{"id":"ce2f7dd0-3278-4b2a-8611-4540c3f1fcda","arxiv_id":"2412.02572","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Tensorial free convolution and an additive R-transform are defined for measures coming from tensor freeness, with high-order semicircular and free Poisson laws and convergence theorems.","lead":"This paper extends free probability, a framework that describes sums of large random matrices, to random tensors by defining a tensorial free convolution and an R-transform. It also proves that Wishart-type random tensors converge to a high-order free Poisson law, generalizing the classical Marchenko-Pastur theorem.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5's proof uses a false bound F_p(n) ≤ (2p)^n, invalidating the key existence step for the tensorial free convolution as written.","rationale":"The paper has real strengths: the moment-cumulant formula (Theorem 1) is proved combinatorially, and the Wishart convergence proof (Propositions 3 and 4) is substantive and mostly self-contained. The reader's concern about the unproved freeness criterion from [12] is legitimate and contributes to a conditional verdict. However, my stress-test found a more concrete, internally checkable flaw: the proof of Proposition 5 uses the inequality F_p(n) ≤ (2p)^n, which is asymptotically false for all p≥3. This matters because Proposition 5 is the bridge that lets the paper conclude that the distribution a+b has exponentially bounded moments and therefore defines a compactly supported measure. Without a valid bound, Proposition 6—the existence of μ⊕_pν—is unsupported as written. The flaw is repairable (any base larger than p^p/(p-1)^{p-1} works), so it does not warrant rejection, but it does reinforce the need for a conditional verdict until the proof is corrected. The false inequality is independent of the freeness-criterion concern, so my read is partial agreement with the reader's assessment.","tokens_in":20969,"tokens_out":23799,"duration_ms":228767,"concrete_test":"Check the inequality in Proposition 5: use the closed form F_3(n)=binom(3n+1,n)/(3n+1) and the asymptotic F_3(n) ∼ C (27/4)^n n^{-3/2} to show that F_3(100) > 6^100, i.e. F_3(n)/(6^n) → ∞. This settles that the bound F_p(n) ≤ (2p)^n is false for p≥3, so the proof of Proposition 5 is invalid as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 5 (Section 4.1) is used in Proposition 6 to show that the distribution a+b of two free distributions has exponentially bounded moments, hence defines a compactly supported probability measure μ⊕_pν. The proof of (ii)⇒(i) asserts |m_n(a)| ≤ F_p(n) M^n ≤ (2pM)^n. The second inequality is false for p≥3. The Fuss-Catalan numbers satisfy the standard asymptotic F_p(n) ∼ sqrt((p-1)/(2π p)) (p^p/(p-1)^{p-1})^n n^{-3/2} (for p=2 the base is 4=2p), and p^p/(p-1)^{p-1} > 2p for every p≥3 (e.g., p=3: 27/4=6.75>6). Hence F_p(n)/(2p)^n → ∞, so the claimed exponential bound with base 2p is not available. Since Proposition 6 relies on Proposition 5 to conclude that the moments of a+b are exponentially bounded, the existence of the convolved measure is not established by the argument as written. The statement may be repairable by using a larger base, but the paper does not provide such a bound. This is a concrete internal gap, distinct from the reliance on the freeness criterion from [12], and should be corrected before the construction of ⊕_p is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a tensorial analogue of free additive convolution in the framework of \"tensorial freeness\" introduced in the author's earlier work with Bordenave. It defines moments and free cumulants of a tensor as sums over connected p-regular trace maps, proves the analytic moment-cumulant formula M_T(z) = C_T(z M_T(z)^{p/2}) (Theorem 1), and introduces higher-order semicircular and free Poisson laws with moments given by Fuss-Catalan and Fuss-Narayana numbers. It also proves convergence of Wigner and Wishart tensors to these laws, states a free central limit theorem, and defines a tensorial free convolution \\oplus_p, an R-transform, and a Q-transform, with additivity of the R-transform and examples for semicircular and free Poisson laws.","tokens_in":21242,"tokens_out":9270,"duration_ms":92188,"significance":"If the main results hold, the paper gives a substantial new extension of free convolution to the tensorial setting and connects several known higher-order laws. The combinatorial proof of Theorem 1 via Lemma 2 is a genuine strength: the moment-cumulant relation is derived from the poset rather than assumed, and the convergence results for Wigner and Wishart tensors are substantive. The identification of the semicircular and free Poisson laws with Fuss-Catalan and Fuss-Narayana objects is also useful. However, the construction of \\oplus_p rests on a freeness criterion imported from [12] and on the exponential-boundedness argument in Proposition 5, which contains a false bound; the variance estimate in Proposition 4 is only sketched. These gaps concern the central claims and require repair.","major_comments":[{"comment":"The proof of (ii)⇒(i) asserts |m_n(a)| ≤ F_p(n) M^n ≤ (2pM)^n. The second inequality is false for p≥3: the Fuss-Catalan numbers satisfy F_p(n) ∼ sqrt((p−1)/(2πp)) (p^p/(p−1)^{p−1})^n n^{-3/2}, and p^p/(p−1)^{p−1} > 2p for every p≥3 (e.g., for p=3 the base is 27/4 = 6.75 > 6). Since Proposition 6 relies on this proposition to conclude that the moments of a+b are exponentially bounded, the existence of the measure μ⊕_pν is not established by the argument as written. The result can likely be repaired by replacing 2p with the correct growth constant, but the paper must supply a correct bound.","section":"§4.1, Proposition 5"},{"comment":"The proof of (i)⇒(ii) is also incomplete: it bounds |κ_n(a)| by a sum involving the same unknown quantities |κ_{b_j}(a)| and then writes \"≤ M^n\" without an explicit induction hypothesis, and the final constant (4p)^n is not derived. This direction is needed in Proposition 6 to know that the free cumulants of μ and ν are exponentially bounded before adding them, so it must be made rigorous.","section":"§4.1, Proposition 5"},{"comment":"The variance estimate is only sketched and contains a non-obvious, unproved assertion: when G1 and G2 share a coincident edge and the graph of noncoincident edges has no cycle, the number of noncoincident vertices is claimed to be at most np via a parity/coincidence argument that is not spelled out. Since Var[m_n(W_N)] = O(N^{-2}) is the step that upgrades moment convergence to convergence in probability in Theorem 5, this gap is load-bearing; the proof should either be completed or replaced by a detailed counting argument.","section":"§3.3.2, Proposition 4"},{"comment":"The convolution is defined only for pairs (μ,ν) for which there exist distributions a,b with μ_a=μ, μ_b=ν and a,b free, and Remark 4 concedes that not every distribution corresponds to a real tensor distribution. The paper does not show that all compactly supported measures, or even all measures arising from tensor distributions, admit such a free pair. Consequently the statements in the abstract and in Corollary 3 that the convolution and R-transform are defined for arbitrary compactly supported measures go beyond what is proved. Please either restrict the claims to the constructed class or prove a realization theorem.","section":"§4.1, Proposition 6 and Definition 5"},{"comment":"The additivity of free cumulants for sums and the characterization of freeness by vanishing mixed cumulants for even families are imported from [12] without stating the exact theorem and its hypotheses. Because the entire convolution construction in Section 4 and the free CLT in Theorem 6 depend on this criterion, the paper should state the precise result from [12] (or prove it in an appendix) so the reader can verify that the families of distributions used here satisfy it.","section":"§2.4, §4 opening, and Theorem 6"}],"minor_comments":[{"comment":"There are several typographical errors: the heading \"Propostion 3\" in §3.3.1 should be \"Proposition 3\"; the abstract has \"f reeness\"; Remark 6 has \"notataions\" and \"concernd\"; §2.3 has \"We remind thet Bn\".","section":"Throughout"},{"comment":"In the final display, the summation index runs from b=0 to n−1, while F^b_{p/2}(n) is defined for 1≤b≤n and the moment formula in Definition 4 sums b=1,...,n; this appears to be a shift typo, and the limit should be ∑_{b=1}^n F^b_{p/2}(n) t_N^b + O(N^{-1}).","section":"§3.3.1, proof of Proposition 3"},{"comment":"In item (2), \"µ({a,b}/2\" should read \"µ({a,b})/2\". In item (3), the support is denoted [−M,M], but the text says \"for some R>0\"; the notation should be made consistent.","section":"§4.3, Proposition 7"},{"comment":"In the last display, the notation \"M_T(s)^{sp/2}\" should be \"M_T(z)^{sp/2}\"; the argument of the moment generating function is z, not the summation index s.","section":"§2.5, proof of Theorem 1"},{"comment":"The text says that \"δ0 has all zero moments and free cumulants for n≥1\", but then the distribution t.1_p is described with cumulants concentrated on bouquet maps and its measure Δ_t has cumulant t at n=1; the relation between δ0 and Δ_0 should be stated explicitly to avoid confusion.","section":"§4.1, after Proposition 6"}],"recommendation":"major_revision","confidential_remarks":"The paper depends heavily on the companion paper [12]; if [12] is not yet published or not accepted, the editor should verify its claims before relying on it. The false Fuss-Catalan bound in Proposition 5 is local and easily repairable, but the sketchy variance estimate in Proposition 4 and the unproved existence of free pairs for arbitrary measures are more substantial and should be addressed in revision. The paper fits Math.OA/random matrix theory, but the requested revisions concern central steps of the construction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth reading if you work on random tensors or free probability. It defines a tensorial version of free additive convolution, gives a clean moment-cumulant formula, and proves a Wishart-type convergence to a high-order free Poisson law. Theorem 1 is proved combinatorially and is solid. The Wishart proof is a genuine combinatorial argument, though the variance estimate (Proposition 4) is more sketched than proved. The high-order semicircular and free Poisson laws are correctly computed, and the examples in Lemmas 8 and 9 check out.\n\nThe soft spots are real but likely fixable. First, the stress-test note is correct: in Proposition 5, the bound F_p(n) ≤ (2p)^n is false for p ≥ 3, since the Fuss-Catalan numbers grow like (p^p/(p−1)^{p−1})^n, which exceeds 2p for all p ≥ 3. The proof that the moments of a+b are exponentially bounded therefore has a gap, but it is repairable by using a larger base; the statement of Proposition 5 is true. Second, Theorem 6 (free CLT) as stated does not assume the tensors are freely independent, yet the proof immediately invokes free independence. Without that assumption the statement is false, as the trivial case T_i = T_1 shows. The theorem needs a freeness hypothesis. Third, the convolution ⊕_p is defined only for measures that admit free distributions, and only for even p. The abstract and Corollary 3 phrase things as if all compactly supported measures are covered, which overstates the domain. The reliance on the freeness criterion from the author's previous paper is legitimate but should be flagged more clearly.\n\nWho is this for? People working on tensorial free probability, random tensor models, or melonic quantum gravity. The core construction is sound and the convergence results are meaningful. It deserves a serious referee, but not acceptance in its current form. I would ask for: fixing the Proposition 5 bound, adding the freeness assumption to Theorem 6, and scoping the abstract and corollaries to what is actually proved. I would engage with a revised version.","headline":"The tensorial free convolution and Wishart convergence are real contributions, but the proof has a false bound in Proposition 5, the free CLT lacks its freeness hypothesis, and the scope is overstated.","tokens_in":21789,"tokens_out":5758,"would_cite":true,"duration_ms":61024,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46L54","15A69","60B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Even-order tensors admit a free additive convolution whose R-transform is additive.","keywords":["tensorial freeness","free convolution","R-transform","semicircular law","free Poisson law","Fuss-Catalan numbers","Wigner tensor","Wishart tensor"],"falsifier":"Compute the mixed free cumulants of an explicit pair of even-order tensor families that satisfy the paper's freeness definition; any nonzero mixed cumulant would invalidate the additivity $\\kappa_n(\\mu\\oplus_p\\nu)=\\kappa_n(\\mu)+\\kappa_n(\\nu)$ on which the convolution is built.","tokens_in":20738,"feed_emoji":"➕","tokens_out":11147,"duration_ms":100437,"temperature":0.7,"pith_summary":"This paper aims to build a tensorial analogue of the classical free additive convolution for compactly supported probability measures. The central claim is that if two measures arise from freely independent tensor distributions of even order $p$, their sum defines a measure $\\mu \\oplus_p \\nu$ whose free cumulants are the sums of the individual free cumulants; equivalently, the $R$-transform (the generating function of the free cumulants) satisfies $R_{\\mu\\oplus_p\\nu}(z)=R_\\mu(z)+R_\\nu(z)$. The operation is commutative and associative, with $\\delta_0$ as neutral element. The paper also identifies high-order semicircular and free Poisson laws as the tensor counterparts of the Wigner and Marcenko-Pastur laws, proves Wigner and Wishart tensors converge to them, and gives a free central limit theorem. If correct, this extends free probability's main convolution tool beyond matrices.","feed_headline":"New tensor convolution makes free cumulants additive","feed_subtitle":"Even-order tensor sums get an additive R-transform, plus Wigner and Wishart limit theorems.","key_machinery":"The machinery is the non-crossing poset on $p$-regular trace maps, graphs whose vertices are decorated by tensors and whose edges are contracted according to the tensor indices. The poset order is generated by switches that exchange the endpoints of two edges, with a switch increasing the number of connected components; Moebius inversion on this poset defines the free cumulants $\\kappa_b$. The analytic moment-cumulant formula $M_T(z)=C_T(z M_T(z)^{p/2})$ packages moments and cumulants into formal power series, and the additivity of $\\oplus_p$ rests on the tensorial freeness criterion for even families: two families are freely independent exactly when all mixed free cumulants vanish, so the free cumulants of a sum are the sums of the free cumulants.","core_discovery":"The paper's main discovery is that the operation $\\oplus_p$ is well defined on compactly supported measures that admit freely independent tensor-distribution lifts: for even $p$, the free cumulants of the sum add, $\\kappa_n(\\mu\\oplus_p\\nu)=\\kappa_n(\\mu)+\\kappa_n(\\nu)$ for $n\\ge 1$, and the $R$-transform of the convolution is the sum of the $R$-transforms. The construction is carried by the analytic moment-cumulant formula $M_T(z)=C_T(z M_T(z)^{p/2})$ together with the vanishing of mixed free cumulants for even tensor families. The paper then computes the basic examples: the high-order semicircular law satisfies $\\mu_p\\oplus_p\\mu_p = \\mu_p^{(\\sqrt{2})}$, a $\\sqrt{2}$-dilated semicircular, and the high-order free Poisson laws satisfy $\\nu_{p,t}\\oplus_p\\nu_{p,t'}=\\nu_{p,t+t'}$. It also proves that a Wigner tensor converges to $\\mu_p$, a Wishart-type tensor converges to $\\nu_{p,t}$, and a free central limit theorem holds in this setting.","pith_inferences":["If the analytic machinery can be extended to odd $p$ by the multiplicity-corrected cumulants the paper sketches, the restriction to even $p$ in the convolution section could be lifted, giving a uniform $\\oplus_p$ for all tensor orders.","The additive $R$-transform suggests that the standard free-probability toolbox, such as subordination functions and free multiplicative convolution, could be developed for tensor models and applied to spectral statistics of spiked or noisy tensors.","A direct numerical simulation of small even-order Wishart tensors (for instance $p=4$ with moderate $N$) should reproduce the Fuss-Narayana moments; such a simulation would test the convergence theorems in a regime the paper does not explore.","The $Q$-transform, which inverts the subordination function $G(z)=z^{p/2-1}g(z)^{p/2}$, may be the right object for studying sums of non-compactly supported tensor-valued random variables, a case the paper leaves open."],"forward_implications":["The high-order semicircular law is stable under $\\oplus_p$ up to dilation: $\\mu_p\\oplus_p\\mu_p = \\mu_p^{(\\sqrt{2})}$.","Free Poisson parameters add: $\\nu_{p,t}\\oplus_p\\nu_{p,t'}=\\nu_{p,t+t'}$, and as $t\\to\\infty$ the reshaped free Poisson law tends to the high-order semicircular law.","For compactly supported measures, the $R$-transform of a $\\oplus_p$-convolution is the sum of the $R$-transforms, giving a direct computational tool.","Wigner and Wishart tensors converge in probability to the high-order semicircular and free Poisson laws, and a free central limit theorem holds for even $p$.","A $Q$-transform, $Q_\\mu(z)=(C_\\mu(z)^{p/2}-1)/z$, is introduced as a tensorial substitute for the usual transform, with a subordination theory left for later work."],"supporting_citations":[{"why":"The preceding paper on tensorial freeness; supplies the freeness definition and the even-family criterion that mixed free cumulants vanish, the load-bearing input for the additivity of the convolution.","marker":"[12]"},{"why":"Establishes the free Bessel laws whose moments are Fuss-Narayana polynomials, used here to define the high-order free Poisson law as a compactly supported measure.","marker":"[8]"},{"why":"Proves that a real symmetric tensor admits a probability measure with moments given by trace invariants, and establishes the Wigner-Gurau limit for Gaussian tensors.","marker":"[21]"},{"why":"The companion universality paper; supplies convergence of Wigner tensors to the high-order semicircular law and the moment bounds used in Theorem 4.","marker":"[11]"},{"why":"The random matrix monograph whose moment-method and graph classification are adapted to prove convergence of Wishart tensors to the free Poisson law.","marker":"[5]"},{"why":"Lecture notes on free probability that provide the framework of free convolution, R-transform and Cauchy transforms generalized in Section 4.","marker":"[31]"}],"fun_headline_variants":["Tensor free convolution makes cumulants additive","Even-order tensor sums: free cumulants add","Wigner and Wishart tensors converge to free laws","Tensorial free convolution defined via R-transform","High-order semicircular and Poisson laws for tensors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction of $\\oplus_p$ rests on an earlier theorem, not proved in this paper, that for even-order tensor families freeness is exactly the vanishing of all mixed free cumulants; if that theorem failed for the families being summed, the additivity that defines the convolution would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Tensor free convolution makes cumulants additive","Even-order tensor sums: free cumulants add","Wigner and Wishart tensors converge to free laws","Tensorial free convolution defined via R-transform","High-order semicircular and Poisson laws for tensors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3189,"prompt_tokens":896,"completion_tokens":2293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2235}},"tokens_in":512,"tokens_out":2293,"duration_ms":17228,"temperature":1.0,"reasoning_tokens":2235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:19:13.223076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the mixed free cumulants of an explicit pair of even-order tensor families that satisfy the paper's freeness definition; any nonzero mixed cumulant would invalidate the additivity $\\kappa_n(\\mu\\oplus_p\\nu)=\\kappa_n(\\mu)+\\kappa_n(\\nu)$ on which the convolution is built.","supporting_citations":[{"cited_title":"Bonnin and C","cited_arxiv_id":null,"evidence_quote":"The preceding paper on tensorial freeness; supplies the freeness definition and the even-family criterion that mixed free cumulants vanish, the load-bearing input for the additivity of the convolution."},{"cited_title":"Banica, S","cited_arxiv_id":null,"evidence_quote":"Establishes the free Bessel laws whose moments are Fuss-Narayana polynomials, used here to define the high-order free Poisson law as a compactly supported measure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves that a real symmetric tensor admits a probability measure with moments given by trace invariants, and establishes the Wigner-Gurau limit for Gaussian tensors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The companion universality paper; supplies convergence of Wigner tensors to the high-order semicircular law and the moment bounds used in Theorem 4."},{"cited_title":"Bai and J","cited_arxiv_id":null,"evidence_quote":"The random matrix monograph whose moment-method and graph classification are adapted to prove convergence of Wishart tensors to the free Poisson law."},{"cited_title":"Speicher","cited_arxiv_id":null,"evidence_quote":"Lecture notes on free probability that provide the framework of free convolution, R-transform and Cauchy transforms generalized in Section 4."}],"review_version":1}