{"id":"03739a28-f7ad-4f56-94e2-7f201fa13c92","arxiv_id":"2411.13098","paper_version":5,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper argues that aliased replica functions in undersampled digital holograms encode high-frequency information, enabling super-resolution after upsampling and learning-based denoising.","lead":"A single-author optics study claims that a low-resolution, undersampled hologram can be reconstructed into a high-resolution image by upsampling and a neural-network denoiser, a result the author says overcomes the fundamental space-bandwidth limit of digital holography.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (18) is load-bearing but not generally valid: pixel-duplication upsampling cancels odd-order spectral replicas only at isolated comb peaks, not across the finite spectral support of a non-constant object, so the claimed recovery beyond the sampling limit is not established.","rationale":"Read in good faith, the paper aims to show that aliasing in lensless holography is not pure information loss: the replica fringes are claimed to be phase-modulated high-frequency components, and upsampling plus denoising is claimed to unmix them. For that central claim to hold, two conditions are necessary: (i) the upsampling operation must actually separate or cancel the aliased spectral copies, and (ii) residual errors must be removable without relying on object-class-specific knowledge. Condition (i) fails at Eq. (18). The factor (1+e^(−iπfΔx)) in the duplicated-comb spectrum cancels the odd-order sampling deltas, but the convolution with G spreads each delta into a band; the factor's zeros do not kill those bands. Thus the ±1st-order diffraction fields are not generally suppressed, consistent with the residual artifacts the paper itself shows in Fig. 6. Because of this, the DnCNN step becomes load-bearing for the demonstration, but it is trained on rescaled MNIST digits, so the reported PSNR values establish only that the denoiser removes artifacts for digits, not that arbitrary wide-field objects are recovered. The information-theoretic degree-of-freedom argument reinforces this: for a complex object, the m-fold undersampled measurement has fewer complex degrees of freedom than the claimed high-resolution reconstruction, so some prior must supply the missing information. The paper deserves credit for careful numerical simulations, a qualitative optical experiment, and an explicit acknowledgment of residual artifacts; the issue is not effort but the gap between the required mathematical identity and the finite-bandwidth case. A more limited claim, such as super-resolution from undersampled holograms with strong object priors or sparse objects, may be defensible, but the stated conclusion in Section 7 overreaches. I therefore keep the reader's REJECT verdict, with the main supporting concern being the invalid general cancellation in Eq. (18) rather than only the infinite-aperture orthogonality derivation.","tokens_in":15428,"tokens_out":19598,"duration_ms":211195,"concrete_test":"Generate a random complex object whose angular spectrum is white up to the properly sampled Nyquist frequency m/(2Δx) for m=2. Synthesize the undersampled Fresnel hologram at pitch Δx and a reference properly sampled hologram at pitch Δx/2. Apply the paper's pixel-duplication upsampling and angular-spectrum reconstruction. Compute the normalized error in the spectral band [1/(2Δx), 1/Δx]; if the high-frequency content is not recovered to the same accuracy as the properly sampled reference, Eq. (18) is falsified. As a direct check, take a non-constant G, e.g., a single off-axis plane-wave component at 0.75/Δx, compute the 2N-point DFT of the duplicated sequence, and verify whether the odd-order replica bin is exactly zero; it will not be.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central mechanism is the claim in Section 3.2 that two-fold upsampling by pixel duplication removes the ±1st-order spectral replicas, leading to Eq. (18): FT[...] = (2/Δx)Σ_q G(f − 2q/Δx). The cancellation argument in Eq. (17) is exact for the Dirac comb alone, because the factor (1+e^(−iπfΔx)) vanishes at f=(2q+1)/Δx. But for a sampled hologram the spectrum is G(f) convolved with the comb, not the comb itself. The odd-order replica bands are Σ_q G(f−(2q+1)/Δx) multiplied by the same frequency-dependent factor; that factor is zero only at the discrete replica centers, not on the spectral support of G. When the object's angular spectrum extends toward the full-aperture bandwidth (exactly the m-fold undersampled configurations in Sections 5.1 and 5.2), the q=±1 replica bands remain partially inside the passband. The paper's own Fig. 6 confirms residual high-frequency artifacts, and the subsequent DnCNN stage is trained on rescaled MNIST, so the reported recovery is not evidence that arbitrary objects can be reconstructed. More fundamentally, N complex samples cannot uniquely determine an mN-pixel complex image without a prior; the low-resolution hologram does not contain the same information as a properly sampled one for general objects.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that an undersampled digital hologram can be used to reconstruct an image with the same spatial resolution and angular field of view as a properly sampled hologram, thereby overcoming the finite space-bandwidth limitation of digital holography. The proposed method has three steps: two-fold (or m-fold) upsampling by pixel duplication, reconstruction with the angular-spectrum method, and denoising with a DnCNN network trained on synthetically generated noisy reconstructions. The central theoretical argument is that high-order diffraction fields are mutually orthogonal and that pixel-duplication upsampling suppresses the ±1st-order spectral replicas, leaving only replicas at multiples of 2/Δx.","tokens_in":15749,"tokens_out":3108,"duration_ms":32368,"significance":"If the central claim were correct, it would overturn the conventional pixel-Nyquist limit for lensless holography and enable wide-field high-resolution imaging without optical magnification, which would be highly significant. The paper contains a useful qualitative description of aliased Fresnel replicas as phase-modulated carrier waves in the complex plane, and it includes both numerical simulations and an optically captured USAF-target hologram. However, the load-bearing mathematical step in Section 3.2 is not generally valid, and the reported recovery results are obtained under conditions (in-distribution MNIST training and evaluation) that do not support the claimed generality. As a result, the central claim is not established by the evidence presented.","major_comments":[{"comment":"The derivation of Eq. (18) is not valid for a finite-bandwidth object. Equation (17) shows cancellation of odd-order peaks for the Dirac comb alone, but for a sampled hologram the spectrum is G(f) convolved with the comb. The odd-order replica terms are multiplied by the factor (1 + e^{-iπ f Δx}), which vanishes only at the isolated frequencies f = (2q+1)/Δx, not across the spectral support of G. Consequently, when the object's angular spectrum extends toward the full-aperture bandwidth, the q = ±1 replica bands remain partially inside the passband, and Eq. (18) does not follow. The paper's own Figure 6 confirms residual high-frequency artifacts, yet the conclusion in Section 7 asserts full space-bandwidth recovery. This flaw undermines the central claim.","section":"Sec. 3.2, Eqs. (16)–(18)"},{"comment":"The orthogonality of high-order diffraction fields is derived using integrals over an infinite plane, but the hologram and the simulations in the paper are finite. Over a finite aperture, the modes are not exactly orthogonal, and high-order diffraction terms do interfere with the primary image. The residual artifacts visible in Figure 6 and the incomplete suppression described in Section 6 are consistent with this limitation. The claim that high-order terms can be removed without corrupting the primary image is therefore not supported by the derivation as presented.","section":"Sec. 3.1, Eq. (15)"},{"comment":"The learning-based demonstration does not establish that arbitrary objects can be reconstructed from low-resolution holograms. The DnCNN network is trained on rescaled MNIST images and evaluated on MNIST test images; the reported PSNR values of approximately 38, 42, and 41 dB are in-distribution results. These numbers show that the network can denoise the specific residual pattern seen in MNIST-like reconstructions, but they do not demonstrate recovery of general object information. Moreover, N complex samples cannot uniquely determine an mN-pixel complex image without a strong prior, so the paper's information-theoretic claim that a low-resolution hologram contains the same information as a properly sampled one is not justified.","section":"Sec. 5.2 and Supplement 1"},{"comment":"The optically captured USAF-target experiment is presented without a quantitative resolution comparison. The paper states that the upsampled reconstruction improves the vertical direction but not the horizontal direction, and no resolution metric, line-pair analysis, or comparison against a properly sampled reference is provided. This result is therefore insufficient to support the broad claim that the method overcomes the space-bandwidth limitation in practice.","section":"Sec. 5.1, Fig. 11"}],"minor_comments":[{"comment":"There are several typographical errors: 'reconsturction' in the Figure 10 caption, 'increse' and 'Similarily' in Section 2.1, 'normlization' in the Figure 12 caption, 'Acqusition' in the Figure 12 caption, and 'apper' in the Introduction. These should be corrected.","section":"Throughout"},{"comment":"The text refers to a 'commercial spectrum calculation program' without identifying the software or algorithm; specifying the method would improve reproducibility.","section":"Sec. 2.1"},{"comment":"The derivation of the sampling condition uses the parameters T and S without a clear definition of their allowed ranges; a sentence defining T in (0,1) and S as a positive real number would help the reader.","section":"Supplement 1, Eq. (6)"},{"comment":"The horizontal-direction degradation is attributed to preprocessing loss of Fourier components, but the paper does not quantify this loss; a spectral plot of the cropped region would clarify the extent to which the experimental demonstration is limited by the off-axis geometry.","section":"Sec. 5.1, Fig. 11"}],"recommendation":"reject","confidential_remarks":"The paper is built on a central mathematical claim that appears to be incorrect as stated, and the remaining demonstration relies on a learned denoiser trained and evaluated on the same data distribution. Given the scope of the claimed result, the load-bearing errors are not local and would require a substantially different argument to establish the conclusion. I therefore recommend rejection, notwithstanding the interesting qualitative observations about aliased Fresnel replicas."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper makes a strong claim: an undersampled hologram can reconstruct an image with the same resolution and field of view as a properly sampled one, beating the pixel-Nyquist limit. That claim, as stated, is not supported. The load-bearing step is Eq. (18), where pixel-duplication upsampling supposedly removes the odd-order spectral replicas so that the Fourier space is simply doubled. The stress-test note gets this right: the cancellation factor (1+e^{-iπfΔx}) vanishes only at isolated replica centers, not across the finite spectral support of a real object. When the angular spectrum extends toward the full aperture bandwidth—exactly the m-fold undersampled cases the paper highlights—the ±1st-order replica bands leak into the passband. The paper's own Fig. 6 admits residual artifacts. So the central mechanism does not establish information recovery beyond the sampling limit for general objects. The orthogonality argument in Section 3.1 also assumes infinite-plane integration, which sidesteps the finite-aperture reality the paper elsewhere emphasizes.\n\nThat said, the paper is not empty. The angular-spectrum interpretation of aliased replica fringes as phase-modulated carriers is genuinely instructive, and the numerical observation that two-fold upsampling suppresses the dominant high-order diffraction peaks is a real effect, even if incomplete. The optical experiment with a USAF target is a serious attempt, and the improvement in the vertical direction is consistent with the upsampling idea, though it is qualitative and limited. The MNIST denoising demonstration is the weakest link: DnCNN trained on rescaled MNIST and then tested on rescaled MNIST means the network is effectively filling in the missing high-frequency content from the training distribution. That is circular as evidence for the general claim, and the paper concedes residual noise in Section 6.\n\nNovelty is also thin. The author's own prior work already reports high-resolution reconstruction from low-resolution holograms and viewing-angle expansion via upsampling. This paper extends that line with a new theoretical framing and a denoiser, but the incremental step does not justify the sweeping conclusion. On the other hand, the treatment of the replica functions is the kind of thing that could be useful if properly bounded.\n\nFor a reader, treat this as a provocative preprint that overclaims. The topic matters—lensless wide-field high-resolution imaging is a real bottleneck—and a serious referee should be let loose on it, because the core derivation deserves a rigorous check and the paper may contain a salvageable weaker claim about super-resolution with strong priors. I would not cite it in its current form, but I would bring it to a reading group for a lively discussion.\n\nRecommendation: send to peer review, with the expectation that the central claim will need major qualification or substantial new evidence.","headline":"A provocative but undersupported claim about beating the space-bandwidth limit in digital holography; the core math does not hold for general objects, yet the paper has enough real content to merit referee scrutiny.","tokens_in":16232,"tokens_out":1221,"would_cite":false,"duration_ms":14204,"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":"The paper claims that an undersampled hologram captured with a coarse pixel pitch can reconstruct an image at the same spatial resolution and angular field of view as a properly sampled hologram, because aliased replica fringes are…","keywords":["digital holography","space-bandwidth product","undersampled hologram","aliasing","angular spectrum","Fresnel diffraction","image reconstruction","learning-based denoising"],"falsifier":"Take a simulated undersampled point-object hologram at half $z_c$, apply two-fold pixel-duplication upsampling with no denoiser, and measure the reconstructed point-spread function: the central image must show the full-aperture resolution (4 $\\mu$m with 8-$\\mu$m pixels, per the paper's geometry) before any learning-based cleanup. If residual high-order artifacts still overlap the primary image, as the paper's own Fig. 6 indicates, the orthogonality-based cancellation is incomplete.","tokens_in":1766,"feed_emoji":"🔬","tokens_out":4093,"duration_ms":95424,"temperature":0.7,"pith_summary":"This paper tries to prove that a lensless digital hologram captured with pixels too coarse to satisfy the Nyquist condition can still reconstruct an image as sharp and as wide-angle as a properly sampled hologram. The key move is to read aliased replica fringes not as noise but as phase-modulated carriers that encode the hologram's high spatial frequencies. The paper then suppresses the unwanted replica images by upsampling through pixel duplication and cleans the remainder with a learned denoiser. If the claim holds, wide-field high-resolution microscopy would not need magnifying optics or nanoscale sensors.","feed_headline":"Aliased fringes hide a full-resolution holographic image","feed_subtitle":"Pixel-duplication upsampling plus a denoiser recovers detail finer than the sensor's sampling pitch.","key_machinery":"The central object is the angle-modulated replica function: an aliased copy of the Fresnel hologram written as $\\exp[i2\\pi(n x/\\Delta x + x^2/2\\lambda z)]$, which is a carrier wave at frequency $n/\\Delta x$ phase-modulated by a quadratic signal. This representation turns aliased fringes into a continuous high-frequency spectrum. The companion mechanism is the orthogonality relation $\\langle\\varphi_m|\\varphi_n^*\\rangle = c_o^2\\delta_{mn}$ for the high-order diffraction modes, the spatial image states associated with individual replica zones; that orthogonality justifies canceling the replica images with pixel-duplication upsampling.","core_discovery":"The paper claims that aliasing in an undersampled Fresnel hologram is not information loss. Each shifted replica is a carrier-wave phase modulation of the original quadratic-phase field, so the spatial frequency continues to rise across replica zones up to the full aperture bandwidth $N\\Delta/\\lambda z$. Because the resulting high-order diffraction modes are orthogonal, a two-fold upsampling by pixel duplication cancels the $\\pm 1$st-order replica images, and a learning-based denoiser removes the residual high-frequency artifacts. The reconstructed image then has the same spatial resolution and angular field of view as a properly sampled hologram, obtained from a hologram captured at a coarser pixel pitch. This is demonstrated in numerical simulations and in an optically captured hologram of a resolution target.","pith_inferences":["Editorial inference: if the orthogonality of high-order diffraction modes is only approximate on a finite hologram aperture, the upsampling cancellation will leave residual replica noise that the learned denoiser must cover; this can be tested by measuring artifacts with upsampling alone and no denoiser.","Editorial inference: the denoiser is trained on MNIST digits, so the claimed wide-field high-resolution pipeline should be validated on biological or industrial samples outside that training distribution before the generalization is trusted.","Editorial inference: the optical USAF demonstration showed little horizontal resolution gain because the off-axis preprocessing discarded horizontal Fourier components; an on-axis or common-path geometry might restore isotropic resolution gains.","Editorial inference: the angle-modulation reading of aliased replicas could apply to any coarsely sampled quadratic-phase signal, including synthetic aperture radar or acoustic imaging, wherever Fresnel-like phase curvature is present."],"forward_implications":["Lensless holographic microscopes could resolve features at the full aperture numerical aperture of the sensor rather than at the pixel-pitch numerical aperture, without magnification.","The space-bandwidth product of the captured hologram, not the pixel size, would become the practical resolution limit for lensless holography.","Short-distance capture with a large image sensor becomes an explicit route to wide-field high-resolution imaging, provided the reference-object geometry keeps the complex hologram measurable.","The same undersampling analysis extends to incoherent holography and to the Rayleigh-Sommerfeld diffraction regime, so the mechanism is not confined to ideal monochromatic Fresnel fields.","Because the method claims recovery of actual high-frequency signal content rather than interpolation, it differs from conventional super-resolution and should be testable by direct point-spread-function measurement."],"supporting_citations":[{"why":"Provides the Fresnel diffraction formula and angular-spectrum method used throughout the derivations and reconstructions.","marker":"[1]"},{"why":"Supplies the prior demonstration that high-resolution images can be reconstructed from low-resolution lensless holograms and frames the imaging-performance question.","marker":"[12]"},{"why":"Defines the space-bandwidth product $B_w = N\\Delta/\\lambda z$ that the paper claims to overcome.","marker":"[13]"},{"why":"Establishes the numerical-aperture dependence of the angular field of view and spatial resolution in holographic imaging that the paper argues can be surpassed.","marker":"[17]"},{"why":"Provides the angle-modulation formalism of carrier frequency and instantaneous frequency used to identify aliased replicas as phase-modulated high-frequency components.","marker":"[24]"},{"why":"Derives the sampled quadratic-phase function representation used in Eq. (3) for the undersampled hologram field.","marker":"[25]"},{"why":"Reports optical experiments showing that replica fringes in undersampled holograms correspond to higher spatial-frequency components.","marker":"[26]"},{"why":"Establishes that high-order diffraction terms can be removed as noise using the orthogonality property, the basis for the upsampling suppression step.","marker":"[27]"},{"why":"Supplies the convolutional denoising network architecture used to remove residual high-frequency artifacts after upsampling.","marker":"[31]"}],"fun_headline_variants":["Undersampled holograms upsampled to dodge the space-bandwidth limit","Pixel-duplication trick beats the hologram sampling limit","Aliasing in holograms isn't loss: upsampling recovers full resolution","Low-res hologram in, high-res image out via upsampling and denoising"],"cache_read_input_tokens":18304,"weakest_assumption_plain":"The method assumes that the high-order diffraction modes stay exactly non-interfering on a real, finite hologram, so that upsampling by pixel duplication cancels them completely; if that exactness fails, leftover replica noise remains and the denoiser must cover the gap.","fun_headline_variants_meta":{"raw":{"variants":["Undersampled holograms upsampled to dodge the space-bandwidth limit","Pixel-duplication trick beats the hologram sampling limit","Aliasing in holograms isn't loss: upsampling recovers full resolution","Low-res hologram in, high-res image out via upsampling and denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3074,"prompt_tokens":900,"completion_tokens":2174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":2090}},"tokens_in":516,"tokens_out":2174,"duration_ms":17265,"temperature":1.0,"reasoning_tokens":2090,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:51:09.937633+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a simulated undersampled point-object hologram at half $z_c$, apply two-fold pixel-duplication upsampling with no denoiser, and measure the reconstructed point-spread function: the central image must show the full-aperture resolution (4 $\\mu$m with 8-$\\mu$m pixels, per the paper's geometry) before any learning-based cleanup. If residual high-order artifacts still overlap the primary image, as the paper's own Fig. 6 indicates, the orthogonality-based cancellation is incomplete.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Fresnel diffraction formula and angular-spectrum method used throughout the derivations and reconstructions."},{"cited_title":"Spatial resolution enhancement in holographic imaging via angular spectrum expansion,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior demonstration that high-resolution images can be reconstructed from low-resolution lensless holograms and frames the imaging-performance question."},{"cited_title":"Space-bandwidth product of optical signals and systems,","cited_arxiv_id":null,"evidence_quote":"Defines the space-bandwidth product $B_w = N\\Delta/\\lambda z$ that the paper claims to overcome."},{"cited_title":"Inverted Gabor holography principle for tailoring arbitrary shaped three-dimensional beams,","cited_arxiv_id":null,"evidence_quote":"Establishes the numerical-aperture dependence of the angular field of view and spatial resolution in holographic imaging that the paper argues can be surpassed."},{"cited_title":"Analysis on image recovery for on-axis digital Fresnel hologram with aliased fringe generated from self-similarity of point spread function,","cited_arxiv_id":null,"evidence_quote":"Provides the angle-modulation formalism of carrier frequency and instantaneous frequency used to identify aliased replicas as phase-modulated high-frequency components."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the sampled quadratic-phase function representation used in Eq. (3) for the undersampled hologram field."},{"cited_title":"Some mathematical properties of the uniformly sampled quadratic phase function and associated issues in Fresnel diffraction simulations,","cited_arxiv_id":null,"evidence_quote":"Reports optical experiments showing that replica fringes in undersampled holograms correspond to higher spatial-frequency components."},{"cited_title":"Wide viewing-angle holographic display based on enhanced-NA Fresnel hologram,","cited_arxiv_id":null,"evidence_quote":"Establishes that high-order diffraction terms can be removed as noise using the orthogonality property, the basis for the upsampling suppression step."},{"cited_title":"Deep learning on image denoising: An overview,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional denoising network architecture used to remove residual high-frequency artifacts after upsampling."}],"review_version":1}