{"id":"bf3861e4-acd4-482b-b959-98a2e7752f64","arxiv_id":"2505.17381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A programmable photonic unitary processor, optimized through a differentiable fiber model, is shown in a 1300-km three-mode fiber experiment to cut modal dispersion by more than half.","lead":"This paper demonstrates a long-haul optical transmission link in which a programmable photonic chip reshapes the fiber channel itself to reduce signal distortion, rather than letting the receiver's digital processor fix everything. A 1300-km three-mode fiber experiment shows that the optimized chip cuts modal delay from about 25 to 10 nanoseconds, which would lighten the digital equalization load.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Recirculating-loop experiment reuses one 51.2-km fiber per pass, so it tests U against a periodic operator (U M)^K, not the independent random spans of Eq. 4; the claimed long-haul DMD reduction may not generalize to real multi-span links.","rationale":"The paper is a serious proof-of-concept: the PLC Clements mesh has genuinely low loss (2.1 dB fiber-to-fiber), C-band fidelity R2>0.96, and the authors are transparent that strong-coupling fibers erase the advantage. The differentiable optimization and the measured landscapes agree, and the NGMI stays above the FEC threshold. The single most load-bearing weakness is not the absolute DMD value or the metric choice; it is the mismatch between the model's independent-span statistics and the experiment's single-fiber periodic loop. The central claim is about reducing DSP load in long-haul SDM, i.e., in a chain of many different fiber segments. The only long-haul experimental evidence is a recirculating loop that reuses the same 51.2-km fiber and the same fixed optical block K times. Repeated multiplication of one fixed transfer matrix is mathematically different from multiplying independent random matrices, and the optimized U was found for the latter. If the 25-to-10 ns reduction is partly an artifact of periodicity, the experiment overstates the practical benefit. This is testable without new hardware by simulating the periodic model with the measured fiber transfer matrix. It does not justify rejection: the framework may still work, and the authors' honest treatment of strong coupling is a point in their favor. It does justify keeping the verdict CONDITIONAL and adding this specific check to the requested revisions. The reader's weakest assumption already points at the digital twin's fiber model; this concern is a concrete, unaddressed way that the model-experiment correspondence fails, hence partial agreement.","tokens_in":23069,"tokens_out":12863,"duration_ms":113851,"concrete_test":"Use the authors' own differentiable simulator to evaluate the exact periodic-chain model M_total = M_{K+1}(U_opt M)^K for K=1...26, with M taken from the measured 3×3 transfer matrix of the actual loop fiber (estimated at 1550 nm), and compare this DMD curve with the i.i.d.-span model used for optimization and with the experimental points in Fig. 6(a). If the periodic model reproduces the 25 ns to 10 ns reduction while the i.i.d. model does not, the recirculating-loop experiment is not representative of the long-haul claim; if the two models agree, the concern is resolved. An experimental alternative is to insert a fast mode scrambler into the loop so that S_k,V_k are decorrelated on each pass and repeat the measurement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 4 models K statistically independent spans, with each M_k(ω)=S_k D_k(ω) V_k† drawn from random S_k,V_k; the optimization in Fig. 2 averages over 100 such draws. The experimental setup is a recirculating loop: the same 51.2-km three-mode fiber and the same DEMUX/EDFA/U/MUX block is traversed K times, so the end-to-end operator is approximately M_{K+1}(U M)^K with a single fixed M. The unitary U was chosen to minimize the ensemble-averaged modal-delay spread of Eq. 9 over independent M_k; applying it to a periodic chain is a different optimization problem. Agreement between one periodic-loop realization and a 100-trial i.i.d.-span average in Figs. 4-6 is therefore not, by itself, evidence that the same U would suppress DMD in a deployed link with physically distinct spans. The observed 25 ns to 10 ns reduction could instead reflect a special property of repeatedly multiplying the same fiber's transfer matrix. No mode scrambler, alternating spool, or measured per-pass decorrelation is reported; the paper only states that the setup 'corresponds to testing' span-independent U, which is not the same as having independent spans.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'parameterized SDM transmission,' in which programmable photonic unitary processors placed at intermediate nodes of a spatial-division-multiplexed fiber link are configured, through a differentiable digital twin, to minimize accumulated differential mode delay and thereby reduce the MIMO-DSP equalization burden. The authors fabricate a silica-PLC 8×8 Clements-style unitary processor with 2.1 dB fiber-to-fiber insertion loss and R2>0.96 fidelity across the C-band, and they demonstrate a 1300-km three-mode recirculating-loop experiment in which an optimized unitary transformation reduces the measured 95% modal-dispersion window from about 25 ns to about 10 ns, with NGMI staying above the 0.836 error-free threshold. The paper also reports numerical optimization studies for three-mode and ten-mode fibers under different coupling scenarios and a series of supplementary analyses (calibration method, span-length dependency, comparison with simulated annealing and fixed permutations).","tokens_in":23305,"tokens_out":7437,"duration_ms":53893,"significance":"If the generalization claims hold, this is a notable step toward optical-domain MIMO preprocessing for long-haul SDM: it demonstrates a telecom-grade programmable unitary processor with state-of-the-art loss and bandwidth, and it provides a gradient-based framework for optimizing physical-layer transformations against a differentiable channel model. The deterministic-Θ sweep (Figs. 4) and the optimized-U experiment (Fig. 6) exhibit internally consistent agreement between the digital twin and measured DMD, and the device characterization in Fig. 3 is careful and convincing. The main significance is tempered by two validation weaknesses: the fiber model parameters are chosen to 'well explain' the same experimental results they are then used to validate, and the recirculating-loop experiment does not reproduce the statistically independent spans assumed in Eq. (4).","major_comments":[{"comment":"Eq. (4) models K statistically independent fiber spans, with each M_k(ω) drawn from independent random S_k and V_k, and the optimization in Fig. 2 averages over 100 such draws. In the recirculating loop, the same 51.2-km fiber and the same DEMUX/EDFA/U/MUX block are traversed K times, so the end-to-end operator is approximately M_{K+1}(U M)^K with a single fixed M. The Methods statement that the setup 'corresponds to testing the span-independent unitary matrix U' addresses only the span-independence of U, not the statistical independence of the spans. The agreement between a periodic-loop realization and an i.i.d.-span simulation average in Figs. 4–6 therefore does not establish that the same U would suppress DMD in a deployed link with physically distinct spans. Please provide a per-pass decorrelation measurement or a periodic-operator simulation demonstrating that the optimized U remains near-optimal, or explicitly limit the claims to periodic/recirculating-loop systems.","section":"Methods, Eq. (4) and experimental setup (Results, Fig. 4(a))"},{"comment":"The key fiber-model parameters (LP11 modal delay of 55 ps/km, strong-intra-group/weak-inter-group coupling) are introduced with the statement that this choice 'well explains the experimental results,' but no independent characterization of the fiber is reported. Because the same experimental dataset is then used to validate the digital twin in Figs. 4–6, the agreement is partially circular and cannot serve as a predictive test of the model. Please provide an independent DMD/coupling measurement (for example OFDR or cutback characterization) or, if that is not available, a sensitivity analysis over the assumed parameters showing that the optimized unitary's DMD reduction is robust to the chosen values.","section":"Methods, 'Numerical simulation and optimization' (three-mode case)"},{"comment":"The printed derivation of the gradient in Eq. (10) is incomplete: the group-delay operator is G(ω) = -j (dM/dω) M^{-1}, so dL/dU_k should involve the derivative of that full expression, including the M^{-1} factor and the -j; the finite-difference expression S'† (1/Δω)[dM(ω+Δω)/dU_k - dM(ω)/dU_k] V' lacks those terms. In practice the authors state that they use PyTorch autograd, which would differentiate through the full model and may be correct, but the equation as written does not support the method. Additionally, the cost function in Eq. (9) contains a maximum over modes, which is non-differentiable at ties; the manuscript does not discuss how autograd or the finite-difference approximation handles this. Please correct the derivation or state explicitly that autograd differentiates through G(ω) and describe the subgradient treatment, and confirm that the reported convergence is unaffected.","section":"Methods, Eq. (10) and gradient-based optimization"}],"minor_comments":[{"comment":"The caption contains a garbled fragment: '(e) Implemented unitary matrices for different Θ setups. θ.' should be a proper sentence describing panel (b) or (e), and the panel labels in the caption do not match the figure panels.","section":"Figure 4 caption"},{"comment":"'slid lines' should be 'solid lines' in both captions.","section":"Figure captions 4(e) and 6(a)"},{"comment":"The caption states 'we sent 2^XXX symbols,' which appears to be an unresolved placeholder; the main text says 33,360 symbols.","section":"Figure 6 caption"},{"comment":"The section title contains a typo: 'Controlability of mode dependent loss uisng photonic unitary processor' should be 'Controllability ... using'.","section":"Supplementary Information S9"},{"comment":"The heading 'SDM-WDM-PDM compatible programable photonic unitary processor' misspells 'programmable'.","section":"Heading before Fig. 3"},{"comment":"The matrix expressions for U_det and U_rot are misaligned in the rendered text; please ensure the row/column structure is typeset correctly.","section":"Eqs. (1) and (2)"}],"recommendation":"major_revision","confidential_remarks":"The device work and the core optimization idea are solid and would be a good fit for the journal, but the validation currently rests on a circular parameter choice and a recirculating-loop experiment that does not match the model's independent-span assumption. The requested additions (independent fiber characterization or sensitivity analysis; per-pass decorrelation evidence or periodic-operator simulation) are within the scope of a major revision. If those cannot be supplied, the authors should scale back the generalization claims and frame the work strictly as a proof-of-concept for a single fixed recirculating loop."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The paper has two substantial pieces: a PLC-based programmable unitary processor with genuinely good specs (2.1 dB fiber-to-fiber loss, C-band, polarization insensitive, R2>0.96), and a differentiable model of SDM transmission used to train that processor to reduce differential mode delay. The headline experiment shows DMD dropping from ~25 ns to ~10 ns after 1300 km in a three-mode fiber loop, with NGMI above threshold. That is a real first: prior work used passive scrambling or fixed permutations.\n\nThe things I like: the processor is a serious piece of engineering, the optimization framework is clearly laid out, and the authors are honest about the regime where the method gives no advantage (strong coupling). The gradient-based optimizer is compared against simulated annealing, and the convergence advantage is plausible.\n\nThe soft spots are also real. The experimental setup is a recirculating loop, so every pass encounters the same 51.2-km fiber span with transfer matrix M; the end-to-end operator is essentially (U M)^K with one fixed M. The simulation model, by contrast, averages over independent random spans M_k. These are different optimization problems, and the paper simply asserts that the setup 'corresponds' to the span-independent case without justifying it. So the 25-to-10 ns reduction is proven for that particular loop, but not for a deployed link with distinct spans. That is a limitation, not a refutation.\n\nSecond, the model's delay parameter (55 ps/km) is calibrated against the same fiber's baseline DMD, which makes the 'prediction' partly a consistency check. The experiment does independently measure the MIMO equalizer window, so the main result isn't manufactured, but the digital twin isn't blind.\n\nThird, experimental points come without error bars, and the code/data are only 'available on request.' For a paper that leans on simulation-experiment agreement, that's a bit thin.\n\nBottom line: this is a solid proof-of-concept for the subfield, with genuine engineering value. The generalization to real long-haul links is open. I'd send it to peer review with a request to address the recirculating-loop issue, add an independent delay measurement, and release artifacts. It's not a desk reject.","headline":"A real first for programmable photonic DMD reduction, but the recirculating-loop setup leaves the multi-span generalization unproven.","tokens_in":23919,"tokens_out":5994,"would_cite":true,"duration_ms":47089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Programmable photonic processors in a multimode fiber link, tuned by gradient descent through a differentiable transmission model, cut the modal delay spread from ~25 ns to ~10 ns over 1300 km and shrink the receiver's MIMO equalizer load.","keywords":["parameterized SDM transmission","differential mode delay","programmable photonic unitary processor","Clements MZI mesh","differentiable transmission model","spatial division multiplexing","MIMO-DSP complexity","silica PLC"],"falsifier":"Characterize the installed three-mode fiber's actual coupling and delay statistics from its measured transfer matrix and repeat the optimization on that measured model: if the resulting unitary no longer outperforms a random mode scrambler in shrinking the 95% equalizer window over 1300 km, the central claim fails. The paper's own simulations show that in the strong-coupling regime optimization gives no advantage over random scrambling, so checking which regime the real fiber sits in is the decisive observation.","tokens_in":22813,"feed_emoji":"🔆","tokens_out":20451,"duration_ms":134057,"temperature":0.7,"pith_summary":"In long-haul transmission over multimode fiber, the spread of arrival times between spatial modes — differential mode delay — forces the receiver's digital MIMO equalizer to cover a long temporal window, and the computational cost grows with that window. This paper tries to establish that the optical channel itself can be made trainable: programmable photonic unitary processors placed at intermediate nodes, configured by gradient descent through a fully differentiable model of the fiber link, can suppress the delay spread in the optical domain before the receiver sees the signal. The authors built an 8×8 Clements-type silica planar-lightwave-circuit (PLC) mesh with 2.1 dB fiber-to-fiber loss, polarization independence, and fidelity above $R^2 = 0.96$ across the C-band; in a 1300-km three-mode-fiber experiment, the optimized matrix cut the measured delay spread from roughly 25 ns to roughly 10 ns at three wavelengths while the normalized generalized mutual information (NGMI) stayed above the 0.836 error-free threshold. If the claim holds, receiver DSP complexity, energy per bit, and post-processing latency in SDM networks can all shrink, and the same programmable mesh could later combine dispersion shaping with routing or mode-dependent-loss control.","feed_headline":"Photonic processor trims fiber delay from 25 to 10 ns over 1300 km","feed_subtitle":"A differentiable fiber model trains the mesh, shrinking the window the receiver's digital equalizer must cover.","key_machinery":"The argument is carried by two coupled objects. The first is the differentiable digital twin of the link: each span is $M_k(\\omega) = S_k D_k(\\omega) V_k^\\dagger$, with $D_k$ holding the modal delays as frequency-dependent phases, and the group-delay operator $G(\\omega) = -j \\frac{d}{d\\omega} \\log M_{\\mathrm{total}}(\\omega)$ converts the full chain into a Hermitian matrix whose eigenvalues are the accumulated delays; the maximum eigenvalue spread (Eq. 9) is the training loss. The second is the Clements decomposition, which factorizes $U_k$ into 2×2 Mach–Zehnder blocks parameterized by physical phase shifts $\\xi_1, \\xi_2$, so that backpropagation updates realizable hardware settings while $U_k$ stays exactly unitary — and the same differentiable model doubles as the twin for machine-learning calibration of the fabricated mesh. The physical enabler is the 8×8 silica PLC Clements circuit: equalized path lengths remove the delay-line wavelength dependence typical of interferometric meshes, dual-arm thermo-optic MZIs suppress thermal crosstalk, and calibration infers residual coupler splitting ratios and external phase errors, lifting fidelity from $R^2 = 0.53$ to $0.99$ at 1550 nm and holding $R^2 > 0.96$ across the C-band — above the $R^2 \\approx 0.95$ level the paper identifies as the degradation threshold.","core_discovery":"The paper's central claim is that differential mode delay — the arrival-time spread among spatial modes that sets the equalizer window length $L$ and hence the $O(N^2 L)$ cost of multiple-input multiple-output digital signal processing (MIMO-DSP) — can be suppressed in the optical domain by applying a wavelength-insensitive unitary matrix $U_k$ at each span, and that the right $U_k$ can be found by gradient-based optimization through a differentiable model rather than by heuristic or combinatorial search. The model writes each fiber span as $M_k(\\omega) = S_k D_k(\\omega) V_k^\\dagger$, chains the spans with the trainable unitaries into $M_{\\mathrm{total}}(\\omega) = M_{K+1}(\\omega) \\prod_{k=1}^{K} U_k M_k(\\omega)$, and reads the accumulated modal delays as the eigenvalues of the Hermitian group-delay operator $G(\\omega) = -j \\frac{d}{d\\omega} \\log M_{\\mathrm{total}}(\\omega)$, computed by singular value decomposition so the whole pipeline stays differentiable. Training then updates the MZI phase parameters of the Clements mesh by backpropagation, preserving unitarity exactly. The paper reports that the procedure converges in three-mode and ten-mode simulations, beats the random-mode-scrambling square-root-of-distance baseline in no-coupling and weak-coupling regimes, and — when the optimized matrix is transferred to the fabricated PLC mesh — reduces the experimentally measured 95% impulse-response window from about 25 ns to about 10 ns over 1300 km at 1541.349, 1550.057, and 1558.173 nm, with NGMI above the 0.836 threshold throughout.","pith_inferences":["The fidelity budget the paper measures (roughly $R^2 > 0.95$ for acceptable DMD degradation) doubles as a portability criterion: any low-loss broadband mesh platform that meets it could host the same optimization, making the result a system-level recipe rather than a single-device demonstration.","Because the optimizer needs only a differentiable model, the same train-the-channel pattern should transfer to other slowly varying impairments — span-wise mode-dependent loss, gain ripple, or slow drift of the mesh itself — wherever a surrogate model can be written.","The paper observes many near-equivalent, initial-condition-dependent optima, which implies the cost landscape is flat along several directions; that degeneracy could be exploited as a free degree of freedom to choose unitaries that also improve routing or MDL without sacrificing delay suppression."],"forward_implications":["Reducing the measured delay spread from ~25 ns to ~10 ns over 1300 km shortens the temporal window the receiver's MIMO-DSP must equalize, so digital post-processing load falls roughly in proportion to the window length.","Because DMD is a slow, statistical property of the link, the optimized unitary remains valid over long timescales and needs no fast tracking, keeping control overhead low in deployment.","Simulations show the suppression strengthens with denser processor deployment (shorter spans) and extends to ten-mode fibers with group-wise coupling, so the three-mode experiment is a scaled proof of a general mechanism.","The same differentiable model can target impairments beyond delay: the paper shows the optimized unitaries also shift mode-dependent loss and its NGMI penalty, opening the way to joint DMD/MDL optimization.","Idle ports of the Clements mesh can be repurposed for spatial-mode routing, so a single programmable node could eventually combine dispersion shaping, MDL control, and switching."],"supporting_citations":[{"why":"Supplies the mesh decomposition that parameterizes each unitary matrix into 2×2 MZI phase shifts, keeping trained matrices exactly unitary.","marker":"[33]"},{"why":"Provides the fiber-span model and the group-delay operator whose eigenvalues define the modal delays used as the cost function.","marker":"[34]"},{"why":"Supplies the statistics of group delays in multimode fiber, grounding the SVD-based delay calculation and the square-root-of-distance random-coupling baseline.","marker":"[35]"},{"why":"The prior 10-spatial-mode 1300-km transmission whose delay profile and results this work extends; its delay values seed the ten-mode simulations.","marker":"[11]"},{"why":"Establishes the random mode-scrambling baseline and modal statistics that optimized unitaries must outperform in the DMD comparison.","marker":"[12]"},{"why":"Supports the claim that random mode coupling gives only square-root-of-distance delay growth, the benchmark the optimization aims to beat.","marker":"[13]"},{"why":"Defines the deterministic 3×3 rotation family used to validate the digital twin against the experimental equalizer window.","marker":"[43]"},{"why":"Defines the NGMI threshold of 0.836 used as the error-free criterion in the transmission experiment.","marker":"[44]"},{"why":"Provides the automatic differentiation engine used to backpropagate through the differentiable SDM model and to calibrate the mesh.","marker":"[50]"}],"fun_headline_variants":["Optical mesh cuts equalizer window from 25 to 10 ns","Differentiable photonic mesh trims modal delay over 1300 km","Programmable photonic chip shrinks fiber delay spread by 60%","Spatial mode transmission boosted by trained photonic processor","Gradient-optimized photonic mesh reduces DSP load in SDM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the digital twin of the fiber is faithful: the three-mode link really is weakly coupled between the LP01 and LP11 groups with negligible mode-dependent loss, and the delay spread the optimizer minimizes in simulation is the same quantity as the 95% pulse-energy window the real receiver's equalizer must cover — if either assumption fails, the optimized photonic setting will not cut the DSP load as predicted.","fun_headline_variants_meta":{"raw":{"variants":["Optical mesh cuts equalizer window from 25 to 10 ns","Differentiable photonic mesh trims modal delay over 1300 km","Programmable photonic chip shrinks fiber delay spread by 60%","Spatial mode transmission boosted by trained photonic processor","Gradient-optimized photonic mesh reduces DSP load in SDM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3180,"prompt_tokens":1168,"completion_tokens":2012,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":784,"completion_tokens_details":{"reasoning_tokens":1919}},"tokens_in":784,"tokens_out":2012,"duration_ms":11673,"temperature":1.0,"reasoning_tokens":1919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:48:15.015160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Characterize the installed three-mode fiber's actual coupling and delay statistics from its measured transfer matrix and repeat the optimization on that measured model: if the resulting unitary no longer outperforms a random mode scrambler in shrinking the 95% equalizer window over 1300 km, the central claim fails. The paper's own simulations show that in the strong-coupling regime optimization gives no advantage over random scrambling, so checking which regime the real fiber sits in is the decisive observation.","supporting_citations":[{"cited_title":"R., Humphreys, P","cited_arxiv_id":null,"evidence_quote":"Supplies the mesh decomposition that parameterizes each unitary matrix into 2×2 MZI phase shifts, keeping trained matrices exactly unitary."},{"cited_title":"Mode-dependent loss and gain: statistics and effect on mode-division multiplexing","cited_arxiv_id":null,"evidence_quote":"Provides the fiber-span model and the group-delay operator whose eigenvalues define the modal delays used as the cost function."},{"cited_title":"Statistics of group delays in multimode fiber with strong mode coupling","cited_arxiv_id":null,"evidence_quote":"Supplies the statistics of group delays in multimode fiber, grounding the SVD-based delay calculation and the square-root-of-distance random-coupling baseline."},{"cited_title":"10 -spatial-mode 1300-km transmission over 6 -LP graded index few -mode fiber with 36 -ns modal dispersion","cited_arxiv_id":null,"evidence_quote":"The prior 10-spatial-mode 1300-km transmission whose delay profile and results this work extends; its delay values seed the ten-mode simulations."},{"cited_title":"Modal statistics in mode-division-multiplexed systems using mode scramblers","cited_arxiv_id":null,"evidence_quote":"Establishes the random mode-scrambling baseline and modal statistics that optimized unitaries must outperform in the DMD comparison."},{"cited_title":"Long-period fiber gratings for mode coupling in mode -division- multiplexing systems","cited_arxiv_id":null,"evidence_quote":"Supports the claim that random mode coupling gives only square-root-of-distance delay growth, the benchmark the optimization aims to beat."},{"cited_title":"Long-haul three-mode-multiplexed transmission employing on-chip mode permutation technique based on a programable photonic unitary processor","cited_arxiv_id":null,"evidence_quote":"Defines the deterministic 3×3 rotation family used to validate the digital twin against the experimental equalizer window."},{"cited_title":"Normalized generalized mutual information as a forward error correction threshold for probabilistically shaped QAM","cited_arxiv_id":null,"evidence_quote":"Defines the NGMI threshold of 0.836 used as the error-free criterion in the transmission experiment."},{"cited_title":"PyTorch: An Imperative Style, High -Performance Deep Learning Library","cited_arxiv_id":null,"evidence_quote":"Provides the automatic differentiation engine used to backpropagate through the differentiable SDM model and to calibrate the mesh."}],"review_version":1}