{"id":"47dda214-2fc6-4510-b2de-47ba9d1a4419","arxiv_id":"2507.20470","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An affine map learned from scattering data reconstructs 16-bit RZ-DPSK pulse trains from the continuous Lax spectrum alone, with validation errors around 10^-4.","lead":"This paper tests whether a simple linear (affine) mapping can replace the complicated inverse scattering transform for a class of optical signals, using only continuous spectral data. The authors report accurate reconstruction of the transmitted field for 16-bit Gaussian pulse trains and a low-rank structure tied to the bit count, but the experiments do not yet include fiber propagation or noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Continuous-spectrum sufficiency is asserted, not proved; an exhaustive injectivity test on the full 16-bit pattern family is needed before accepting either the z=0 reconstruction claim or the fiber-output bit-recovery claim.","rationale":"The reader's weakest assumption—that the continuous reflection coefficient uniquely and sufficiently encodes the initial potential within this family—is exactly the load-bearing point. My independent reading of Sections 3–5 confirms that the paper provides no analytic argument for this uniqueness, only an intuitive claim and a finite-sample numerical demonstration. The standard theory of the focusing Zakharov–Shabat problem makes the reflection coefficient and the discrete spectral data independent in general, so the burden is on the authors to show that this particular family has no aliasing. The proposed exhaustive or Hamming-neighbor injectivity check would settle this directly: if two distinct bit patterns yield near-identical Re r, the affine map cannot generalize to the full codebook, and the communications claim fails. If the check shows all 65,536 patterns are well separated in input space, the z=0 claim is materially supported, though the fiber-output claim would still require an actual propagation/back-propagation experiment. I agree with the reader's conditional verdict: the paper should not be rejected outright because the numerical evidence is coherent and reproducible in principle, but it should not be accepted as a demonstration of fiber-output bit recovery until the injectivity and propagation gaps are closed.","tokens_in":8931,"tokens_out":11466,"duration_ms":159122,"concrete_test":"Carry out an exhaustive injectivity check on the full 16-bit family. Using the same FNFT solver, grid (ξ∈[-5,5], Mξ=4096, then n=1024) and preprocessing, compute x(a)=Re r(0,ξ) for all 65,536 patterns a∈{±1}^16 (or, if too costly, for the 4500 patterns at Hamming distance 1 from the validation set plus all-same-sign and alternating patterns). Compute the smallest normalized pairwise distance d_min = min_{i≠j} ‖x_i−x_j‖_2 / mean_k‖x_k‖_2. If d_min is comparable to the numerical noise level (estimated by shifting the ξ-grid by one step, per the grid sensitivity acknowledged in Section 5), distinct bit patterns are aliased and the continuous spectrum is insufficient. For the closest pair, retrain the affine map on the other 65,534 patterns and check that the predicted potential decodes to the correct bit pattern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Standard IST treats the reflection coefficient r(ξ) and the discrete spectral data (eigenvalues and norming constants) as independent scattering data. The paper's Section 5 asserts that 'the reflection coefficient implicitly encodes information about discrete part,' but this is an intuitive statement with no derivation, and it is not a standard property of the focusing Zakharov–Shabat problem. If two 16-bit RZ-DPSK potentials in the tested family had nearly identical real parts Re r(0,ξ) but different bit patterns, no inverse map—affine or otherwise—could separate them. The reported validation error (7.5e-5) on 300 random held-out patterns out of 65536 does not rule this out: near-collisions, if they exist, would be rare and would not appear in a random 30% split. The rank-16 SVD observation (Section 4.2) also does not establish uniqueness; it is consistent with the training inputs lying in a 16-dimensional subspace (one dimension per bit), which is what any differentiable parameterization by 16 independent bits would produce. Finally, the abstract's claim about recovery 'at the fiber output' has no supporting experiment: all numerical results use z=0 scattering data, and the linear back-propagation stage of Figure 2 is never implemented or tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes approximating the inverse scattering transform for the focusing nonlinear Schrödinger equation with an affine map learned from simulated data, using only the continuous part of the Lax spectrum (the reflection coefficient). For 16-bit RZ-DPSK Gaussian pulse trains, the authors compute scattering data at z=0, fit an affine map from Re r(0,ξ) to the initial potential, observe that a rank-16 SVD truncation suffices, and report average relative L2 errors of 1.9e-6 (training) and 7.5e-5 (validation). The abstract and conclusion extrapolate to recovery of the transmitted bit sequence at the fiber output after linear back-propagation of the scattering data, but no propagation experiment is performed.","tokens_in":9198,"tokens_out":3457,"duration_ms":39232,"significance":"If the extrapolation to fiber output were validated, the results would suggest a fast, hardware-friendly surrogate for the inverse scattering step in nonlinear optical communications, with the appealing observation that the learned operator's rank matches the number of bits. The z=0 numerical experiment is internally consistent: training and validation are properly separated, normalization statistics are taken from training data, and the held-out error is low. However, the central generalization—continuous-spectrum sufficiency and transferability to propagated signals—rests on assertions rather than proof, and the paper's own Discussion acknowledges the empirical nature of the simplification. The significance is conditional on addressing these gaps.","major_comments":[{"comment":"The abstract and conclusion claim that accurate recovery of the transmitted bit sequence can be achieved from the continuous part of the Lax spectrum at the fiber output, but every numerical experiment in Section 4 uses scattering data computed at z=0, with the map G defined from r(0,ξ) to u(0,t) in Section 3.2. The linear back-propagation stage of Figure 2 is never implemented or tested. The manuscript should either add an end-to-end fiber-propagation experiment or restrict the claims to the z=0 reconstruction.","section":"Abstract; Section 6; Figure 2"},{"comment":"Section 5 states that the continuous spectrum (reflection coefficient) is sufficient to reconstruct the initial potential for Gaussian RZ pulse trains and that the reflection coefficient implicitly encodes the discrete part. This is an empirical assertion without derivation or exhaustive verification; standard focusing Zakharov-Shabat theory treats discrete eigenvalues and norming constants as independent scattering data. The reported validation on 300 random patterns out of 65536 does not rule out near-collisions between different bit patterns in the map to Re r(0,ξ). The authors should either prove injectivity (e.g., by an exhaustive test over all 2^16 patterns or a rigorous argument) or explicitly label this as a hypothesis and quantify the risk of collision.","section":"Section 5; Section 4.1"},{"comment":"The observation that the singular values decay sharply after the sixteenth and the choice r=16 are consistent with any model in which 16 independent scalar parameters determine the sequence; it does not independently establish that the continuous spectrum alone encodes the bit pattern. This is not an error, but the claim that the numerically evaluated rank equals the number of bits should be phrased as a data-dependent model-selection result, not as a structural property of the inverse map.","section":"Section 4.2; Abstract"}],"minor_comments":[{"comment":"The sentence 'This work present an affine map approximation' should be 'This work presents an affine map approximation' or 'This paper presents...'.","section":"Abstract"},{"comment":"The word 'surogate' in the introductory paragraph of Section 3.1 should be 'surrogate'.","section":"Section 3.1"},{"comment":"In the sentence 'Thus, to find a discreetized version of the affine operator,' the word 'discreetized' should be 'discretized'.","section":"Section 3.3"},{"comment":"The phrase 'These samples are splitted into training' should be 'These samples are split into training'.","section":"Section 4.1"},{"comment":"Reference [27] appears to contain a typographical artifact: '3 _J N Elgin' should likely be 'J N Elgin'.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable proof-of-concept for the special issue. The main risk is that the abstract overclaims what is tested; the authors should either conduct a propagation simulation or temper the title and abstract. I would also encourage the authors to situate the work against existing data-driven approaches to nonlinear Fourier transform inversion and to discuss the computational cost of the direct scattering transform that is used to generate training data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, know this: the paper is a clean numerical demonstration, not a theory breakthrough. What is actually new is showing that for 16-bit RZ-DPSK Gaussian pulse trains, a rank-16 affine map from the real part of the reflection coefficient reconstructs the initial potential with relative L2 validation error around 7.5e-5, and that the singular values of the fitted map drop sharply after 16. That is a useful empirical observation for anyone building nonlinear Fourier transform receivers, and the experiment is reproducible: 1000 samples, 70/30 split, standard least squares plus SVD. The paper is honest enough to mention grid sensitivity in Section 5.\n\nThe major soft spot is the continuous-spectrum sufficiency premise. The paper asserts in Section 5 that the reflection coefficient implicitly encodes the discrete spectrum, but that is not proven and is not standard for the focusing Zakharov-Shabat problem. The 300-sample validation does not rule out rare near-collisions among the 65536 patterns, so an exhaustive injectivity test—or at least a sweep over all patterns—would be the right check before accepting the z=0 claim. Relatedly, the abstract's statement about 'fiber output' is not supported: all numerics use z=0 scattering data, and the linear back-propagation step in Figure 2 is never implemented. That is a scope overclaim, not a fatal error, but it needs fixing.\n\nThe rank-16 observation, as the stress-test note says, is exactly what you'd expect from 16 independent bits; it is consistent with the map being a linear readout of the bit pattern, not evidence of a new spectral property. Still, it is mathematically nice and motivates the reduced model.\n\nOne more missing piece: no comparison against using the full scattering data (reflection coefficient plus discrete data) or against a nonlinear surrogate. Without that baseline, the paper cannot claim the continuous spectrum alone is sufficient; it only shows the affine map does well on the data used.\n\nBottom line: for an audience working on fast NFT approximations, this is a worth-reading methods paper. It deserves peer review, but a serious referee should send it back for a propagation simulation and an injectivity check. I would not cite it in my own work; it is more of a building block than a result I would lean on.","headline":"A clean, modest numerical study of an affine-map surrogate for the inverse scattering transform; the z=0 reconstruction works, but the continuous-spectrum sufficiency premise and the fiber-output claim are asserted, not tested.","tokens_in":9676,"tokens_out":2205,"would_cite":false,"duration_ms":26331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q55","37K15","78A60"],"pacs":["42.65.-k","42.79.Sz"],"model":"deepseek-v4-flash","headline":"One low-rank affine map reconstructs 16-bit optical pulse trains from the continuous Lax spectrum alone.","keywords":["inverse scattering transform","affine map","nonlinear Schrödinger equation","optical fiber communications","reflection coefficient","reduced order modeling","low-rank approximation","differential phase-shift keying"],"falsifier":"Compute the real part of the reflection coefficient $r(0,\\xi)$ on the same 1024-point grid for all 65,536 possible 16-bit DPSK patterns; if any two distinct patterns yield reflection coefficients whose $\\ell^2$ distance is below the numerical noise floor, the affine map cannot separate them, and the claim that the continuous spectrum alone encodes the bit sequence fails for that pair. Alternatively, propagate a held-out pattern to fiber length $\\ell$, back-propagate its scattering data linearly, and check whether the affine map trained at $z=0$ reconstructs the transmitted field.","tokens_in":8734,"feed_emoji":"📡","tokens_out":5731,"duration_ms":58431,"temperature":0.7,"pith_summary":"This paper tries to establish that, for return-to-zero Gaussian pulse trains carrying 16 bits of differential phase-shift keying, the inverse scattering transform of the nonlinear Schrödinger equation can be replaced by a single affine map learned from data. The map takes samples of the reflection coefficient, the continuous part of the Lax spectrum, and outputs the initial optical field; the discrete soliton part of the spectrum is discarded. If true, this gives a fast, integrability-based way to undo nonlinear fiber distortions at the receiver, with a reduced-order model whose rank equals the number of bits per sequence. The paper reports mean relative $L^2$ reconstruction errors around $7.5\\times10^{-5}$ on held-out patterns.","feed_headline":"Continuous spectrum alone rebuilds 16-bit optical pulses","feed_subtitle":"A rank-16 affine map from the reflection coefficient to the initial field reaches ~7.5e-5 relative L2 error on held-out DPSK patterns.","key_machinery":"The central object is the affine map operator $\\mathcal{G}_\\theta(x_j)=x_j A + b$, discretized as real matrices $X=\\big[\\mathrm{Re}(X_c)\\ \\mathrm{Im}(X_c)\\big]$ and $Y$, with parameters fitted by $A_{\\mathrm{aug}}=X_{\\mathrm{aug}}^+ Y$ using the Moore-Penrose pseudo-inverse. The input is the real part of the reflection coefficient $x(\\xi)=\\mathrm{Re}\\,r(0,\\xi)$ sampled on a spectral grid, and the output is the temporal profile of the initial potential $y(t)=u(0,t)$. A truncated singular value decomposition with $r=16$ projects the scattering data into a 16-dimensional subspace before the output map is applied, which is what turns the surrogate into a reduced-order model.","core_discovery":"The central claim is that the map $\\mathcal{G}$ sending the continuous reflection coefficient $r(0,\\xi)$ at the transmitter to the initial potential $u(0,t)$ is, within this pulse family, well approximated by an affine operator with low-rank structure. Numerically, the paper learns the affine map by least squares with a Moore-Penrose pseudo-inverse on 700 training 16-bit RZ-DPSK patterns, and finds that its matrix has a sharp singular-value drop after the sixteenth singular value. Setting the rank to 16, the reduced map reconstructs the 300 held-out initial potentials with average relative $L^2$ error about $7.5\\times10^{-5}$, with the worst validation case around $1.3\\times10^{-3}$. The authors interpret the rank-$16$ signature as matching the information content of a 16-bit input, and they argue that the reflection coefficient implicitly encodes enough of the discrete spectrum that the soliton sector can be neglected for this signal class.","pith_inferences":["If the rank-equals-bits relation holds beyond 16 bits, the affine surrogate would need the projection dimension to grow linearly with sequence length, which would also raise the number of training patterns needed to cover the pattern space; the paper does not test this.","The claim that the continuous spectrum alone suffices is empirical and is not backed by a uniqueness proof; standard inverse scattering theory treats discrete eigenvalues and norming constants as independent data, so a future counterexample with distinct bit patterns sharing nearly identical reflection coefficients would bound the method's validity.","The workflow assumes the affine map learned at $z=0$ transfers to the fiber-output case after linear back-propagation; since the numerical experiments are all conducted at $z=0$, that transfer is an untested extrapolation.","A natural testable extension would be to apply the same affine surrogate to non-return-to-zero or quadrature-phase formats; the rank structure might then reveal how many effective degrees of freedom those formats impose on the inverse map."],"forward_implications":["A receiver could reconstruct a transmitted 16-bit RZ-DPSK pattern directly from the back-propagated reflection coefficient, without solving Gelfand-Levitan-Marchenko or Riemann-Hilbert equations.","Because only the real part of the reflection coefficient is needed, the required scattering-data acquisition and storage are roughly halved.","The rank-16 reduced model needs only a $1024\\times16$ projection and a $16\\times2048$ output map, making real-time or hardware-implemented decoding plausible.","The observed rank-equals-bits relation suggests that the number of significant singular values may act as a proxy for the information dimension of a modulation format."],"supporting_citations":[{"why":"Supplies the numerical direct scattering transform used to compute the reflection coefficient for each generated bit pattern.","marker":"[31]"},{"why":"Establishes the least-squares Moore-Penrose pseudo-inverse formula used to fit the affine map $A_{\\mathrm{aug}}=X_{\\mathrm{aug}}^+Y$.","marker":"[29, 30]"},{"why":"Introduces the Lax pair formalism on which the nonlinear-Schrödinger inverse scattering framework rests.","marker":"[20]"},{"why":"Provides the Zakharov-Shabat Lax pair representation of the NLSE used to define the scattering problem.","marker":"[21]"},{"why":"Shows the linear evolution of scattering data along the fiber, which underwrites the linear back-propagation step in the proposed workflow.","marker":"[22]"}],"fun_headline_variants":["Continuous spectrum alone reconstructs 16-bit optical pulses","Rank-16 affine map from reflection coefficient recovers bits","Soliton-free inverse scattering for high-speed fiber links","Affine map approximation makes inverse scattering practical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the empirical premise that, for these Gaussian 16-bit RZ-DPSK pulse trains, the continuous reflection coefficient (in fact its real part alone) encodes the initial potential completely, so the discrete soliton spectrum can be discarded without loss.","fun_headline_variants_meta":{"raw":{"variants":["Continuous spectrum alone reconstructs 16-bit optical pulses","Rank-16 affine map from reflection coefficient recovers bits","Soliton-free inverse scattering for high-speed fiber links","Affine map approximation makes inverse scattering practical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1192,"prompt_tokens":840,"completion_tokens":352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":456,"tokens_out":352,"duration_ms":4116,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:06:25.124249+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the real part of the reflection coefficient $r(0,\\xi)$ on the same 1024-point grid for all 65,536 possible 16-bit DPSK patterns; if any two distinct patterns yield reflection coefficients whose $\\ell^2$ distance is below the numerical noise floor, the affine map cannot separate them, and the claim that the continuous spectrum alone encodes the bit sequence fails for that pair. Alternatively, propagate a held-out pattern to fiber length $\\ell$, back-propagate its scattering data linearly, and check whether the affine map trained at $z=0$ reconstructs the transmitted field.","supporting_citations":[{"cited_title":"Introducing the fast nonlinear fourier transform","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical direct scattering transform used to compute the reflection coefficient for each generated bit pattern."},{"cited_title":"Exact theory of two-dimensional self-focusing and one-dimensional self-modulation of waves in nonlinear media","cited_arxiv_id":null,"evidence_quote":"Provides the Zakharov-Shabat Lax pair representation of the NLSE used to define the scattering problem."},{"cited_title":"Nonlinear inverse synthesis and eigenvalue division multiplexing in optical fiber channels","cited_arxiv_id":null,"evidence_quote":"Shows the linear evolution of scattering data along the fiber, which underwrites the linear back-propagation step in the proposed workflow."}],"review_version":1}