{"id":"68e6f100-90cf-48de-a3d7-7d2315745d3c","arxiv_id":"1909.02078","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Vector-valued, quaternion, and complex phase retrieval are characterized by the non-existence of decompositions f=u+v with orthogonal measurements, and graph vector fields are recoverable up to rotation when their associated simplex graph is connected.","lead":"This paper characterizes when complex, quaternion, or vector-valued functions can be reconstructed from magnitude-only measurements, up to the natural symmetries. It also gives conditions under which a vector field on a graph is determined by absolute and neighboring-relative magnitudes, which matters for flocking robots and networked sensing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3.1 is not merely unproved in infinite dimensions; explicit point-evaluation measurements on a vector-valued RKHS violate it, so the central vector-valued characterization is conditional on a hypothesis with no verified infinite-dimensional instances.","rationale":"The reader's weakest_assumption identifies exactly the point on which the central claim balances. The proof of Theorem 3.2 is a short algebraic equivalence; the only place where the definition of phase retrieval (magnitudes only) connects to the pairwise-inner-product condition (3.8) is Assumption 3.1. Sufficiency needs the contrapositive: a candidate g with equal magnitudes but not a unitary image must produce a nonzero pairwise inner-product discrepancy, from which the perpendicular split u=(f+g)/2, v=(f-g)/2 is manufactured. Without Assumption 3.1, equal magnitudes alone can fail to detect such a g; the example above gives exactly that failure, and the theorem's split condition holds vacuously while phase retrieval fails. Hence the theorem's applicability in infinite dimensions is the central scientific risk. This is not a disagreement with consensus; it is an internal gap between the theorem's hypothesis and the paper's advertised scope ('(in)finite dimensions', vector-valued RKHS). The scalar Corollary 3.5 remains safe because H=R makes Assumption 3.1 equivalent to injectivity. The missing proof of Theorem 4.1 (Section 5 lists only Theorem 4.3) is a real but secondary editorial defect; it does not change the assessment. No issue with the graph results, which use finite-dimensional geometry directly. Verdict remains CONDITIONAL: the paper should either prove Assumption 3.1 for a natural infinite-dimensional class or state the vector-valued characterization as finite-dimensional/conditional.","tokens_in":22747,"tokens_out":18811,"duration_ms":205083,"concrete_test":"Verify the following one-line family: for H=ℓ²(N), S=ℓ²(N;H), Φ={δ_n}, f(n)=2^{-n}e_{n+1}, g(n)=2^{-n}e_{n+2}. Check (1) injectivity of TΦ; (2) ⟨f(n),f(m)⟩=4^{-n}δ_{nm}=⟨g(n),g(m)⟩; (3) no unitary U satisfies Uf(n)=g(n) for all n, because such U would map e_{n+1}↦e_{n+2} and fail surjectivity. If the example verifies, it settles that Assumption 3.1 is not automatic and that Theorem 3.2 cannot cover natural infinite-dimensional point-evaluation measurements without an additional argument. If the authors can instead prove Assumption 3.1 in this case, the objection is answered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.2 and its space-level version Theorem 3.3 hinge on Assumption 3.1: equality of all pairwise measurement inner products must force g=Uf for a global unitary U. In finite-dimensional H this is equivalent to injectivity of TΦ (Remark 3.4), but in infinite-dimensional H it is strictly stronger, and the paper supplies no example satisfying it. It is not merely unproved for natural measurement families; it fails. Let H=ℓ²(N), S=ℓ²(N;H) with point evaluations Φ={δ_n}, which is a unitary-invariant vector-valued RKHS. Set f(n)=2^{-n}e_{n+1}, g(n)=2^{-n}e_{n+2}. Then TΦ is injective and ⟨g(n),g(m)⟩=4^{-n}δ_{nm}=⟨f(n),f(m)⟩, so (3.8) holds. Yet any U∈U(H) with g=Uf would satisfy Ue_{n+1}=e_{n+2} for all n≥1; such a U cannot be unitary (its range would omit e_1 or e_2 and so not be onto H). Thus Assumption 3.1 fails exactly at the advertised infinite-dimensional point-evaluation setting, and f is not phase retrieval even though the 'no perpendicular split' condition of Theorem 3.2 holds. The infinite-dimensional vector-valued part of the abstract is therefore not established; the characterization is conditional on a hypothesis whose only proved instances are finite-dimensional fibers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops phase-retrieval characterizations for complex-valued, quaternion-valued, and vector-valued functions from the magnitudes of linear measurements. Section 2 characterizes complex conjugate phase retrieval in a conjugate-invariant linear space by the non-existence of a split f=u+v satisfying the orthogonality condition (2.8) and the non-vanishing cross-correlation condition (2.9), with applications to matrix range spaces, shift-invariant spaces, and quaternion-valued functions. Section 3 introduces unitary-invariant spaces of vector-valued functions and, under Assumption 3.1, characterizes phase retrieval of a single function and of the whole space by the absence of a perpendicular split (Theorems 3.2 and 3.3), together with a projection theorem and a graph-vector-field reconstruction theorem. Section 4 treats affine phase retrieval through injectivity of the maps T_u. The paper states that all proofs are collected in Section 5.","tokens_in":23040,"tokens_out":6612,"duration_ms":60380,"significance":"If the results hold as stated, the paper provides a useful unifying framework: the scalar characterizations recover the classical complement-property equivalence, the quaternion characterization is new, and the graph-vector-field theorem gives a concrete connectivity condition for reconstruction up to an orthogonal matrix. The proofs that are actually supplied are generally clean, and the isomorphism argument in Theorem 2.10 is elegant. However, the infinite-dimensional vector-valued results rest on Assumption 3.1, for which the paper proves no infinite-dimensional instance and for which a natural point-evaluation family fails; moreover, Theorem 4.1 is missing from the proof section despite the paper's declared policy. The finite-dimensional and scalar parts appear sound, but the advertised infinite-dimensional vector-valued claims are not established.","major_comments":[{"comment":"The central vector-valued characterization is conditional on Assumption 3.1, but the paper only proves this assumption in the finite-dimensional case (Remark 3.4) and supplies no infinite-dimensional example satisfying it. In infinite dimensions the assumption is strictly stronger than injectivity of T_Φ, and the natural point-evaluation family on the unitary-invariant space S=ℓ²(N;ℓ²(N)) violates it: let f(n)=2^{-n}e_{n+1} and g(n)=2^{-n}e_{n+2}. Then ⟨g(n),g(m)⟩=4^{-n}δ_{nm}=⟨f(n),f(m)⟩ for all n,m, so (3.8) holds for Φ={δ_n}, but any U with g=Uf would satisfy Ue_{n+1}=e_{n+2} for all n≥1; such a U is not onto H, contradicting unitarity. Thus the abstract's promise of an infinite-dimensional vector-valued characterization is not established; the theorems hold only under a hypothesis whose verified instances are finite-dimensional. The authors should either prove Assumption 3.1 for a nontrivial infinite-dimensional measurement family or explicitly restrict the vector-valued theorems to the finite-dimensional case.","section":"Section 3.1 (Assumption 3.1; Theorems 3.2–3.3)"},{"comment":"Theorem 4.1 is stated as a characterization of affine phase retrieval of a linear space, yet Section 5—which the paper says collects all proofs—does not prove it. The paragraph preceding the theorem only derives the identity ⟨φ(f−g),φ(f+g)+2b_{φ,i}⟩=0 from equal affine magnitudes; it does not establish the asserted equivalence, in particular the necessity direction that injectivity of every T_u follows from affine phase retrieval. Please provide a complete proof of both directions, or explicitly include the missing argument in Section 5.","section":"Section 4 (Theorem 4.1)"}],"minor_comments":[{"comment":"In the displayed expansion after (5.37), the final term '2µ_iµ_j a_1 a_1^T' should presumably be '2µ_iµ_j a_2 a_2^T'.","section":"Section 5.6 (proof of Proposition 2.8)"},{"comment":"The notation Q_8 for the quaternions is nonstandard and is easily confused with the quaternion group of order 8; consider using H or a different symbol.","section":"Section 2.3"},{"comment":"The caption is hard to parse, especially the clause beginning 'where we assume that the hyperplane containing two vectors ...'; please rephrase it for clarity.","section":"Figure 1 caption"},{"comment":"The statement 'All proofs are collected in Section 5' is inaccurate also for Theorem 3.3, whose proof is not listed there, although it is a direct consequence of Theorem 3.2.","section":"Section 1 and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's finite-dimensional and scalar content is solid, and the quaternion and graph results are genuinely interesting. The main barrier is the unverified Assumption 3.1 in infinite dimensions, which the authors should either prove for a concrete infinite-dimensional family or remove from the claimed scope; the missing proof of Theorem 4.1 is a smaller but real gap in light of the paper's own stated policy. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the paper's headline vector-valued theorem is only proved for finite-dimensional fibers. Assumption 3.1 — equality of all pairwise measurement inner products forces g = Uf — is not an innocent technical condition. It fails for the most natural infinite-dimensional example: H = ℓ²(N), S = ℓ²(N;H), Φ point evaluations. Take f(n) = 2^{-n}e_{n+1}, g(n) = 2^{-n}e_{n+2}. The Gram matrices of the measurements coincide, so (3.8) holds, but no unitary maps f to g: the map e_{n+1} ↦ e_{n+2} for all n is not surjective. So f satisfies the no-perpendicular-split condition of Theorem 3.2 and yet is not phase retrieval. The infinite-dimensional version of the abstract is false as stated. Remark 3.4 only covers dim H < ∞.\n\nWhat's genuinely good: the quaternion conjugate phase retrieval characterization (Theorem 2.10) is new and clean; the graph reconstruction via d-simplex graphs (Theorem 3.10) is a nice result with a clear robotics reading; the projection theorem (3.7) and the affine results are solid. The scalar complement property drops out correctly. Proofs are generally readable, and the paper is honest enough to state Assumption 3.1 explicitly, but it overstates the range of validity in the abstract.\n\nOther soft spots: Theorem 4.1 is announced with no proof in Section 5; Section 5 proves 4.3 only. That is a missing proof, not a fatal problem, but it should be fixed. There is a small typo in the proof of Proposition 2.8 (the term 2µ_iµ_j a1a1^T should presumably be a2a2^T). Neither affects anything else.\n\nBottom line: as a finite-dimensional, quaternion, and graph paper, this deserves attention. The vector-valued infinite-dimensional claims need either to be withdrawn or to come with a verified class of measurements satisfying Assumption 3.1. This deserves peer review, but it needs major revision before acceptance.","headline":"The vector-valued characterization is sound in finite dimensions, but Assumption 3.1 fails for natural infinite-dimensional measurements, so the infinite-dimensional claims overreach; the quaternion and graph results are genuinely worth keeping.","tokens_in":23568,"tokens_out":3597,"would_cite":true,"duration_ms":35607,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["42C15","46C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Phase retrieval reduces to a perpendicular-split test for complex and vector-valued functions.","keywords":["phase retrieval","vector-valued functions","quaternion-conjugate phase retrieval","unitary invariance","complement property","affine phase retrieval","vector fields on graphs","infinite-dimensional Hilbert spaces"],"falsifier":"Construct an infinite-dimensional real Hilbert space $H$, a unitary-invariant space $S$ of $H$-valued functions on a two-point domain, and a unitary-invariant measurement family $\\Phi$ whose measurement map is injective but whose image does not determine all pairwise inner products—for instance, choose the two point evaluations so that the span of one function's measurements has an orthogonal complement of different dimension from the span of a competitor's measurements. If for such $f$ the pairwise-inner-product condition (3.8) holds for some $g$ that is not a unitary image of $f$, and no perpendicular split $u,v$ exists, then the 'if' direction of Theorem 3.2 fails and Assumption 3.1 is pinpointed as the obstruction.","tokens_in":22527,"feed_emoji":"📐","tokens_out":11027,"duration_ms":106707,"temperature":0.7,"pith_summary":"The paper establishes a common geometric criterion for phase retrieval across three settings: complex functions in conjugation-invariant spaces, quaternion-valued functions, and vector-valued functions in unitary-invariant spaces with unitary-invariant measurements. In each case, a function is recoverable from the magnitudes of its linear measurements, up to the natural trivial ambiguity, if and only if it cannot be written as $f=u+v$ where every measurement of $u$ is orthogonal to the corresponding measurement of $v$, yet some pair of cross-measurements has a nonzero symmetric sum. The characterization makes no dimension restriction and, in the real scalar case, reduces to the classical complement property for measurement systems. The same machinery is then applied to reconstruct vector fields on graphs—up to an orthogonal matrix—from absolute magnitudes at vertices and relative magnitudes between neighboring vertices, and to affine phase retrieval of vector-valued functions. A sympathetic reader would care because the result supplies a single structural test that governs uniqueness in a wide family of phaseless-recovery problems.","feed_headline":"A perpendicular-split test settles phase retrieval for vector functions","feed_subtitle":"Magnitude data determine a function up to rotation exactly when it has no perpendicular split; the classical complement property follows.","key_machinery":"The load-bearing mechanism is the perpendicular-split identity. If a rival function $g$ has the same magnitude data as $f$ and is not a unitary image of $f$, then $u=(f+g)/2$ and $v=(f-g)/2$ produce exactly such a split: the measurement orthogonality comes from the equality of magnitudes, and the nonzero symmetric cross-term comes from the difference between the inner products of measurements of $g$ and of $f$. Conversely, any perpendicular split yields a rival $g=u-v$ with identical magnitudes and different pairwise measurement inner products. Assumption 3.1 is the bridge that lets magnitude data control those inner products: it states that equal pairwise inner products for all measurement pairs imply the two functions differ by a unitary, a condition that is automatic in finite dimensions but is an extra hypothesis in infinite dimensions.","core_discovery":"On the paper's own terms, the central claim is Theorem 3.2: in a unitary-invariant space $S$ of $H$-valued functions with unitary-invariant measurements $\\Phi$ satisfying Assumption 3.1, a function $f$ is phase retrieval—determined up to a unitary operator on $H$ by the magnitudes $\\|\\varphi(f)\\|$—if and only if there are no $u,v\\in S$ with $f=u+v$, $\\langle\\varphi(u),\\varphi(v)\\rangle=0$ for every $\\varphi\\in\\Phi$, and $\\langle\\varphi_0(u),\\varphi_1(v)\\rangle+\\langle\\varphi_1(u),\\varphi_0(v)\\rangle\\neq 0$ for some $\\varphi_0,\\varphi_1\\in\\Phi$. The same perpendicular-split condition, with the trivial ambiguity adjusted, characterizes complex conjugate phase retrieval of complex functions and quaternion conjugate phase retrieval of quaternion-valued functions. Direct corollaries are that a real scalar space is phase retrieval exactly when the measurement set has the complement property, that phase retrieval passes to projections onto subspaces, and that a vector field on a complete graph is determined up to an orthogonal matrix by its absolute and neighboring relative magnitudes, with a connected-simplex-graph condition covering non-complete graphs. Affine phase retrieval of vector-valued functions is characterized by injectivity of a family of linear maps built from the reference vectors.","pith_inferences":["If Assumption 3.1 could be established for natural infinite-dimensional measurement systems—sampling at points or frames—the perpendicular-split test would become a practical certificate for uniqueness in those systems; the paper proves the assumption in finite dimensions and leaves the infinite-dimensional case open.","The perpendicular-split condition is algebraic: in the finite-dimensional range-space case it is equivalent to excluding a real matrix of rank at most two with zero diagonal traces, which suggests a low-rank matrix test for conjugate phase retrieval that could be checked by semidefinite relaxation.","The graph reconstruction theorem implies a distributed reconstruction algorithm: each $(d+1)$-clique fixes a local orthogonal frame, and connectivity of the $d$-simplex graph is exactly the condition for gluing these local frames into one global orthogonal transformation; where the graph disconnects, one should expect gauge ambiguities between components.","In affine phase retrieval, the paper's necessary-and-sufficient condition suggests that when the reference differences fail to span the space, the number of reference vectors required will grow with the codimension of their span—a quantitative question the paper does not address."],"forward_implications":["In finite-dimensional Hilbert spaces, the vector-valued characterization is an 'if and only if' once the measurement map is injective, and in the real scalar case it reproduces the complement property as the exact condition for phase retrieval.","Phase retrieval is hereditary: if $f$ is phase retrieval in $S$, then its projection onto any subspace $\\tilde H$ is phase retrieval in the projected space, so uniqueness on a subspace is forced by uniqueness on the whole space.","A velocity field on a complete graph is determined up to an orthogonal matrix by absolute speeds at vertices and relative speeds between every pair of vertices; for a general graph the same holds when the associated $d$-simplex graph is connected and covers all vertices.","With at least two reference vectors per measurement, affine phase retrieval of a vector-valued space is equivalent to injectivity of the measurement map whenever the reference differences span the Hilbert space.","For the shift-invariant space generated by the hat function, the paper identifies precisely which coefficient sequences are complex conjugate phase retrieval: all coefficients in the support interval must be nonzero and at most one adjacent pair may have a nonzero imaginary cross term."],"supporting_citations":[{"why":"introduces the complement property as the criterion for phase retrieval in finite frames, which the scalar case reproduces.","marker":"[9]"},{"why":"provides the scalar phase-retrieval characterization in shift-invariant spaces that underlies the real-part and hat-function results.","marker":"[17]"},{"why":"sets up phase retrieval in infinite-dimensional Hilbert spaces and documents its instability, the backdrop for Assumption 3.1.","marker":"[12]"},{"why":"extends phase retrieval to continuous frames in Banach spaces, the general measurement setting the paper's Phi follows.","marker":"[4]"},{"why":"gives the equivalent rank-at-most-two matrix formulation of conjugate phase retrieval that Corollary 2.4 uses.","marker":"[25]"},{"why":"supplies the affine phase retrieval characterization for scalar spaces that the vector-valued affine results generalize.","marker":"[27]"}],"fun_headline_variants":["No split, no ambiguity: vector phase retrieval solved","Split test resolves vector-valued phase retrieval","Perpendicular-split condition decides vector phase retrieval","Vector functions recovered from magnitudes up to rotation","Phase retrieval for vector functions: a perpendicular-split test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the vector-valued theorems, the argument assumes that any two functions whose measurements have the same pairwise dot products must be related by a single global rotation; this is automatic in finite dimensions but is a strictly stronger, unproved condition in infinite dimensions.","fun_headline_variants_meta":{"raw":{"variants":["No split, no ambiguity: vector phase retrieval solved","Split test resolves vector-valued phase retrieval","Perpendicular-split condition decides vector phase retrieval","Vector functions recovered from magnitudes up to rotation","Phase retrieval for vector functions: a perpendicular-split test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3513,"prompt_tokens":988,"completion_tokens":2525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":604,"tokens_out":2525,"duration_ms":19439,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:02:17.720849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct an infinite-dimensional real Hilbert space $H$, a unitary-invariant space $S$ of $H$-valued functions on a two-point domain, and a unitary-invariant measurement family $\\Phi$ whose measurement map is injective but whose image does not determine all pairwise inner products—for instance, choose the two point evaluations so that the span of one function's measurements has an orthogonal complement of different dimension from the span of a competitor's measurements. If for such $f$ the pairwise-inner-product condition (3.8) holds for some $g$ that is not a unitary image of $f$, and no perpendicular split $u,v$ exists, then the 'if' direction of Theorem 3.2 fails and Assumption 3.1 is pinpointed as the obstruction.","supporting_citations":[{"cited_title":"Balan, P","cited_arxiv_id":null,"evidence_quote":"introduces the complement property as the criterion for phase retrieval in finite frames, which the scalar case reproduces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the scalar phase-retrieval characterization in shift-invariant spaces that underlies the real-part and hat-function results."},{"cited_title":"Cahill, P","cited_arxiv_id":null,"evidence_quote":"sets up phase retrieval in infinite-dimensional Hilbert spaces and documents its instability, the backdrop for Assumption 3.1."},{"cited_title":"Alaifari and P","cited_arxiv_id":null,"evidence_quote":"extends phase retrieval to continuous frames in Banach spaces, the general measurement setting the paper's Phi follows."},{"cited_title":"Conjugate Phase Retrieval on ${\\mathbb C}^M$ by real vectors","cited_arxiv_id":"1709.08836","evidence_quote":"gives the equivalent rank-at-most-two matrix formulation of conjugate phase retrieval that Corollary 2.4 uses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the affine phase retrieval characterization for scalar spaces that the vector-valued affine results generalize."}],"review_version":1}