{"id":"c078fc33-7266-45e3-a6e7-b36b25e28f13","arxiv_id":"2504.20401","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Encoding data in the position of an optical source turns a linear optical medium into a nonlinear input-output map, enabling trainable optical classifiers.","lead":"This paper shows that a linear optical system can perform nonlinear computation when the input is encoded in the position of a light source, because the field at a fixed detector is a nonlinear function of the source location. The authors optimize the material layout so the position-to-field map acts as a classifier, and they demonstrate the idea on MNIST and Fashion MNIST.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Intensity readout is an unisolated nonlinearity: without an amplitude-encoding control, the reported gains cannot be attributed to source-position encoding.","rationale":"The reader's weakest assumption concerns expressivity: that the nonlinearity is elementwise, so tasks requiring feature interactions cannot be handled. That concern is real if the measured quantity is the complex field, but the system actually measures intensity, and |E|^2 contains cross terms between the active sources in different rows. The paper's own limitation statement in Section 4.1 thus does not describe the implemented model, and the empirical gap versus the MLP cannot be cleanly attributed to an elementwise cap. The deeper issue is attribution: the experiment shows that a full system (position encoding + intensity readout + optimized material + linear decoder) beats a linear model, but it does not isolate the position encoding. Because the intensity readout is itself nonlinear, a control with amplitude-only encoding is needed. If that control also beats the linear model, the central claim that source-position encoding is the operative nonlinear mechanism would be unsupported. This is a correctness risk, not a novelty dispute; the mechanism in Eq. (8) is mathematically valid, but the empirical demonstration is confounded. The missing control is easily added to the existing simulation framework, so the appropriate verdict remains CONDITIONAL: accept only after the control is run and reported. I therefore leave the reader's verdict unchanged, while noting the specific experimental gap.","tokens_in":37227,"tokens_out":12552,"duration_ms":138944,"concrete_test":"Run the same topology-optimization pipeline (same FDFD grid, same neural-field or B-spline parameterization, same optimizer hyperparameters, same intensity readout and linear decoder) but encode the input in source amplitudes rather than positions: for each of the p rows, keep one source at a fixed position and set its complex amplitude to the quantized input value (or a fixed linear function of it). Compare final test accuracy on the Table 1 benchmarks. If the amplitude-encoded network matches the position-encoded accuracy (within one point), the reported gains do not require position encoding; if it is substantially worse (e.g., >5 points), the position-encoding mechanism is supported. Additionally, run both encodings with an unoptimized (random) topology and with the decoder trained alone, to bound the contribution of the optimized material.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim is that the position encoding, not some other nonlinearity, is responsible for the gains over the linear baseline. The model in Eq. (11) is f = D R(A^{-1} s(x)), where R is the intensity readout, R(E)_k = |E(z_k)|^2. This is quadratic in the field and therefore quadratic in the source vector b = s(x). With one active source per input row, |E|^2 contains cross terms between different rows, i.e., products of features from different input dimensions, so the map is not elementwise as Section 4.1 states; the stated limitation is inconsistent with the actual readout. More importantly, the comparison against a linear model conflates three nonlinear mechanisms: the one-hot quantized position encoding, the intensity readout, and the optimized material distribution. A system with fixed sources and amplitude-modulated inputs (b_i proportional to quantized x_i) would also be nonlinear through the intensity readout and might beat the linear baseline. The paper certifies only that the decoder is linear (Section 2.2.4), not that the readout is; the assertion that the nonlinearity is 'a result of the source position encoding' is therefore unsupported by the experiments. The reported accuracy gap at high precision (86.7% vs 96.9% on 64-bit MNIST) is also attributed by the authors to expressivity limits, but if the readout creates pairwise interactions, that interpretation is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a method for performing nonlinear computation in fully linear optical media by encoding input data in the spatial position of optical sources. The authors derive a modal expansion showing that the field at a fixed measurement point is a nonlinear function of source position (Eq. 8), and use topology optimization with a differentiable finite-difference frequency-domain solver and Fiber Monte Carlo shape derivatives to design material distributions that map the source position to a desired classifier output. The system is evaluated on MNIST and Fashion-MNIST with PCA-projected, quantized inputs at several bit precisions; the optical network is reported to outperform a linear baseline and to approach an MLP at low precision. The paper also discusses limitations, including the claim that the nonlinearity is elementwise.","tokens_in":37535,"tokens_out":8166,"duration_ms":83003,"significance":"If the attribution to source-position encoding is upheld, the work is a significant contribution to the field of optical neural networks: it offers a low-power nonlinearity mechanism that does not rely on materials-based nonlinearities, and it integrates a full inverse-design pipeline (differentiable full-wave simulation, topology optimization, and decoder co-design) that is of independent interest. The modal derivation leading to Eq. (8) is a clear strength, and the use of Fiber Monte Carlo for exact-Heaviside topology differentiation is technically interesting. However, the experimental validation as reported does not isolate the proposed mechanism from the square-law intensity readout, and the stated elementwise limitation is inconsistent with the implemented readout. The central idea is plausible, but the evidence base needs strengthening before the claims can be accepted.","major_comments":[{"comment":"The intensity readout R(E) = |E|^2 defined in §2.2.3 is quadratic in the field and therefore in the source vector b = s(x); since the fields from the p active sources (one per row) superpose before detection, R contains cross terms between different rows. This means the aggregate map f in Eq. (11) is not an elementwise nonlinearity of the input coordinates, and the certification in §2.2.4 that the decoder is linear does not isolate the source-position encoding as the source of the nonlinearity. The comparison against the linear baseline conflates at least three mechanisms: the one-hot quantized position encoding, the square-law readout, and the optimized material distribution. A system with fixed sources and amplitude-modulated inputs would also be nonlinear through the same readout, so the reported gains cannot be attributed to source-position encoding without a control experiment (e.g., amplitude encoding with fixed positions, or a field-amplitude readout). Please add such a control or otherwise analyze the separate contributions.","section":"§2.2.3–2.2.4, Eq. (11)"},{"comment":"The stated limitation that 'the nonlinearities hold only elementwise with respect to the data' is contradicted by the actual system because the intensity readout creates products of contributions from different rows of the source array, i.e., nonlinear mixtures of variables from different input dimensions. Consequently, the explanation of the high-precision gap in Table 1 (MNIST 64-bit: 86.7% vs 96.9%) as a capacity limit of elementwise nonlinearities is not supported by the implemented model. The authors should either correct the limitation statement to reflect the readout-induced cross terms, or analyze the functional form of the map and re-interpret the accuracy gap.","section":"§4.1"},{"comment":"The experimental results are reported as single runs without error bars, multiple random initializations, or statistical tests. At the smaller margins (e.g., MNIST 64-bit, +2.2% relative improvement), the claimed 'significant improvements' over the linear baseline are not statistically substantiated. Please report the variance across multiple optimization runs and seeds and, if appropriate, a significance test.","section":"§3, Table 1"}],"minor_comments":[{"comment":"The inner product in Eq. (4) includes the weight ε_r(z), so the modal coefficient for a point source should read c_i = -j ω_s/(ω_i^2 - ω_s^2) ε_r(z_s) E_i^*(z_s); the ε_r(z_s) factor is missing in Eq. (8). If the source array is placed in a region of constant ε_r the factor is a global constant, but the general formula should be corrected or the constant-ε_r assumption stated.","section":"§2.1, Eq. (8)"},{"comment":"Dirichlet boundary conditions correspond to a perfect electric conductor enclosing the domain; this is a strong idealization for an open scattering device and should be mentioned as an approximation in the text, since it may introduce cavity resonances that are not present in a free-space device.","section":"§2.1, Eqs. (1)–(2)"},{"comment":"Section 2.2.1 states 'log(k) bits of precision' while Section 3.1 correctly writes 'log2 q'; please use consistent notation for the number of quantization bits.","section":"§2.2.1 and §3.1"},{"comment":"The parenthetical values are labeled as relative percentage improvement in the caption, but the text does not define how the percentages are computed; please provide the definition in the caption or text.","section":"Table 1 caption"},{"comment":"The decoder box labels 'ω_d → R^d → (k→k)' are unclear; clarifying the map from the d-dimensional input to the K readout channels would improve readability.","section":"Figure 2"},{"comment":"In the notation list, the units for the electric current density J are given as W/m^2, which is a power density; the units should be A/m^2.","section":"Appendix A.2"},{"comment":"The cross-reference for the neural implicit field parameterization is missing a target: the text says '(cref appendix)' without an actual reference; please fix the cross-reference.","section":"§2.3.1 and Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's optics/computing scope and the underlying idea is attractive. The main risk is the attribution of the nonlinearity to source-position encoding: the square-law readout is an unisolated nonlinear mechanism, and the paper's own limitation statement is internally inconsistent with the implemented system. I would ask the authors to add a control experiment or a careful analytic decomposition before publication. Also consider requesting code/data release to strengthen reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for its core idea: encoding data in the source position to obtain a nonlinearity from a linear optical medium is new, and the modal derivation is clean. The problem is that the experiments don't actually isolate that nonlinearity from the quadratic intensity readout, so the headline claim is under-supported.\n\nWhat's good: Equation (8) is the load-bearing result, and it's correct in spirit—the field at a fixed detector is a mode-weighted sum in which the source position enters through the mode amplitudes, which can be made complicated. The differentiable FDFD solver plus Fiber Monte Carlo topology optimization is a real engineering contribution, and comparing against a linear model trained on the same quantized data is more fair than most baselines in this area. The discussion of prior work (Wanjura-Marquardt, Xia, Li, Yildirim) is accurate, and the distinctions are genuine.\n\nThe soft spot is the attribution. The readout R(E)=|E|^2 is itself nonlinear. With one active source per row, the intensity at a detector contains cross terms between different rows of the source array, so the output is not elementwise in the input coordinates; the Section 4.1 claim that \"the nonlinearities hold only elementwise\" is incorrect as written. More importantly, the paper never runs the control that would separate the two mechanisms: fix the sources, encode the data in drive amplitudes (a linear encoding), and keep the same intensity readout. That system is also nonlinear and might beat the linear baseline. Without it, the reported gains cannot be cleanly assigned to source-position encoding. The paper only certifies that the decoder is linear, not the readout.\n\nThere are also standard reproducibility gaps: no error bars, no code, and the main table reports only the neural-field parameterization. The high-precision gap versus the MLP is real, but the interpretation as an expressivity limit is not established if the readout creates pairwise interactions.\n\nOverall, the idea is worth careful refereeing, but the experimental section needs major revision before the central claim is credible. I'd send it to review with the expectation of a revision that adds an amplitude-modulation control, multiple seeds, and a corrected limitation paragraph.","headline":"Source-position encoding is a genuinely new idea with a sound modal derivation, but the experimental section never isolates it from the quadratic intensity readout, so the central attribution is not yet supported.","tokens_in":38018,"tokens_out":5025,"would_cite":true,"duration_ms":53188,"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":"This paper claims that encoding data in the spatial position of a light source makes a fully linear optical medium compute nonlinearly, and that topology-optimized material designs built on this encoding outperform linear models and rival…","keywords":["source-position encoding","nonlinear optics","linear media","topology optimization","optical neural networks","differentiable simulation","inverse design","modal expansion"],"falsifier":"Simulate the fully optimized optical classifier on a two-input binary XOR task using the same p=2, q=2 source-position encoding, and compare accuracies of the optical design, a linear model (chance), and a single-hidden-layer MLP (near 100%). If the optical design does not exceed chance, the source-position nonlinearity cannot produce the nonlinear feature interactions that the paper acknowledges are missing; this directly tests whether the elementwise ceiling is real or softened by the intensity readout's cross terms.","tokens_in":37034,"feed_emoji":"💡","tokens_out":10392,"duration_ms":106170,"temperature":0.7,"pith_summary":"This paper proposes a way to perform nonlinear computation in an optical system made entirely of linear materials. The idea is to encode the input data in the spatial position of a light source rather than in its amplitude: the field measured at a fixed detector is a nonlinear function of the source's position, because it is a superposition of the medium's eigenmodes evaluated at that position. The paper shows this relationship explicitly, and introduces a topology-optimization framework that designs the material distribution so that the source-position-to-field map approximates a target classifier. On quantized MNIST and Fashion-MNIST, the resulting simulated optical network beats a linear model by a wide margin and matches a standard neural network, though it falls behind at high input precision.","feed_headline":"Moving the light source turns linear optics into a nonlinear classifier.","feed_subtitle":"Encoding data in source position lets passive optical hardware rival neural networks on MNIST and Fashion-MNIST.","key_machinery":"The load-bearing mechanism is the source-position-to-field modal expansion, Equation (8): $E(z_R) = \\sum_{i} \\frac{-j\\omega_s}{\\omega_i^2-\\omega_s^2} E_i^*(z_s)\\,E_i(z_R)$. The term $E_i^*(z_s)$ is the source-position-dependent mode amplitude, and in a heterogeneous engineered medium the eigenmodes $E_i$ can be complex and shaped by the material distribution, which makes the transfer function from source position to measured field highly nonlinear and tunable. The companion mechanism is an end-to-end differentiable design pipeline: a finite-difference frequency-domain (FDFD) discretization of Maxwell's equations, a level-set topology optimization parameterization (B-spline, triangular mesh, or neural implicit field), and Fiber Monte Carlo to differentiate through the level-set boundary. This allows co-optimization of the material distribution and the linear decoder.","core_discovery":"The central claim is that the map from source position to measured field is nonlinear even when the optical medium is entirely linear, because the field at a fixed readout point is a weighted sum of the medium's eigenmodes, and the weights depend on the source position through the mode functions. Equation (8) expresses this as $E(z_R) = \\sum_{i} \\frac{-j\\omega_s}{\\omega_i^2-\\omega_s^2} E_i^*(z_s)\\,E_i(z_R)$. Since the eigenmodes $E_i$ are generally not affine functions of position, varying the source position $z_s$ produces a nonlinear response at a fixed detector. The paper claims that by optimizing the spatial distribution of permittivity and permeability through topology optimization, one can shape these mode functions so that the source-position-to-field map approximates a desired transfer function, such as a classifier. The full system encodes each input coordinate by activating one source in a row of a source array, propagates the field through the engineered medium, reads intensities at fixed locations, and applies a linear decoder before a softmax. The paper reports that on the evaluated tasks this optical classifier significantly improves over a linear model and is competitive with a standard multilayer perceptron.","pith_inferences":["The intensity readout's squared-magnitude operation creates pairwise cross terms between input coordinates, so the paper's stated elementwise limitation is likely softened: the effective function class is quadratic in the per-coordinate features, which can represent interactions such as XOR that a purely elementwise nonlinearity cannot.","Because Equation (8) also depends on source frequency through $\\omega_s$, frequency could be used as a second encoding axis alongside position, potentially increasing capacity without additional spatial hardware; the paper flags this as future work.","The topology-optimization pipeline built on Fiber Monte Carlo is not optics-specific and could be applied to design other physical computing elements, such as acoustic or microwave classifiers, wherever the forward physics is linear and inputs can be position-coded.","A sharper baseline than the numerical linear model would be an optical linear network using the same source array but a random or freespace material distribution, which would isolate the contribution of the optimized topology to the classification gain."],"forward_implications":["Optical neural networks could be built from passive, linear materials with no optoelectronic conversion, because the nonlinearity arises from position encoding rather than material response.","The design methodology produces extremely task-specific hardware: the material distribution is optimized for a single function and is not reused across tasks.","The approach requires only the ability to drive a source at a data-dependent position, so it works at low power and should generalize to other linear wave systems such as microwaves and acoustics.","The performance gap between the optical network and a standard neural network widens as input precision increases (from +15.2% over linear at 2-bit to +2.2% at 64-bit on MNIST), indicating that the encoding precision is a limiting factor."],"supporting_citations":[{"why":"Derives the mode-excitation coefficient formula used in Equation (8), the foundation of the source-position nonlinearity.","marker":"(Gilbert, 1971; Ulrich et al., 2022)"},{"why":"Supplies the level-set topology optimization formulation used to parameterize and optimize the material distribution.","marker":"(Van Dijk et al., 2013)"},{"why":"Provides Fiber Monte Carlo, the differentiable estimator of area-of-intersection that enables exact level-set differentiation during optimization.","marker":"(Richardson et al., 2024)"},{"why":"Establishes the prior approach of inverse-designing a nanophotonic neural network with topology optimization, serving as the direct design baseline.","marker":"(Khoram et al., 2019)"},{"why":"Demonstrates nonlinear computation via linear wave scattering, the closest competing method that the present work contrasts with.","marker":"(Wanjura & Marquardt, 2024)"},{"why":"Provides adjoint and forward-mode differentiation of Maxwell's equations, underpinning the differentiable field solver used for co-design.","marker":"(Hughes et al., 2018b; 2019)"}],"fun_headline_variants":["Source position encoding makes linear optics nonlinear","Linearity bypassed: nonlinear optics via source placement","Nonlinearity from linear media by moving the source","Data-dependent source position yields nonlinear optical computing","Linear optics, nonlinear computation via source encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that an elementwise nonlinearity applied independently to each input coordinate, followed by a linear readout, is expressive enough for the target classification tasks.","fun_headline_variants_meta":{"raw":{"variants":["Source position encoding makes linear optics nonlinear","Linearity bypassed: nonlinear optics via source placement","Nonlinearity from linear media by moving the source","Data-dependent source position yields nonlinear optical computing","Linear optics, nonlinear computation via source encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1278,"prompt_tokens":918,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":534,"tokens_out":360,"duration_ms":3832,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:30:29.968441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the fully optimized optical classifier on a two-input binary XOR task using the same p=2, q=2 source-position encoding, and compare accuracies of the optical design, a linear model (chance), and a single-hidden-layer MLP (near 100%). If the optical design does not exceed chance, the source-position nonlinearity cannot produce the nonlinear feature interactions that the paper acknowledges are missing; this directly tests whether the elementwise ceiling is real or softened by the intensity readout's cross terms.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides adjoint and forward-mode differentiation of Maxwell's equations, underpinning the differentiable field solver used for co-design."}],"review_version":1}