{"id":"dbb4082d-5f43-4111-a109-1cfa748ae2b4","arxiv_id":"2507.18610","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Parallax imaging is derived as a quadratic approximation to the direct ptychography aperture overlap kernel, enabling a Fourier-based upsampling algorithm that relaxes the scan Nyquist limit in direct ptychography.","lead":"This paper shows that parallax imaging is a first-order Taylor approximation of the direct ptychography kernel, and uses that link to build an algorithm that up-samples direct ptychography reconstructions beyond the scan step size. The method could make dose-efficient, non-iterative phase retrieval practical for beam-sensitive materials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upsampling limits rest on an unproven, empirically calibrated D_vBF ≳ 1 criterion (Eq. 10), tested on only three defocus values; the algorithm may work while the predicted range of validity is wrong.","rationale":"The reader's weakest assumption is the same one I would flag: the empirically calibrated D_vBF criterion in Eq. (10) is the load-bearing rule for when upsampling works, and it is neither derived nor tested beyond three defocus values. The paper's core insight—that parallax is a truncated Taylor expansion of the direct ptychography kernel and that the BF-pixel loop with phase diversity can be exploited for upsampling—is well motivated and supported by the simulated and experimental reconstructions shown. The concern does not attack the algorithm's mechanics in the tested regime; it attacks the generality of the predicted limits. Because the paper presents itself as establishing relaxed sampling requirements generally, and the quantitative boundary is a heuristic calibrated on a narrow set, this is a real but addressable limitation rather than a fatal flaw. The incomplete data availability statement ('[ToDo: Add Zenodo link]') and the qualitative experimental comparisons are secondary issues; they do not change the central assessment. I therefore recommend keeping the reader's CONDITIONAL verdict unchanged, with the condition being a more rigorous characterization of the upsampling validity boundary.","tokens_in":13245,"tokens_out":13394,"duration_ms":149384,"concrete_test":"Simulate a weak-phase object at fixed defocus Δf = 100 Å and scan step s chosen so D_vBF = 0.5, 1.0, 1.5, and 2.0, with Poisson noise at fluences matching Figure 4 (e.g., 10, 100, 1000 e/Ų). For each setting, run Algorithm 1 at 2× and 4× upsampling and compute the Fourier ring correlation of the reconstruction against the known ground-truth phase. Then repeat the same protocol with a pure coma term C_{3,1} whose max|∇χ| matches the defocus case. If the D_vBF ≈ 1 threshold shifts by more than ~30% with upsampling factor, noise level, or aberration type, Eq. (10) is not the general validity limit claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that direct ptychography can be upsampled by 2–4× and that the achievable limit grows with aberration magnitude—is governed by Eq. (10), the vBF diversity rule D_vBF = λ|k_BF||Δf|/s ≳ 1. The paper calls this a 'rule of thumb', gives no derivation, and calibrates it on only three defocus values (50, 100, 200 Å) in Table 1. The physical idea is plausible: defocus creates parallax shifts between BF detector pixels, and D_vBF measures the shift spread relative to the scan step. But the threshold of 1, the linear scaling with |Δf|, and the omission of the upsampling factor f, detector pixel count, noise level, and aberration type are all assumed, not demonstrated. Algorithm 1's upsampling step tiles the sub-sampled FFT and applies a matched filter; it contains no explicit alias-inversion step. Suppression of aliased spectral copies therefore relies on phase diversity across BF pixels, yet the paper provides no conditioning or null-space analysis showing when the tiled matched filter actually inverts the alias mixing. If the threshold shifts with f or with noise, or fails for non-defocus aberrations of comparable magnitude, the abstract's statement that resolution is no longer tied to the scan step size is broader than what is demonstrated. This is a concern about the claimed range of validity, not a demonstrated failure of the implementation in the tested cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a theoretical connection between direct ptychography and parallax imaging (tcBF-STEM) by showing that the parallax kernel is a first-order Taylor expansion of the aperture overlap function in the aberration phase. It derives the parallax CTF as the aperture autocorrelation times -sin[χ(q)] and proposes Algorithm 1, which upsamples direct ptychography reconstructions by tiling the Fourier-transformed virtual bright-field images and applying a matched filter constructed from the full aperture overlap kernel. The authors claim that, under the weak-phase object approximation, direct ptychography can be upsampled by 2-4x beyond the scan Nyquist limit provided the vBF diversity ratio D_vBF = λ|k_BF||Δf|/s is ≳ 1 (Eq. 10), and they support this with simulated CTF recovery, low-fluence simulations on apoferritin, and experimental reconstructions of MOFs, gold nanoparticles, twisted MoS2, and virus-like particles.","tokens_in":13576,"tokens_out":2667,"duration_ms":28660,"significance":"If the central claim holds, the paper would make a useful contribution by decoupling direct ptychography resolution from scan step size in the weak-phase regime, with immediate applicability to dose-sensitive biological and organic materials. The strengths of the manuscript include a clean analytical derivation of the parallax CTF from the aperture overlap function (Eqs. 4, 7, 8), reproduction of the known TEM CTF modulation, a transparent algorithmic description, and a public implementation in the quantem package with notebooks for the CTF results. The conceptual framing of parallax imaging as a quadratic approximation to direct ptychography is illuminating and likely to be pedagogically valuable. The numerical and experimental demonstrations are broadly consistent with the theory, although the quantitative range of validity of the upsampling criterion is not established with the same rigor as the CTF derivation.","major_comments":[{"comment":"The upsampling limit is governed by the vBF diversity rule D_vBF = λ|k_BF||Δf|/s ≳ 1, which the paper calls a 'rule of thumb' and presents without derivation. The criterion is calibrated on only three defocus values (50, 100, 200 Å, Table 1) and for a single aberration type (defocus). The central quantitative claim—that upsampling factors of 2x-4x are achievable and grow with aberration magnitude—depends on this threshold. The manuscript does not show that the threshold is invariant to the upsampling factor f, detector pixel count, noise level, or aberration type. Please either provide a derivation of D_vBF from the aliasing/conditioning analysis of the reconstruction operator, or substantially broaden the numerical validation to include variations in f, noise, detector geometry, and non-defocus aberrations. As written, the abstract's statement that 'resolution is no longer tied to the scan step size' is broader than what Eq. (10) and Table 1 demonstrate.","section":"Relaxing Sampling Requirements, Eq. (10) and Table 1"},{"comment":"Algorithm 1 tiles the sub-sampled FFT (line 3) and applies the matched filter Γ*/|Γ| (line 6), but the paper provides no conditioning or null-space analysis showing that this operation actually inverts the alias mixing caused by sub-Nyquist scan sampling. The algorithm has no explicit alias-inversion step; suppression of aliased copies must rely on phase diversity across the BF detector pixels. Without an analysis of when the linear map from the upsampled object to the measured vBF images is invertible, or at least a numerical conditioning study as a function of D_vBF, f, and noise, it is unclear whether the observed success in Figure 3 is robust or a property of the specific simulation parameters. This is load-bearing because the algorithm's advertised capability is precisely to recover information beyond the scan Nyquist limit.","section":"Algorithm 1 and 'Relaxing Sampling Requirements'"},{"comment":"The experimental validation is qualitative. For Figure 5, the text states that 'many of the salient features of the scan-sampled reconstruction persist' in the upsampled reconstruction, but no quantitative metric (e.g., Fourier ring correlation against the scan-sampled reference, resolution estimates, or SNR) is provided. Figure 6 similarly shows side-by-side images without quantitative comparison. Since the paper's central claim is about recovering information beyond the scan sampling limit, quantitative evaluation of the upsampled experimental reconstructions—especially for the 4x and 1/16-dose case using 6 e/Ų—would substantially strengthen the claim and is needed to support the stated limits.","section":"Experimental Reconstructions (Figures 5 and 6)"}],"minor_comments":[{"comment":"The second exponential in Eq. (2) appears to be missing a closing parenthesis: 'ei[𝜒(𝒒+𝒌−𝜒(𝑘))]' should likely read 'ei[𝜒(𝒒+𝒌)−𝜒(𝑘)]'. Please check and correct.","section":"Equation (2)"},{"comment":"Typo: 'In the present of higher-order aberrations' should be 'In the presence of higher-order aberrations'.","section":"Page 4, caption of Figure 2"},{"comment":"The Data Availability statement contains the placeholder '[ToDo: Add Zenodo link]'. This must be resolved before publication, and the statement should specify the license and access conditions for the processed datasets and code.","section":"Data Availability"},{"comment":"The tiling operation 'tile_f' is not fully specified: it is not stated whether the tiling is performed in the detector-frequency grid after rotation or in the scan-frequency grid, nor how the edges of the sub-sampled FFT are handled (zero-padding versus periodic replication). A precise definition would improve reproducibility.","section":"Algorithm 1, line 3"},{"comment":"The hyper-parameter optimization procedure (θ, C1,0, C1,2, φ1,2) is described only in one sentence ('we optimize ... against a self-consistent loss function given by the mean variance across the aligned BF images stack'). Additional details—initialization, bounds, number of iterations, and sensitivity to initial values—should be provided to allow others to reproduce the experimental reconstructions.","section":"Supplementary Table 1 and text following Figure 6"},{"comment":"The caption states 'Scalebars are 10 nm and 0.5 nm−1 respectively', but the figure shows multiple rows and columns with potentially different scalebars; please clarify which scalebar corresponds to which panel or indicate that all real-space scalebars are 10 nm and all reciprocal-space scalebars are 0.5 nm−1.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is engaging and the connection between parallax imaging and direct ptychography is well motivated. My main concern is that the empirical D_vBF criterion is the linchpin of the practical upsampling claim, and it is not supported by derivation or sufficient parametric study. The manuscript also contains a placeholder link in the Data Availability statement, which should be resolved before publication. The experimental demonstrations are promising but would benefit from quantitative resolution metrics. I believe the central idea is sound and the manuscript can be made acceptable with focused additional analysis and validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's real contribution is the conceptual link: parallax imaging is a truncated Taylor expansion of the direct ptychography kernel, and this factorization enables an FFT-based upsampling algorithm for direct ptychography. The CTF derivations are crisp and reproduce known results (aperture autocorrelation times -sin chi). The simulations validate that Algorithm 1 recovers the analytical CTF from 2-4x subsampled scans when defocus is sufficient. That is a genuinely useful advance for dose-sensitive 4D-STEM.\n\nThe experimental demonstrations on MOFs and other datasets are welcome, though they are qualitative. I'd have liked a resolution metric or a side-by-side with the Einstein-summation approach of Lalandec Robert et al., but the visual evidence of lattice recovery from 1/16th dose is suggestive.\n\nThe soft spots, in order of importance. First, the vBF diversity ratio D_vBF = λ|k_BF||Δf|/s ≳ 1 (Eq. 10) is presented as a rule of thumb, not derived. It's calibrated on three defocus values (50, 100, 200 Å) and Table 1 shows it correlates with the iterative overlap ratio, but the threshold is empirical. The stress-test worry that the limit might shift with upsampling factor, noise, or non-defocus aberrations is legitimate, but it's a range-of-validity issue, not a demonstrated failure. The algorithm works in the tested cases.\n\nSecond, the upsampling procedure tiles the sub-sampled FFT and applies a matched filter; there's no explicit alias-inversion step. The paper doesn't analyze conditioning or null-space of this inversion, so we don't know exactly when aliased copies are suppressed beyond the heuristic D_vBF criterion. Missing analysis, not a fatal flaw.\n\nThird, the data release is incomplete: the Zenodo link is a ToDo placeholder. The CTF notebooks are available, but the processed datasets are not. For a methods paper, that's a fixable but real gap.\n\nI disagree with the reader a bit on the stress-test: the concern is valid but proportionate. The central claim that upsampling works in the weak-phase regime with sufficient defocus holds up. The paper deserves a serious referee. I'd recommend accepting it for peer review with requests for (a) a derivation or more systematic calibration of the D_vBF rule, (b) an analysis of the alias-inversion step, and (c) complete data release.\n\nFor anyone working on 4D-STEM phase retrieval, it's worth reading. I'd bring it to reading group.","headline":"A clean theory linking parallax imaging to direct ptychography enables a working FFT-based upsampling scheme; the main soft spot is the empirically calibrated diversity threshold that sets the claimed resolution limits.","tokens_in":14130,"tokens_out":2327,"would_cite":true,"duration_ms":22655,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.30.Rx","68.37.Ma"],"model":"deepseek-v4-flash","headline":"The paper shows that direct ptychography, a non-iterative electron-microscopy phase-retrieval method, can be numerically upsampled by 2x–4x beyond its scan-step Nyquist limit using an insight borrowed from parallax imaging, as long as the…","keywords":["direct ptychography","parallax imaging","tilt-corrected bright-field STEM","scan Nyquist limit","contrast transfer function","upsampling","weak phase object approximation","low-dose electron microscopy"],"falsifier":"Acquire (or simulate) a 4D-STEM dataset with a pure higher-order aberration such as coma or spherical aberration at zero defocus, choose the scan step so that an effective $\\mathcal{D}_{\\mathrm{vBF}}$ computed from the aberration magnitude exceeds one, and run Algorithm 1 at 4x upsampling; if the recovered CTF does not match the analytical direct-ptychography CTF of the paper, then Equation 10 is not a general rule and the upsampling limits are specific to defocus-dominated acquisitions.","tokens_in":13060,"feed_emoji":"🔬","tokens_out":10342,"duration_ms":93150,"temperature":0.7,"pith_summary":"Direct ptychography recovers the phase of a thin electron-microscopy sample from 4D-STEM data without iteration, but its traditional resolution is capped by the scan step size through the Nyquist condition. This paper claims that cap can be lifted: it identifies parallax imaging as a truncated Taylor expansion of the direct ptychography aperture-overlap kernel $\\Gamma(\\mathbf{q},\\mathbf{k})$, and uses that identification to reformulate direct ptychography as a loop over bright-field detector frequencies. The resulting algorithm tiles and shifts Fourier-transformed virtual bright-field images, applies the full kernel, and thereby recovers the analytical direct-ptychography contrast transfer function from scans subsampled by 2x–4x, up to the numerical aperture, when the vBF diversity ratio $\\mathcal{D}_{\\mathrm{vBF}} = \\lambda |\\mathbf{k}_{\\mathrm{BF}}| |\\Delta f|/s$ is about one or larger. If true, dose-sensitive materials could be acquired with large, defocused scan steps and later computationally sharpened, without giving up the accuracy of full direct ptychography. The same analysis explains why parallax imaging is so robust at low dose and why direct ptychography should win at high dose or when higher-order aberrations dominate.","feed_headline":"Coarse scans get 4x sharper with a parallax ptychography trick","feed_subtitle":"A reformulation of direct ptychography's kernel lets defocused low-dose scans reach numerical-aperture resolution.","key_machinery":"The load-bearing object is the complex-valued aperture overlap function $\\Gamma(\\mathbf{q},\\mathbf{k}) = \\psi^{*}(\\mathbf{k})\\psi(\\mathbf{q}-\\mathbf{k}) - \\psi(\\mathbf{k})\\psi^{*}(\\mathbf{q}+\\mathbf{k})$, with probe wavefunction $\\psi(\\mathbf{k}) = A(\\mathbf{k})e^{-i\\chi(\\mathbf{k})}$; it describes interference between the probe and its shifted copies at each spatial frequency $\\mathbf{q}$. The mathematical bridge to parallax imaging is the Taylor expansion $\\chi(\\mathbf{q}\\pm\\mathbf{k}) \\approx \\chi(\\mathbf{q}) \\pm \\nabla\\chi(\\mathbf{q})^{T}\\cdot\\mathbf{k}$, which turns $\\Gamma$ into $\\mathcal{B}(\\mathbf{q},\\mathbf{k})e^{i\\nabla\\chi(\\mathbf{q})\\cdot\\mathbf{k}}$; truncating after the linear term gives parallax, while keeping the full expansion gives direct ptychography. Algorithm 1 implements the full kernel in a loop over bright-field detector frequencies $\\mathbf{k}_{\\mathrm{BF}}$, tiling and shifting each vBF image Fourier transform and filtering by $\\Gamma/|\\Gamma|$ before accumulation.","core_discovery":"On the paper's own terms, the central discovery is that parallax imaging is not a separate phase-retrieval trick but the first term of a controlled approximation to the direct ptychography kernel. Factoring the phase ramp $e^{i\\nabla\\chi(\\mathbf{q})\\cdot\\mathbf{k}}$ out of the aperture overlap function $\\Gamma(\\mathbf{q},\\mathbf{k}) = \\psi^*(\\mathbf{k})\\psi(\\mathbf{q}-\\mathbf{k}) - \\psi(\\mathbf{k})\\psi^*(\\mathbf{q}+\\mathbf{k})$ leaves an amplitude term whose phase contains $\\chi(\\mathbf{q})$, making the parallax CTF $-[\\mathcal{A} \\star \\mathcal{A}](\\mathbf{q})\\sin[\\chi(\\mathbf{q})]$; direct ptychography, by contrast, keeps the full kernel and implicitly performs the $\\mathrm{sgn}[\\sin[\\chi(\\mathbf{q})]]$ phase-flipping that removes parallax's zero crossings. Because the factorization remains accurate for quadratic aberrations, Algorithm 1 can upsample a coarsely scanned dataset by tiling the Fourier transform of each virtual bright-field image onto a finer grid, shifting it by the detector frequency, filtering with the full $\\Gamma$ kernel normalized by its modulus, and summing. The paper validates on simulated apoferritin and on experimental metal-organic frameworks, twisted MoS2, gold nanoparticles, and virus-like particles, including a 4x-thinned MOF dataset reconstructed from 1/16 of the original dose.","pith_inferences":["Because the paper calibrates Equation 10 only for defocus, a natural extension would be to test whether the same diversity ratio governs upsampling for astigmatism or coma when the aberration magnitude is expressed through an effective transverse phase gradient; if it does, aberration tuning could replace defocus as the diversity source for near-focus acquisitions.","The tiled-shift picture treats each bright-field detector pixel as an independent aperture view, which suggests a synthetic-aperture interpretation; one could derive a Cramér–Rao bound on the reconstructed phase versus fluence and upsampling factor, quantifying when the zero-crossing recovery of direct ptychography actually becomes visible above shot noise.","The empirical rule $\\mathcal{D}_{\\mathrm{vBF}} \\gtrsim 1$ likely follows from a support argument on the tiled Fourier grid—the shifted apertures must overlap in the upsampled $\\mathbf{q}$-space—so deriving Equation 10 from first principles would make it portable to arbitrary aberrations and detector geometries.","If the low-dose convergence of parallax and direct ptychography holds for real detectors, then for frozen-hydrated biological specimens the extra computation of direct ptychography buys little; a testable crossover dose could be identified above which zero-crossing recovery matters."],"forward_implications":["For weak-phase samples with $\\mathcal{D}_{\\mathrm{vBF}} \\gtrsim 1$, direct ptychography can be performed on 2x–4x subsampled, defocused scans, meaning the resolution no longer has to be tied to the scan step size.","The practical resolution ceiling for upsampled direct ptychography becomes the numerical aperture of the probe, the same ceiling parallax imaging already achieved.","At low electron fluence, direct ptychography and parallax imaging reconstructions converge, so for dose-sensitive biological specimens the substantially simpler parallax algorithm is sufficient.","At high fluence or with higher-order aberrations, direct ptychography outperforms parallax by recovering the contrast-transfer zero crossings that parallax imaging misses.","The same tiling-and-shifting upsampling recipe transfers directly to related linear phase-retrieval variants such as OBF STEM and SSB-ICOM."],"supporting_citations":[{"why":"Supplies the transfer-of-information formalism and the complex CTF integral (Eq. 2) on which both the direct-ptychography and parallax analyses are built.","marker":"Hammel and Rose (1995)"},{"why":"Establishes Wigner-distribution deconvolution as a direct ptychographic technique, the foundation the paper extends.","marker":"Rodenburg and Bates (1992)"},{"why":"Shows analytical sample-phase recovery for weak phase objects in direct ptychography, justifying the linear CTF framework.","marker":"Nellist et al. (1995)"},{"why":"Demonstrates single-sideband direct ptychography and is the reference for the scan-step Nyquist limitation that the paper relaxes.","marker":"Pennycook et al. (2015)"},{"why":"Defines the aperture overlap function and the full-kernel CTF; Algorithm 1 filters with this kernel.","marker":"Yang et al. (2016a)"},{"why":"Provides the comparative transfer-of-information analysis, the parallax CTF expression, and the zero-crossing behavior.","marker":"Bekkevold et al. (2025)"},{"why":"Discusses where information resides in 4D-STEM and the implicit link between parallax imaging and direct ptychography.","marker":"Ma et al. (2025)"},{"why":"Provides iterative phase-retrieval implementations and the experimental gold-nanoparticle and virus-like-particle datasets used for validation.","marker":"Varnavides et al. (2023)"},{"why":"Supplies the dose-efficient tilt-corrected BF-STEM demonstration and the empirical low-dose behavior the paper explains.","marker":"Yu et al. (2024)"},{"why":"Proposes an alternative algorithm for relaxing direct-ptychography scan sampling via Einstein summations, which the tiling-FFT approach is set against.","marker":"Lalandec Robert et al. (2025)"}],"fun_headline_variants":["Parallax insight upsamples direct ptychography scans","Direct ptychography gets 4x upsample via parallax","Parallax approximation relaxes ptychography sampling","4x scan thinning possible in ptychography via parallax","Coarse ptychography scans sharpened with parallax math"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the empirically calibrated rule $\\mathcal{D}_{\\mathrm{vBF}} = \\lambda|\\mathbf{k}_{\\mathrm{BF}}||\\Delta f|/s \\gtrsim 1$: the paper derives the algorithm but not this threshold, calibrating it on just three defocus values (50, 100, and 200 Å), so if the threshold does not generalize to other aberration types, detector geometries, or noise levels, the predicted upsampling factors would not hold even if the algorithm's mechanics are sound.","fun_headline_variants_meta":{"raw":{"variants":["Parallax insight upsamples direct ptychography scans","Direct ptychography gets 4x upsample via parallax","Parallax approximation relaxes ptychography sampling","4x scan thinning possible in ptychography via parallax","Coarse ptychography scans sharpened with parallax math"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1951,"prompt_tokens":987,"completion_tokens":964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":880}},"tokens_in":603,"tokens_out":964,"duration_ms":8983,"temperature":1.0,"reasoning_tokens":880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:09:38.549855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire (or simulate) a 4D-STEM dataset with a pure higher-order aberration such as coma or spherical aberration at zero defocus, choose the scan step so that an effective $\\mathcal{D}_{\\mathrm{vBF}}$ computed from the aberration magnitude exceeds one, and run Algorithm 1 at 4x upsampling; if the recovered CTF does not match the analytical direct-ptychography CTF of the paper, then Equation 10 is not a general rule and the upsampling limits are specific to defocus-dominated acquisitions.","supporting_citations":[],"review_version":2}