{"id":"3f0a069a-af6e-459a-af50-eda8e4c89ba2","arxiv_id":"1908.01131","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper restates the existing tensor normal distribution in tensor-product notation and derives its characteristic function, but the definitions contain inconsistencies and the claimed novelty is overstated.","lead":"This paper rewrites the tensor normal distribution, a known extension of the matrix normal distribution, using tensor products and derives its characteristic function. It is a notation-focused reformulation that does not introduce a new probability model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.12's proof switches from U_k U_k^T = Sigma_k in Definition (4.5) to U_k^T U_k = Sigma_k; the central formula only follows under the latter convention, so the stated theorem is unsupported as written.","rationale":"The reader's rationale already notes the U_k U_k^T versus U_k^T U_k mismatch, so my concern is not new to the review. However, the reader's stated weakest assumption was the unproved equivalence between the fiber-wise and affine definitions; I find the covariance-convention switch to be the more direct and load-bearing defect, because it blocks the central theorem even if every equivalence theorem were supplied. The m=2 counterexample shows the theorem as stated is false under the paper's Definition (4.5), not merely underproved. This strongly supports the reader's REJECT verdict: the paper's central reference formula is not reliable as written. The underlying mathematics is likely salvageable by correcting the convention or the definition, and the characteristic-function formula itself is standard, so this is a fixable correctness flaw rather than a refutation of the entire approach. But as submitted, the inconsistency invalidates the proof of the main theorem, so the verdict should remain REJECT.","tokens_in":21673,"tokens_out":10462,"duration_ms":104037,"concrete_test":"Set m = 2, U1 = U2 = [[1,1],[0,1]], and let Z be a 2x2 SND tensor. Compute X = Z x1 U1 x2 U2 directly: X_11 = Z_11, so Var(X_11) = 1. The theorem's formula, using Definition (4.5) with Sigma_i = U_i U_i^T (so Sigma_i[1,1] = 2), predicts Var(X_11) = Sigma_1[1,1] Sigma_2[1,1] = 4. Recomputing with Sigma_i = U_i^T U_i (so Sigma_i[1,1] = 1) gives 1, matching the direct calculation. This settles that (4.11) is a consequence of U_k^T U_k = Sigma_k, not of the U_k U_k^T = Sigma_k stated in (4.5).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 4.12 quietly changes the covariance convention. Definition (4.5) defines a Gaussian tensor through X = mu + Z x1 U1 ... xm Um with U_k U_k^T = Sigma_k, but the proof of (4.11) requires Sigma_k = U_k^T U_k and explicitly invokes 'Sigma_s = U_s^T U_s' in the final identity. For a nonsymmetric U_k these two factorizations differ, so the theorem does not follow from the paper's own definition. Concretely, for m = 2, X = Z x1 U1 x2 U2 has covariance (U2^T U2) tensor (U1^T U1); the stated formula with Sigma_k = U_k U_k^T would give (U2 U2^T) tensor (U1 U1^T), a different distribution unless each U_k is normal. Thus the central characteristic-function claim is established only under a convention opposite to the one in the formal definition. This is more directly blocking than the unproved equivalence theorems, because Theorem 4.12 does not even follow from the affine definition as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a tensor-valued analog of the matrix normal distribution. It introduces a standard normal tensor, defines a Gaussian tensor as an affine image X = µ + Z ×1 U1 ... ×m Um of a standard normal tensor, and states several equivalent characterizations in terms of fiber-wise normality, mode-k flattenings, and characteristic functions. The main result, Theorem 4.12, claims that the characteristic function of a Gaussian tensor is exp{i⟨T, µ⟩ − 1/2⟨T^[2], Σ⟩} with Σ = Σ1 × ... × Σm. Sections 2 and 3 set up tensor calculus and matrix-normal facts that are then extended to tensors.","tokens_in":21905,"tokens_out":7628,"duration_ms":73418,"significance":"If carried out rigorously, the tensor-product form of the characteristic function would be a clean coordinate-free summary of the tensor normal distribution and a useful reference for high-order extensions of matrix normal theory. The elementary parts, such as the characterization of SND tensors in Theorem 4.2 and the computation of the SND characteristic function in Lemma 4.8, are mostly correct and clearly presented. However, the paper's positive contribution is currently limited: the key equivalence statements are either unproved or asserted, and the main characteristic-function theorem is entangled with an inconsistent covariance convention. The paper therefore does not yet provide a reliable foundation for the claimed results.","major_comments":[{"comment":"Definition (4.5) requires U_k U_k^⊤ = Σ_k, but Theorem 4.5 states Σ_k = U_k^⊤ U_k, and the proof of Theorem 4.12 explicitly uses 'Σ_s = U_s^⊤ U_s' in the final identity. For nonsymmetric U_k these are different factorizations: under (4.5) the second-order structure of X is (U_1 U_1^⊤) × ... × (U_m U_m^⊤), while the algebra in the proof of Theorem 4.12 computes (U_1^⊤ U_1) × ... × (U_m^⊤ U_m). Consequently the displayed characteristic function does not follow from the paper's own definition as written. The authors must reconcile the convention, either by changing (4.5) and recomputing the covariance of the affine construction or by correcting the proof to use U_k U_k^⊤ consistently.","section":"§4, Eq. (4.5), Theorem 4.5, Theorem 4.12 proof"},{"comment":"The fiber-wise definition of a Gaussian tensor uses an undefined scalar λ_k and writes A^(k)(:, j) ∼ N_{n_k, m_k}(M^(k)(:, j), λ_k Σ_k); the notation N_{n_k, m_k} with a single covariance matrix is not defined for matrix normal distributions. Without fixing these objects, the equivalence between the fiber-wise and affine definitions (Theorems 4.4 and 4.9) has no precise content.","section":"§4, paragraph before Theorem 4.4"},{"comment":"Theorem 4.6 is stated as an 'iff' result, but it essentially restates the definition of a Gaussian tensor, and the converse direction is not proved. Theorem 4.9, which is supposed to supply the converse from the unfolded matrices, is dismissed with 'can be proved by the same technique as in the case of order three.' Since Theorem 4.12 is proved only for the affine definition, the paper needs an actual proof that the fiber-wise or unfolded class coincides with the affine class before the characteristic-function result can be claimed for all Gaussian tensors.","section":"Theorems 4.6 and 4.9"},{"comment":"Theorem 4.7 claims an equivalence between tensor normality and the normality of every mode-k unfolding, but its proof is a one-line appeal to induction with no details. Because this equivalence is load-bearing for Theorem 4.9 and for the interpretation of the main definition, a real proof or a complete sketch should be supplied.","section":"Theorem 4.7"}],"minor_comments":[{"comment":"There is an unresolved equation reference '(??)' in the proof of Theorem 3.10, and several references to 'Lemma 3' should be 'Lemma 3.1' or 'Lemma 3.2' as appropriate.","section":"Theorem 3.10 and surrounding text"},{"comment":"The text contains numerous typos, including 'initialised', 'dist ri-', 'matric es', and 'eq uiv-'; a careful proofreading pass is needed.","section":"Abstract and body"},{"comment":"Theorem 4.4 refers to Definition 4.1 for a general m-order tensor, but Definition 4.1 is stated only for third-order tensors; the general definition should be stated before the theorem.","section":"Theorem 4.4"},{"comment":"Item (1) of Proposition 2.1, as written, appears dimensionally inconsistent; the intended product notation should be clarified.","section":"Proposition 2.1"},{"comment":"Reference [30] is incomplete, lacking volume and page numbers, and should be completed for the reader's convenience.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript overlaps substantially with the existing multilinear normal distribution literature, especially [23] and [19], and the authors should be asked to clarify what is genuinely new beyond notation. The covariance-convention error in §4 is fixable in principle, but if the missing equivalence proofs cannot be supplied, the paper's central claim remains unsubstantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a competent but flawed reformulation of the tensor normal distribution in tensor notation. The main characteristic-function formula is correct only after changing the covariance convention from the one in the definition, and that inconsistency breaks the paper as a reference.\n\nWhat's genuinely useful: the tensor derivative calculus in Sections 2–3 is put together carefully. Lemma 4.10 and 4.11 give clean adjoint relations for the tensor inner product. The exposition of the matrix normal distribution in tensor form is fine for someone wanting that particular notation.\n\nThe soft spots are serious. Definition (4.5) fixes U_k U_k^T = Sigma_k. Theorem 4.12's proof instead uses U_k^T U_k = Sigma_k, and the identity <UT,UT> = <T^[2], Sigma> only holds in the latter convention. For non-orthogonal U_k these describe different distributions. So Theorem 4.12 does not follow from the paper's own definition. The stress-test note is correct. Also the fiber-wise definition in the paragraph before Theorem 4.4 uses an undefined lambda_k and writes N_{n_k,m_k} with one covariance; the equivalence with the affine definition is asserted, with Theorem 4.9 dismissed by 'same technique.' Since the whole paper depends on these definitions, these are not cosmetic typos.\n\nOn novelty: the tensor normal distribution is already in Ohlson et al. [23], Manceur and Dutilleul [19], Basser and Pajevic [6]. What this paper adds is notation, not a new model or estimator. That's a slim contribution even absent the errors.\n\nBottom line: this is a paper for a reader who wants to see the matrix-normal machinery translated into tensor products. Such a reader would have to fix the convention themselves. With the convention corrected, the characteristic function is standard. But as written, it cannot serve as a reliable reference. I'd send it to a referee if the authors promised to fix the convention, but I wouldn't cite it in its current form.","headline":"Correct after a change of convention, but as written the central theorem contradicts the paper's own definition; a useful notation exercise, not a reliable reference.","tokens_in":22434,"tokens_out":4001,"would_cite":false,"duration_ms":35463,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53A45","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"A closed-form characteristic function for Gaussian random tensors, proved via per-mode covariance matrices.","keywords":["random tensor","Gaussian tensor","tensor normal distribution","standard normal distribution","characteristic function","tensor moments","multilinear normal distribution","tensor product"],"falsifier":"A concrete check: for a $2 \\times 2 \\times 2$ random tensor, test whether the fiber-wise standard-normal conditions (all mode-1, mode-2, and mode-3 fibers standard normal) force the eight entries to be independent standard normals. Any joint distribution that satisfies the fiber conditions but is not iid, or any affine Gaussian tensor whose second-moment tensor is not $\\Sigma_1 \\times \\Sigma_2 \\times \\Sigma_3$, would invalidate the equivalence behind Theorem 4.12.","tokens_in":21453,"feed_emoji":"📊","tokens_out":13210,"duration_ms":142642,"temperature":0.7,"pith_summary":"This paper introduces Gaussian random tensors as a higher-order version of matrix normal distributions. A standard normal tensor is one whose fibers are standard normal along every mode, and the paper proves this is equivalent to a tensor with independent standard normal entries and to a tensor whose vectorization is standard normal. A general Gaussian tensor is defined as an affine transformation of a standard normal tensor, with one covariance matrix per mode. The central result gives the characteristic function as $\\varphi_X(T) = \\exp\\{i\\langle T, \\mu\\rangle - \\tfrac{1}{2}\\langle T^{[2]}, \\Sigma\\rangle\\}$, where $\\Sigma$ is the tensor product of the mode covariance matrices. This is the tensor analogue of the matrix normal characteristic function and would let statisticians work with high-order normal arrays without flattening them.","feed_headline":"Gaussian tensors now have a closed-form characteristic function","feed_subtitle":"A normal m-way tensor's distribution is fixed by per-mode covariances, extending matrix normals.","key_machinery":"The carrying object is the mode-wise tensor product action, written $A \\times_1 U_1 \\times_2 \\cdots \\times_m U_m$, together with the square tensor $T^{[2]}$ and the inner product identity $\\langle T, ZU \\rangle = \\langle UT, Z\\rangle$ that lets covariance tensors be pulled out one mode at a time. Lemma 4.10 identifies the mode-wise action with multiplication by the tensor product $U = U_1 \\times \\cdots \\times U_m$, and Lemma 4.11 supplies the adjoint move used in the characteristic-function proof. The final object that carries the formula is $\\Sigma = \\Sigma_1 \\times \\cdots \\times \\Sigma_m$, a $2m$-order tensor built by multiplying one covariance factor per mode.","core_discovery":"The paper's central claim is that a Gaussian distribution on an $m$-order tensor is specified by a mean tensor plus one covariance matrix per mode, and that the resulting characteristic function has the tensor-product form $\\varphi_X(T) = \\exp\\{i\\langle T, \\mu\\rangle - \\tfrac{1}{2}\\langle T^{[2]}, \\Sigma_1 \\times \\cdots \\times \\Sigma_m\\rangle\\}$. The proof treats a Gaussian tensor as $X = \\mu + ZU$ for a standard normal tensor $Z$ and a mode-wise product of matrices $U_k$ with $\\Sigma_k = U_k^\\top U_k$; Lemma 4.10 and Lemma 4.11 move the covariance tensors across the inner product, reducing the calculation to the known characteristic function of a standard normal tensor. Along the way the paper gives equivalent characterizations of standard normal tensors and derives tensor-form moments for Gaussian matrices, including $m_2[X] = \\Sigma_1 \\times \\Sigma_2$.","pith_inferences":["If the fiber-wise and affine classes differ, the characteristic-function formula still defines a coherent 'affine Gaussian tensor' class; the paper's results would remain valid with that narrower definition.","The same adjoint inner-product trick could be applied to any tensor distribution built from an affine transformation of a base law, yielding characteristic functions for tensor elliptically contoured or skew-normal families.","The moment formulas suggest a tensor analogue of the covariance test: in a Gaussian tensor model, the empirical second-moment tensor should be close to a tensor product, so a statistically significant deviation is a direct lack-of-fit test."],"forward_implications":["If the formula holds, every Gaussian tensor is described by $m$ covariance matrices of sizes $n_k \\times n_k$ rather than one huge covariance matrix, making high-order models parsimonious.","Moments of Gaussian tensors can be computed by differentiating the characteristic function in tensor form, avoiding the index ambiguity of vectorized derivatives.","Simulation of a Gaussian tensor reduces to drawing an iid standard normal array and applying mode-wise linear transformations, one per mode.","The standard-normal equivalences in Theorem 4.2 give a simple route to checking normality: a tensor is standard normal exactly when every fiber in every mode is standard normal."],"supporting_citations":[{"why":"Defines the matrix normal distribution whose characteristic function is the base case being generalized.","marker":"[16]"},{"why":"Supplies the proof that a Gaussian matrix is an affine transformation of a standard normal matrix, the step generalized in Theorem 4.5.","marker":"[2]"},{"why":"Provides the tensor product and mode-wise multiplication conventions used throughout the characteristic-function proof.","marker":"[15]"}],"fun_headline_variants":["Gaussian tensors: per-mode covariances fix the distribution","Closed-form characteristic function for normal tensors","Normal m-way tensors, one covariance per mode","Random tensors go Gaussian with per-mode structure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formula is proved for tensors that are affine images of a standard normal tensor, but the paper assumes without a general-order proof that this class coincides with tensors whose fibers are Gaussian along every mode.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian tensors: per-mode covariances fix the distribution","Closed-form characteristic function for normal tensors","Normal m-way tensors, one covariance per mode","Random tensors go Gaussian with per-mode structure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1361,"prompt_tokens":882,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":498,"tokens_out":479,"duration_ms":5170,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:23:42.642331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: for a $2 \\times 2 \\times 2$ random tensor, test whether the fiber-wise standard-normal conditions (all mode-1, mode-2, and mode-3 fibers standard normal) force the eight entries to be independent standard normals. Any joint distribution that satisfies the fiber conditions but is not iid, or any affine Gaussian tensor whose second-moment tensor is not $\\Sigma_1 \\times \\Sigma_2 \\times \\Sigma_3$, would invalidate the equivalence behind Theorem 4.12.","supporting_citations":[{"cited_title":"Kollo and D","cited_arxiv_id":null,"evidence_quote":"Defines the matrix normal distribution whose characteristic function is the base case being generalized."},{"cited_title":"Bilodeau and D","cited_arxiv_id":null,"evidence_quote":"Supplies the proof that a Gaussian matrix is an affine transformation of a standard normal matrix, the step generalized in Theorem 4.5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the tensor product and mode-wise multiplication conventions used throughout the characteristic-function proof."}],"review_version":1}