{"id":"023eaee2-d279-4a64-a6a0-58e3e5892a49","arxiv_id":"2412.01908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"KRISP uses kernel regression to interpolate the unmeasured spatial frequencies of sparse interferometric arrays, reconstructing images without training data or user tuning.","lead":"Interferometers such as the Event Horizon Telescope measure only part of the spatial-frequency map of a source, leaving gaps that make image reconstruction hard. KRISP, a new kernel-based algorithm, fills those gaps statistically from the measured data alone and recovers high-fidelity images without training on prior images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ghost-point zero-padding imposes a band-limited prior; compact components with power near bmax can bias the reconstruction inside bmax, and the formal error bars from Eq. (6) do not reflect this model misspecification.","rationale":"The reader's weakest assumption correctly identifies the stationary Gaussian-process smoothness and the ghost-point zero-forcing as the key priors. My concern sharpens this into a concrete failure mode: the ghost points are not data and are not optimized; they encode a band-limited prior. For sources with Fourier power that does not decay by bmax—e.g., a compact component whose visibility is still substantial at the maximum baseline—the reconstruction inside the observed domain will be biased by the zero-valued ghost points. Because GP posterior variance depends only on kernel and sampling locations, not on the data values, the formal error bars provide no warning of this bias. The paper's own condition for success (gaps smaller than a correlation length) does not protect against this failure, since a compact component has a long correlation length in Fourier space while also having significant power at bmax. This undercuts the abstract's 'agnostic to the underlying image' and 'without tuning by a user' statements. The recommended verdict remains CONDITIONAL: the method may work for the tested smooth, compact-source images, but the paper must either remove the overclaim, test the compact-component regime, or explicitly restrict the applicability. I also note a secondary technical issue: equations (5)–(6) appear to omit the noise term (the predictive mean should use (K + σ_n^2 I)^{-1}, and the predictive variance should include k** - k^T (K + σ_n^2 I)^{-1} k), which makes the reported uncertainty quantification inconsistent with the stated 'formal errors at data points'. This is not the central claim but should be corrected for reproducibility.","tokens_in":9780,"tokens_out":6893,"duration_ms":82957,"concrete_test":"Simulate a ground-truth image consisting of the crescent of §6.1 plus a 2%-flux, 0.5-µas Gaussian component. Run KRISP with the same ghost-point and kernel settings on the rose-pattern coverage of Fig. 10. Measure the Fourier-domain rms error within 0.9 bmax and compare the true visibility against the 1-σ posterior band from Eq. (6). If the true visibility is outside the band at >3σ or the rms error exceeds the paper's quoted few percent, the ghost-point prior is a load-bearing source of bias and the 'entirely from data' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that KRISP reconstructs the complete Fourier map up to bmax 'based entirely on the data' is load-bearing on a prior that is not data-driven: the user-specified Matérn kernel (ν=1.5) and, especially, the ghost points—random locations in an annulus [bmax, 2bmax] set to zero. These ghost points enforce a band-limitation that actively removes any true Fourier power near bmax. For images with significant high-baseline power (e.g., a compact unresolved component superposed on the ring, whose visibility decays slowly), the GP will trade off between the true visibility data inside bmax and the zero ghost values outside, pulling the reconstruction down inside bmax. The resulting bias is not reflected in the formal posterior variance of Eq. (6)—that variance depends only on the kernel and on the data locations, not on the measured visibility values—so the reported error bars do not capture model misspecification. The paper's 'agnostic to underlying image' claim therefore fails exactly in a case that is relevant to black hole imaging (hot spots, photon-ring substructure).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9987,"tokens_out":6256,"duration_ms":73277,"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":[{"comment":"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.","section":"Abstract; §3"},{"comment":"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.","section":"§2, Eqs. (5)–(6)"},{"comment":"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.","section":"§5–§6, Figures 3–12"}],"minor_comments":[{"comment":"The number of ghost points and the random seed used in the reconstructions are not stated; specifying them would improve reproducibility.","section":"§3"},{"comment":"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.","section":"§5"},{"comment":"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.","section":"§4.1"},{"comment":"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.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal scope and the core idea is promising, but the overstatement of the data-only nature of the method and the incorrect posterior-variance formula are load-bearing issues. I recommend major revision with emphasis on a compact-source bias test, correction of the GP variance, and quantitative comparisons to existing algorithms."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere'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.\n\nNow the soft spots, in order of size.\n\nFirst, 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.\n\nSecond, 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.\n\nThird, 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.\n\nThese 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.\n\nWho'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.\n\nNet: worth engaging, not ready as is.","headline":"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.","tokens_in":10541,"tokens_out":2811,"would_cite":true,"duration_ms":27988,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["interferometric imaging","kernel regression","Gaussian process","sparse u-v coverage","Event Horizon Telescope","Fourier-domain reconstruction","black hole imaging","Matern kernel"],"falsifier":"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.","tokens_in":1608,"feed_emoji":"🔭","tokens_out":2072,"duration_ms":57888,"temperature":0.7,"pith_summary":"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.","feed_headline":"KRISP rebuilds the full Fourier map from sparse telescope data","feed_subtitle":"No tuning, no training images: errors stay near 1 percent when baseline gaps stay within one correlation length.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the Gaussian process regression framework, including the posterior mean and variance equations and the hyperparameter optimization algorithm that KRISP uses.","marker":"Rasmussen & Williams 2006"},{"why":"Establishes that a compact source has a unique correlation length in the Fourier domain, which is the condition that makes gap interpolation possible.","marker":"Psaltis et al. 2024"},{"why":"Introduced the goal of reconstructing the full Fourier domain up to the maximum baseline through learned building blocks; KRISP replaces that training step with kernel regression.","marker":"Medeiros et al. 2023"},{"why":"Provides the analytic crescent model used as a ground-truth image for synthetic tests of the reconstruction.","marker":"Kamruddin & Dexter 2013"},{"why":"Gives the kernel trick derivation, showing that the fitted function depends only on the scalar product kernel and justifying the regression formulation.","marker":"Bishop 2006"}],"fun_headline_variants":["Kernel method rebuilds Fourier map from sparse telescope data","No-tuning KRISP reconstructs full Fourier map from sparse baselines","KRISP: kernel regression fills gaps in interferometric data","Sparse interferometry? KRISP interpolates the Fourier map","KRISP reconstructs complete Fourier map without user tuning"],"cache_read_input_tokens":12672,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Kernel method rebuilds Fourier map from sparse telescope data","No-tuning KRISP reconstructs full Fourier map from sparse baselines","KRISP: kernel regression fills gaps in interferometric data","Sparse interferometry? KRISP interpolates the Fourier map","KRISP reconstructs complete Fourier map without user tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000388,"raw_usage":{"total_tokens":2012,"prompt_tokens":878,"completion_tokens":1134,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1061}},"tokens_in":494,"tokens_out":1134,"duration_ms":9096,"temperature":1.0,"reasoning_tokens":1061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:50:58.237641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"E., & Williams, C","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process regression framework, including the posterior mean and variance equations and the hyperparameter optimization algorithm that KRISP uses."},{"cited_title":"R., &¨Ozel, F","cited_arxiv_id":null,"evidence_quote":"Introduced the goal of reconstructing the full Fourier domain up to the maximum baseline through learned building blocks; KRISP replaces that training step with kernel regression."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the kernel trick derivation, showing that the fitted function depends only on the scalar product kernel and justifying the regression formulation."}],"review_version":1}