{"id":"07380ce5-ce10-46fb-8ddb-363e3e1cfc82","arxiv_id":"1908.04403","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The scaling limits of uniform connected graphs with fixed surplus and of random unicellular maps are built from tilted Brownian trees by identifying points at common heights sampled via local time.","lead":"The paper constructs limiting random graph spaces by starting with a random tree and gluing pairs of leaves at equal heights. This gives a new breadth-first view of the critical Erdős-Rényi graph limit and of random unicellular maps, yielding formulas for their radii and distance profiles.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5.5's joint local-time convergence is the load-bearing step; Appendix A only sketches it, so Theorem 3.1 is conditional on completing that proof.","rationale":"The central claim is the distributional identity H(s) = H^BF(s). The proof in Section 5.2 factors through bijections between maps and admissible-corner pairs, and then proves convergence of the BF model via Propositions 5.1 and 5.2. The key probabilistic input in Section 5.3 is Proposition 5.5, which gives the joint convergence of the rescaled contour process and the two-parameter local time field. This is genuinely load-bearing: the sampling of heights and times in Construction 2.2 is expressed in terms of the local time field, and the convergence of the quotient spaces uses the joint convergence in the sup-norm topology. The Appendix A proof is explicitly only an outline. The strong-invariance step for the unconstrained walk is standard, but the bridge conditioning, the endpoint tightness (A.11), and the Vervaat transform of the local-time field are asserted rather than proved. In particular, the Vervaat transform is not continuous in the sup-norm topology at all paths, although it is continuous at paths with a unique minimum; this requires a separate argument. These are technical gaps, not contradictions: the statements are plausible, and the authors' overall strategy is coherent. The paper includes independent support, such as the enumeration in Section 5.7 and the known convergence (2.3) from prior work. Therefore the appropriate verdict remains CONDITIONAL: the result is likely correct, but the proof of Proposition 5.5 must be completed or cited before Theorem 3.1 can be regarded as fully proved.","tokens_in":37063,"tokens_out":13418,"duration_ms":133302,"concrete_test":"Complete the proof of Proposition 5.5 by writing out two steps currently only asserted in Appendix A: (i) prove (A.11) using the stated time-reversal symmetry and the local central limit theorem, with explicit estimates showing the limsup is arbitrarily small; (ii) show that the first argmin of the bridge converges in distribution to the unique argmin of Brownian bridge and that the Vervaat transform of the pair (Sbr_n, ℓbr_n) converges to (e, η(e;·,·)) in C[0,1]×C([0,1]×R), for instance by establishing continuity of the Vervaat map at paths with a unique minimum. If these cannot be completed, locate a published theorem giving exactly (5.14) for conditioned plane trees. A numerical check that (2n)^{-3/2}B(C_n) converges in distribution to ∫ η(e;1,y)^2 dy for n=10^5 and 10^6 would provide supporting evidence but would not settle the proof gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.1 asserts H(s) d= H^BF(s). The proof passes through Proposition 5.1, whose proof in Section 5.3 uses Proposition 5.5 to get joint convergence of the rescaled contour process and its two-parameter local-time field, (C̄_n, L̄_n) ⇒ (e, η(e;·,·)) in C[0,1]×C([0,1]×R). This joint convergence is used at (5.18), (5.21), (5.22), and (5.24) to identify the limiting heights and times of the horizontal identifications in Construction 2.2, and in the final GHP convergence argument after (5.25). Without it, the breadth-first construction's discrete approximation has no demonstrated limit. Appendix A is an outline: it cites strong invariance [19, Thm F] for the unconstrained random walk, derives (A.9) for the bridge restricted to [0,1−ε], asserts (A.11) to pass to the full bridge, and then says the Vervaat transform 'follows' from (A.10). The Vervaat step needs joint convergence of the argmin with the path, and the endpoint tightness (A.11) is not proved. These are fillable gaps, but they are load-bearing because the local-time field is the exact input to the sampling of identifications in Construction 2.2. The paper also relies on the unproved convergence (2.9) for CRUM(g), but Theorem 3.1 does not depend on that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes continuum 'breadth-first' constructions for two families of random measured R-graphs: the scaling limits H(s) of uniform connected graphs with fixed surplus s, which include the critical Erdős-Rényi scaling limit, and the continuum random unicellular maps CRUM(g) of fixed genus. In Construction 2.2, a tilted Brownian excursion e^BF_(s) is sampled, s heights are drawn from the squared total local time, and pairs of leaves at those heights are identified using the local-time measure; in Construction 2.3 an analogous procedure is carried out with a random permutation in S(g). The main theorems assert H(s) = H^BF_(s) and CRUM(g) = CRUM^BF_(g). The proof proceeds through discrete breadth-first and depth-first explorations of uniform maps (Section 5.1), the discrete approximations in Propositions 5.1 and 5.2, and a joint convergence of the rescaled contour process with its two-parameter local-time field (Proposition 5.5, with proof outlined in Appendix A). The paper also derives corollaries for the radius, the two-point function, and the distance profile, together with a result on convergence of distance profiles (Theorem 3.4).","tokens_in":37325,"tokens_out":10241,"duration_ms":94744,"significance":"If the main theorems are valid, the paper gives the first breadth-first construction of the critical Erdős-Rényi scaling limit, answering a question raised in [2], and provides explicit descriptions of the radius, two-point function, and distance profile of H(s) and CRUM(g) that are not immediate from the depth-first or core-decomposition constructions. The combinatorial encodings in Section 5.1, the tilted discrete models (5.7)-(5.8), and the uniform-integrability arguments around (5.18)-(5.19) are careful and plausible. The main analytic input, namely the joint convergence of contour process and local-time field, is, however, only sketched, and the convergence (2.9) defining CRUM(g) is assumed without a written proof; consequently the theorems as stated are conditional on completing these parts.","major_comments":[{"comment":"Proposition 5.5 is load-bearing for Proposition 5.1: equations (5.18), (5.21), (5.22), and (5.24) all use the joint convergence (Cbar_n, Lbar_n) to (e, eta(e;.,.)) to identify the limiting heights and times of the identifications in Construction 2.2. The proof in Appendix A is an outline: after (A.9), the passage to the full bridge is deferred to (A.11), which is asserted as immediate from time reversal without proof, and the Vervaat transform step says only that (A.1) follows from (A.10). Please supply a complete proof of (A.11), including control of sup_y (ellbar^br_n(1,y)-ellbar^br_n(1-epsilon,y)), and of the joint convergence of the argmin with the path and the local-time field under the Vervaat transform, or cite a published result containing the full statement.","section":"Appendix A, Eq. (A.11) and the Vervaat step"},{"comment":"The convergence (2.9) that defines CRUM(g) is not proved in the manuscript; the text states that a proof 'seems' not to be in the literature and 'can be deduced' by following [1] or [2]. Since Theorem 3.5 is a statement about this space, the equality CRUM(g) = CRUM^BF(g) is conditional on an unproved existence and identification result. Please either prove (2.9) or restate Theorem 3.5 and Corollary 3.6 explicitly as conditional on (2.9).","section":"Section 2.2 and Theorem 3.5, Eq. (2.9)"},{"comment":"The completion of the proof of Theorem 3.5 is omitted with 'We omit the details as no new idea is involved here.' The convergences in (5.71)-(5.72) concern the tilted contour process, the permutation, the heights, and the times, but the theorem is about pointed GHP convergence of the glued metric measure spaces. Please provide the quotient-space convergence argument, including an analogue of the correspondence argument after (5.25) and control of the measure under the identifications.","section":"Section 5.7, after (5.71)-(5.72)"},{"comment":"In the proof of Lemma 5.3, the claim that the pointed GHP distance between (2n)^{-1/2} G^o_{n,s} and Gbar_{n,s} tends to 0 is declared routine and the details are omitted. This step converts the line-measure quotient convergence into the convergence of the vertex-measure graph spaces, and it is part of the proof of Proposition 5.1; please spell it out or refer to an existing lemma that covers it.","section":"Section 5.3, Eq. (5.27)"}],"minor_comments":[{"comment":"The display 'We set bf(G)=t and bf(G)=bar{t}' uses the same symbol bf for both the breadth-first tree and its symmetrization; please use distinct notation, for example overline{bf}(G), throughout that argument.","section":"Section 5.6"},{"comment":"The normalization constant in (2.11) is derived only later in (5.70); a forward reference would help the reader understand why the expression defines a probability measure.","section":"Section 2.2, Eq. (2.11)"},{"comment":"In the use of Jeulin's identity, the equality in distribution in (5.50) and the statement 'jointly with' (5.51) should specify the coupling used; the current wording is slightly ambiguous about whether the two identities hold jointly with the same e.","section":"Section 5.5, Eq. (5.49)"},{"comment":"In the sentence after (5.4), the symbol M_{n,s} is used both for the uniform map and for the rooted metric measure space with the non-root vertices carrying mass 1/n; please make the passage to the metric measure space explicit, for instance by writing (M_{n,s}, d, root, mu).","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is attractive and the combinatorial bijections appear sound, but the current version is not fully self-contained at the analytic core. In particular, Proposition 5.5 is the pivot for Theorem 3.1 and its proof is only sketched; the Vervaat and endpoint-tightness steps in Appendix A need to be supplied. The status of (2.9) also needs to be made explicit. I would be willing to look at a revision that fills these gaps. The paper is within the scope of the journal and would be a significant contribution once the conditional parts are either proved or clearly flagged as assumptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper does something new: it gives breadth-first constructions of the H(s) spaces and of CRUM(g), answering an explicit question from Addario-Berry, Broutin and Goldschmidt. Second, the main theorems are very likely true, but the proof leans on two results that are not written down in full, so I would send it to a careful referee rather than accept on faith.\n\nWhat I like: the tilted excursion plus local-time gluing is a clean idea, and the paper makes good on the promise of the breadth-first view by reading off radii, distance profiles, and two-point functions from functionals of the tilted excursion. The route through maps and contour functions is a clever way to get the local-time convergence that would be missing for labeled trees. The discrete arguments in Section 5 are detailed and, modulo the input results, the coupling in Lemmas 5.4 and 5.9 is convincing. The authors are also honest that (2.9) for CRUM(g) lacks a published proof.\n\nThe soft spots are the two inputs just mentioned. Proposition 5.5 -- joint convergence of the rescaled contour process and its local-time field -- is the step that carries the discrete approximation in Lemma 5.3. Appendix A is only a sketch. In particular, the Vervaat transform step needs joint convergence of the argmin with the path, and the endpoint tightness (A.11) is asserted rather than proved. These are probably routine for the experts, but they are load-bearing: without them, the sampling of heights and times in Construction 2.2 has no demonstrated limit. The second gap is (2.9); it underpins Theorem 3.5. The authors note the issue but do not supply a proof. If the referee asks for those two pieces to be completed or precisely referenced, the paper would be in good shape.\n\nWho this is for: anyone working on critical random graphs, continuum random trees, or maps. It will be cited. My verdict: deserve a real peer review. I would not desk-reject it.","headline":"A genuinely new breadth-first construction with two load-bearing technical gaps that deserve a careful referee.","tokens_in":37896,"tokens_out":2847,"would_cite":true,"duration_ms":29406,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60C05","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"Uniform connected graphs with a fixed surplus admit a breadth-first scaling limit built from a tilted Brownian tree by gluing leaves at random heights.","keywords":["Erdős-Rényi random graph","critical random graphs","scaling limit","continuum random tree","breadth-first construction","depth-first construction","unicellular maps","Gromov-Hausdorff distance"],"falsifier":"Compute the joint limit in $C([0,1]\\times\\mathbb{R})$ of $((2n)^{-1/2}C_n(2nt), (2n)^{-1/2}L_n(2nt,y\\sqrt{2n}))$ for uniform plane trees; if this pair fails to converge to $(e,\\eta(e;\\cdot,\\cdot))$ along some subsequence, Proposition 5.5—and with it the convergence of the gluing heights and times—is false. A simpler check: simulate large critical Erdős-Rényi components, measure the distance profile around a uniformly chosen vertex, and compare with the claimed limit $\\tfrac12\\eta(e^{\\mathrm{BF}}_{(s)};1,r/2)$; a systematic mismatch would refute Corollary 3.2(iii) and Theorem 3.4.","tokens_in":36815,"feed_emoji":"🌳","tokens_out":10386,"duration_ms":90392,"temperature":0.7,"pith_summary":"Every connected graph with n vertices and s extra edges is a tree plus s cycles, and this paper proves that the scaling limit of a uniformly random such graph, including the critical Erdős-Rényi component scaling limit, can be built by a breadth-first rule. The rule starts from a Brownian excursion tilted by the s-th power of its total squared local time, samples s heights according to that squared local time, and glues pairs of leaves sitting at each sampled height. The same style of construction gives the continuum random unicellular map of genus g by gluing leaves in pairs according to a random permutation. If the main theorems hold, the radius, the distance between two uniform points, and the distance profile of these limiting random spaces are read off directly as explicit functionals of the tilted excursion. The paper's central distributional equalities are $H(s) \\stackrel{d}{=} H^{\\mathrm{BF}}_{(s)}$ and $\\mathrm{CRUM}(g) \\stackrel{d}{=} \\mathrm{CRUM}^{\\mathrm{BF}}_{(g)}$.","feed_headline":"Random graph limits get a breadth-first construction","feed_subtitle":"Uniform connected graphs with s surplus edges arise by gluing leaves of a tilted Brownian tree at random heights.","key_machinery":"The load-bearing machinery pairs a discrete exploration with a continuum convergence statement. On the discrete side, the breadth-first exploration of a uniformly random map with fixed surplus produces a uniform plane tree together with admissible corner pairs, counted by the breadth-first weight $B(f)=\\sum_i B(f;i)$, where $B(f;i)$ is the number of later corners at height equal to or one below the current height. The continuum limit of this weight is $2\\int_0^\\infty \\eta(e;1,y)^2\\,dy$, exactly the tilt used in Construction 2.2. The passage to the limit rests on Proposition 5.5, the joint convergence in $C([0,1]\\times\\mathbb{R})$ of the rescaled contour process $(2n)^{-1/2}C_n(2nt)$ and the two-parameter local-time field $(2n)^{-1/2}L_n(2nt,y\\sqrt{2n})$ to $(e,\\eta(e;\\cdot,\\cdot))$, together with Gaussian tail bounds that make the tilted discrete weights uniformly integrable. Jeulin's local-time identity $\\int_0^\\infty \\eta(e;1,y)^2\\,dy \\stackrel{d}{=} 2\\int_0^1 e(t)\\,dt$ connects this breadth-first tilt to the area tilt of the depth-first Construction 2.1, which is why the two constructions can describe the same space.","core_discovery":"Theorem 3.1 asserts that for every $s \\ge 0$, the scaling limit $H(s)$ of uniform connected labeled graphs with $s$ surplus edges has the same distribution as a space $H^{\\mathrm{BF}}_{(s)}$ built as follows: sample a Brownian excursion $e^{\\mathrm{BF}}_{(s)}$ tilted by the $s$-th power of $\\int_0^\\infty \\eta(e;1,y)^2\\,dy$, sample heights $H_1,\\dots,H_s$ with density proportional to squared local time, and at each sampled height identify two independent leaves of the encoded tree, then double all distances. Theorem 3.5 proves the analogous statement for the continuum random unicellular map $\\mathrm{CRUM}(g)$: it equals the space obtained by gluing together $4g$ leaves in pairs dictated by a random transposition structure. The breadth-first spanning tree of the limiting space is then simply the tilted Brownian tree itself. Consequently the radius of $H(s)$ is $2\\|e^{\\mathrm{BF}}_{(s)}\\|_\\infty$, the two-point distance is $2e^{\\mathrm{BF}}_{(s)}(U)$ for an independent uniform $U$, and the distance profile is half the local time of the tilted excursion.","pith_inferences":["Beyond the paper, the equality $H(s) \\stackrel{d}{=} H^{\\mathrm{BF}}_{(s)}$ suggests a continuum self-duality between depth-first and breadth-first explorations of the same random metric space; comparing the two codings could yield new identities for Brownian excursion functionals.","Because the construction glues only leaves at equal heights, the diameter of $H(s)$ should be expressible as twice the largest height at which two independent local-time samples lie in different branches of the tilted excursion, a quantity the paper does not compute.","The assumed convergence (2.9) defining $\\mathrm{CRUM}(g)$ is an input to Theorem 3.5; if an independent proof of that convergence appeared, the breadth-first description of $\\mathrm{CRUM}(g)$ would follow without further work.","The same gluing scheme could be attempted for the $\\alpha$-stable analogues of random graphs discussed in the paper's final section, provided a two-parameter local-time convergence of the kind conjectured there holds."],"forward_implications":["If Theorem 3.1 is correct, the scaling limit of critical Erdős-Rényi random graphs—and the wider family of mean-field random graph models it governs—admits a breadth-first construction, answering the question posed in the paper's introduction.","The radius of $H(s)$ has the law of $2\\|e^{\\mathrm{BF}}_{(s)}\\|_\\infty$; for $s \\ge 1$ this also equals $\\int_0^1 dt/e^{\\mathrm{DF}}_{(s)}(t)$, extending the classical height identity for the Brownian continuum random tree.","The two-point function of $H(s)$ is $2e^{\\mathrm{BF}}_{(s)}(U)$ with $U$ uniform and independent, and the distance profile around the root is $\\tfrac12\\eta(e^{\\mathrm{BF}}_{(s)};1,r/2)$.","The rescaled number of vertices at distance $\\lfloor r\\sqrt{n}\\rfloor$ from the root of $H_{n,s}$ converges in Skorokhod $J_1$ topology to that local-time profile, extending the known height-profile convergence for random trees to graphs with surplus.","The same radius, two-point, and distance-profile formulas hold for $\\mathrm{CRUM}(g)$ with the tilted excursion $e^{\\mathrm{UM}}_{(g)}$ replacing $2e^{\\mathrm{BF}}_{(s)}$."],"supporting_citations":[{"why":"Introduces the space $H(s)$ and the depth-first construction of the critical Erdős-Rényi scaling limit, and poses the breadth-first question answered here.","marker":"[2]"},{"why":"Introduces the continuum random unicellular map $\\mathrm{CRUM}(g)$ and its core-decomposition construction, which Theorem 3.5 re-derives breadth-first.","marker":"[3]"},{"why":"Supplies sub-Gaussian tail bounds for the height and width of conditioned Galton-Watson trees, used to prove uniform integrability of the tilted discrete weights.","marker":"[5]"},{"why":"Provides the coding of the Brownian continuum random tree by Brownian excursion and the convergence of conditioned tree contour functions to the excursion.","marker":"[8]"},{"why":"Provides the strong-invariance approximation for local times of random walks that underlies the joint convergence claim in Proposition 5.5.","marker":"[19]"},{"why":"Gives the height-profile convergence for random trees that Theorem 3.4 extends to graphs with surplus.","marker":"[22]"},{"why":"Shows that depth-first processes of Galton-Watson trees converge to the same Brownian excursion, used in the proof of the depth-first proposition.","marker":"[35]"},{"why":"Provides the generating-function enumeration of rooted maps by genus, used to count unicellular maps and to compute the normalization in Construction 2.3.","marker":"[38]"}],"fun_headline_variants":["Tilted Brownian trees glue to form graph limits","Scaling limits built by gluing Brownian tree leaves","Erdos-Renyi limit gets breadth-first construction","Random unicellular maps from tilted Brownian trees","Gluing leaves at random heights yields graph limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the joint convergence of the rescaled contour process and the two-parameter local-time field of uniform plane trees to the Brownian excursion and its local time (Proposition 5.5), a convergence the paper only sketches; if it fails in the required topology, the heights and times of the identifications in Construction 2.2 need not converge to the claimed limit, and the existence of the $\\mathrm{CRUM}(g)$ limit is assumed as a separate input.","fun_headline_variants_meta":{"raw":{"variants":["Tilted Brownian trees glue to form graph limits","Scaling limits built by gluing Brownian tree leaves","Erdos-Renyi limit gets breadth-first construction","Random unicellular maps from tilted Brownian trees","Gluing leaves at random heights yields graph limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001179,"raw_usage":{"total_tokens":4871,"prompt_tokens":941,"completion_tokens":3930,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":3853}},"tokens_in":557,"tokens_out":3930,"duration_ms":24176,"temperature":1.0,"reasoning_tokens":3853,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:43:15.104198+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the joint limit in $C([0,1]\\times\\mathbb{R})$ of $((2n)^{-1/2}C_n(2nt), (2n)^{-1/2}L_n(2nt,y\\sqrt{2n}))$ for uniform plane trees; if this pair fails to converge to $(e,\\eta(e;\\cdot,\\cdot))$ along some subsequence, Proposition 5.5—and with it the convergence of the gluing heights and times—is false. A simpler check: simulate large critical Erdős-Rényi components, measure the distance profile around a uniformly chosen vertex, and compare with the claimed limit $\\tfrac12\\eta(e^{\\mathrm{BF}}_{(s)};1,r/2)$; a systematic mismatch would refute Corollary 3.2(iii) and Theorem 3.4.","supporting_citations":[{"cited_title":"Addario-Berry, N","cited_arxiv_id":null,"evidence_quote":"Introduces the space $H(s)$ and the depth-first construction of the critical Erdős-Rényi scaling limit, and poses the breadth-first question answered here."},{"cited_title":"Addario-Berry, O","cited_arxiv_id":null,"evidence_quote":"Introduces the continuum random unicellular map $\\mathrm{CRUM}(g)$ and its core-decomposition construction, which Theorem 3.5 re-derives breadth-first."},{"cited_title":"Addario-Berry, L","cited_arxiv_id":null,"evidence_quote":"Supplies sub-Gaussian tail bounds for the height and width of conditioned Galton-Watson trees, used to prove uniform integrability of the tilted discrete weights."},{"cited_title":"Aldous, The continuum random tree III, Ann","cited_arxiv_id":null,"evidence_quote":"Provides the coding of the Brownian continuum random tree by Brownian excursion and the convergence of conditioned tree contour functions to the excursion."},{"cited_title":"Csáki and P","cited_arxiv_id":null,"evidence_quote":"Provides the strong-invariance approximation for local times of random walks that underlies the joint convergence claim in Proposition 5.5."},{"cited_title":"Drmota and B","cited_arxiv_id":null,"evidence_quote":"Gives the height-profile convergence for random trees that Theorem 3.4 extends to graphs with surplus."},{"cited_title":"Marckert and A","cited_arxiv_id":null,"evidence_quote":"Shows that depth-first processes of Galton-Watson trees converge to the same Brownian excursion, used in the proof of the depth-first proposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generating-function enumeration of rooted maps by genus, used to count unicellular maps and to compute the normalization in Construction 2.3."}],"review_version":1}