{"id":"67c162d1-f071-47b8-90cf-ac8176e2830b","arxiv_id":"2508.17440","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A proposal to make one spatial light modulator implement programmable higher-order Ising terms and all-optical KAN nonlinearities, but the core polynomial mechanism is not compatible with linear propagation.","lead":"This paper proposes a single optical add-on that would let existing photonic computers handle higher-order Ising optimization problems and trainable neural layers called Kolmogorov-Arnold networks. The central mechanism, however, is inconsistent with linear optics: the measured intensity of a field that is linear in the input amplitude cannot produce the higher powers the scheme relies on.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central optical claim is inconsistent with linear propagation: a phase-only second pass cannot produce intensity powers of the clique sum S_q above quadratic, so Eq. (11) and the k-local/KAN conclusions are unsupported.","rationale":"The reader's weakest_assumption is the same concern I identify, and I agree with it. The mathematical interpolation of k-spin products (Eq. 9) is correct, and the architecture could plausibly provide quadratic and linear responses; the problem is exclusively in the optical mapping from S_q to measured intensity. This is load-bearing because Eq. (11) is used in the calibration loss (Eq. 24), the closed-form seeds (Eqs. 20-23), and the KAN ridge model (Eq. 32); if only quadratic and linear terms in S_q are physically available, the claimed k-local Ising couplings (k>2) and all-optical KAN ridge polynomials (degree >2) cannot be implemented by the proposed two-pass relay. The numerical simulations do not rescue the claim because they assume Gamma = A theta (Eq. 25) rather than deriving it from scalar diffraction or Maxwell's equations. This is an internal-consistency problem, not a disagreement with external consensus, and it does not depend on platform details; the VCSEL and AOC variants inherit the same field-linear transfer. I therefore keep the reader's REJECT verdict, requiring no change.","tokens_in":20896,"tokens_out":9082,"duration_ms":95571,"concrete_test":"Run a scalar-diffraction simulation of the folded 4f relay for the k=4 worked example: set the first-pass window field to E_in(y) = S_q (e^{i k0 y} + e^{-i k0 y}) for S_q in {-4,-2,0,2,4}, apply the second-pass patch phi_q(y) = alpha cos(2 k0 y) + beta cos(4 k0 y) with |alpha|, |beta| << 1, integrate the output intensity over the window, and fit I_q = sum_r Gamma_r S_q^r. The fitted Gamma_4 and Gamma_3 will be zero to numerical precision, with only Gamma_2 nonzero (plus Gamma_1 and Gamma_0 if a local oscillator is added), directly contradicting Eq. (20). The analytic form of this check is the identity integral |S_q A(y) exp(i phi(y))|^2 dy = |S_q|^2 integral |A(y)|^2 dy, which holds for every phase profile and every relay geometry consistent with the stated linearity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Eq. (11): I_q(S_q; theta_q) = sum_{r parity-matched} Gamma_{r,q}(theta_q) S_q^r, with r up to k. Under the paper's own linear scattering relation (Eq. 5) and the described two-pass relay, this cannot hold for k>2. The field entering window q on the second pass is E_in(y) = S_q A(y) (Sec. II: 'the on-axis complex field in that window is proportional to the coherent sum ... proportional to S_q'). The second pass multiplies by a phase-only mask exp(i phi_q(y; theta_q)); all propagation to the detector is linear. Therefore E_out(y) = S_q A(y) exp(i phi_q(y; theta_q)) plus, at most, an S_q-independent local oscillator. Since |exp(i phi)| = 1, the window-integrated intensity is exactly |S_q|^2 integral |A|^2 dy, plus a homodyne linear term 2 Re[S_q integral A exp(i phi) E_LO^* dy] and an LO-only constant. No S_q^3 or S_q^4 term can appear for any phase profile. The k=4 derivation in Eqs. (18)-(20) shows the error: Eq. (20) assigns the S_q^4 coefficient to alpha^2 beta, where alpha and beta are second-pass phase depths; in the true intensity those depths multiply |S_q|^2, not S_q^4. The numerical experiments (Figs. 2, 3, 6) presuppose the contested map Gamma = A theta (Eq. 25) and optimize within it, so they test linear algebra rather than the optical mechanism. Structural nonlinearity in theta (Ref. [39]) does not create nonlinearity in the input amplitude S_q: for fixed theta, the output field is affine in the input, so detected intensity is at most quadratic in S_q. The proposed hardware therefore realizes only |S_q|^2 (plus S_q with a local oscillator), and cannot natively program k-local couplings or KAN ridge polynomials of degree greater than two.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-pass folded 4f relay in which a second programmable phase-only pass through the same SLM is used to convert nominally linear optical propagation into per-window polynomial nonlinearities of the clique sum S_q or projection amplitude z. The authors claim this delivers native k-local Ising couplings and all-optical KAN ridge functions on SPIM, VCSEL, and AOC platforms, with M(k)=ceil(k/2) phase depths per window and two-frame physical gradients for training. The mathematical interpolation of a k-spin product by a parity-matched univariate polynomial in S_q is correct, and the numerical simulations are clearly described, but the physical mechanism underlying Eq. (11) is the core question.","tokens_in":21299,"tokens_out":4473,"duration_ms":46573,"significance":"If the central claim were correct, the paper would unify two important photonic computing tasks in a simple architecture, which is a genuinely significant prospect. The paper also makes its methods reproducible: the interpolation construction, the pseudo-code for calibration (poke tests, diagonal calibration, closed-form seeds), and the gradient-descent simulations are concrete and easy to follow. However, the significance is conditional on the existence of an optical mechanism producing intensity terms of degree higher than two in the input amplitude. As detailed in the major comments, the described linear-optical relay cannot produce such terms, and the numerical experiments assume the contested map rather than validating it. The real contribution is therefore a correct algebraic reduction plus a learning algorithm for a hypothetical device, not a supported optical primitive.","major_comments":[{"comment":"The load-bearing assertion is Eq. (11), which states that the window-averaged intensity admits a polynomial expansion in S_q of degree up to k. This is inconsistent with the paper's own linear-optics model. For fixed second-pass phases θ_q, the optical system described in Section II is linear, so the field at the detector is affine in the input amplitude: E_out(y) = S_q A(y;θ_q) + E_LO (at most adding a local oscillator). The detected intensity is then |E_out|^2 = |S_q|^2 ∫|A|^2 dy + 2 Re[S_q ∫ A E_LO^* dy] + const, which contains only powers 0, 1, and 2 of S_q. No mechanism in Section II generates S_q^3 or S_q^4. The structural nonlinearity of Eq. (5) is a nonlinear dependence of the output on the programmable parameters θ, not on the input amplitude; for fixed θ the output remains linear (or affine) in the input, so the intensity is at most quadratic in S_q. This invalidates Eq. (11) for k>2 and with it the central claims of k-local Ising couplings and KAN ridge nonlinearities.","section":"Section II, Eq. (11)"},{"comment":"The worked k=4 example makes the error concrete. The second-pass field is E_in(y) ∝ S_q(e^{ik0 y}+e^{-ik0 y}) multiplied by exp(iφ(y)); the field amplitude returning to the ±k0 spatial order is proportional to S_q times a function of α and β through Bessel functions. The paper computes terms 2J1(α)^2, 3J1(β)^2, and 2J2(α)J1(β), which are coefficients of the spatial Fourier decomposition of the output field. The intensity at the detector is proportional to |S_q|^2 times |c(θ)|^2, so those coefficients multiply |S_q|^2, not S_q^4. Equation (20) therefore incorrectly assigns the quartic coefficient to α^2β; physically, the α^2β term appears in a function of the phase depths multiplying the quadratic intensity |S_q|^2. The same conflation of field-coefficient order with input-amplitude polynomial order appears throughout the calibration section.","section":"Section II, Eqs. (18)-(20)"},{"comment":"The numerical experiments do not validate the optical mechanism because they presuppose the contested map Γ = Aθ (Eq. (25)) from phase depths to polynomial coefficients of S_q or z. In Fig. 2 the device map is drawn as a random lower-triangular matrix A and the loss is ||wAθ - c||^2, which optimizes within the assumed linear relationship. In the teacher-student KAN experiment, the student model is defined through the same calibrated linearization Γ_m(θ_m)=A_m θ_m (Eq. (37)), so the experiment tests whether gradient descent can fit coefficients in the assumed model; it cannot confirm the existence of an optical device whose intensity is a polynomial of degree higher than 2 in the input amplitude. The two-frame FD proxy (Eq. (41)) likewise measures gradients of the assumed model. Thus Figs. 2, 3, and 6 demonstrate a linear-algebra learning procedure, not the proposed optical primitive.","section":"Section III, Eq. (32) and Fig. 6; Section II, Eqs. (24)-(27)"}],"minor_comments":[{"comment":"The abstract says that 'measured intensity in each window becomes a freely programmable polynomial of the clique sum or projection amplitude'; this is precisely what is unsupported. A clarifying sentence stating that the nonlinearity is only in the programmable parameters θ, not in the input amplitude, would have made the fundamental issue evident to the reader.","section":"Abstract and Section II"},{"comment":"The homodyne expression is correct as written for a local oscillator, but the accompanying text states E_sig ≈ α_q S_q + O(S_q^3); for a linear optical system the field is exactly linear in S_q (no higher-order terms), so the O(S_q^3) term should be removed. This is a direct consequence of the same linearity issue.","section":"Section II, Eq. (14)"},{"comment":"The paper would benefit from stating explicitly that no experimental data are presented; the claim rests entirely on simulations of the assumed model. This is not a problem per se, but it underscores that the physical validity of Eq. (11) needs independent scrutiny.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central flaw is unambiguous: a phase-only second pass in a linear optical system cannot generate intensity terms of degree higher than two in the input field amplitude. This is not a matter of experimental detail but of the fundamental input-output relation of the proposed device. The reader's stress-test analysis correctly identifies the issue, and I concur with the reject verdict. The mathematical interpolation and the optimization simulations are fine as abstract exercises, but they do not constitute a valid physical mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the per-window, per-clique addressing of structural nonlinearity is a genuinely new architectural idea, and the paper presents it clearly. But the load-bearing optical claim doesn't hold up, and the numerical experiments never test it.\n\nEq. (11) asserts that the window-averaged intensity after the two-pass relay is a parity-matched polynomial in the clique sum S_q of degree up to k. The paper's own linear scattering setup (Eq. 5) says the output field is linear in the input amplitude. The field entering the second-pass patch is proportional to S_q times a fixed spatial profile; the patch is phase-only; everything downstream is linear. So for any fixed mask, the detected intensity is exactly |S_q|^2 times a theta-dependent gain, plus at most a homodyne linear term in S_q and a constant. No S_q^3 or S_q^4 term can appear for any phase profile. The k=4 worked example shows the error: the alpha^2 beta contribution the authors assign to S_q^4 is, in the true intensity, just another coefficient multiplying S_q^2. The scheme has the known structural nonlinearity in the control phases theta (Refs. [39,42]); the paper mistakes that for nonlinearity in the input amplitude S_q. The LO added for odd parity buys a linear term in S_q; nothing buys a cubic. So the primitive realizes at most quadratic per-window responses, which is the status quo for SPIMs, not native k>2 or KAN ridge polynomials.\n\nCredit where earned: the interpolation reduction of a k-spin product to a univariate polynomial of the clique sum is correct, including the k=4 coefficients and the M(k)=ceil(k/2) parameter count. The per-window architecture is a real departure from single-global-mode (Wanjura-Marquardt) and whole-aperture (nPOLO) structural nonlinearity. The writing is careful, and the calibration sections — poke tests, closed-form seeds, two-frame adjoint training — are concrete. The citation pattern is solid.\n\nThe soft spots trace to the same root error. Figures 2, 3, and 6 all presuppose the map Gamma = A theta that the paper needs to establish; they show gradient descent can fit coefficients inside an assumed linearized model, which is linear algebra, not optical validation. The teacher-student KAN test has the same circularity. The paper itself flags the even-only symmetry of phase-only detection; the asserted Esig = alpha S_q + O(S_q^3) is exactly the term linear optics cannot generate.\n\nFor the photonic Ising / optical KAN crowd: the polynomial-reduction formulation is worth reading, and the per-window idea is worth knowing, but the proposed hardware cannot deliver the headline capability. If this is submitted, I'd send it to review — the error is subtle enough and the area active enough to justify a referee report — but the expected outcome is rejection or a major rewrite confined to per-window quadratic responses.","headline":"A novel per-window structural-nonlinearity architecture, clearly presented, but the central optical claim fails: a phase-only second pass makes window intensity at most quadratic in the clique sum, so the k>2 Ising and KAN-polynomial conclusions are unsupported.","tokens_in":21865,"tokens_out":18064,"would_cite":false,"duration_ms":170925,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a folded 4f relay with a second programmable pass through the same spatial light modulator turns each Fourier-plane window into a freely programmable polynomial of the spin-sum or projection amplitude, thereby…","keywords":["k-local Ising machine","structural nonlinearity","folded 4f relay","Kolmogorov-Arnold network","spatial light modulator","spatial photonic Ising machine","two-frame physical gradients","analog optical computer"],"falsifier":"For one k=4 clique with both second-pass harmonics enabled and no local oscillator, record the window-integrated intensity for each admissible value S_q in {-4,-2,0,2,4} while all other parameters are fixed; if the dependence is exactly constant plus quadratic in S_q, Eq. (11) with r=4 is not realized, whereas a quartic component would support the polynomial mechanism.","tokens_in":20634,"feed_emoji":"🔁","tokens_out":7198,"duration_ms":76765,"temperature":0.7,"pith_summary":"Two open problems in photonic computing—native k-local Ising couplings and the many independent nonlinearities needed by Kolmogorov-Arnold networks—are addressed with one optical primitive: a folded 4f relay that sends the Fourier-plane field back through a second programmable pass on the same spatial light modulator. The paper's central claim is that each disjoint window carries a clique sum or a linear projection, and that a small-depth, parity-matched phase patch on that window makes the measured intensity a freely programmable polynomial of the amplitude, of degree up to k. If true, this would let spatial-photonic Ising machines, VCSEL arrays, and analog optical computers embed k-body Ising terms without nonlinear media, and train all-optical KAN ridge functions with two-frame physical gradients. The construction is developed through a worked k=4 example, closed-form seeds, a locally lower-triangular calibration map, and numerical calibration-refinement simulations for clique orders 3 to 15.","feed_headline":"Folded 4f relay turns SLM windows into programmable polynomials","feed_subtitle":"A second pass through the same SLM is claimed to deliver k-local Ising couplings and KAN ridge functions on one platform.","key_machinery":"Folded 4f relay with second-pass phase patches. Lens L1 Fourier-transforms the spin-encoded field so the k macropixels of a clique overlap in one window whose on-axis amplitude is proportional to S_q; lens L2 re-images that plane onto a dedicated patch of the same SLM carrying a small-depth cosine phase profile φ_q(y;θ_q)=Σ θ cos(n k_0 y). The paper's argument is that the small-phase expansion of exp(iφ) sends products of these harmonics back to the ±k_0 bank, that only DC terms survive window averaging, and that parity-matched harmonics n=2m−δ with δ=k mod 2 select even or odd powers of S_q. The map from phase depths to polynomial coefficients is locally lower-triangular, so M(k)=ceil(k/2) depths suffice, and a two-frame adjoint protocol supplies in-situ physical gradients for calibration and refinement.","core_discovery":"The paper proposes that every k-spin product over a clique can be replaced by a parity-restricted univariate polynomial of the clique sum S_q, and that this polynomial can be manufactured optically as the measured intensity of a window after a second pass through the same SLM. Formally, Eq. (11) asserts I_q(S_q; θ_q) = Σ_{r≡k mod 2} Γ_{r,q}(θ_q) S_q^r, with M(k)=ceil(k/2) phase depths per window controlling all non-constant coefficients. The same per-window response is then reused as a KAN ridge function: holographic fan-out computes projections z_{j,m}=w^T_{j,m}x, the relay applies an independent polynomial Φ_{j,m}(z), and banked optical summation forms the output. The paper supports this with a closed-form k=4 seed from Cardano's formula, a locally lower-triangular Jacobian argument, and simulations of gradient-refinement training, and it sketches identical insertion into VCSEL and AOC feedback paths.","pith_inferences":["A direct experimental test of Eq. (11) would measure I_q(S_q) at all admissible S_q for one k=4 window; if only S_q^2 and a constant appear without a local oscillator, the polynomial-degree claim is not realized by the linear relay as written.","The windowed-addressing idea is separable from the degree claim: any per-window nonlinear element, such as a saturable absorber or electro-optic modulator, could slot into the same relay and supply true higher-order terms, making the relay's durable contribution the per-window programmability rather than the polynomial order.","Parity-matched harmonics imply a hardware ceiling: the maximum representable degree k is bounded by the number of grating periods that fit across a patch, so scaling to very high k will demand either finer SLM pixels, more bounces, or segmentation.","The bank-wise local-oscillator scheme sets one odd-order coefficient per bank; independent odd coefficients for every ridge would require per-window local-oscillator phase control, which the paper leaves as an engineering target."],"forward_implications":["Any k-body Ising term can be represented as a parity-restricted polynomial of the clique sum and embedded optically without χ^(2) or Kerr media.","The same relay provides per-window, independently programmable univariate polynomials, giving all-optical KAN layers whose ridge functions are trainable with two optical frames per update.","The hardware insertion is one lens and a fold, or an on-chip 4f loop, and the recipe carries over from free-space SPIMs to injection-locked VCSEL arrays and to the analog optical computer.","Quadratization is bypassed: higher-order Ising formulations keep their native variable count instead of expanding into ancilla spins, in line with reported reductions in required variables.","Numerical calibration-refinement on clique orders k=3 to 15 reaches relative coefficient errors around 10^-3 to 10^-4 within a few hundred gradient steps, with weak dependence on k."],"supporting_citations":[{"why":"Supplies the structural-nonlinearity mechanism and the two-frame adjoint physical-gradient protocol that the paper adopts.","marker":"[39]"},{"why":"Establishes the spatial-photonic Ising machine platform with large-scale spin encoding on an SLM, the baseline hardware for the relay insertion.","marker":"[8]"},{"why":"Demonstrates focal-plane division for fully programmable SPIMs, the direct predecessor of the per-window separation used here.","marker":"[34]"},{"why":"Shows multi-bounce structural nonlinearity (nPOLO) in free space, which the paper extends from a shared global activation to per-window local addressing.","marker":"[42]"},{"why":"Defines the KAN ridge-function architecture that the optical KAN implementation targets.","marker":"[30]"},{"why":"Provides the prior photonic KAN implementation whose claimed energy and latency advantages motivate an all-optical KAN.","marker":"[31]"},{"why":"Supports the paper's motivation that native higher-order Ising interactions reduce required variables compared with quadratization.","marker":"[26]"},{"why":"Documents efficiency gains of higher-order Ising machines, the algorithmic rationale for native k-local hardware.","marker":"[27]"}],"fun_headline_variants":["Folded 4f relay makes SLM windows compute polynomials","Two passes through one SLM unify Ising and KAN","One lens and a fold: same SLM does Ising and KAN","Programmable polynomials from SLM second pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each window's measured intensity is a polynomial of degree up to k in the summed spin amplitude; but the optics described are linear in that amplitude, so intensity detection yields at most a quadratic dependence, and nothing in the equations produces cubic or quartic powers.","fun_headline_variants_meta":{"raw":{"variants":["Folded 4f relay makes SLM windows compute polynomials","Two passes through one SLM unify Ising and KAN","One lens and a fold: same SLM does Ising and KAN","Programmable polynomials from SLM second pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1688,"prompt_tokens":1074,"completion_tokens":614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":690,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":690,"tokens_out":614,"duration_ms":6281,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:05:20.319554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For one k=4 clique with both second-pass harmonics enabled and no local oscillator, record the window-integrated intensity for each admissible value S_q in {-4,-2,0,2,4} while all other parameters are fixed; if the dependence is exactly constant plus quadratic in S_q, Eq. (11) with r=4 is not realized, whereas a quartic component would support the polynomial mechanism.","supporting_citations":[{"cited_title":"Pierangeli, G","cited_arxiv_id":null,"evidence_quote":"Establishes the spatial-photonic Ising machine platform with large-scale spin encoding on an SLM, the baseline hardware for the relay insertion."},{"cited_title":"Veraldi, D","cited_arxiv_id":null,"evidence_quote":"Demonstrates focal-plane division for fully programmable SPIMs, the direct predecessor of the per-window separation used here."},{"cited_title":"Yildirim, N","cited_arxiv_id":null,"evidence_quote":"Shows multi-bounce structural nonlinearity (nPOLO) in free space, which the paper extends from a shared global activation to per-window local addressing."},{"cited_title":"Bybee, D","cited_arxiv_id":null,"evidence_quote":"Documents efficiency gains of higher-order Ising machines, the algorithmic rationale for native k-local hardware."}],"review_version":2}