and the paper's fidelity claims are tested against known ground-truth Fourier maps (analytic rings, crescents, and GRMHD images), so the benchmark is not circular. The one self-citation that touches the central premise is Psaltis et al. (2024), cited for the correlation-length sufficiency condition. That citation overlaps with the present authors, but the present paper does not treat it as a black box: Figures 3, 4, and 6 recompute the condition for KRISP on analytic and GRMHD images, so the self-citation is not load-bearing in a circular sense. The ghost points are an explicitly disclosed prior (zero visibilities beyond bmax) rather than a hidden fitted input; they can bias reconstructions for sources with substantial power near bmax, and the formal variance from Eq. (6) does not capture that model misspecification, but this is a modeling/calibration limitation, not a circular derivation. The paper also states its assumptions (equal uncorrelated errors, measurable complex phases) as caveats for future work rather than disguising a fitted quantity as an independent prediction. Overall, no claim in the paper reduces by construction to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central reconstruction rests on a Gaussian process prior, a fixed smoothness, a learned lengthscale, an imposed zero-taper beyond bmax, and an equal-variance direct-phase data model. The paper introduces no new physical entities; the ghost points are the most consequential non-data element.

free parameters (3)

Matern kernel lengthscale rho = not reported; optimized by scikit-learn on each dataset
This hyperparameter sets the correlation length in the u-v plane and is learned from the visibility data via Gaussian process marginal likelihood (Section 3). It controls how interpolation behaves between sparse samples.
Matern smoothness nu = 1.5 (fixed)
Chosen by hand rather than inferred; it determines the smoothness of the interpolated Fourier map and is an input assumption of the algorithm (Section 3).
Ghost point count, placement, and zero-value range = not specified precisely; visibilities set to zero for baseline lengths in [bmax, 2bmax]
The algorithm adds user-specified random pseudo-data beyond bmax to enforce a low-pass constraint. The number and distribution of ghost points are not quantified in the paper, and results may depend on them.

assumptions (5)

domain assumption The measured complex visibilities are samples from a zero-mean stationary Gaussian process with a Matern covariance kernel.
This is the central modeling prior for KRISP; it is assumed in Section 3 without independent justification beyond the image-compactness argument.
domain assumption For any compact source, the Fourier map has a finite correlation length inversely proportional to the image size, and gaps smaller than this length permit accurate interpolation.
The paper invokes Psaltis et al. 2024 for this criterion (Sections 1 and 5), and it is the validity boundary of the algorithm.
domain assumption Real and imaginary parts of visibilities are Gaussian, uncorrelated, and homoscedastic, and can be measured directly.
Stated in Section 3; the authors acknowledge this holds for space or long-wavelength interferometers but not current ground-based millimeter arrays, and the results depend on it.
standard math Mercer's theorem and Gaussian process regression formulas from Rasmussen and Williams (2006) hold as used.
Standard background used in Section 2 to construct equations (5) and (6).
ad hoc to paper Visibilities vanish smoothly beyond the maximum baseline, so setting ghost points to zero in [bmax, 2bmax] is appropriate.
This is an invented regularization choice (Section 3), not derived from data or first principles; it directly affects the reconstructed image's resolution taper.

invented entities (1)

Ghost points (zero-valued pseudo-visibilities beyond the maximum baseline)
purpose: Suppress high spatial frequency content and enforce a smooth taper to zero outside the observed u-v domain, so that the inverse Fourier transform yields a real image without ringing.
These are user-inserted, not measured, data. They impose an assumption that the source has no significant power beyond bmax; the paper does not validate this assumption against real data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Kernel Methods for Interferometric Imaging." pith.science (2026). https://pith.science/paper/P77UGO7R

@misc{pith2026241201908,
  author       = {Pith},
  title        = {Pith review of: Kernel Methods for Interferometric Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P77UGO7R}},
  note         = {Machine review of arXiv:2412.01908}
}

read the original abstract

Increasing the angular resolution of an interferometric array requires placing its elements at large separations. This often leads to sparse coverage and introduces challenges to reconstructing images from interferometric data. We introduce a new interferometric imaging algorithm, KRISP, that is based on kernel methods, is statistically robust, and is agnostic to the underlying image. The algorithm reconstructs the complete Fourier map up to the maximum observed baseline length based entirely on the data without tuning by a user or training on prior images and reproduces images with high fidelity. KRISP works efficiently for many sparse array configurations even in the presence of significant image structure as long as the typical baseline separation is comparable to or less than the correlation length of the Fourier map, which is inversely proportional to the size of the target image.

Figures

Figures reproduced from arXiv: 2412.01908 by the authors.

**Figure 1.** Three examples of sparse interferometric coverage: (left) uniformly spaced, (middle) EHT array for 2017 observations of Sgr A∗, (right) simple representation of two satellites at different orbits. In each panel, the red points represent the locations of the measurements, whereas the blue points show the randomly located ghost points used to suppress the high spatial frequencies. The blue circles identify the size of… view at source ↗

**Figure 2.** (Top left) The 2D Fourier transform of an infinitesimally thin ring with and an example of a uniform 2D grid used for sampling the function. (Top right) The approximate function reconstructed using Kernel methods throughout the observed domain. (Bottom left) The difference between the exact and the approximate function. The black horizontal line shows the rms error calculated on the observed domain. (Bottom right) T… view at source ↗

**Figure 3.** The typical rms error between the Fourier map of an infinitesimally thin ring and its reconstruction obtained from sampling on a uniform grid plotted against the spacing between successive baselines divided by the correlation length of each Fourier map. The blue and red curve correspond to reconstructions with maximum baseline length of 12 Gλ and 24 Gλ, respectively. The error remains ≲1% for baseline spacings sma… view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Same panels as [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Same panels as [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Same as [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: The ground truth and KRISP reconstruction for an example asymmetric crescent image with data uncertainties and Fourier maps shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: The three panels in the top row show the ground-truth and reconstructed Fourier maps as well as the difference between the two for an asymmetric crescent image described in the text and shown in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: (Left) Example image from a snapshot of a GRMHD simulation; (Middle) the same image filtered at the resolution of the maximum baseline of an array with sparse sampling shown in [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

**Figure 10.** Figure 10: Same as Fig, 8 for the example shown in [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

**Figure 11.** Figure 11: Same as [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

**Figure 12.** Figure 12: Same as [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 5 canonical work pages

[1]

2017, AJ, 153, 159, doi: 10.3847/1538-3881/aa6302

Akiyama, K., Ikeda, S., Pleau, M., et al. 2017, AJ, 153, 159, doi: 10.3847/1538-3881/aa6302

doi:10.3847/1538-3881/aa6302 2017
[2]

D., & Kloppenborg, B

Baron, F., Monnier, J. D., & Kloppenborg, B. 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7734, Optical and Infrared Interferometry II, ed. W. C. Danchi, F. Delplancke, & J. K. Rajagopal, 77342I, doi: 10.1117/12.857364 Ben Zineb, Y., Ozel, F., & Psaltis, D. 2024, ApJ, submitted

doi:10.1117/12.857364 2010
[3]

Bishop, C. M. 2006, Pattern Recognition and Machine Learning (Springer)

work page 2006
[4]

C., Talbot, N

Cawley, G. C., Talbot, N. L., Foxall, R. J., Dorling, S. R., & Mandic, D. P. 2004, Neurocomputing, 57, 105, doi: https://doi.org/10.1016/j.neucom.2004.01.005

work page doi:10.1016/j.neucom.2004.01.005 2004
[5]

A., Johnson, M

Chael, A. A., Johnson, M. D., Narayan, R., et al. 2016, ApJ, 829, 11, doi: 10.3847/0004-637X/829/1/11

doi:10.3847/0004-637x/829/1/11 2016
[6]

Clark, B. G. 1980, A&A, 89, 377

1980
[7]

L., Ravi, V., & Hallinan, G

Connor, L., Bouman, K. L., Ravi, V., & Hallinan, G. 2022, MNRAS, 514, 2614, doi: 10.1093/mnras/stac1329

doi:10.1093/mnras/stac1329 2022
[8]

D., & Pfuhl, O

Eisenhauer, F., Monnier, J. D., & Pfuhl, O. 2023, ARA&A, 61, 237, doi: 10.1146/annurev-astro-121622-045019 Event Horizon Telescope Collaboration. 2019a, ApJL, 875, L2, doi: 10.3847/2041-8213/ab0c96 —. 2019b, The Astrophysical Journal Letters, 875, L1, doi: 10.3847/2041-8213/ab0ec7 —. 2022, ApJL, 930, L16, doi: 10.3847/2041-8213/ac6672

doi:10.1146/annurev-astro-121622-045019 2023

Show all 20 references

[9]

2023, AandA, 677, A167, doi: 10.1051/0004-6361/202347073 H¨ ogbom, J

Geyer, F., Schmidt, K., Kummer, J., et al. 2023, AandA, 677, A167, doi: 10.1051/0004-6361/202347073 H¨ ogbom, J. A. 1974, A&AS, 15, 417

2023 doi
[10]

B., & Dexter, J

Kamruddin, A. B., & Dexter, J. 2013, MNRAS, 434, 765, doi: 10.1093/mnras/stt1068

2013 doi
[11]

S., Khartov, V

Kardashev, N. S., Khartov, V. V., Abramov, V. V., et al. 2013, Astronomy Reports, 57, 153, doi: 10.1134/S1063772913030025

2013 doi
[12]

Kulkarni, S. R. 1989, AJ, 98, 1112, doi: 10.1086/115202

1989 doi
[13]

2024, arXiv e-prints, arXiv:2411.12233, doi: 10.48550/arXiv.2411.12233

Hoefs, L. 2024, arXiv e-prints, arXiv:2411.12233, doi: 10.48550/arXiv.2411.12233

2024 doi arXiv
[14]

R., &¨Ozel, F

Medeiros, L., Psaltis, D., Lauer, T. R., &¨Ozel, F. 2023, ApJ, 943, 144, doi: 10.3847/1538-4357/acaa9a

2023 doi
[15]

2020, arXiv e-prints, arXiv:2004.06210, doi: 10.48550/arXiv.2004.06210

Ozel, F. 2020, arXiv e-prints, arXiv:2004.06210, doi: 10.48550/arXiv.2004.06210

2020 doi arXiv
[16]

Psaltis, D., Ozel, F., Medeiros, L., & Lauer, T. R. 2024, arXiv e-prints, arXiv:2408.10322, doi: 10.48550/arXiv.2408.10322

2024 doi arXiv
[17]

E., & Williams, C

Rasmussen, C. E., & Williams, C. K. I. 2006, Gaussian Processes in Machine Learning (MIT Press) 12

2006
[18]

Rogers, A. E. E., Doeleman, S. S., & Moran, J. M. 1995, AJ, 109, 1391, doi: 10.1086/117371

1995 doi
[19]

2022, AandA, 664, A134, doi: 10.1051/0004-6361/202142113

Schmidt, K., Geyer, F., Fr¨ ose, S., et al. 2022, AandA, 664, A134, doi: 10.1051/0004-6361/202142113

2022 doi
[20]

R., Moran, J

Thompson, A. R., Moran, J. M., & Swenson, George W., J. 2017, Interferometry and Synthesis in Radio Astronomy, 3rd Edition, doi: 10.1007/978-3-319-44431-4

2017 doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

2017, AJ, 153, 159, doi: 10.3847/1538-3881/aa6302

Akiyama, K., Ikeda, S., Pleau, M., et al. 2017, AJ, 153, 159, doi: 10.3847/1538-3881/aa6302

doi:10.3847/1538-3881/aa6302 2017

[2] [2]

D., & Kloppenborg, B

Baron, F., Monnier, J. D., & Kloppenborg, B. 2010, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 7734, Optical and Infrared Interferometry II, ed. W. C. Danchi, F. Delplancke, & J. K. Rajagopal, 77342I, doi: 10.1117/12.857364 Ben Zineb, Y., Ozel, F., & Psaltis, D. 2024, ApJ, submitted

doi:10.1117/12.857364 2010

[3] [3]

Bishop, C. M. 2006, Pattern Recognition and Machine Learning (Springer)

work page 2006

[4] [4]

C., Talbot, N

Cawley, G. C., Talbot, N. L., Foxall, R. J., Dorling, S. R., & Mandic, D. P. 2004, Neurocomputing, 57, 105, doi: https://doi.org/10.1016/j.neucom.2004.01.005

work page doi:10.1016/j.neucom.2004.01.005 2004

[5] [5]

A., Johnson, M

Chael, A. A., Johnson, M. D., Narayan, R., et al. 2016, ApJ, 829, 11, doi: 10.3847/0004-637X/829/1/11

doi:10.3847/0004-637x/829/1/11 2016

[6] [6]

Clark, B. G. 1980, A&A, 89, 377

1980

[7] [7]

L., Ravi, V., & Hallinan, G

Connor, L., Bouman, K. L., Ravi, V., & Hallinan, G. 2022, MNRAS, 514, 2614, doi: 10.1093/mnras/stac1329

doi:10.1093/mnras/stac1329 2022

[8] [8]

D., & Pfuhl, O

Eisenhauer, F., Monnier, J. D., & Pfuhl, O. 2023, ARA&A, 61, 237, doi: 10.1146/annurev-astro-121622-045019 Event Horizon Telescope Collaboration. 2019a, ApJL, 875, L2, doi: 10.3847/2041-8213/ab0c96 —. 2019b, The Astrophysical Journal Letters, 875, L1, doi: 10.3847/2041-8213/ab0ec7 —. 2022, ApJL, 930, L16, doi: 10.3847/2041-8213/ac6672

doi:10.1146/annurev-astro-121622-045019 2023