{"id":"de6921ff-eb9b-4b3f-9439-055e23d4d3fe","arxiv_id":"2411.09321","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"An expository Bourbaki survey presenting the 2023 proof that r(k) ≤ (4−δ)^k via the CGMS book algorithm and the Balister et al. geometric refinement lemma, with explicit marking of every non-rigorous step.","lead":"This Bourbaki seminar paper surveys the 2023 proof that diagonal Ramsey numbers r(k) are at most (4−δ)^k, the first improvement of the exponential base 4 in about 90 years. The paper presents both the book algorithm by Campos, Griffiths, Morris, and Sahasrabudhe and the newer geometric proof by Balister and collaborators, while carefully flagging which steps are rigorous and which are sketches.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 6.7's claimed non-positivity is false: f(-1,1)=1, so the proof of the geometric refinement lemma in Section 6 is invalid as written.","rationale":"I read the paper as a Bourbaki survey whose value lies in giving correct conceptual explanations of two proofs of r(k) <= (4-delta)^k. Theorems 1.5, 4.6, 4.7, and 5.11 are attributed to external works, so the survey's own mathematical claims are the expositions. The reader's weakest assumption was the 'bogus' regularity assumptions, which the author explicitly labels as unjustified; I agree these are survey-level gaps because the author discloses them. However, my check of Section 6 turned up a hidden mathematical error that the author does not disclose: Lemma 6.7 is false. This is a more serious problem because the entire geometric proof of Lemma 5.3 rests on it. It is not merely a missing technicality; a displayed assertion is wrong, and the subsequent argument explicitly relies on that assertion. I would not reject the survey outright, since the underlying theorems are established elsewhere and much of the exposition is sound, but the paper should be conditionally revised: either correct Lemma 6.7 and the surrounding proof, or explicitly state that the proof of Theorem 6.2 as presented is flawed and refer to the original source for a correct argument. For this reason my recommendation is CONDITIONAL rather than UNCHANGED.","tokens_in":35798,"tokens_out":4857,"duration_ms":49889,"concrete_test":"Directly evaluate f at (-1,1) with f defined in Eq. (6.8). The result is 1, contradicting the asserted conclusion f<=0 on {y<=-1 or z<=-1}. If a corrected version of the lemma exists, replace Lemma 6.7 with the corrected function and re-check the two required properties: non-positivity on the complement of the good event E and the growth bound (6.9). Without such a fix, the derivation of Eq. (6.11) and hence Theorem 6.2 fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The survey's central exposition includes presenting Balister et al.'s refinement lemma (Lemma 5.3) and its geometric proof. Section 6.3 reduces this to Theorem 6.2 and then to Lemma 6.7. Lemma 6.7 asserts that f(y,z)=1+y(2+cosh(sqrt(2)z))+z(2+cosh(sqrt(2)y)) is non-positive on the set {y<=-1 or z<=-1}. This is false: f(-1,1)=1-(2+cosh(sqrt(2)))+(2+cosh(sqrt(2)))=1. The proof immediately uses the claimed non-positivity to discard the complement of E={both inner products >= -1}, obtaining E[V 1_E] <= 0 and hence E[V] <= E[V 1_E]. With f(-1,1)>0, V can be positive on the complement of E, so the transfer of E[V]>=1 to the desired probability is unjustified. This is not one of the self-flagged 'bogus' assumptions (Assumptions 3.1, 4.4, 5.1, 5.2), nor is it listed as a sketch; Section 6 presents the geometric lemma as the one remaining missing piece and omits only a 'calculus exercise.' The theorem of Balister et al. may still be true, but this survey's proof of the key refinement lemma is not correct as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This Bourbaki survey presents the history of upper bounds on diagonal Ramsey numbers and explains the recent breakthrough of Campos, Griffiths, Morris, and Sahasrabudhe that r(k) ≤ (4−δ)^k, together with the later alternative proof by Balister et al. The paper gives three views of the Erdős–Szekeres bound, introduces the book algorithm with density-boost steps, sketches the off-diagonal refinement that yields r(k) ≤ 2^{(2−3/200+o(1))k}, and then describes the symmetric book algorithm of Balister et al., whose key refinement lemma is reduced to a high-dimensional geometric statement about inner products. The exposition is explicitly conditional: Assumptions 3.1, 4.4, 5.1, and 5.2 are admitted to be unjustified, and several numerical inequalities are verified by plots rather than rigorous computations.","tokens_in":36273,"tokens_out":6547,"duration_ms":63295,"significance":"If the proofs were correct, this would be a valuable and readable account of a major recent development in Ramsey theory. The paper is unusually honest about which parts are rigorous and which are sketches; the proof of the Erdős–Szekeres theorem, the one-color geometric lemma (Proposition 6.3), and Ramsey's book proof are clean and self-contained. The survey also gives useful context, including Thomason's conjecture and Conlon's theorem on book Ramsey numbers. However, the proof of the central geometric lemma contains a false statement, and the main theorems of Sections 4 and 5 are derived under assumptions that the paper itself labels bogus. As a result, the survey's account of the new proofs is not correct as written.","major_comments":[{"comment":"The asserted non-positivity of f is false. For y = −1 and z = 1, f(−1,1) = 1 − (2 + cosh(√2)) + (2 + cosh(√2)) = 1, which is positive. The proof of Theorem 6.2 uses this non-positivity to conclude that V is non-positive on the complement of E and hence E[V 1_E] ≤ 0. With f(−1,1) > 0, that conclusion is unjustified, and the derivation of the key estimate (6.11) collapses. This is not one of the explicitly flagged 'bogus' assumptions; it is presented as a calculus exercise. The lemma must be corrected or the proof of Theorem 6.2 must be supplied by a different argument, since this is the load-bearing step in the presented proof of the refinement lemma and hence of Theorem 5.11.","section":"Section 6.3, Lemma 6.7"},{"comment":"The derivations of Theorems 4.6, 4.7, and 5.11 depend on the bi-regularity and initial-density assumptions that the paper itself describes as 'bogus' and 'completely unjustified.' The promised cleaning argument is never given. For example, Lemma 4.5 and the proof of Theorem 4.6 rely on |Y| ≥ (1−μ)^{t+s+o(k)}N, which in turn requires that blue steps keep p constant and that red steps shrink Y by a factor of p; both facts are consequences of Assumption 3.1. In an expository survey, a clearly labeled conditional sketch would be acceptable, but the paper states Theorem 4.7 as a theorem and Theorem 5.11 as a proof, with only parenthetical remarks saying the proof is incomplete. Please either supply or cite the exact cleaning argument from the original papers, or restate the central theorems as conditional on these assumptions.","section":"Sections 3–5, Assumptions 3.1, 4.4, 5.1, 5.2"},{"comment":"The proof of Lemma 5.3 uses the identity |Y'| = p_R|Y| and the analogous identity for Z, citing Assumption 5.1. However, Lemma 5.3 is stated without that assumption, and Lemma 6.1 is also stated as a general statement about arbitrary sets X,Y,Z in a coloring. Thus Lemma 5.3 is not proved as stated; it is proved only under an additional regularity hypothesis that is not part of the lemma's hypotheses. Either the lemmas should be stated with the required regularity assumptions explicitly included, or a proof without those assumptions should be supplied.","section":"Section 6.1, proof of Lemma 5.3"},{"comment":"Step 4(a) says to perform a red density-boost step if d_R(X',Y') ≥ (κ^2−1)p_R, but Lemma 5.3(c) and Table 5.1 require the condition d_R(X',Y') ≥ p_R + (κ^2−1)α_R. As written, the algorithm's condition is inconsistent with the later analysis: for large κ, (κ^2−1)p_R can be much larger than p_R, and the stated update rule would not follow from the refinement lemma. This appears to be a typo, but it is load-bearing because the subsequent Lemma 5.9 uses the density increase guaranteed by the correct condition.","section":"Section 5.2, Algorithm 5.4, step 4(a)"}],"minor_comments":[{"comment":"The final sentence says the coloring contains a monochromatic K_t, but Lemma 2.4 with m = r(k−t,k) produces a monochromatic K_k; the displayed bound is for r(k), so 'K_t' should be 'K_k'.","section":"Section 5.4, proof of Theorem 5.11"},{"comment":"The numerical verification that max min{F̂, G_{2/5}} < 0.985 is presented via contour plots. For a survey this may be acceptable, but the theorem is stated with a definite numerical constant; a reference to the rigorous computation in the original papers would make the claim verifiable.","section":"Section 4.3, Theorem 4.7"},{"comment":"The claim that this recovers Theorem 2.2 'up to the subexponential error term' is slightly loose, since the bound obtained is N < (1−γ)^{−k}γ^{−ℓ}, which matches (k+ℓ choose ℓ) only up to polynomial factors; the statement could be made more precise.","section":"Section 2.3, Algorithm 2.6"}],"recommendation":"major_revision","confidential_remarks":"The false Lemma 6.7 is the main technical problem; it is a concrete counterexample, not a matter of interpretation. The paper's self-flagged unjustified assumptions might be tolerable for a Bourbaki survey if clearly labeled, but the central geometric proof is presented as complete and is not. I would not reject outright, because the survey could be repaired by replacing or correcting the function f in Lemma 6.7 and by citing the cleaning arguments from the original papers. However, the current version should not be accepted as an accurate exposition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this survey. First, it is a genuinely good expository piece: the historical framing is crisp, the book algorithm is explained with unusual clarity, and the author is commendably upfront about the places where the sketched proofs rely on unjustified assumptions, flagging them as 'bogus' and 'cheating.' Second, there is a real mathematical error in the proof of Lemma 6.7, and it is not one of the self-flagged soft spots.\n\nThe paper's actual value is expository. It gives a clean proof of the Erdős–Szekeres bound, a readable account of the CGMS book algorithm and the off-diagonal refinement, and a conceptual presentation of the Balister et al. alternative. The author repeatedly and honestly says when a proof is not rigorous, which is exactly what a good survey should do. The self-contained parts (Erdős–Szekeres, the book algorithm analysis, Ramsey's book proof) are correct as far as I checked.\n\nNow the flaw. Lemma 6.7 claims the function f(y,z)=1+y(2+cosh(√2 z))+z(2+cosh(√2 y)) is non-positive on {y≤−1 or z≤−1}. That is false: f(−1,1)=1. The proof of Theorem 6.2 then uses this non-positivity to discard the complement of the event E, and that step is unjustified. This is not a minor typo; it breaks the proof of the geometric lemma as written. The underlying theorem (Balister et al.) may well be true, and the survey elsewhere attributes the result properly, but this particular self-contained proof is invalid.\n\nHow much does this matter? For a Bourbaki survey, the goal is to explain, not to prove every theorem from scratch. The error is localized to one lemma in Section 6, and the author could fix it by either supplying a correct proof (there may be a sign issue in the definition of f) or by explicitly marking Lemma 6.7 as a sketch and referring to the original paper. As it stands, the paper should not be published with this lemma stated as proved.\n\nThe other soft spots—Assumptions 3.1, 4.4, 5.1, 5.2—are honestly labeled as unjustified, so they are not hidden defects. The numerical verifications in Section 4 are admittedly picture-based, but the author says so. The one unflagged error is the real problem.\n\nRecommendation: send this to a serious referee, yes. The survey is valuable and the exposition is mostly excellent, but the referee should insist on correcting or reclassifying Lemma 6.7 before acceptance. I would not cite it in my own work until that is fixed.","headline":"A genuinely useful Bourbaki survey with a real but localized bug in the proof of the geometric lemma—worth fixing, not rejecting.","tokens_in":36726,"tokens_out":1596,"would_cite":false,"duration_ms":19077,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05D10","05C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the first time since 1935, the upper bound on the diagonal Ramsey number has been pushed below $4^k$, and this survey explains both proofs.","keywords":["diagonal Ramsey numbers","upper bounds","book algorithm","book graphs","Ramsey theory","refinement lemma","geometric lemma","multicolor Ramsey numbers"],"falsifier":"A direct way to test the central claim is to search for a two-coloring of $K_N$ with $N = 3.994^k$ and no monochromatic $K_k$; any such coloring for even one large $k$ would falsify the bound. Short of that, the proof mechanism can be tested by simulating the book algorithm on colorings whose red degrees between $X$ and $Y$ are deliberately non-regular: if the cleaning step cannot restore approximate regularity without shrinking $Y$ by more than $2^{-o(k)}$, the parameter tables in the survey no longer support the claimed conclusion.","tokens_in":35490,"feed_emoji":"📖","tokens_out":13563,"duration_ms":116429,"temperature":0.7,"pith_summary":"This survey explains a 90-year-old breakthrough: the diagonal Ramsey number $r(k)$, the least $N$ such that every red/blue coloring of the edges of $K_N$ contains a monochromatic $K_k$, is at most $(4-\\delta)^k$ for some constant $\\delta>0$, concretely $3.993^k$ for large $k$. The first proof route, the book algorithm, is presented as a refined version of the Erdős–Szekeres argument, and the survey shows exactly why the naive version stalls near $4.15^k$ before two corrections bring it below $4$. The second route, due to a later group, is more symmetric and rests on a geometric lemma about correlations of random vectors; it yields $r(k) \\le 2^{(2-\\eta^2/10)k}$ with $\\eta=1/8000$. The survey is explicit that its sketches rely on regularity assumptions it calls bogus and that a rigorous proof needs a cleaning step that is not carried out here.","feed_headline":"Ramsey upper bound finally breaks the 4^k barrier","feed_subtitle":"A 90-year-old barrier falls: a survey walks through the book algorithm and the geometric proof of r(k) ≤ 3.993^k.","key_machinery":"The load-bearing object is the monochromatic book graph $B_{t,m}$: a $K_t$ spine with $m$ pages, all connected to the spine. Both proofs grow such books through an algorithm that moves vertices between sets $A,B,X,Y$ (and $Z$ in the symmetric proof) and tracks how the sizes and the red/blue densities evolve; the step-by-step parameter tables are what convert the algorithm into a bound on $N$. In the symmetric proof the crucial mechanism is the refinement lemma, proved by a geometric argument: centered indicator vectors of red and blue neighborhoods are fed into the function $f(y,z)=1+y(2+\\cosh\\sqrt{2z})+z(2+\\cosh\\sqrt{2y})$, chosen so that all Taylor coefficients are nonnegative, so the expectation of $f$ over independent random vertices is at least $1$; the lemma then converts a large value of $f$ into a simultaneous lower bound on the two inner products. The quadratic dependence of the density boost on $\\kappa$ is what makes the density-boost steps negligible in the final bookkeeping.","core_discovery":"The central claim, stated on the paper's own terms, is that $r(k) \\le (4-\\delta)^k$ is a theorem and that both known proofs are sound in structure: the book algorithm of the 2023 proof and the symmetric book algorithm of the later proof. The first proof maintains a red book $(A,Y)$ and a blue book $(B,X)$, sacrificing $X$ to build $Y$; its success reduces to a numerical maximum of two functions $F$ and $G$, and the missing margin is supplied by an off-diagonal bound $r(k,\\ell) \\le 2^{-2\\ell/9+o(k)}\\binom{k+\\ell}{\\ell}$ for $\\ell \\le k/4$. The second proof adds a third set $Z$ and applies a refinement lemma that, for any two families of vectors, finds a vertex whose red and blue neighborhoods have a large simultaneous correlation; the density boost from this lemma is quadratic in the shrinkage parameter, which is exactly what makes the parameter tracking close. The survey presents both proofs as sketches and repeatedly flags that the regularity assumptions used to write the parameter tables are unjustified as stated.","pith_inferences":["The unproved cleaning step is the most delicate point to formalize: any cleaning that shrinks $Y$ by more than $2^{-o(k)}$ between steps would degrade the constant below $3.993$.","A polynomial strengthening of the geometric lemma, where only an exponential bound is currently known, would likely improve the exponent in the final bound, since the present proof loses a factor $e^{-4\\kappa}$ in the probability estimate.","The numerical-inequality route invites a computational search over the cutoff $\\mu$; the paper notes that optimizing the constants yields substantially stronger bounds, so a systematic optimization could push the base below the current values."],"forward_implications":["The exponential growth rate of $r(k)$ is strictly below the base $4$, so the Erdős–Szekeres constant is not the asymptotic truth.","For $\\ell \\le k/4$, the off-diagonal bound improves the binomial estimate by an exponential factor $2^{-2\\ell/9+o(k)}$, which is the ingredient that closes the gap in the diagonal argument.","The symmetric proof extends to multicolour Ramsey numbers, giving new upper bounds for any fixed number of colors.","Because the book algorithm either finds a monochromatic $K_k$ or produces a book whose parameters force one, the proof is in principle algorithmic."],"supporting_citations":[{"why":"Supplies Theorem 1.5 and the book algorithm, the first proof that $r(k) \\le (4-\\delta)^k$.","marker":"Campos, Griffiths, Morris, and Sahasrabudhe (2023)"},{"why":"Supplies the symmetric book algorithm, the geometric refinement lemma, and the bound $r(k) \\le 2^{(2-\\eta^2/10)k}$.","marker":"Balister et al. (2024)"},{"why":"Establishes the baseline bound $r(k) \\le 4^k$ and the binomial off-diagonal bound used throughout the argument.","marker":"Erdős and Szekeres (1935)"},{"why":"Gives rigorous bounds that validate the survey's unrigorous numerical claims in Theorems 4.6 and 4.7.","marker":"Gupta, Ndiaye, Norin, and Wei (2024)"},{"why":"Introduces book graphs and Lemma 2.4, the framework through which both proofs convert book existence into a Ramsey bound.","marker":"Thomason (1982)"},{"why":"Provides the original theorem and the first upper bound $r(k) \\le k!$, and the epilogue's book-based proof.","marker":"Ramsey (1929)"}],"fun_headline_variants":["Ramsey upper bound finally drops below 4^k","Survey explains 3.993^k Ramsey breakthrough after 90-year gap","Two proofs crack the 4^k barrier for diagonal Ramsey numbers","New bound 3.993^k for Ramsey numbers ends decades-long standstill","How the 4^k barrier fell: a survey of the Ramsey breakthrough"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that at every step the red edges between $X$ and $Y$ are exactly regular, meaning every vertex in $X$ has the same number $p|Y|$ of red neighbors in $Y$ and every vertex in $Y$ has the same number $p|X|$ of red neighbors in $X$. The survey calls this assumption completely unjustified and bogus, and says that a cleaning step deleting outlier vertices can make it approximately true, but the cleaning is never actually proved.","fun_headline_variants_meta":{"raw":{"variants":["Ramsey upper bound finally drops below 4^k","Survey explains 3.993^k Ramsey breakthrough after 90-year gap","Two proofs crack the 4^k barrier for diagonal Ramsey numbers","New bound 3.993^k for Ramsey numbers ends decades-long standstill","How the 4^k barrier fell: a survey of the Ramsey breakthrough"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1836,"prompt_tokens":975,"completion_tokens":861,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":770}},"tokens_in":591,"tokens_out":861,"duration_ms":9210,"temperature":1.0,"reasoning_tokens":770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:46:53.006915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct way to test the central claim is to search for a two-coloring of $K_N$ with $N = 3.994^k$ and no monochromatic $K_k$; any such coloring for even one large $k$ would falsify the bound. Short of that, the proof mechanism can be tested by simulating the book algorithm on colorings whose red degrees between $X$ and $Y$ are deliberately non-regular: if the cleaning step cannot restore approximate regularity without shrinking $Y$ by more than $2^{-o(k)}$, the parameter tables in the survey no longer support the claimed conclusion.","supporting_citations":[],"review_version":1}