{"id":"69259ffe-2e81-4334-8e80-9b1a43fbdbe9","arxiv_id":"2505.24032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A supervised learning method linearizes the phase-dependence of a linear optical interferometer, so its unitary response can be modeled and programmed by least squares instead of non-convex optimization.","lead":"The paper trains a linear model of a reconfigurable optical interferometer using random phase settings and measured unitary matrices. Once trained, the model predicts the phase shifts needed to program the device for a desired unitary, without an analytic circuit decomposition.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The constant pseudoinverse F asserted in Eqs. (9)-(11) is not constant across matrix elements because the features in Eq. (3) include boundary phases; as written, the training formula is inconsistent.","rationale":"The reader's conditional verdict is appropriate: the core linearization idea is sound, but the paper contains a concrete internal inconsistency in the derivation of the training step. I focused on the constant-pseudoinverse claim because it directly affects whether the presented algorithm, as specified, produces the correct model. The feature matrix in Eq. (3) explicitly includes boundary phases φ_i^(L) and φ_j^(1), so the design matrices for different output elements are related by diagonal scaling; their Moore-Penrose inverses are generally different. Equation (11) therefore does not follow from Eq. (10). This is not a matter of loose notation: using a single F for all (i,j) minimizes the wrong objective, so a literal implementation of the paper's formulas would not recover the interferometer model. The sample-complexity mismatch (M_min = N^L in the figures versus N^{L-2} features per output element) reinforces the ambiguity. At the same time, the error is repairable: one can solve the least-squares problem separately for each (i,j), or redefine the feature set to include the boundary phases and use one large design matrix with N^L columns. The numerical results suggest the authors may have implemented a corrected variant, but without code or an explicit correction the paper cannot be accepted as written. I do not see grounds to reject the underlying approach, so the verdict remains conditional: revision should clarify the feature indexing, correct Eq. (11), and provide code or a precise description of the implemented estimator.","tokens_in":11451,"tokens_out":13950,"duration_ms":142572,"concrete_test":"Re-implement the training step as written: for N=4, L=5, generate random Haar basis matrices U_l and random training phases; build the per-output design matrices Θ_{ij} from Eq. (3). Compute the pseudoinverse once for a reference output (e.g., i=j=1) and apply it via Eq. (11) to estimate all w_{i\\bar{k}j}; then measure the Frobenius prediction error on a held-out phase set. Compare against solving the least-squares problem separately for each (i,j). If the single-F model fails to reach the noise floor while the per-output solution succeeds, Eq. (11) is invalid as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eqs. (2)-(3) the feature for output element (i,j) is θ_{i\\bar{k}j}^{(m)} = φ_i^{(L,m)} φ_{k_1}^{(L-1,m)} ... φ_j^{(1,m)}. Thus for fixed (i,j) the design matrix is Θ_{ij} = D_{ij} Θ_0 with D_{ij} = diag(φ_i^{(L,m)} φ_j^{(1,m)}) and Θ_0 (M × N^{L-2}) common to all outputs. The Moore-Penrose inverse in Eq. (9) is then F_{ij} = (Θ_0^* D_{ij}^2 Θ_0)^{-1} Θ_0^* D_{ij}, which depends on (i,j) unless the boundary products are equal for all training samples. Hence the assertion in Section III A that F remains constant for each index i and j is false, and Eq. (11) — using one F_{\\bar{k} m} for every i,j — does not solve the least-squares functional (10). The numerical experiments use a minimum sample size M_min = N^L (Figs. 4-6), while each output has only N^{L-2} features; the paper never reconciles this with the per-element formulation of Eqs. (5)-(6). This makes the exact training procedure ambiguous and, taken literally, incorrect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a supervised-learning framework for programming reconfigurable linear-optical interferometers. It observes that for an interferometer of the layered form U = Φ^L U^{L-1} ... U^1 Φ^1, each matrix element u_{ij} is a sum of monomials in the phase factors φ_k^{(ℓ)}, with coefficients (weights) composed of the fixed basis-matrix elements. This leads to two methods: an 'all-layers' method that trains the full weight tensor from a set of tomographed unitaries by solving a linear least-squares problem, and a 'layer-wise' method (alternating least squares) that trains a local linear model for each phase layer using O(N) samples per layer. Numerical simulations for Haar-random basis matrices show that the all-layers model fits after a minimum number M_min of training samples and that the test Frobenius error decreases as O(M^{-1}) and linearly with the tomography noise amplitude; the layer-wise method converges for full-depth interferometers with and without noise. The paper discusses limitations due to path-dependent loss and phase-shifter crosstalk.","tokens_in":11586,"tokens_out":10275,"duration_ms":100453,"significance":"If the proposed formulation were fully correct, the paper would offer a useful architecture-agnostic route to interferometer programming that avoids non-convex optimization for shallow circuits and uses cheap linear algebra. The algebraic expansion in Eq. (4) is exact and the numerical demonstrations are extensive, including noise scaling and the layer-wise ALS variant; the authors also state the main physical limitations explicitly. However, the central least-squares derivation contains a load-bearing inconsistency between the per-element feature definition and the claimed common pseudoinverse, and the sample-complexity threshold reported in the experiments does not follow from the equations as written. These issues need to be resolved before the central claim is established.","major_comments":[{"comment":"The assertion that the matrix F in Eq. (11) is the same for all matrix elements (i,j) is not supported by the feature definition in Eq. (3). Since θ^{(m)}_{i\\bar{k}j} contains the boundary phase factors φ^{(L,m)}_i and φ^{(1,m)}_j, the design matrix for fixed (i,j) is Θ_{ij} = D_{ij} Θ_0, where D_{ij}=diag_m(e^{i(φ^{(L,m)}_i+φ^{(1,m)}_j)}) and Θ_0 is common to all (i,j). The Moore–Penrose inverse is then F_{ij}= (Θ_0^*Θ_0)^{-1} Θ_0^* D_{ij}^*, which depends on (i,j) through D_{ij}; hence Eq. (11), which uses one F_{\\bar{k}m} for every (i,j), is not the minimizer of the functional (10). If the authors instead solved a single multi-output least-squares problem with a global feature vector of length N^L, then Eq. (10) is not the objective being minimized, and the unconstrained solution does not enforce that each output (i,j) depends only on the N^{L-2} features θ_{i\\bar{k}j}. The authors must state which formulation is intended and give the correct solution for that formulation.","section":"Section III A, Eqs. (3)–(11)"},{"comment":"The reported minimum training sample size M_min=N^L (e.g., 1024 for N=4, L=5) is inconsistent with the per-element feature count from Eq. (3) and the functional (10), which contain N^{L-2} features per output and hence require only M ≥ N^{L-2} for generic phases. The numerical threshold suggests that the simulations actually solve a global least-squares problem with N^L features shared across outputs. The paper needs to reconcile this discrepancy, because the claimed sample-complexity scaling O(N^L) is a central quantity in the discussion of when all-layers learning is practical.","section":"Section IV A, Figs. 4–6"}],"minor_comments":[{"comment":"The Moore–Penrose solution for the overdetermined least-squares problem should be written as W=(Θ^*Θ)^{-1}Θ^*U (or W=Θ^†U); as displayed, (Θ^†Θ)^{-1}Θ^† equals Θ^† because Θ^†Θ is the identity, which is correct but obscures the intended normal equations. Please rewrite to avoid confusion.","section":"Eq. (9)"},{"comment":"The phrase 'the number of these features is O(N L)' should read O(N^L), or should explicitly distinguish the N^{L-2} features per output from the N^L features when counting over all (i,j).","section":"Section III A, near Eq. (11)"},{"comment":"M_min is used to mark the gray vertical line in Fig. 4 but is never formally defined; please state the criterion for 'successful learning' used to determine this threshold.","section":"Fig. 4 caption and Section IV A"},{"comment":"The notation 'x ∈ N(0,1)' should be 'x ∼ N(0,1)' for random variables; the same applies to y.","section":"Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely of interest to the quantum-optics and integrated-photonics community. The main issue is the mismatch between the theoretical derivation in Section III A and the numerical implementation underlying Figures 4–6; the authors should be asked to provide the exact objective, design matrix, and solution used in the simulations. The paper's own Section V already lists important physical limitations, so the scope of the claims is partly delimited, but the linear-algebra inconsistency is load-bearing and must be fixed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jordan,\n\nQuick take: the paper's core idea is real—writing the unitary response of a layered interferometer as a linear model in phase monomials turns training into a linear least-squares problem. That's a genuine advance over the non-convex fitting in ref. [19]. The algebraic expansion in Eq. (4) is correct, and the layer-wise ALS in Sec. III.B is a sensible way to handle deeper circuits. The numerics show the method converges, and the authors are honest about limitations (path-dependent loss, crosstalk) in Sec. V.\n\nBut there's a load-bearing error in the derivation. The features in Eq. (3) include boundary phases φ_i^(L) and φ_j^(1), so for fixed output element (i,j) the design matrix is Θ_ij = D_ij Θ_0 with a diagonal D_ij that varies over training samples and over i,j. The Moore-Penrose inverse is then F_ij = (Θ_0^* D_ij^2 Θ_0)^{-1} Θ_0^* D_ij, which depends on (i,j). The paper's claim that F is constant across i,j is false, and Eq. (11) does not minimize functional (10). The stress-test note is correct.\n\nThat doesn't kill the method—you can solve the least-squares separately for each (i,j) with M ≥ N^{L-2} samples, or move the boundary phases into the weights. But then the reported M_min = N^L in Figs. 4-6 is unexplained; the per-element feature count is N^{L-2}. The paper never reconciles this, and it doesn't compare against refs [19]-[21] or release code. The experiments are all in-distribution, generated from the same decomposition, so they don't test model mismatch.\n\nThe paper deserves serious peer review because the idea is publishable if the derivation is fixed and the claims are made precise. A referee should ask the authors to rewrite Section III.A, clarify the sample complexity, and add a baseline comparison. As written it shouldn't be accepted without revision.\n\nRecommendation: send it out, but with a clear request for revision.","headline":"Promising linearization idea for interferometer control, but the shared-pseudoinverse claim is wrong and the sample-complexity figures don't match the equations.","tokens_in":12224,"tokens_out":4066,"would_cite":false,"duration_ms":37679,"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":"Machine learning turns interferometer control into linear algebra.","keywords":["linear optical interferometer","machine learning","phase shift programming","unitary reconstruction","least squares","photonic circuits","tomography"],"falsifier":"Apply the training procedure to a device with deliberately coupled phase shifters, where setting $\\varphi_k$ also shifts a neighboring phase. If the model error on held-out phases cannot be reduced below a floor no matter how large the training sample is, the monomial-feature linearization is violated.","tokens_in":11135,"feed_emoji":"⚛️","tokens_out":4820,"duration_ms":44501,"temperature":0.7,"pith_summary":"This paper claims that the response of a layered linear-optical interferometer can be turned into a linear model. Each entry of the device matrix is written as a weighted sum of monomial phase features, so learning the device reduces to a least-squares problem rather than nonconvex optimization. The trained model then lets a user find the phase settings that implement a desired unitary by optimizing the model instead of the physical chip. If this holds, any interferometer that fits the alternating phase-mixer decomposition can be programmed without an analytic decomposition, including architectures that resist one.","feed_headline":"Machine learning turns interferometer control into linear algebra","feed_subtitle":"Layered optical circuits can be programmed from tomographed data without analytic decomposition.","key_machinery":"The key object is the monomial feature expansion of Eq. (4), $u_{ij} = \\sum_{\\bar{k}} w_{i\\bar{k}j}\\,\\theta_{i\\bar{k}j}$, with $\\theta$ formed by products of the phase exponentials. It converts device identification into a linear least-squares fit with feature matrix $\\Theta$, whose pseudoinverse is computed once. For layer-wise training, the features shrink to the single-layer phase exponentials, giving $O(N)$ tomography samples per layer.","core_discovery":"The central discovery is that for an interferometer written as a product $U = \\Phi_L U_{L-1} \\cdots U_1 \\Phi_1$, every matrix element $u_{ij}$ is exactly a scalar product of a weight vector $w_{i\\bar{k}j}$ and a feature vector $\\theta_{i\\bar{k}j}$ whose entries are products of phase exponentials $e^{i\\varphi}$ along paths from input $i$ to output $j$. The weights are products of the fixed basis-matrix elements and fully characterize the device; the features depend only on the phase settings. Given $M$ tomographed unitaries at random phase settings, the weights follow from one Moore-Penrose pseudoinverse, and phase programming of the model is then a small optimization. The paper also shows a layer-wise variant that applies to deeper circuits.","pith_inferences":["A testable extension is to add coupling or crosstalk features to the model; if the monomial basis is expanded with interaction terms, the same linear framework could absorb small phase-shifter crosstalk.","Because the model is linear, gradient-based programming could be replaced by closed-form updates for the single-layer case, potentially removing iterative optimization inside layer-wise training.","Intensity-only characterization methods could be combined with the layer-wise scheme to avoid full phase-resolved tomography, at the cost of phase ambiguities."],"forward_implications":["Any interferometer of the layered form can be programmed from data without an analytic decomposition, so new or imperfect architectures become usable.","For shallow circuits, a single tomography campaign suffices: after training, all future target unitaries are programmed by optimizing the model.","For deep circuits, layer-by-layer alternating least squares works with $O(N)$ tomography samples per layer.","Training error on the model falls as $1/M$ beyond a minimum sample size, so noise in tomography is diluted by more measurements.","The programming phase no longer touches the physical device repeatedly, reducing wear and measurement overhead."],"supporting_citations":[{"why":"Establishes that the general layered phase-mixer expansion (1) covers robust universal interferometer architectures, providing the framework the algorithm is built on.","marker":"[14]"},{"why":"Earlier architecture-agnostic machine-learning programming method that this paper linearizes and improves upon.","marker":"[19]"},{"why":"Unitary reconstruction from tomography, supplying the training data $U^{(m)}$ with phases.","marker":"[18]"},{"why":"Fast linear-algebraic reconstruction approach, cited as an alternative that the method complements.","marker":"[20]"},{"why":"Original alternating least squares formulation used for the layer-wise training scheme.","marker":"[22]"},{"why":"Method for generating Haar-random unitary matrices, used to simulate basis matrices and target unitaries.","marker":"[25]"},{"why":"Waveguide-lattice implementation of the mode-mixing layers shown in the device schematic.","marker":"[23]"},{"why":"Low-depth, error-tolerant photonic architectures where the all-layers learning method is most effective.","marker":"[16]"}],"fun_headline_variants":["ML solves interferometer tuning with one pseudoinverse","Program optical circuits from data, no analytics needed","Interferometer control: learn features, set phases","Tomography-trained model steers optical interferometers","From data to unitary: ML for interferometer control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the physical device is exactly a sequence of fixed mode-mixing layers and independent, phase-only shifters, so every phase enters only as $e^{i\\varphi}$ and the tomographed unitaries are faithful.","fun_headline_variants_meta":{"raw":{"variants":["ML solves interferometer tuning with one pseudoinverse","Program optical circuits from data, no analytics needed","Interferometer control: learn features, set phases","Tomography-trained model steers optical interferometers","From data to unitary: ML for interferometer control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1522,"prompt_tokens":823,"completion_tokens":699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":625}},"tokens_in":439,"tokens_out":699,"duration_ms":6904,"temperature":1.0,"reasoning_tokens":625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:38:56.767109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the training procedure to a device with deliberately coupled phase shifters, where setting $\\varphi_k$ also shifts a neighboring phase. If the model error on held-out phases cannot be reduced below a floor no matter how large the training sample is, the monomial-feature linearization is violated.","supporting_citations":[{"cited_title":"Burgwal, W","cited_arxiv_id":null,"evidence_quote":"Establishes that the general layered phase-mixer expansion (1) covers robust universal interferometer architectures, providing the framework the algorithm is built on."},{"cited_title":"Kondratyev, V","cited_arxiv_id":null,"evidence_quote":"Earlier architecture-agnostic machine-learning programming method that this paper linearizes and improves upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Unitary reconstruction from tomography, supplying the training data $U^{(m)}$ with phases."},{"cited_title":"Tillmann, C","cited_arxiv_id":null,"evidence_quote":"Fast linear-algebraic reconstruction approach, cited as an alternative that the method complements."},{"cited_title":"Bantysh, K","cited_arxiv_id":null,"evidence_quote":"Original alternating least squares formulation used for the layer-wise training scheme."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Method for generating Haar-random unitary matrices, used to simulate basis matrices and target unitaries."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Waveguide-lattice implementation of the mode-mixing layers shown in the device schematic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Low-depth, error-tolerant photonic architectures where the all-layers learning method is most effective."}],"review_version":1}