{"id":"b53a275f-b005-4395-84b7-818c78ef9fda","arxiv_id":"2606.16724","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For intensity-only detectors, optimal incoherent transfer matrices for Shannon and Fisher objectives are permutation matrices, so each source must focus onto a distinct detector.","lead":"This paper proves that for intensity-only imaging systems, many information-based design objectives are optimized by a transfer matrix that routes each source to exactly one distinct detector—a condition the authors call generalized focusing. The result gives a fundamental limit for computational meta-imagers and suggests that focusing, not speckle, is optimal with intensity-only detection.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Partially coherent extension rests on an omitted SM proof; if the Schur-complement analysis fails for known or nuisance coherence, the claim that the intensity bottleneck is independent of source coherence is unsupported.","rationale":"The incoherent core of the paper is mathematically sound: the Sec. III proof correctly shows that for nonnegative T with column sums ≤ 1, minimizing a convex decreasing sum of singular-value functions forces T^T T = I and hence T is a permutation matrix. The Shannon and Fisher objectives both fit this framework. The examples are consistent with the theory, and the coherent extension in Sec. V is a clean argument. However, the paper explicitly defers the partially coherent analysis to a supplementary material that is not present in the manuscript, despite the abstract and Sec. V claiming the result as a key conclusion. Since the reviewing rules require flagging omitted proofs, this is the most load-bearing soft spot. The reader's conditional verdict already captures this concern, so no change in verdict is needed; the stress test identifies the precise missing step and a way to settle it.","tokens_in":15893,"tokens_out":16512,"duration_ms":192560,"concrete_test":"For N=2 partially coherent sources, take a generic mixing field transfer matrix H (e.g., a real rotation by angle θ) and numerically or analytically compute the effective Fisher information for the source amplitudes after marginalizing over unknown coherence phases via the Schur complement. If the effective Fisher information is full rank for θ ≠ 0, then a mixing T can be optimal for partially coherent sources, contradicting the claimed persistence of generalized focusing; if it is rank-deficient exactly when |H|^2 is not a permutation matrix, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The broadest advertised result is that generalized focusing persists for coherent and partially coherent sources (abstract; Sec. V). The Sec. V proof covers only fully coherent fields: for a lossless unitary T, phase-independent amplitude recovery from intensity-only measurements forces T = D_out Π D_in. The partially coherent case is then dismissed with a single sentence: 'A complementary Fisher-information analysis confirms the same conclusion (cf. SM),' treating unknown phases as nuisance parameters via the Schur complement. That SM is not included here, so the central claim that the bottleneck is independent of source coherence is unverified. The conclusion is not obvious: if the coherence matrix is partially known, the intensity model is linear in the off-diagonal coherence entries, and mixing might still permit amplitude recovery after marginalizing over the unknown coherence parameters, which would destroy the permutation necessity. The main text provides no analysis for this case.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two deterministic, data-free objectives for designing the photonic stage of an intensity-only computational imager: a Shannon-capacity surrogate (Eq. 5) and a Fisher-information/CRB surrogate (Eq. 9), both closed-form functions of the incoherent transfer matrix T. It proves that under nonnegativity and column-sum constraints (passive optics), the optima of both objectives—and a broader family of singular-value cost functions—are permutation matrices, i.e., each source is focused onto a single distinct detector ('generalized focusing'). This is validated in three inverse-design settings (two-way imager, random-scatterer imaging, Hermite-Gauss mode sorting), with the Fisher objective matching end-to-end optimization. The paper further claims that this focusing constraint is independent of source coherence, arguing via a coherent-field amplitude-recovery proof and a Fisher-information Schur-complement argument deferred to the SM.","tokens_in":16088,"tokens_out":15229,"duration_ms":163126,"significance":"If the claims hold, the paper provides a strong, counterintuitive design principle: for unbiased, prior-free intensity imaging, the information-optimal photonic response is non-mixing and delta-like, regardless of source/detector geometry. The incoherent proof is elegant and essentially correct: the column-sum constraint confines the eigenvalues of T^T T to a simplex, convexity forces equality, and nonnegative orthonormal columns are necessarily permutation vectors. The data-free objectives are practical, and the numerical comparison to end-to-end optimization is a genuine strength. However, the advertised coherence-independence result is not established in the submitted manuscript, and a stated general-family theorem has a sign/direction error; these must be fixed before the broadest claims can be accepted.","major_comments":[{"comment":"The abstract and Sec. IV rely on the claim that generalized focusing persists for partially coherent sources, but the only support in the main text is the sentence 'A complementary Fisher-information analysis confirms the same conclusion (cf. SM)' with no SM included. This is load-bearing: the coherent proof preceding it treats only fully coherent fields through a unitary T, and the partially coherent intensity model is linear in the coherence-matrix elements. If some coherence or phase parameters are treated as unknown nuisances, the Schur-complement argument is not immediate; non-permutation T could in principle allow amplitude recovery after marginalization. Please include the proof in the main text or SM, or clearly restrict the claim to fully coherent sources.","section":"§V, after Eq. (14)"},{"comment":"The theorem for the wider family is stated as 'every function of the singular values of the form F = Σ_j f(λ_j²) with f smooth, strictly convex, and monotonically decreasing is optimized by a permutation matrix.' As written, this is false for maximization: if f is decreasing, maximizing Σ f(µ_j) over the simplex pushes all µ_j to 0, not to 1. The proof sketch actually establishes that such F is minimized at µ_j=1. The Shannon capacity objective enters as the negative of such an F. Please restate the theorem as a minimization claim and clarify the sign convention for the Shannon case.","section":"§III, 'One can widen the class...'"},{"comment":"The proof optimizes over the full set of nonnegative matrices with column sums ≤1 and concludes that 'any permutation' is optimal. Physical passive reciprocal systems may not realize arbitrary permutations: when source and detector ports coincide, reciprocity forces T symmetric, and a symmetric permutation matrix is an involution. The paper's examples use distinct input/output ports, so they are consistent, but the universal 'regardless of geometry' claim needs a caveat or a proof that arbitrary permutations are feasible in the considered geometries. Otherwise the mathematical optimum and the physically realizable optimum may differ.","section":"§III proof; §VI discussion"}],"minor_comments":[{"comment":"The symbol T is used both for the incoherent intensity transfer matrix (Eq. (1)) and for the coherent field transmission matrix (Eq. (12)); these are different objects and should be denoted separately (e.g., U for the field matrix).","section":"§V"},{"comment":"In the S<N extension, 'forces each row to be a distinct canonical basis vector' should read 'each column'; in the S>N extension, the Gram matrix whose off-diagonals are forced to vanish is T T^T (N×N), not T^T T (S×S). Please correct the dimensions and wording.","section":"§VI"},{"comment":"The notation N_n and N_g is used without explicit definition; please define the ensemble sizes in the text or caption.","section":"Eq. (11)"},{"comment":"The high-SNR capacity formula is attributed to Ref. [50] and also to an SM derivation; since the SM is not included, please ensure the derivation is either provided or the attribution and conditions (peak vs. average power, Gaussian noise) are made explicit.","section":"Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's main theorem for incoherent sources is sound and the numerical validation is convincing. The unresolved partially coherent claim and the sign/direction error in the general-family theorem are the reasons for major revision. If the authors provide the SM proof and correct the theorem statement, I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The main result, generalized focusing—that with intensity-only detection and passive nonnegative transfer matrices, the optimal T for both Shannon and Fisher objectives (and any convex decreasing function of squared singular values) is a permutation matrix—is correct. The proof is the clean eigenvalue-simplex argument: column sums cap Tr[T^T T], decreasing convexity pushes the µ_j to the simplex face, equality forces T^T T=I, and nonnegativity forces permutations. I checked the steps against the text and they hold. The coherent-input extension is also solid for the unitary lossless case: if amplitude recovery must be phase-independent for all phases, the interference cross-terms force at most one nonzero per row, and unitarity then gives T=D_out Π D_in. That is a nice complement to the incoherent result.\n\nThe paper also does something genuinely useful: it packages these objectives as data-free, closed-form surrogates that avoid expensive end-to-end training, and the random-scatterer example shows the Fisher objective performing as well as (slightly better than) full end-to-end optimization. The two-way imager and HG mode sorter are nice demonstrations that the condition is not tied to planar geometry.\n\nSoft spots, in proportion. The broadest advertised claim—that generalized focusing persists for partially coherent sources—is not proved in the main text. Sec V says a complementary Fisher-information analysis in the SM confirms it, but the SM is not included. That is a real gap for the headline, not for the core incoherent result. The stress-test concern is legitimate: when coherence is unknown and treated as a nuisance, the intensity model is not the simple incoherent linear model, and it is not obvious that mixing can't help after marginalizing. I don't know whether the Schur-complement argument works; we simply can't verify it from what's in front of us. The authors should include that proof or qualify the abstract.\n\nTwo smaller quibbles. The \"matches end-to-end\" claim rests on one example; the other two examples don't include e2e comparison, so the claim is stronger than the evidence. And the analysis implicitly assumes no source prior and unbiased estimation; the authors acknowledge this in the Discussion, so it's not a hidden flaw.\n\nCitation pattern looks fine. Self-references are contextual, and the MIMO optical wireless capacity literature is credited.\n\nBottom line: the paper deserves a serious referee. The core theorem is correct, the examples support it, and the partially coherent claim needs checking with the actual supplement. I'd send it out and ask for the SM proof.","headline":"The core optimality theorem is sound and the contribution is real; the partially coherent extension is asserted on a supplement we can't see.","tokens_in":16568,"tokens_out":3364,"would_cite":true,"duration_ms":36543,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For intensity-only imaging, the optimal optical response sends each source to a single distinct detector — generalized focusing — for both Shannon-capacity and Fisher-information objectives.","keywords":["computational imaging","information theory","permutation matrix","Fisher information","Shannon capacity","intensity bottleneck","metasurface inverse design","point spread function"],"falsifier":"Run a numerical search over all N×N matrices with nonnegative entries and column sums ≤ 1 for N ≥ 3, evaluating the regularized Shannon capacity (Eq. 5) and the Fisher objective (Eq. 9); if any matrix beats the best permutation matrix for the same SNR and noise level, the theorem is false. Alternatively, fabricate or simulate a metasurface whose intensity transfer matrix is demonstrably non-permutation and show its reconstruction error falls below the best permutation-matrix design under identical noise and unbiased reconstruction.","tokens_in":15780,"feed_emoji":"📷","tokens_out":4354,"duration_ms":42835,"temperature":0.7,"pith_summary":"This paper tries to establish a fundamental limit on what the optics in a computational imager can do when detectors measure only intensity. It develops two closed-form, data-free objectives — one from Shannon capacity, one from Fisher information — that score a photonic stage without any training data. It then proves that both objectives, and any objective in a broad convex family, are maximized by a permutation matrix: each source's light must land on exactly one distinct detector. The authors call this generalized focusing, and show numerically that inverse-designed scatterers converge to it in geometries where conventional focusing intuition fails, including two-way imagers, random scattering media, and Hermite–Gauss mode sorting. They also show the same constraint holds for coherent and partially coherent sources, because the bottleneck is the nonnegativity of intensity measurements, not source coherence.","feed_headline":"For intensity cameras, best optics send each source to its own detector","feed_subtitle":"Data-free information theory shows non-mixing, delta-like responses beat speckle in any geometry.","key_machinery":"The central object is the incoherent transfer matrix T mapping source intensities to detector intensities, with nonnegative entries and column sums ≤ 1. The argument runs through the singular values of T: both proposed objectives (regularized Shannon capacity C = ½ log det(I + (P²/2πeσ²)TTᵀ) and Fisher/CRB objective E = σ² Tr[(TTᵀ + SNR⁻²I)⁻¹]) are functions of the eigenvalues of TᵀT. The column-sum constraint implies ∥μ∥₁ ≤ N, so the eigenvalues lie in a simplex; strict convexity and monotonicity of f push the optimum to μ = (1,…,1), i.e., TᵀT = I. A nonnegative orthonormal matrix must be a permutation matrix. The 'intensity bottleneck' — the fact that nonnegative vectors cannot form a comp","core_discovery":"The paper's central claim is that for intensity-only detection, the optimal incoherent transfer matrix T — for both the Shannon-capacity objective and the Fisher-information/CRB objective, and for every objective of the form Σ f(λ_j²) with f smooth, strictly convex and decreasing — is a permutation matrix. Equivalently, each source's emission is concentrated on a single, distinct detector, a condition the authors call generalized focusing. The proof relies on the observation that passive optics constrain the columns of T to be nonnegative with column sums at most 1, which places the eigenvalues of TᵀT on a simplex; strict convexity then forces all eigenvalues to 1, so TᵀT = I, and only permu","pith_inferences":["The theorem suggests a concrete design rule for computational metasurface imagers: unless prior information is exploited, the photonic stage should be engineered to be as close to permutation-like as possible; attempts to design informative spread-out point-spread functions for intensity cameras are fighting a fundamental limit.","If the intensity bottleneck is the true root, interferometric or phase-preserving detection at the focal plane should be able to circumvent generalized focusing and unlock mixing designs for coherent sources — a testable prediction that extends the paper's discussion.","The convexity argument also implies that the optimal transfer matrix is independent of the specific noise level (as long as it is Gaussian and the regularization is set by SNR), so designs should transfer across noise conditions; this is a corollary worth verifying in practice.","The same nonnegativity obstruction suggests that any intensity-only measurement scheme claiming to recover more degrees of freedom than the permutation optimum must be introducing priors or bias, which the paper's framework would classify as prior-aware extensions rather than violations of the limit."],"forward_implications":["Data-free objectives can replace full end-to-end optimization: optimizing Shannon capacity or Fisher information alone yields reconstruction performance matching end-to-end training, verified through a random scattering medium.","Any speckled or mixing point-spread function carries a scaling penalty: if a single input produces M hot spots, output resolution scales as N/M rather than N.","The generalized-focusing optimum holds regardless of source/detector geometry, including non-planar arrangements such as the two-way imager.","The constraint persists for coherent and partially coherent inputs with intensity-only detection: only permutation-like unitary matrices (up to output/input phases) allow phase-independent full amplitude recovery.","With unequal source/detector counts, generalized focusing persists (with unused detectors for S<N, or multiple sources sharing each detector for S>N), so only N independent channels can be resolved without priors."],"fun_headline_variants":["Proof: For intensity imaging, optimal optics focus each source","Intensity optics optimal when each source hits distinct detector","Generalized focusing: Intensity cameras need permutation optics","Data-free proof: Best optics for intensity are permutation matrices","Intensity-only imaging: Optimal optics always focus (permutation)"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof assumes a passive optical system — every column of the transfer matrix is nonnegative and sums to at most 1 — and that the imaging is done by unbiased estimation with no prior information about the scene; if any of these fails, permutation optimality need not hold.","fun_headline_variants_meta":{"raw":{"variants":["Proof: For intensity imaging, optimal optics focus each source","Intensity optics optimal when each source hits distinct detector","Generalized focusing: Intensity cameras need permutation optics","Data-free proof: Best optics for intensity are permutation matrices","Intensity-only imaging: Optimal optics always focus (permutation)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001096,"raw_usage":{"total_tokens":4428,"prompt_tokens":774,"completion_tokens":3654,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":3588}},"tokens_in":518,"tokens_out":3654,"duration_ms":26981,"temperature":1.0,"reasoning_tokens":3588,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T11:09:32.130662+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a numerical search over all N×N matrices with nonnegative entries and column sums ≤ 1 for N ≥ 3, evaluating the regularized Shannon capacity (Eq. 5) and the Fisher objective (Eq. 9); if any matrix beats the best permutation matrix for the same SNR and noise level, the theorem is false. Alternatively, fabricate or simulate a metasurface whose intensity transfer matrix is demonstrably non-permutation and show its reconstruction error falls below the best permutation-matrix design under identical noise and unbiased reconstruction.","supporting_citations":[],"review_version":1}