{"id":"ad45a194-ca13-40ec-aa7a-3b21db825f9f","arxiv_id":"2507.18601","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Global height-function fluctuations in three random partition models converge in the sense of moments to a conditioned Gaussian Free Field through a new Young generating function framework.","lead":"The paper introduces a Young generating function for probability measures on integer partitions and uses it to give general conditions for laws of large numbers and central limit theorems, including a multilevel version for random standard Young tableaux.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Process-level convergence to the conditioned GFF is not established; §7.2–7.5 only prove moment convergence of polynomial observables, with no tightness or approximation argument.","rationale":"I read the paper in good faith and find its core contribution substantial: the Young generating function, the CLT/LLN criteria, and the multilevel covariance computations are new and intricate, and the explicit formulas in §7.2–7.4 are impressive. The reader's weakest-assumption diagnosis matches my own reading: the theorems state process-level convergence, but the supplied proofs establish only moment convergence of the fixed-time polynomial observables M^P_{α,k}, M^S_{α,k}, M^Fix_{α,k}. This is exactly the place where the central claim 'the limiting fluctuations are a conditioned GFF' needs an additional argument. A careful reader can see the gap in §7.5, which is explicitly described as a brief discussion and only verifies the Plancherel case. The concern is specific and addressable: one needs to add a tightness result or a functional-analytic approximation lemma showing that moment convergence of these observables implies convergence of the whole field. This does not invalidate the moment computations, and the paper already contains most of the ingredients; hence the reader's CONDITIONAL verdict is appropriate. My read does not change that verdict.","tokens_in":78251,"tokens_out":15702,"duration_ms":170896,"concrete_test":"Pick a smooth bump φ(t), a mollifier ρ_ε(x−x_0), and form Φ_ε(x,t)=φ(t)ρ_ε(x−x_0). Using the covariance formulas of §7.2–7.4, compute an explicit bound on limsup_{n→∞} E[(√π ∫ Φ_ε (H−EH) dx dt)^2] as a function of ε. If this second moment does not converge, as ε→0 after n→∞, to the covariance of C applied to Φ_ε, or if the bound diverges unless an additional regularity condition on the height process is imposed, then the moment computations do not yield the claimed field convergence and the theorems require a separate tightness or approximation argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Theorems 3.6, 3.9, and 3.12 is that the rescaled height process converges to the conditioned Gaussian Free Field C. The proofs, however, establish a strictly weaker statement: the collection of observables M^P_{α,k}, M^S_{α,k}, M^Fix_{α,k} — integrals of the height fluctuation against x^k at fixed times — converges in the sense of moments (via Theorem 2.22 and Lemma 7.3). To conclude that the height process itself converges to C as a random distribution, one must pass from these finite-dimensional, polynomial, fixed-time observables to convergence of ∫ f (H − EH) for a dense family of test functions f(x,t), uniformly in n. No such passage is supplied. Sections 7.2–7.4 compute the limiting covariance kernel explicitly, and Section 7.5 'briefly discusses' the identification with a conditioned GFF, treating only the Plancherel case and declaring the other two 'equivalent.' The unstated premise — that moment convergence of these observables determines the field — is exactly the weak link. In particular, there is no tightness criterion and no estimate controlling the difference between the field tested against a general smooth function and its finite polynomial approximation. The covariance kernel alone does not identify the limit of the random height functions without such control.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a 'Young generating function' for probability measures on integer partitions and uses it to give necessary and sufficient conditions for a sequence of random partitions to satisfy a law of large numbers and a central limit theorem. It then proves a multilevel CLT for random increasing sequences of partitions, which is applied to the Plancherel growth process, random standard Young tableaux of fixed shape, and measures induced by extreme characters of S_∞. In all three applications, the claimed limit of the rescaled height-function fluctuations is a conditioned Gaussian free field C on the upper half-plane. The proofs are based on explicit expansions of operators in the Gelfand–Tsetlin algebra, with detailed combinatorial arguments involving non-crossing partitions, Kreweras complements, and cumulant expansions.","tokens_in":78531,"tokens_out":4163,"duration_ms":47074,"significance":"If the main claims are fully established, this would be a substantial contribution: it provides a unified framework covering LLN, CLT, and multilevel fluctuations for a broad class of partition measures, and it gives explicit covariance formulas in three previously inaccessible two-dimensional fluctuation regimes. The operator expansions in the Gelfand–Tsetlin algebra and the related combinatorial identities are technically impressive and are of independent interest. The paper also resolves a Pittel–Romik conjecture and offers falsifiable, explicit covariance formulas that can be checked numerically. However, the advertised process-level convergence to a conditioned GFF is not actually proved: the text establishes moment convergence of a restricted family of polynomial observables, and the missing tightness/approximation step is load-bearing for the central claim.","major_comments":[{"comment":"The theorems are stated as process-level convergence, e.g. √π(H(√n x, nt) − E H(√n x, nt)) → C(x,t), but the proofs only establish convergence in the sense of moments of the observables M^P_{α,k}, M^S_{α,k}, and M^Fix_{α,k}, which are integrals of the height fluctuation against x^k at fixed times. No tightness, no uniform estimate, and no approximation by a dense set of test functions f(x,t) is supplied. Moment convergence of these fixed-time polynomial observables is strictly weaker than convergence of the generalized random field tested against smooth compactly supported functions; the covariance kernel alone does not identify the law of the random height process without an additional argument. This gap affects all three applications and is not addressed in §7.5, which only discusses the Plancherel case.","section":"§7.2–7.4, Theorems 3.6, 3.9, 3.12"},{"comment":"The identification of the limiting object C with a conditioned Gaussian free field is verified only in the Plancherel case: the projection P[G](f) = G(f − f(0)) is checked for semicircular contours, and the other two models are dismissed as 'equivalent.' The contour systems s_F and ŝ_F in Theorems 3.9 and 3.12 are genuinely different, and the orthogonality of the projection and the covariance identity must be checked separately for those systems. Since the central claim of the paper is that the limiting fluctuations are exactly this conditioned GFF, this is a load-bearing omission rather than a cosmetic one.","section":"§7.5, Proposition 3.5"},{"comment":"The conditioning 'C(1)=0' is used as the defining property of the conditioned field, but C is a generalized field and C(1) is not defined by the covariance kernel given in §3.3. A rigorous definition of the conditioning subspace K, the projection P, and the sense in which C = G − P[G] is a Gaussian process indexed by the admissible test functions is needed. As written, Proposition 3.5 equates two objects without specifying the common space on which they are defined.","section":"§3.3 and §7.5"}],"minor_comments":[{"comment":"The text contains a duplicated phrase: 'in section in section 7.1' should read 'in section 7.1.'","section":"§3.1"},{"comment":"The displayed covariance formula is difficult to parse because the large parenthesis around the logarithmic term appears unbalanced; adding an extra closing parenthesis or restructuring the display would improve readability.","section":"§2.4, Theorem 2.19"},{"comment":"The computation of A_{ρ_n} is written with an approximate equality '≈'; since CLT-appropriateness is defined through exact derivatives of ln A_{ρ_n}, it would be clearer to state the precise asymptotic expansion and explain why the derivatives of the error term are negligible.","section":"§3.5, Example 3.10"},{"comment":"In the displayed formula for h(x), the condition '|t| ≤ √2' uses t rather than x; the variable of the limiting diagram should be x.","section":"§3.6, Remark 3.13"},{"comment":"The phrase 'C can be identified as a Gaussian free field G conditioned to be 0 when integrated over the curves C_α' is used before the rigorous definition of the conditioning; it would help to state this as a proposition rather than as an informal identification.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper contains substantial and credible combinatorial work, but the advertised process-level convergence to the conditioned GFF is not established by the supplied arguments. The moment computation is a necessary component, but it is not sufficient. I recommend major revision rather than rejection, because the gap appears fixable with additional technical work, and the rest of the paper is valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, this is a serious and substantial paper: the Young generating function framework, the LLN/CLT characterizations, the multilevel CLT, and the three applications (Plancherel growth, fixed-shape SYT, extreme characters) are genuinely new, and the resolution of the Pittel-Romik conjecture is a real win. Second, the central claim as sold in the abstract and the informal sample theorem—that the rescaled height process converges to the conditioned Gaussian free field—is not fully supported by the proofs. The precise statements (Theorems 3.6, 3.9, 3.12) say the moments of the observables converge, and that is exactly what the proofs establish. What is missing is the passage from fixed-time polynomial observables to a field-level limit. There is no tightness argument and no estimate controlling the difference between the height fluctuation tested against a general smooth function and its polynomial approximation. Section 7.5, which identifies the limit object, treats only the Plancherel case and waves at the other two as 'equivalent.' So the stress-test note is on target.\n\nWhat the paper does well deserves emphasis. The expansion machinery in the Gelfand-Tsetlin algebra is built carefully from scratch, with explicit leading coefficients and new combinatorial identities. The covariance computations are concrete and checkable, and the paper is honest about which parts are restatements of Biane's results. The multilevel LLN/CLT is a genuine technical contribution, not an incremental tweak. The resolution of Pittel-Romik is clearly derived, not assumed. I found no circularity: the characterizations really go both ways.\n\nThe soft spots are real but localized. The main one is the moment-to-field gap, which is not a matter of taste: without tightness, the covariance kernel alone does not identify the limit of the random height functions. There are also a few lemmas (e.g., Lemma 6.4 and parts of Lemma 7.5) that are only sketched, though the sketches look plausible. Given the length of the paper, some compression is expected, but the tightness issue is load-bearing for the GFF claim.\n\nWho is this for? Probabilists and combinatorialists working on random partitions, Young tableaux, and asymptotic representation theory. They will get genuine value from the machinery and formulas. It deserves a serious referee: this is exactly the kind of paper where referee time is justified because the central claims are important and the gap is plausibly fixable. I would send it to peer review with a strong recommendation to either add the missing functional-analytic step or rewrite the claims to match the moment convergence that is actually proved.","headline":"Strong new machinery and real results, but the advertised convergence to a conditioned GFF runs ahead of the proofs, which establish moment convergence of polynomial observables only.","tokens_in":79029,"tokens_out":1537,"would_cite":true,"duration_ms":21254,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","60G15","05E10","20C30","60C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Random standard Young tableaux height fluctuations converge to a conditioned Gaussian Free Field, in three distinct models.","keywords":["Young generating function","Plancherel growth process","standard Young tableaux","Gaussian Free Field","integer partitions","central limit theorem","Gelfand–Tsetlin algebra","extreme characters of S∞"],"falsifier":"Simulate the Plancherel growth process via RSK for $n=10^4$ and times $t=0.25, 0.5, 0.75$, compute the empirical covariance of $\\sqrt{\\pi}(H(\\sqrt{n}x,nt)-EH(\\sqrt{n}x,nt))$ integrated against $x^k$, and compare with the kernel of Section 3.4; a discrepancy beyond Monte Carlo error would refute Theorem 3.6. Equivalently, for fixed $k$, compute the fourth cumulant of the observable $M^P_{\\alpha,k}$ at $n=10^6$: the paper predicts it vanishes as $n^{-2}$, and any slower decay falsifies the CLT part of the claim.","tokens_in":78049,"feed_emoji":"🎲","tokens_out":6670,"duration_ms":64833,"temperature":0.7,"pith_summary":"This paper introduces the Young generating function, a power-series object attached to any probability measure on integer partitions, and proves that simple analytic conditions on this function are necessary and sufficient for the measure to satisfy a law of large numbers and a central limit theorem. A multilevel version of the same criteria controls random increasing sequences of partitions, which are exactly random standard Young tableaux, and yields explicit covariance formulas for their height functions. Applying this machinery to three models — the Plancherel growth process, random tableaux of fixed shape, and measures induced by extreme characters of the infinite symmetric group — the paper identifies all three two-dimensional fluctuation limits as one universal object: the Gaussian Free Field conditioned to have zero mass on every horizontal slice. The conditioning is forced by the deterministic identity $\\int H(x,t)\\,dx = t$, and it distinguishes these partition models from the unconditioned GFF limits familiar from random matrix theory.","feed_headline":"Random tableaux heights converge to a conditioned Gaussian field","feed_subtitle":"A new generating function reduces partition asymptotics to one calculation, unifying three models.","key_machinery":"The load-bearing object is the Young generating function $A_\\rho(x_1,x_2,\\ldots) := M_\\rho(U_\\infty)$, defined as the image under the central character $M_\\rho$ of a universal element $U_\\infty$ of the group ring of $S_\\infty$; its logarithm is the generating function of permutation-cumulants, making it the exact partition analog of the characteristic function. The technical core is a new expansion of Biane's operator $D_k$ — realized as the trace of powers of a transposition-weighted matrix, equivalently the conditional expectation of powers of Jucys–Murphy elements — and of products of such operators inside the Gelfand–Tsetlin algebra of $S_n$. The leading coefficients of these expansions are counted by non-crossing set partitions and the Kreweras complement, via new summation identities for generalized falling factorials. These expansions convert multilevel moment computations into contour integrals whose integrands produce exactly the covariance kernel of the conditioned GFF.","core_discovery":"The paper's central claim is that the scaled height fluctuations of (i) the Plancherel growth process, (ii) uniformly random standard Young tableaux of a fixed deterministic shape, and (iii) distributions induced by rescaled extreme characters of $S_\\infty$, all converge as $n\\to\\infty$ to the same Gaussian field $\\mathcal{C}(x,t)$, whose covariance kernel is the ordinary Gaussian Free Field kernel minus a deterministic term $-\\frac{\\min(t(z),t(w))}{\\pi}\\Im(1/z)\\Im(1/w)$. In the Plancherel case the statement takes the explicit form $\\sqrt{\\pi}\\big(H(\\sqrt{n}x,nt)-EH(\\sqrt{n}x,nt)\\big) \\to \\mathcal{C}(x,t)$, where convergence holds in the sense of moments of the integrated observables $M^P_{\\alpha,k}$. Because the subtracted term is exactly what forces $\\int \\mathcal{C}(x,t)\\,dx = 0$, the paper interprets these fluctuations as those of a GFF conditioned on a single linear constraint — a constraint already present at the level of the height function itself. The theorems thereby contradict the prior expectation, based on random-matrix analogies, that these models would exhibit unconditioned GFF fluctuations.","pith_inferences":["Editorial inference: the mechanism behind the conditioning — a conserved integral of the height function surviving in the limit — suggests that any growth model whose height function has a deterministic linear constraint will produce a conditioned, not free, field; this could be tested on other Young-graph random walks or on Jack–Plancherel measures.","Editorial inference: the paper proves convergence in the sense of moments of the observables $M^P$, $M^S$, $M^{\\mathrm{Fix}}$, but not tightness of the rescaled height processes; adding tightness would upgrade the theorems to full weak convergence of the random fields, and the explicit covariance formulas make that a concrete next step.","Editorial inference: because the covariance kernel is written in closed form, one can formally compute the distribution of the field integrated against arbitrary test functions; this yields testable predictions, e.g. the variance of the field's integral over a wedge in space-time should equal a specific number that numerical RSK simulations could check."],"forward_implications":["Any probability measure on partitions whose Young generating function satisfies two analytic conditions automatically obeys an LLN and a CLT, so the criterion can be checked without constructing a coupling to particles or a determinantal process.","The multilevel CLT yields central limit theorems for statistics such as the content of the box containing $n$ in a random tableau, resolving a 2007 conjecture of Pittel and Romik.","Sending the intermediate time scale $\\alpha\\to 0$ recovers a semicircle law and the Vershik–Kerov–Logan–Shepp limit shape for sublinear random tableaux, unifying edge and bulk behavior in one formula.","The Gelfand and Schur–Weyl distributions also fall into the framework, so the conditioned GFF is not tied to the Plancherel measure but is a shared fluctuation law for a whole family of representation-theoretic distributions.","The explicit covariance kernels provide a blueprint for numerical simulation of the limiting fluctuations through standard RSK or hook-walk algorithms."],"supporting_citations":[{"why":"Introduces the operator $D_k$ and free-probability framework that the paper's operator expansions extend.","marker":"[Bi98]"},{"why":"Kerov's CLT for the Plancherel measure is the single-level baseline that the multilevel theorem generalizes.","marker":"[IO02]"},{"why":"Provides the asymptotic factorization criterion that the paper's cumulant-based criteria extend and compare against.","marker":"[Śn06b]"},{"why":"Lemma 7.5 in the paper is a deformation of their Proposition 3.13, supplying the diffeomorphism used for the fluctuation domains.","marker":"[BG18]"},{"why":"The Markov–Krein correspondence is used throughout to translate between Young diagrams and transition measures.","marker":"[Ke93]"},{"why":"Malyshev's formula for cumulants of products of permutations is used in Lemma 6.5 to decompose cumulants.","marker":"[Le04]"},{"why":"Rosas' identities for falling and rising factorial expansions are reproved and extended in Lemma 6.9.","marker":"[Ro02]"},{"why":"The Pittel–Romik conjecture on the box containing $n$ is resolved as an application of the multilevel CLT.","marker":"[PR07]"}],"fun_headline_variants":["Height fluctuations of tableaux converge to a constrained Gaussian field","Young tableaux: three models, one conditioned Gaussian limit","Tableau height fields: a single constrained GFF after scaling","All roads lead to a conditioned Gaussian free field for tableaux","Conditioned Gaussian field unifies three tableau models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proofs establish convergence of the moments of the integrated observables to the moments of the conditioned GFF; the unstated load-bearing premise is that this moment convergence forces the rescaled height process itself to converge to that field — a tightness step that the paper does not carry out.","fun_headline_variants_meta":{"raw":{"variants":["Height fluctuations of tableaux converge to a constrained Gaussian field","Young tableaux: three models, one conditioned Gaussian limit","Tableau height fields: a single constrained GFF after scaling","All roads lead to a conditioned Gaussian free field for tableaux","Conditioned Gaussian field unifies three tableau models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000689,"raw_usage":{"total_tokens":3097,"prompt_tokens":893,"completion_tokens":2204,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":2123}},"tokens_in":509,"tokens_out":2204,"duration_ms":15490,"temperature":1.0,"reasoning_tokens":2123,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:10:08.049458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the Plancherel growth process via RSK for $n=10^4$ and times $t=0.25, 0.5, 0.75$, compute the empirical covariance of $\\sqrt{\\pi}(H(\\sqrt{n}x,nt)-EH(\\sqrt{n}x,nt))$ integrated against $x^k$, and compare with the kernel of Section 3.4; a discrepancy beyond Monte Carlo error would refute Theorem 3.6. Equivalently, for fixed $k$, compute the fourth cumulant of the observable $M^P_{\\alpha,k}$ at $n=10^6$: the paper predicts it vanishes as $n^{-2}$, and any slower decay falsifies the CLT part of the claim.","supporting_citations":[],"review_version":2}