{"id":"8f38bf7b-e268-403e-90e5-d5d82486340d","arxiv_id":"2607.24104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Schur-Horn-based optimization recovers the eigenvalues of a light field's spatial coherence matrix from output power measurements on any programmable photonic circuit.","lead":"This paper proposes a way to measure the spatial coherence of light using any programmable photonic chip, reading only output brightnesses and using the Schur-Horn theorem to recover the coherence-matrix eigenvalues. It could turn existing programmable chips into coherence analyzers without knowing their internal design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Loss-resilience claim is an overclaim: with non-unitary A, Eq. (2) recovers output eigenvalues, not input eigenvalues; arbitrary losses can irreversibly change the spectrum.","rationale":"The reader's weakest assumption was the unproven equivalence between minimizing Eq. (2) and diagonalizing, including global convergence. That is a genuine concern, but it is a gap in proof that numerical experiments partly mitigate. The loss claim is stronger: even with perfect convergence and an ideal Schur-Horn objective, the object being diagonalized is the output coherence matrix; for non-unitary A its eigenvalues differ from the input's. The diagonal-loss counterexample isolates the issue from optimization, network topology, and parameterization, showing it is an information-theoretic limitation. This does not invalidate the lossless unitary method, so the reader's conditional verdict (accept after proof and qualification) is unchanged, but the 'arbitrary losses' statements must be removed or sharply qualified.","tokens_in":6303,"tokens_out":10127,"duration_ms":93117,"concrete_test":"Take N=8, gamma_n sampled uniformly in [0.8,1], A = diag(gamma_1,...,gamma_8), and input rho = I_8/8 (fully incoherent). Compute rho_out = A rho A† / tr(A rho A†); it is already diagonal. Evaluate the paper's error metric d(Lambda, diag(rho_out)) from Eq. (3) using Lambda = (1/8,...,1/8), sorting the diagonal entries of rho_out. This is the smallest error any optimizer could achieve, since rho_out is already diagonal. If this lower bound is positive (it will be unless all gamma_n are equal), then the lossy-network claim in Fig. 2b cannot be attributed to imperfect optimization; it is an information-theoretic lower bound. Compare with the FI curve in Fig. 2b at gamma_min = 0.8: the plotted error must be at least this value. If the plotted error is below this bound, the simulation or comparison is inconsistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core ideal-unitary procedure (Eq. (2) plus Schur-Horn) is plausible, but the advertised resilience to 'arbitrary optical losses' cannot hold. The Results define rho_out = A rho A† for A in GL(N). For non-unitary A, the eigenvalues of rho_out are not the eigenvalues of rho. The Schur-Horn objective can at best diagonalize rho_out; the sorted powers it returns are the eigenvalues of rho_out, not of the input field. Thus losses do not merely reduce precision, they change the target spectrum. Concretely, take an N-port diagonal lossy network A = diag(gamma_1,...,gamma_N), gamma_n in (0,1), and a fully incoherent input rho = I/N. Then rho_out = diag(gamma_n^2)/tr(diag(gamma_n^2)), whose eigenvalues are gamma_n^2 / sum gamma_k^2 rather than 1/N. No choice of phases or subsequent unitary can restore equal eigenvalues, because the input spectral information has been multiplicatively scrambled by unknown losses. Figure 2b's growing error with decreasing gamma_min is therefore not a convergence artifact but a fundamental limit. The abstract's statement that the approach is 'inherently resilient to arbitrary optical losses' and the conclusion's 'highly adaptable... in the presence of arbitrary optical losses' are unsupported; at best the method characterizes the coherence of the field transformed by the lossy device.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an architecture-agnostic method for extracting the eigenvalues of the spatial coherence matrix of a partially coherent field using a programmable photonic network treated as a black box. The method is based on the Schur-Horn theorem: by defining a scalar objective from cumulative output power measurements (Eq. (2)) and minimizing it with respect to the network's phase parameters, the authors claim that the output coherence matrix becomes diagonal and the measured powers directly yield the coherence eigenvalues. The approach is validated numerically on Reck, Clements, and interlaced multiport-coupler topologies for 500 random fully coherent, partially coherent, and fully incoherent coherence matrices. The paper also reports performance for truncated (under-parameterized) interlaced networks and for networks with random layer-wise losses, claiming robustness and only minor precision loss.","tokens_in":6683,"tokens_out":9059,"duration_ms":76808,"significance":"If the ideal-unitary core of the method is correct, the proposal is practically significant: it replaces phase-sensitive interferometric coherence characterization with output power measurements from a programmable photonic circuit, independent of the circuit's internal architecture. The numerical study across three topologies and random ensembles is a strength, as is the absence of fitted parameters. The Schur-Horn connection is elegant and the black-box treatment is appealing for integration with imperfect devices. However, the central theoretical equivalence behind Eq. (2) is not proven, and the loss-resilience claim is mathematically overbroad. These issues must be corrected before the paper can be recommended for publication.","major_comments":[{"comment":"The claim that minimizing ||eP(Phi)||^2 is 'equivalent' to diagonalizing rho_out is not established. First, eP is defined using P_k in fixed output-port order, whereas the Schur-Horn theorem applies to sorted diagonal entries. If the protocol sorts the measured powers before forming cumulative sums, this must be stated; otherwise the objective is not permutation-invariant and the theorem does not directly apply. Second, the sentence 'normalized power measurements will never exceed the normalized coherence matrix eigenvalues' is incorrect as written: for a 2x2 matrix with eigenvalues 0.8, 0.2 and diagonal entries 0.6, 0.4, the second power measurement exceeds the second eigenvalue. The correct statement is majorization of the sorted diagonal vector. Third, even accepting the objective, the paper provides no proof that gradient descent on this non-convex landscape reaches a global minimum,","section":"Results, Eq. (2)"},{"comment":"The claim of resilience to 'arbitrary optical losses' is unsupported and, as stated, cannot hold. For A in GL(N) non-unitary, rho_out = A rho_in A† does not preserve eigenvalues. The minimization in Eq. (2) operates on output powers and can at best diagonalize rho_out; the eigenvalues it returns are those of the field after the lossy transformation, not the input field. For example, take A = diag(gamma_1,...,gamma_N) with unequal gamma_n and rho = I/N. Then rho_out has eigenvalues gamma_n^2 / sum_k gamma_k^2, not 1/N, and no unitary post-processing can restore the equal input spectrum. The growing error in Fig. 2b is therefore a fundamental spectral change, not a precision or convergence artifact. The abstract's statement that the approach is 'inherently resilient to arbitrary optical losses' and the conclusion's 'highly adaptable... in the presence of arbitrary optical losses' should be","section":"Results, loss paragraph and Fig. 2b"},{"comment":"The paper claims that under-parameterized, non-universal architectures achieve coherence analysis with 'only minor loss in precision' and that the objective can extract eigenvalues with as few as three layers. For M < N+1 the feasible set of unitaries is a proper subset, and there is no argument that the minimizer of Eq. (2) yields a diagonal rho_out or eigenvalues close to the input eigenvalues. The numerical averages over 500 random samples are useful, but no worst-case or theoretical bound is provided. Since the abstract and conclusions highlight non-universal architectures as a selling point, this should be either supported analytically or explicitly framed as an empirical observation with stated limitations.","section":"Results, under-parameterized networks"}],"minor_comments":[{"comment":"'as few as three layers for PI light' appears to be a typo; the text elsewhere defines FC, PC, and FI ensembles. Please clarify which ensemble is meant, and whether the singular 'eigenvalue' should be 'eigenvalues'.","section":"Results, under-parameterized networks"},{"comment":"'Clemments' should be 'Clements'.","section":"Fig. 1b"},{"comment":"The phrase 'which is an splits the light among its output ports' contains a grammatical error; should be 'which splits the light'.","section":"Results, interlaced network"},{"comment":"'power measuresd' should be 'power measured'.","section":"Eq. (1)"},{"comment":"References [14] and [17] are identical (Markowitz, Zelaya, Miri, Opt. Express 31, 37673 (2023)); they should be merged to avoid duplicate citation.","section":"References"},{"comment":"The statement that 'exact gradients of the objective function can be readily obtained' is only true in simulation where the network transfer matrix is known. In an experimental black-box setting, gradient-free optimization would be needed; this should be clarified so the reader is not misled.","section":"Results, numerical validation"}],"recommendation":"major_revision","confidential_remarks":"The ideal-unitary, universal-topology core of the paper is plausible and the numerical validation is reasonably thorough. However, the loss-resilience claim is mathematically overbroad and cannot be fixed by rewording alone; the authors need to either restrict the claim to characterization of the output coherence matrix or add a calibration procedure for losses. The Eq. (2) equivalence also needs a proof or an explicit downgrade to a heuristic supported by numerical evidence. These are substantial but addressable revisions, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zelaya et al. propose treating a programmable photonic circuit as a black box and using output power measurements only to recover the eigenvalues of the input coherence matrix. The core move is to drive the circuit until the output coherence matrix is diagonal, which they formulate via Schur-Horn as minimizing cumulative power shortfalls. That's a neat reformulation: it removes the need to know the circuit topology or to do layer-by-layer tuning as in Ref. [19]. The simulations on Reck, Clements, and interlaced networks are clean and the accuracy looks good for universal, lossless circuits. I believe the ideal-unitary part of the claim.\n\nThe paper has two problems.\n\nFirst, the loss-resilience sales pitch is not supported. For a non-unitary transfer matrix A, the output coherence matrix is AρA†, whose eigenvalues are not the eigenvalues of ρ except in trivial cases. The optimization can at best diagonalize the output; the powers you read are the eigenvalues of the lossy output field, not the input. Their own Figure 2b shows the error growing as γ_min shrinks—they attribute that to convergence, but it's a fundamental information loss. The abstract's 'inherently resilient to arbitrary optical losses' is an overclaim. The method is robust to component deviations that are still unitary, but not to amplitude losses that scramble the spectrum.\n\nSecond, the Schur-Horn objective as written in Eq. (1) uses the output powers in fixed port order. But Schur-Horn majorization applies to the sorted diagonal entries. To know whether you've diagonalized, you need to compare the cumulative sums of the sorted powers, not the natural port order. Without sorting, minimizing ||eP||² could get stuck or might not correspond to diagonalization. The text says 'normalized power measurements will never exceed the normalized coherence matrix eigenvalues'—that's only true after sorting. If the authors intend to sort but didn't state it, fine; but as written, the math doesn't line up.\n\nThe equivalence between the scalar minimization and actual diagonalization is also asserted without proof. For universal networks the numerics are convincing, but for under-parameterized networks the claim that the objective recovers eigenvalues is heuristic. A proof or at least a convergence analysis would be needed.\n\nThe paper is worth a serious referee—the core idea is pragmatic and could be a useful tool for integrated photonics. But the loss-resilience claim should be toned down or reframed as characterizing the output field, and the sorting issue and equivalence need clarification. I'd recommend conditional acceptance after revision, not desk rejection.","headline":"A genuinely useful black-box trick for extracting coherence eigenvalues from any universal unitary PIC, but the loss-resilience claim is mathematically wrong and the Schur-Horn objective as written misses a sorting step.","tokens_in":7081,"tokens_out":3364,"would_cite":false,"duration_ms":29265,"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 the eigenvalues of a light field's spatial coherence matrix can be read directly from output power meters after a blind optimization of any programmable photonic network, bypassing interferometry and knowledge of the","keywords":["spatial coherence","coherence matrix","Schur-Horn theorem","programmable photonics","power measurements","black-box optimization","eigenvalue extraction","partially coherent light"],"falsifier":"Pick an 8-port universal network, feed a partially coherent state with known eigenvalues, and run the power-only optimization from many random initializations. If the best-found objective value is nonzero, or if the recovered eigenvalues differ from the known ones beyond numerical precision, then the equivalence between minimizing ‖eP‖² and diagonalizing the output coherence matrix does not hold for that architecture.","tokens_in":6235,"feed_emoji":"💡","tokens_out":7494,"duration_ms":68529,"temperature":0.7,"pith_summary":"This paper tries to establish that the eigenvalues of the spatial coherence matrix of a partially coherent light field can be extracted from simple power measurements at the output of any programmable photonic circuit, without knowing the circuit's internal design. The trick is a scalar optimization, derived from the Schur-Horn theorem, that drives the output coherence matrix to diagonal form by maximizing cumulative output powers. If true, this removes the need for phase-sensitive interferometry and per-topology calibration, making coherence analysis practical on integrated platforms. The authors back the claim with numerical tests across several universal topologies and show graceful degradation for non-universal and lossy networks.","feed_headline":"Coherence eigenvalues read from power meters alone","feed_subtitle":"Any programmable photonic chip can analyze spatial coherence with no interferometry and no knowledge of its internal layout.","key_machinery":"The central mechanism is the Schur-Horn theorem, which says that the diagonal entries of a Hermitian matrix are majorized by its eigenvalues: sorted cumulative sums of the diagonal never exceed the corresponding cumulative eigenvalue sums, with equality only for diagonal matrices. The paper packages this into the objective ‖eP(Φ)‖², where eP collects the deficits (1 minus cumulative normalized output power) at each output port; minimizing this drives the device to the point where output powers saturate the Schur-Horn bound, i.e., the output coherence matrix is diagonal. Because the cost uses only power readings, the internal topology of the photonic circuit never needs to be known; the devic","core_discovery":"The paper's central claim is that for any programmable unitary photonic network — triangular, rectangular, or interlaced multiport-coupler — one can extract the eigenvalues of an unknown input coherence matrix by minimizing a scalar cost built from output power measurements alone. The cost is the squared norm of the vector of tail sums of normalized output powers. By the Schur-Horn theorem, a Hermitian matrix's sorted diagonal is majorized by its eigenvalues, and equality of the cumulative sums holds only when the matrix is diagonal; hence pushing the cumulative output powers to their maximum forces the output coherence matrix to diagonalize, making the measured powers equal to the input's c","pith_inferences":["The identical cost function could be pushed further to reconstruct the full coherence matrix, not just its eigenvalues, by combining power readings from multiple random unitary settings — the paper deliberately stops at eigenvalues.","Since the method never inspects the device, it should transfer as-is to other programmable unitary platforms, such as free-space spatial light modulators or fiber-based meshes, and possibly to classical wave systems that obey the same algebra.","The loss-error curves in the paper suggest a quantitative design rule: if the minimum per-component transmission stays above some threshold, reconstruction error stays below tolerance; deriving that threshold analytically would turn the numerical observation into an engineering bound."],"forward_implications":["Spatial coherence analysis becomes a plug-and-play task: any programmable photonic chip, regardless of internal layout, can act as a coherence meter after a blind optimization run.","The same circuit can be re-programmed back to the identity after analysis, so the light field is available downstream unaffected.","Lower-depth and under-parameterized networks are sufficient for useful accuracy, shrinking device footprint and fabrication cost.","Component-level losses and phase errors are absorbed by the black-box optimization, relaxing fabrication tolerances for integrated coherence analyzers."],"fun_headline_variants":["Coherence eigenvalues via power meters only","Any photonic chip can extract coherence eigenvalues","Schur-Horn theorem enables power-based coherence analysis","No interferometry needed for coherence eigenvalues","Programmable photonics: coherence from power readings"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the global minimum of the cumulative-power cost is always zero and is reached only when the output coherence matrix is diagonal, and that gradient descent reliably finds that minimum; for lossy networks this exact equivalence cannot hold because a non-unitary transform changes the eigenvalues.","fun_headline_variants_meta":{"raw":{"variants":["Coherence eigenvalues via power meters only","Any photonic chip can extract coherence eigenvalues","Schur-Horn theorem enables power-based coherence analysis","No interferometry needed for coherence eigenvalues","Programmable photonics: coherence from power readings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1320,"prompt_tokens":682,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":571}},"tokens_in":426,"tokens_out":638,"duration_ms":5935,"temperature":1.0,"reasoning_tokens":571,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:03:37.872431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick an 8-port universal network, feed a partially coherent state with known eigenvalues, and run the power-only optimization from many random initializations. If the best-found objective value is nonzero, or if the recovered eigenvalues differ from the known ones beyond numerical precision, then the equivalence between minimizing ‖eP‖² and diagonalizing the output coherence matrix does not hold for that architecture.","supporting_citations":[],"review_version":1}