{"id":"fd1e22c1-1828-46f6-9988-6eeaf35a2f53","arxiv_id":"2506.15362","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Gaussian random tensors of order D >= 3 admit multi-trace observables whose joint cumulants scale in N as strongly as the product of their one-trace expectations, so large N factorization fails in general.","lead":"For random tensors with three or more indices, the large N limit does not always factorize multi-trace expectations, contrary to a recent conjecture. The authors prove, with combinatorial and probabilistic arguments, that certain invariants have connected parts that are not suppressed at large N.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 1's counterexample is established by the direct double-trace argument and Proposition 3, not by the compressed converse of Lemma 1.","rationale":"The reader's conditional verdict is reasonable, but the specific weakest_assumption (the converse of Lemma 1) is not the load-bearing step. Theorem 1 is proven through Proposition 3 plus the explicit copy-pairing construction, which is a direct argument: a component Mρ with max F < Dnρ/2 yields a double-trace whose connected expectation has exponent at least Dnρ, strictly larger than the product exponent 2 max F. I verified the key estimates: Proposition 2's expectation bound and the monotonicity follow via p_1 m ≥ 1; Proposition 3's union bound yields a fraction 1 - O((2/e)^n) of admissible graphs; the condition A < Dn/2 is met for all large n when D≥3. The only concrete issues are a typo in the displayed bound for |M_n| and the compressed proof of Lemma 1, neither of which affects the central claim. Hence no significant objection; the verdict stays conditional pending minor corrections.","tokens_in":14091,"tokens_out":46089,"duration_ms":427231,"concrete_test":"As an independent check, recompute the number of perfect matchings |M_n| = (2n)!/(2^n n!) using Stirling's formula and confirm the union-bound fraction decays like (2/e)^n; then for a small connected 3-colored graph, explicitly construct the copy-pairing on M⊔M and verify that G(M0,M⊔M) is connected and F equals Dn. This would settle both the numerical typo and the key structural step of the counterexample.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 1) is that for D≥3 there exist connected observables whose joint cumulant has larger N-scaling than the product. The proof does not depend on the converse of Lemma 1. Proposition 3 supplies a graph M with max_{M0} F_n(M0,M) < A < Dn/2 for all M0; taking a connected component Mρ (which must satisfy max F_{nρ} < Dnρ/2 by the union-of-matchings argument), the copy-pairing of the two copies contributes exactly Dnρ alternating cycles and makes G(M0,Mρ⊔Mρ) connected. Hence the connected cumulant has exponent at least Dnρ, while the product has exponent 2 max F_{nρ} < Dnρ, so the cumulant dominates. This route avoids Lemma 1 entirely. I checked Proposition 2's induction (the monotonicity of E[m^{F_r}] follows from p_1 m ≥ 1) and Proposition 3's union bound; the only blemish is the typographical bound |M_n|, which should be sqrt(2) 2^n e^{-n} n^n, not with e^{+n}. This does not alter the (2/e)^n decay in the union bound. Thus no load-bearing flaw was found.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Gaussian random tensors of order D >= 3 and the large-N scaling of connected expectations of multi-trace invariants. It proves that the Gaussian scaling is not strictly subadditive: for n large, almost every edge D-colored graph M on 2n vertices contains a connected component M_rho with 2n_rho vertices such that the connected expectation <Tr_{M_rho}(T) Tr_{M_rho}(T)>_con has scaling exponent at least D n_rho, while the product <Tr_{M_rho}(T)>^2 has scaling exponent strictly smaller. The proof uses the probabilistic method: Proposition 2 bounds the number of alternating cycles between a fixed and a random perfect matching, and Proposition 3 shows that for almost every D-tuple of random perfect matchings the maximal cycle count is (1+o(1))n, hence below Dn/2 for D >= 3. The paper also verifies that the melonic family does satisfy large-N factorization, contrasting with the general failure.","tokens_in":14333,"tokens_out":18175,"duration_ms":170321,"significance":"If the result stands, it refutes a conjecture in the random-tensor/free-probability literature and establishes a sharp qualitative separation from random matrices, where multi-trace expectations always factor at large N. The argument is self-contained, uses no fitted parameters, and gives explicit quantitative bounds on the fraction of counterexamples; Propositions 2 and 3 are proven in detail and the main theorem follows from them by a direct copy-pairing argument that bypasses the more speculative converse of Lemma 1. This is a compact but substantial negative result that will be of interest to the random tensor and physics communities.","major_comments":[],"minor_comments":[{"comment":"The displayed upper bound on the number of perfect matchings is numerically false as written: it states |M_n| < sqrt(2e) 2^n e^n / n^n, but for n=4 this gives about 7.95 while the true value is |M_4|=105. The intended bound is |M_n| < sqrt(2e) 2^n n^n e^{-n}, and the subsequent union bound in Proposition 3 uses this corrected form. Please fix the display and the surrounding Stirling estimate.","section":"Section 3, displayed bound before Proposition 3"},{"comment":"The sufficiency direction of Lemma 1 is only sketched. The key step asserts that re-pairing each boundary graph produces more than Ds/2 alternating cycles and that this implies F_n(M0,M) < sum_rho F_{n_rho}(M0_rho union M0'_rho, M_rho), but the counting argument behind this claim is not given. Since Theorem 1 is subsequently proved directly in the last subsection via the copy-pairing of a component from Proposition 3, this gap does not affect the main theorem, but the lemma should either be proved in full or stated with only the necessary direction and a remark that the direct argument is used.","section":"Section 3, Lemma 1 proof"},{"comment":"The line beginning with 'n^D prod_{t1+t2+...+tD=A}' is not well-formed; after choosing the maximizing tuple and using independence, the bound should be written as n^D (7^n)^D (2n)^{-A}. Please correct this display.","section":"Section 3, Proposition 3 proof"},{"comment":"In the right-hand side of (2.5), the index of F is written as F_{n1}(M0_rho, M_rho); it should be F_{n_rho}(M0_rho, M_rho).","section":"Section 2, Eq. (2.5)"},{"comment":"The sentence 'for D=3 for example, starting at an n of order ~7 6' is unclear; the exponent appears to be missing and the exact threshold for the inequality (D-2)n/2 > (D ln 7 - ln 2)n/ln n + D should be stated explicitly.","section":"Section 3, last paragraph"},{"comment":"There are several small typographical slips, such as 'connect M' instead of 'connected M' shortly after Eq. (2.4), and the repeated use of n1 instead of n_rho in the proof of Lemma 1. These should be cleaned up in revision.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is solid on its main claim, but the authors may wish to clarify the precise status of the conjecture in [41] (e.g., whether it was formulated for complex unitarily invariant tensors) and to avoid relying on the underproved converse of Lemma 1 anywhere in the main logical path. Neither point undermines the central result as presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The main claim is real: for Gaussian random tensors of order D>=3, multi-trace expectations do not universally factorize. That refutes the conjecture from Collins-Gurau-Lionni, and the paper gives a clean mechanism: for a typical graph M, every matching gives F_n(M0,M) < n + o(n), but the copy-pairing of M with itself yields Dn alternating cycles, so the connected double-trace scales at least N^{Dn}, beating the product of single-trace scalings. The probabilistic counting is the right tool, and I think it works.\n\nWhat is genuinely new is the negative statement. That melonic observables factorize is known; the counterexample for arbitrary observables is not. Proposition 2 is proved carefully and Proposition 3's union bound is the heart. The main theorem follows from those two plus the copy-pairing idea, without needing the converse of Lemma 1. The paper is also honest about the restricted family: melonic check is direct, and planarity is noted as insufficient.\n\nSoft spots are minor but real. The displayed bound on |M_n| is numerically wrong as written: it should be sqrt(2) 2^n n^n e^{-n}, not sqrt(2e) 2^n e^n / n^n (or whatever the OCR is doing). The preceding Stirling line gives the correct value, so it is a typo, but it appears in the union bound and should be fixed, otherwise the 1-epsilon conclusion doesn't follow from the displayed line. Also, Lemma 1's converse is compressed. The boundary graph re-pairing argument is plausible but the 'no edge of S0 belongs to a 2-edge cycle' and the count Ds/2 need more detail. As I said, the main theorem doesn't rely on that direction, so it's a polish issue, not a correctness issue.\n\nWho is this for? Anyone working on tensor freeness, tensor model large-N limits, or invariant observables of random tensors. It changes what you can assume: factorization has to be proved for a restricted family, not taken as generic. I would bring it to a reading group and I'd cite it. For peer review: yes, send it to a serious referee. The theorem is important enough and the proof is mostly self-contained; the fixes are small.","headline":"Solid, important counterexample: large-N factorization fails for Gaussian random tensors in D>=3; the proof is sound despite a Stirling typo and a compressed lemma.","tokens_in":14848,"tokens_out":2716,"would_cite":true,"duration_ms":27321,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","05C80","05C70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Gaussian random tensors of order three or higher fail to factorize multi-trace expectations at large N, and the failure is typical among large observables.","keywords":["random tensors","large N factorization","trace invariants","Gaussian random tensors","melonic graphs","strict subadditivity","classical cumulants","edge D-colored graphs"],"falsifier":"For the smallest $D=3$ graphs produced by the probabilistic construction, compute the exact leading power of $N$ in the connected expectation $\\langle \\mathrm{Tr}_M(T)\\mathrm{Tr}_M(T)\\rangle_{\\mathrm{con}}$ and compare it with $2 \\max_{M_0} F_n(M_0,M)$; if it fails to exceed that product, the claimed non-factorization mechanism would be wrong. More sharply, any single connected graph violating $\\max_{M_0} F_n(M_0,M) > Dn/2$ while still obeying strict subadditivity would disprove the equivalence in Lemma 1 on which the theorem depends.","tokens_in":1858,"feed_emoji":"🎲","tokens_out":2208,"duration_ms":126363,"temperature":0.7,"pith_summary":"This paper proves that for Gaussian random tensors of order $D \\ge 3$, expectations of products of trace invariants do not in general factorize at large $N$ into products of single-trace expectations, contrary to a conjecture in the tensor free probability literature. The root cause is that the Gaussian scaling of joint cumulants is not strictly subadditive: there exist observables for which the connected two-point expectation scales like $N^{Dn}$, strictly larger than the square of the one-point expectation. This marks a sharp contrast with random matrices ($D=2$), where factorization always holds. The paper further shows that factorization does survive for the melonic observables, the dominant family in the large-$N$ limit.","feed_headline":"Random tensors break the large-N factorization rule","feed_subtitle":"Unlike random matrices, D≥3 tensors show connected expectations that out-scale single-trace products.","key_machinery":"The central objects are edge $D$-colored graphs: graphs on $2n$ vertices whose edges are partitioned into $D$ perfect matchings, one per tensor-index color. Each such graph $M$ carries a trace invariant $\\mathrm{Tr}_M(T)$, and the Gaussian expectation of a product of invariants becomes a sum over pairings $M_0$ of tensor entries, with each term contributing $N^{F_n(M_0,M)}$, where $F_n(M_0,M)$ counts the cycles that alternate between the pairing $M_0$ and the colored matchings. The Gaussian scaling of an observable is $\\max_{M_0} F_n(M_0,M)$. Lemma 1 reduces strict subadditivity of this scaling (equivalently, large-$N$ factorization) to the component-wise bound $\\max_{M_0} F_n(M_0,M) > Dn/2$ for every connected graph; its converse uses boundary graphs, which compress the information contained in a partial pairing. The proof then runs a probabilistic argument: a tail bound for the number of alternating cycles of two random matchings (Proposition 2) shows that for a fixed pairing $M_0$, the $D$ random matchings have total $F_n(M_0,M) < (1+o(1))n$ with probability close to 1, and a union bound over all $M_0$ yields a graph violating the bound for every $M_0$.","core_discovery":"For every $D \\ge 3$ and every $\\epsilon > 0$, for $n$ large enough, a fraction at least $1-\\epsilon$ of all edge $D$-colored graphs $M$ on $2n$ vertices have a connected component $M_\\rho$ with $2n_\\rho$ vertices such that $\\max_{M_0 \\in \\mathcal{M}_{2n_\\rho}} F_{2n_\\rho}(M_0, M_\\rho \\sqcup M_\\rho) \\ge D n_\\rho$, while $2 \\max_{M_0 \\in \\mathcal{M}_{n_\\rho}} F_{n_\\rho}(M_0, M_\\rho) < D n_\\rho$. Consequently the connected expectation $\\langle \\mathrm{Tr}_{M_\\rho}(T)\\,\\mathrm{Tr}_{M_\\rho}(T)\\rangle_{\\mathrm{con}}$ scales at least as $N^{D n_\\rho}$, strictly larger than the product $\\langle \\mathrm{Tr}_{M_\\rho}(T)\\rangle\\,\\langle \\mathrm{Tr}_{M_\\rho}(T)\\rangle$. In other words, the Gaussian scaling of cumulants is not strictly subadditive for $D \\ge 3$, so large-$N$ factorization of arbitrary multi-trace expectations fails for tensors, while it holds for the melonic family and for $D=2$.","pith_inferences":["The proof suggests that non-factorizing observables are generic rather than exceptional: for large $n$, almost every graph violates the subadditivity bound, so the failure should persist for any Gaussian tensor ensemble with the same Wick structure.","The same combinatorial mechanism should carry over to complex Gaussian tensors and to tensors with $O(N)^D$ symmetry, since the argument only uses the pairing structure of the Gaussian covariance, not the realness of entries.","For proposed tensorial generalizations of free probability, the failure of factorization implies that asymptotic moment–free cumulant relations derived under a factorization assumption would receive corrections at subleading orders; the counterexample family gives a concrete place to test such corrections.","A testable extension: the typical value of $\\max_{M_0} F_n(M_0,M)$ for random $D$-colored graphs may be close to $(1+o(1))n$, and refining the tail bound could give the exact distribution of non-factorization exponents."],"forward_implications":["The conjecture that Gaussian random tensors satisfy large-$N$ factorization, proposed in the tensor free probability literature, is false for $D \\ge 3$.","Large-$N$ factorization survives for the melonic family of observables, so the leading large-$N$ sector of tensor models still admits a factorization description.","For $D \\ge 3$, generic multi-trace observables can have connected expectations that dominate the product of single-trace expectations, making the large-$N$ structure of random tensor models richer than that of random matrices.","The result leaves untouched the known factorization of matrix moments for $D=2$.","The proof identifies a sharp threshold: non-factorization appears once the Gaussian scaling bound for a connected graph with $2n$ vertices falls at or below $Dn/2$."],"supporting_citations":[{"why":"Defines trace invariants from edge D-colored graphs and supplies the Gaussian scaling bound (Proposition 1) used throughout.","marker":"[3]"},{"why":"Provides the Gaussian scaling bound, boundary graphs, and the connected-pairing expansion of cumulants that Lemma 1 builds on.","marker":"[50]"},{"why":"Formulates the conjecture that Gaussian random tensors factorize at large N, the claim the paper refutes.","marker":"[41]"},{"why":"Gives the combinatorial tail bound for cycles in unions of random matchings that Proposition 2 reproves and uses.","marker":"[53]"}],"fun_headline_variants":["Tensor large-N factorization fails for D ≥ 3","Multi-trace tensors defy large-N factorization","Gaussian tensors: no subadditive scaling for D ≥ 3","Tensors vs matrices: large-N factorization breaks","Why melonic observables alone factor at large N"],"cache_read_input_tokens":17024,"weakest_assumption_plain":"The proof of the main theorem rests on the converse direction of Lemma 1: strict subadditivity of the Gaussian scaling is equivalent to the bound $\\max_{M_0} F_n(M_0,M) > Dn/2$ holding for every connected graph $M$; that direction relies on the claim that a partial pairing of boundary graphs can always be re-paired component-wise to produce more than $Ds/2$ alternating cycles, and if this re-pairing claim fails, the violation of the bound would not imply non-factorization.","fun_headline_variants_meta":{"raw":{"variants":["Tensor large-N factorization fails for D ≥ 3","Multi-trace tensors defy large-N factorization","Gaussian tensors: no subadditive scaling for D ≥ 3","Tensors vs matrices: large-N factorization breaks","Why melonic observables alone factor at large N"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1291,"prompt_tokens":1110,"completion_tokens":181,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":102}},"tokens_in":726,"tokens_out":181,"duration_ms":2680,"temperature":1.0,"reasoning_tokens":102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:37:04.452977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the smallest $D=3$ graphs produced by the probabilistic construction, compute the exact leading power of $N$ in the connected expectation $\\langle \\mathrm{Tr}_M(T)\\mathrm{Tr}_M(T)\\rangle_{\\mathrm{con}}$ and compare it with $2 \\max_{M_0} F_n(M_0,M)$; if it fails to exceed that product, the claimed non-factorization mechanism would be wrong. More sharply, any single connected graph violating $\\max_{M_0} F_n(M_0,M) > Dn/2$ while still obeying strict subadditivity would disprove the equivalence in Lemma 1 on which the theorem depends.","supporting_citations":[{"cited_title":"Gurau,Random Tensors","cited_arxiv_id":null,"evidence_quote":"Defines trace invariants from edge D-colored graphs and supplies the Gaussian scaling bound (Proposition 1) used throughout."},{"cited_title":"Universality for Random Tensors,","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian scaling bound, boundary graphs, and the connected-pairing expansion of cumulants that Lemma 1 builds on."},{"cited_title":"Profiles of Permutations,","cited_arxiv_id":null,"evidence_quote":"Gives the combinatorial tail bound for cycles in unions of random matchings that Proposition 2 reproves and uses."}],"review_version":2}