{"id":"b9a0f625-071f-441b-b2ec-ff272978af4f","arxiv_id":"2506.14716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A cold-atom EIT memory performs edge-detecting convolution on stored images using spiral phase contrast, with readable output extending beyond 320 microseconds.","lead":"This paper shows pictures being stored and processed inside a cloud of cold rubidium atoms, using a vortex-shaped laser to highlight the edges of the stored image. It demonstrates an all-optical processing-in-memory approach that could one day reduce the energy and delay of moving data between memory and processors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central convolution claim relies on treating the cold ensemble as a thin 2D Fourier-plane filter, but the only stated justification is a dimensionally mismatched comparison of the 2 mm cloud diameter to the 100 mm Rayleigh length; a 3D/PSF check is needed to see if Eq. (1) is actually realized.","rationale":"The reader's identified weakest assumption is the same one that I find most load-bearing: the entire convolution claim passes through the 2D Fourier-plane model. I agree that the paper plausibly demonstrates the effect; the images show the expected edge enhancement and the secondary-convolution cancellation in Fig. 3 is a nice internal consistency check. The issue is not an internal contradiction but an unsupported quantitative step: the thin-medium condition is asserted with an irrelevant metric and the derivation/simulation parameters are deferred to a missing supplement. The 320 μs record claim is also stated without benchmarking, but I regard it as secondary: even if the record is not the longest, the convolution-in-memory demonstration would still stand, whereas a breakdown of the Fourier-plane assumption would undermine the central physics. Therefore the reader's CONDITIONAL verdict is the correct one and no adjustment is needed.","tokens_in":8815,"tokens_out":14019,"duration_ms":153839,"concrete_test":"Place a small pinhole (or a tightly focused Gaussian reference) at the object plane of the 4f system and measure the point-spread function of the full store-and-readout channel using the same HyGG write/read controls; compare the measured transverse amplitude and phase profile with F^{-1}[g^2(r) exp(iΔlφ)] predicted by Eq. (1). If the PSF deviates beyond the diffraction limit of the 4f system (e.g., in the radial null structure or azimuthal phase winding), the finite longitudinal extent of the ensemble is corrupting the assumed Fourier-plane convolution kernel.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing step is the identification of the cold ensemble as an ideal thin 2D Fourier-plane filter. This is required for the spin wave to record E_S_in(r) U_W_l(r) and for the readout to realize the convolution kernel g^2(r) exp(iΔlφ) in Eq. (1). Section II justifies the 2D Fourier-plane model by stating: 'Given that the spatial dimensions of the atomic medium (~2 mm in diameter) are significantly smaller than the Rayleigh length of the signal field (~100 mm), we limit the numerical analysis to the 2D Fourier plane for simplicity [XXX].' The comparison is dimensionally mismatched: a transverse diameter says nothing about the longitudinal thickness L of the cloud, which is the quantity that must be small compared with the Rayleigh range for the transverse profiles of signal and control to be z-independent. In a real EIT memory, the spin wave fills a finite cloud, the signal and both control envelopes carry Gouy phases, and the retrieved field is an integral over z of their product. If L is not negligible, the effective transfer function is a z-averaged version of g^2(r) exp(iΔlφ), not the pure multiplicative kernel, and the comparison between measured and simulated images in Figs. 2 and 3 loses its quantitative force. The missing supplement and [XXX] placeholders for simulation parameters mean this assumption cannot currently be checked from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an experiment on all-optical convolution performed during electromagnetically induced transparency (EIT) based storage and retrieval in a cold 85Rb ensemble. The authors place the atomic ensemble at the Fourier plane of a 4f optical system, imprint spiral phases from control fields onto spin waves during both write-in and read-out stages, and demonstrate spiral phase contrast (SPC) edge enhancement of retrieved images. They claim on-demand readout of processed images for over 320 microseconds, stated to be the longest duration reported for an atomic-based memory to date. The experimental images show SPC edge enhancement and a cancellation effect when identical vortices are used in the two stages, qualitatively consistent with simulations; however, the theoretical derivation behind Eq. (1) and the simulation parameters are deferred to a supplementary file and to placeholder references, so the quantitative support for the central claim is currently incomplete.","tokens_in":9108,"tokens_out":3949,"duration_ms":42177,"significance":"If the claims hold, the paper demonstrates a useful integration of storage and computation in a single atomic medium, an appealing realization of optical processing-in-memory. The observed edge enhancement, the nullification of the write-in convolution by an identical read-out vortex, and the high-order SPC results are concrete experimental advances, and the 320-microsecond memory lifetime is notable. The reliance on magnetically insensitive states and a guiding field builds on the group's prior work and is credible. However, the central ideal thin-lens Fourier-plane assumption is not yet rigorously established, and the missing simulation parameters and supplementary derivation prevent the reader from verifying the key quantitative claims. The significance is therefore conditional on completing the theoretical and simulation support.","major_comments":[{"comment":"The justification for treating the cold atomic ensemble as an ideal thin 2D Fourier-plane filter compares a transverse dimension (~2 mm diameter) with the longitudinal Rayleigh length (~100 mm), but the relevant condition is that the longitudinal thickness L of the atomic cloud be much smaller than the Rayleigh range. The manuscript does not report L or show that the transverse profiles of signal and control are z-independent over the cloud; if L is not negligible, the recorded spin wave is a z-averaged product that includes Gouy-phase variations, and the kernel g^2(r) exp(iΔlϕ) in Eq. (1) is not realized exactly. The authors should provide a quantitative criterion with a measured L, or perform a 3D propagation simulation to show that the 2D Fourier-plane model is adequate.","section":"Section II, paragraph beginning 'Given that the spatial dimensions...'"},{"comment":"The statement that the read-out control pulse imprints a spiral phase 'with chirality opposite to that observed during the write-in stage [XXX]' is load-bearing: it determines whether the read-out convolution cancels or adds to the write-in phase and directly defines Δl = |l - l'| in Eq. (1). The manuscript supports this statement only with a placeholder reference rather than a derivation or an explicit experimental calibration. A rigorous derivation of the sign relationship, or at minimum a clearly described measurement that fixes the relative chirality, is required for the central claim to be verifiable.","section":"Section II, chirality statement and Eq. (1)"},{"comment":"The claim that the measured images 'align well with the theoretical simulations' cannot be assessed because the simulations are said to be 'based on our experimental parameters [XXX]' and no parameters are given in the manuscript. The authors should provide the complete parameter list used in the numerical analysis, as well as a defined quantitative metric for image agreement (e.g., normalized cross-correlation or a visibility comparison with error bars), so that the match between experiment and simulation can be independently evaluated.","section":"Section III, Figs. 2 and 3"},{"comment":"Equation (1) mixes the notation E(-r), E, and a convolution with F^{-1}(g^2(r) exp(iΔlϕ)) without defining the field arguments, the convolution variables, or the Fourier-transform convention. As written, it is not a self-contained statement of the convolution theorem for the 4f system. Please rewrite the equation with explicit definitions of all fields, coordinates, and transform conventions, and derive it from the spin-wave recording and the read-out phase-imprinting process.","section":"Section II, Eq. (1)"}],"minor_comments":[{"comment":"The heading 'ACKONWLEDGEMENTS' is misspelled and should be 'ACKNOWLEDGEMENTS'.","section":"Section V heading"},{"comment":"The sentence 'The signal and control light are combined using a PBS and and interact collinearly' contains a duplicated 'and'.","section":"Section II, experimental setup paragraph"},{"comment":"Several '[XXX]' placeholders remain in the text, including the justification of the 2D Fourier-plane approximation, the chirality statement, and the simulation parameters; these must be replaced with actual citations and numerical values before the paper can be considered complete.","section":"Sections II and III"},{"comment":"The figure caption appears garbled in places, including a label '1#4' and inconsistent units ('400µs' vs '400 μs'); the figures and captions should be carefully proofread.","section":"Figure 1"},{"comment":"The definition 'V= (I max-I min)/(Imax +I min)' uses inconsistent subscript formatting; please define I_max and I_min consistently and specify how they are extracted from the images.","section":"Section III, image visibility definition"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an incomplete draft: multiple placeholder citations, a missing supplement, and garbled figure elements suggest it was submitted before the final editing pass. The core experimental idea is promising and the qualitative images are suggestive, but the load-bearing theoretical assumption is under-supported. If the authors can supply the missing derivation, simulation parameters, and a proper justification of the thin 2D Fourier-plane approximation, the paper could become a solid experimental contribution. I would encourage the editor to require the supplementary material to be included in the review copy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine experimental demonstration of in-memory image processing in a cold-atom EIT memory, and the core observation is likely correct. But the manuscript as posted is not complete enough to evaluate quantitatively: the derivation behind Eq. (1), the simulation parameters, and the [XXX] references are in a missing supplement, central figures have no error bars, and the 320 us 'record' claim is unbenchmarked.\n\nWhat's new: the combination of placing the memory at the Fourier plane of a 4f system and using a programmable vortex control beam to imprint spiral phase contrast on images, both at write-in and read-out. The write-in experiment (Fig. 2) shows recognizable 'Ψ' images with edge enhancement at storage times up to ~300 us, with visibility actually higher for the vortex case than the Gaussian case. The read-out sequence (Fig. 3a) shows that applying the same vortex at read-out cancels the edge enhancement, while using different charges gives higher-order SPC (delta-l = 2,3). That internal consistency is a strong point, and the efficiency drop with |l| in Fig. 3c matches the expected effect of the hollow control mode. I find the central claim plausible.\n\nThe soft spot is theoretical. Eq. (1) asserts that the retrieved field is the input convolved with F^-1[g^2(r) exp(i delta-l phi)], but the only justification in the main text is a sentence that compares the 2 mm cloud diameter to the 100 mm Rayleigh length. That comparison is dimensionally wrong for the purpose: the relevant condition is that the longitudinal thickness of the cloud be small compared to the Rayleigh range, so that the signal and control envelopes don't change significantly along z. If the cloud is several Rayleigh ranges long, the spin wave records a z-integrated product of fields with Gouy phases, and the effective kernel is not the clean multiplicative one of Eq. (1). The stress-test note is correct to flag this. That said, I don't think it sinks the paper. The observed edge enhancement and cancellation are robust features of SPC, and the qualitative agreement with simulation in Figs. 2-3 suggests the 2D Fourier-plane model captures the essential physics. But the authors need to either provide a proper justification (e.g., a PSF calculation or a statement of the cloud length) or temper the claim.\n\nAlso, the missing supplement and placeholders are not minor: the derivation is explicitly deferred there, so the main text's central equation is currently unsupported. And error bars on the visibility and efficiency curves would help; right now we only see point estimates.\n\nBottom line: this is a solid proof-of-concept that deserves a serious referee, not a desk-reject. The requested support is addressable. My recommendation: send it to peer review with a strong request that the supplement, parameter values, and a discussion of the thin-medium approximation be included before publication.","headline":"A plausible cold-atom EIT memory that does SPC edge enhancement at the Fourier plane, but the posted manuscript hides the derivation and simulation parameters in a missing supplement.","tokens_in":9641,"tokens_out":2912,"would_cite":true,"duration_ms":29551,"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 demonstrates all-optical convolution inside a cold-atomic-ensemble memory, with spiral-phase edge enhancement stored and read out on demand after more than 320 microseconds.","keywords":["processing in memory","all-optical convolution","cold atomic ensemble","electromagnetically induced transparency","spiral phase contrast","edge enhancement","optical image storage","transverse spatial modes"],"falsifier":"Place a pinhole or a small object at the mask plane and directly measure the retrieved point-spread function at several lateral positions: if it is not the predicted $\\mathcal{F}^{-1}\\{g^2(r)e^{i\\Delta l \\phi}\\}$ — for instance, if the vortex center shifts, the doughnut size changes with input position, or the pattern depends on axial placement of the cloud — then the two-dimensional Fourier-plane filter model is wrong and the agreement with simulation would not generalize.","tokens_in":8608,"feed_emoji":"🌀","tokens_out":10563,"duration_ms":97373,"temperature":0.7,"pith_summary":"The paper sets out to show that a cold rubidium atomic ensemble can serve as computational storage: instead of merely preserving an image, it performs an all-optical convolution on that image during storage and releases the processed result later. The convolution is spiral phase contrast (SPC), a phase-only edge-enhancement filter, implemented by imprinting vortex phases carried by the control light onto the atomic spin wave through electromagnetically induced transparency (EIT). Experimental images of the letter 'Ψ' match simulations at storage times beyond 300 µs, and the authors claim on-demand readout of processed images for over 320 µs, the longest duration reported for an atomic-based memory to date. If the claim is right, placing the atomic ensemble at the Fourier plane of a 4f system merges computation and memory into a single all-optical device, a step toward optical processing-in-memory.","feed_headline":"Images are edge-enhanced inside an atomic memory and stored for 320 µs","feed_subtitle":"The rubidium cloud that stores the image also performs the convolution, releasing edge-enhanced results on demand.","key_machinery":"The load-bearing object is the dark-state polariton (DSP), the quasiparticle formed by a resonant two-photon Raman process that adiabatically converts a combined signal-control field into a purely atomic collective excitation and back. Its transverse profile carries the spatial information: because the control field is a hypergeometric-Gaussian vortex $g(r)e^{-il\\phi}$ rather than a plane wave, the DSP stores not the image itself but the image's Fourier spectrum multiplied by the vortex phase. The second lens of the 4f imaging system performs the inverse Fourier transform, which turns that product into the convolution of the original image with a vortex-shaped point-spread function. The experiment's conceptual move is to place the filter inside the memory, so that computation, storage, and readout happen in one coherent optical path.","core_discovery":"The paper's central discovery is that the EIT storage-and-retrieval sequence in a cold atomic ensemble realizes a programmable phase-only convolution directly in memory. During write-in, the stored spin wave acquires the product $E_S^{\\mathrm{in}}(\\mathbf{r})\\,U_l^W(\\mathbf{r})$ of the signal's Fourier spectrum and the control field's hypergeometric-Gaussian profile $U_l^W(\\mathbf{r})=g(r)e^{-il\\phi}$; during read-out, a second control field imprints the conjugate vortex, so the retrieved field is proportional to $E * \\mathcal{F}^{-1}\\{g^2(r)e^{i\\Delta l \\phi}\\}$, with $\\Delta l=|l-l'|$. The authors verify this convolution theorem experimentally for the edge-enhancement case $\\Delta l=1$, for higher-order SPC with $\\Delta l=2,3$, and for the reversal case $l=l'$ where the second vortex cancels the first. They also show that the processed image remains visible for storage times beyond 300 µs, and on-demand retrieval survives beyond 320 µs, which they attribute to mapping the memory onto magnetically insensitive hyperfine states in a guiding magnetic field.","pith_inferences":["The paper itself limits the model to a 2D Fourier plane (Sec. II), noting that the atomic medium is about 2 mm in diameter versus a Rayleigh length of about 100 mm; this is the main assumption that would need revisiting for larger images, thicker clouds, or higher vortex charges.","The manuscript contains unresolved citation placeholders ([XXX]) at the Fourier-plane assumption and at the simulation comparison, so the numerical support for those steps is not yet fully documented in the text.","The same phase-transfer mechanism should work for any programmable phase mask on the control beam, not only vortices, which would turn the atomic memory into a reconfigurable all-optical convolution engine for arbitrary kernels.","Since the retrieval is already recorded with time-correlated single-photon imaging and cold-ensemble memories have stored single photons, a quantum version of processing-in-memory — storing and convolving nonclassical images — is a natural next step if the phase-imprinting step does not add significant noise."],"forward_implications":["Optical processing-in-memory becomes feasible in practice: the same atomic ensemble stores an image, transforms it, and releases it, bypassing the data movement between memory and processor that dominates latency and energy in conventional architectures.","Edge-enhanced images can be held for more than 320 µs and read out on demand, a timescale corresponding to roughly 60 km of optical-fiber delay and long enough for many network-level synchronization and time-multiplexing tasks.","The read-out control field acts as a second programmable filter: applying the same vortex during write-in and read-out cancels the first convolution, while using different charges produces higher-order spiral phase contrast ($\\Delta l=2,3$), showing that the operation is reversible and tunable.","Because SPC is a phase-only filter, the technique preserves photon throughput; even at $|l|=3$, retrieval efficiency remains above $1/e$ of the case without SPC, so the processing does not destroy the stored signal."],"supporting_citations":[{"why":"supplies the EIT formalism for the lambda-level coherent medium used in the memory.","marker":"[17]"},{"why":"introduces dark-state polaritons, the quasiparticle model for adiabatic storage and retrieval that the write-in/read-out analysis is built on.","marker":"[18]"},{"why":"extends the dark-state polariton model to quantum memory for photons, grounding the retrieval equation for the signal field.","marker":"[19]"},{"why":"proposes spiral-phase transfer between light and atomic excitations in a Raman memory, the idea generalized here to on-resonance EIT.","marker":"[21]"},{"why":"introduces spiral phase contrast imaging, the edge-enhancement technique implemented inside the memory.","marker":"[22]"},{"why":"identifies the spiral phase filter as a radial Hilbert transform, explaining the point-spread function used in Eq. (1).","marker":"[23]"},{"why":"provides the method for suppressing inhomogeneous magnetic dephasing of stored images, credited with enabling the long storage times.","marker":"[24]"},{"why":"defines hypergeometric-Gaussian modes, the vortex profile $g(r)e^{-il\\phi}$ imprinted on the control fields and its far-field form.","marker":"[27]"}],"fun_headline_variants":["Atomic memory performs all-optical convolution and stores result","Cold atom cloud edge-enhances images while storing them","EIT memory does convolution, keeps image for 320 µs","Optical convolution inside atomic storage, lifetime 320 µs","Process and store photon images in a cold atomic ensemble"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiment assumes the cold atomic cloud acts as a perfect thin two-dimensional screen at the Fourier plane, so that the stored spin wave is exactly the product of the image spectrum and the vortex phase; if finite cloud size, diffraction, or phase curvature materially corrupts that multiplication, the claimed convolution kernel is not realized.","fun_headline_variants_meta":{"raw":{"variants":["Atomic memory performs all-optical convolution and stores result","Cold atom cloud edge-enhances images while storing them","EIT memory does convolution, keeps image for 320 µs","Optical convolution inside atomic storage, lifetime 320 µs","Process and store photon images in a cold atomic ensemble"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1397,"prompt_tokens":915,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":531,"tokens_out":482,"duration_ms":5178,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:48:09.425689+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a pinhole or a small object at the mask plane and directly measure the retrieved point-spread function at several lateral positions: if it is not the predicted $\\mathcal{F}^{-1}\\{g^2(r)e^{i\\Delta l \\phi}\\}$ — for instance, if the vortex center shifts, the doughnut size changes with input position, or the pattern depends on axial placement of the cloud — then the two-dimensional Fourier-plane filter model is wrong and the agreement with simulation would not generalize.","supporting_citations":[{"cited_title":"Ding, Z.-Y","cited_arxiv_id":null,"evidence_quote":"supplies the EIT formalism for the lambda-level coherent medium used in the memory."},{"cited_title":"Fleischhauer, A","cited_arxiv_id":null,"evidence_quote":"introduces dark-state polaritons, the quasiparticle model for adiabatic storage and retrieval that the write-in/read-out analysis is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"proposes spiral-phase transfer between light and atomic excitations in a Raman memory, the idea generalized here to on-resonance EIT."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces spiral phase contrast imaging, the edge-enhancement technique implemented inside the memory."},{"cited_title":"F ¨urhapter, A","cited_arxiv_id":null,"evidence_quote":"identifies the spiral phase filter as a radial Hilbert transform, explaining the point-spread function used in Eq. (1)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the method for suppressing inhomogeneous magnetic dephasing of stored images, credited with enabling the long storage times."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines hypergeometric-Gaussian modes, the vortex profile $g(r)e^{-il\\phi}$ imprinted on the control fields and its far-field form."}],"review_version":2}