{"id":"3a27ec98-bbea-46eb-a548-21c3335d7c9d","arxiv_id":"2508.05839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Functions built from intersections of set families satisfy a strong hypergraph regularity lemma, and the two known sources of ternary instability both satisfy this regularity, so strong 2-stability cannot be defined by excluded hypergraphs.","lead":"This paper proves a strong regularity theorem for functions that measure overlaps of sets indexed by multiple coordinates, and shows that two known counterexamples to ternary stability share a hidden regularity property. The result changes how model theorists and combinatorists think about higher-order stability, since it cannot be characterized by a single forbidden pattern.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.1's proof does not establish the strong F(b) error bound: a one-shot hypergraph regularity application can only give a fixed error η(ε'), while b grows like a tower in 1/ε', so for fast-decaying F no choice of ε' works; an iterative refinement step is missing.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the specific weakness named (the infinite Ω gap) is not the most load-bearing: the Ω issue can be repaired by a finite approximation, since for each e there are only finitely many measurable sets P^e_{\\vec{x}}. The stronger problem is that the proof uses a single application of hypergraph regularity and hence can only control the exceptional set by an absolute parameter η(ε'), not by the prescribed function F of the final partition size. Standard regularity bounds make this impossible for arbitrary F without an iterative strong-regularity construction. This directly affects the central claim of Theorem 2.1, and consequently the paper's advertised strong regularity for averages of hypergraphs. Section 3's Theorem 3.9 is a separate construction, but it relies on the framework, so the overall verdict remains CONDITIONAL until the F(b) step is supplied. I agree with the reader's net assessment but differ on where the critical gap lies.","tokens_in":32719,"tokens_out":23154,"duration_ms":268843,"concrete_test":"Specialize to k=2,d=1. Derive from the proof an explicit upper bound η(ε') for the exceptional density on a rectangle after one application of graph regularity with parameter ε' to P^e, and let B(ε') be the standard bound on the number of parts (e.g., the tower bound in [RS04] or [Gow07]). Test whether there exists ε'>0 with η(ε') < 2^{-B(ε')}. If, using any reasonable η(ε') = C ε'^α yielded by the averaging argument, no such ε' exists, then the proof as written cannot establish Theorem 2.1 for F(n)=2^{-n}; the theorem would require an additional iterative refinement step, which the paper does not provide.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 2.1 (Section 2), hypergraph regularity is applied once to each P^e, producing partitions with sizes b_e ≤ B(ε') and with all but ε' of the cells ε'-regular. The subsequent averaging argument shows that on each cylinder intersection C, f is ε-constant outside a set Z of density at most η(ε'), where η(ε') depends only on ε' (a small power of ε' after the quantifier exchanges and Fubini steps). The theorem requires η(ε') < F(max_e b_e), with F an arbitrary prescribed function. Standard hypergraph regularity bounds give B(ε') at least an iterated exponential in 1/ε'. For F(n)=2^{-n}, the target is 2^{-B(ε')}, which is far smaller than any power of ε' as ε'→0. The proof contains no iterative refinement step (unlike the defect argument in [AFN07] or the strong-stability machinery in [MS14]), and the phrase 'with suitable parameters ε'≪ε and F'≪F' is not justified by the cited finite hypergraph regularity lemmas, which do not take an error function F'. Thus the claimed strong dependence on F is not obtained. This is distinct from the infinite-Ω measurability issue: even if Ω were finite, the same F(b) obstruction remains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a higher-arity analogue of the stable regularity lemma for functions defined as measures of intersections of families of measurable sets, e.g. f(x,y,z)=μ(P_{x,y}∩Q_{x,z}∩R_{y,z}). Theorem 2.1 claims a strong hypergraph regularity statement: for every ε>0 and every function F:N→(0,1], there is N such that for any finite sets X_i and any probability space Ω, one can partition each d-fold product into at most N parts so that on every nonexceptional cylinder intersection the function is ε-constant outside a set of density < F(max b_e). Corollary 2.2 extends this to integrals of continuous combinations of arbitrary measurable functions. The second part of the paper addresses higher-arity model-theoretic stability. After relating perfect d-stable regularity to strong d-stable regularity (Prop. 3.2) and to Terry–Wolf's binary disc_{2,3} error (Prop. 3.3), the authors prove Theorem 3.9: every finite 3-hypergraph embedding both into the half-simplex and into GS_3 satisfies perfect 2-stable regularity. Hence strong ternary stability cannot be characterized by a single excluded family of hypergraphs.","tokens_in":33030,"tokens_out":25672,"duration_ms":287922,"significance":"If the results hold, they are significant. Theorem 2.1 would give a regularity lemma far stronger than ordinary hypergraph regularity for a natural class of 'average hypergraph' functions, and Corollary 2.2 would make the statement quite robust. The second part addresses an explicit open question from Terry and Wolf about strong 2-stability; Theorem 3.9 is a substantial direct construction using monotonicity and the GS_3 structure, rather than a routine adaptation of known machinery. The paper also makes a useful conceptual contribution by separating perfect, strong, and ordinary regularities in higher arity. However, the proof of the main finitary theorem is not complete in its current form, and one auxiliary lemma used in Corollary 2.2 is only proved in a special case. The paper is well organized and clearly written in its broad lines, but the missing technical steps are load-bearing.","major_comments":[{"comment":"The asserted dependence on F is not established. The proof applies hypergraph regularity once, obtaining partitions of size b_e ≤ B(ε') and an error η(ε') that is a power of ε'. In standard hypergraph regularity, B(ε') is at least an iterated exponential in 1/ε'. The averaging argument in the proof can at best show that on each cylinder C the bad set has density O(η(ε')). The theorem requires this density to be < F(max_e b_e) for an arbitrary prescribed F. For example, for F(n)=2^{-n}, the requirement is η(ε') < 2^{-B(ε')}, which cannot be satisfied by any power of ε' as ε'→0. The phrase 'with suitable parameters ε'≪ε and F'≪F' is not justified: the cited finite hypergraph regularity lemmas have no functional parameter F'. No iterative refinement step is supplied. This is not a minor bookkeeping gap; the size-dependence of the error is exactly the 'strong' part of the theorem. The proof","section":"Section 2, proof of Theorem 2.1"},{"comment":"The proof applies hypergraph regularity to the (d+1)-uniform hypergraph P^e with vertex set (∏_{i∈e} X_i) × Ω, where Ω is an arbitrary probability space, possibly infinite. The cited regularity lemmas ([RS04], [Gow07], [Tao06]) are for finite hypergraphs. Since the X_i are finite, the family {P^e_x} is finite for each e, so one can reduce Ω to the finite Boolean algebra generated by these sets before invoking the finite regularity lemma. But this reduction is not stated, and without it the claimed application of the finite hypergraph regularity lemma is not justified. This issue is fixable, but it is a genuine missing step in the current proof.","section":"Section 2, proof of Theorem 2.1, definition of P^e"},{"comment":"Corollary 2.2 says that a common refinement of the partitions obtained from Theorem 2.1 is taken 'using Remark 3.4'. However, Remark 3.4 is proved only in the setting of 3-partite 3-hypergraphs with d=1: its proof uses pair sets {u,v}, vertex partitions, and graph regularity. The statement of Remark 3.4 for general d and k is not proved, and it is not a formal consequence of Theorem 2.1 alone. The common refinement of finitely many partitions requires a lower bound on the ratio μ(C)/μ(C_r) for sub-cylinders; without a general form of Remark 3.4, the error bounds F'(b_r)|C_r| do not transfer to the refined cylinder with the desired F(B). Thus Corollary 2.2 currently rests on an unproved generalization.","section":"Corollary 2.2 and Remark 3.4"},{"comment":"The construction of R^{u,v} says we 'add sR sets to R^{u,v} by induction on length n∈N∗'. Since N∗ is a nonprincipal ultrapower of N, it is not well-ordered, so this is not a legitimate external induction principle. If the intended meaning is that the selected sR sets are those minimal under containment, the text should say so and prove that a minimal element exists for every sR set, or explain why chains without minimal elements are covered by the later L sets. Claim 4 proves pairwise disjointness of the selected sets, but it does not address the existence of selected minimal elements in descending chains. This is a load-bearing point in the construction of the partitions P^{u,v} for Theorem 3.9.","section":"Section 3.2, Definition 3.13 and Claim 4"}],"minor_comments":[{"comment":"The notation 'For every 1≤d<k∈N' is nonstandard; it should be written as 'For every k∈N and every 1≤d<k'.","section":"Theorem 2.1 statement"},{"comment":"The symbol μ is used both for the probability measure on Ω and for the normalized counting measure on products of the finite sets X_i. This overload is confusing, especially in the displayed equations involving μ(C^e), μ(C_{[k]}), and μ(P^e∩C^e).","section":"Section 2, proof of Theorem 2.1"},{"comment":"The phrase 'For an ordinal κ∈ω∪{ω}' is confusing; presumably the intended set is κ∈ω+1 (all ordinals ≤ω).","section":"Definition 3.6"},{"comment":"The cross-reference 'Claim 3(3a)' should be 'Claim 3(2a)': Claim 3 has parts (1) and (2), with subparts (a)–(c) under (2).","section":"Remark 3.8(3)"},{"comment":"The paper relies heavily on the authors' unpublished preprints [CT20] and [CT24b] for definitions, the graded probability space framework, and Proposition 3.2. This is acceptable, but the dependence should be flagged clearly, and the referee report should note that the current manuscript is not self-contained in these places.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper has two independent-looking strands. The second strand, Theorem 3.9, is a substantial and inventive construction, though I did not fully verify every combinatorial claim. The first strand, Theorem 2.1, currently has a real gap in the strong F(b) bound and in the use of infinite Ω; these need a substantive new argument or a weakening of the statement. Since the abstract and the claimed significance of the first part rest on that theorem, I cannot recommend acceptance in the present form. I would suggest the editor ask for a repaired proof of Theorem 2.1, or an explicit statement of a weaker fixed-error version, before further review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First take: this paper has a genuine new idea and probably true results, but Theorem 2.1 is not fully proved as written. The gap is load-bearing, not cosmetic.\n\nWhat is actually new: Theorem 2.1 and Corollary 2.2 claim a strong hypergraph regularity statement for measures of intersections, with the exceptional density controlled by an arbitrary function of the partition size. That is genuinely stronger than ordinary hypergraph regularity. Theorem 3.9, saying that strong 2-stability cannot be characterized by excluded hypergraphs, is new and directly addresses Terry and Wolf's question. The embedding argument using monotone hypergraphs and GS3 is clever, and the model-theoretic discussion is honest about what remains open.\n\nWhere the soft spots are. First, the proof of Theorem 2.1 applies hypergraph regularity to the (d+1)-uniform hypergraph P^e whose last vertex class is Ω, an arbitrary probability space. Standard hypergraph regularity lemmas are finite. No measure-theoretic version is stated or cited. That might be patchable, but it is not in the paper. Second, and more seriously: the proof wants error density < F(max b_e) for arbitrary F. Ordinary hypergraph regularity applied once gives a fixed error η(ε′), while b grows like a tower in 1/ε′. For fast-decaying F, say F(n)=2^{-n}, the target is far smaller than any power of ε′. The sentence \"apply hypergraph regularity with suitable parameters ε′≪ε and F′≪F\" is not backed by a stated strong hypergraph regularity lemma. The stress-test concern is correct: even if Ω were finite, this step would not follow from the cited lemmas. The authors either need to prove an iterative/strong hypergraph regularity lemma or substantially revise the argument.\n\nI also want to flag that Proposition 3.2, used in the model-theoretic part, comes from their own unpublished preprint [CT24b]. Self-citation is not a problem in itself, but it means part of the paper depends on unrefereed work. The Section 3 proof is long and intricate; I did not find an obvious flaw in the R-set/C-set partition argument, but it is too dense to fully certify in a desk read.\n\nBottom line: this deserves a serious referee. The right outcome is not a desk reject but a strong referee request, asking for a complete proof of Theorem 2.1 with the infinite-Ω issue handled and the F(b) step either proved or replaced by an explicit strong regularity theorem. I would not cite the current version yet, but I expect the ideas to survive revision.","headline":"A real result and a real gap: the advertised strong regularity theorem is only sketched, and the F(b) error bound does not follow from the cited lemma; still deserves peer review.","tokens_in":33544,"tokens_out":5564,"would_cite":false,"duration_ms":67973,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C65","03C45","05D10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Functions that measure intersections of set families satisfy a strong hypergraph regularity lemma, and the two known counterexamples to ternary stability share a hidden regularity that blocks excluded-substructure characterizations.","keywords":["hypergraph regularity","higher arity stability","strong 2-stability","averages of hypergraphs","perfect regularity lemma","monotone hypergraphs","half-simplex","GS3"],"falsifier":"A direct refutation would exhibit $k=3,d=2$, some $\\varepsilon>0$ and $F:\\mathbb{N}\\to(0,1]$, and families $P_{x,y},Q_{x,z},R_{y,z}$ of measurable sets such that for every $N$ there is some finite $X,Y,Z$ and $\\Omega$ with no partitions of the three pair-products into $\\le N$ parts having exceptional density $<\\varepsilon$, and cylinder intersections on which $f=\\mu(P\\cap Q\\cap R)$ varies by more than $\\varepsilon$ outside any set of density $<F(b)$. For Theorem 3.9, a refutation would be one sequence of finite 3-partite 3-hypergraphs, each an induced subhypergraph of both $GS_3$ and the half-","tokens_in":32583,"feed_emoji":"🔺","tokens_out":9134,"duration_ms":86709,"temperature":0.7,"pith_summary":"Functions such as $f(x,y,z)=\\mu(P_{x,y}\\cap Q_{x,z}\\cap R_{y,z})$ — measures of intersections of families of measurable sets indexed by overlapping coordinates — are not arbitrary high-arity functions: the paper proves they satisfy a very strong hypergraph regularity lemma. The partition of each $d$-fold product can be chosen so that every cylinder intersection set, outside a set of density controlled by an arbitrary prescribed $F(b)$, is nearly constant. The proof views these functions as averages of hypergraphs over a probability space and applies hypergraph regularity to auxiliary $(d+1)$-uniform hypergraphs. The second part shows that the two known sources of failure of ternary stability — the half-simplex and $GS_3$ — cannot jointly exclude anything: every finite 3-hypergraph that embeds into both still satisfies perfect 2-stable regularity. Hence strong 2-stability cannot be characterized simply by forbidden induced subhypergraphs.","feed_headline":"Averages of set intersections obey strong hypergraph regularity","feed_subtitle":"k-ary functions built from intersecting measurable set families are nearly constant on every cylinder cell outside a small exceptional set.","key_machinery":"The key objects are the auxiliary $(d+1)$-uniform hypergraphs $P^e=\\{(x_e,z): z\\in P^e_{x_e}\\}$ on $(\\prod_{i\\in e} X_i)\\times\\Omega$. For Theorem 2.1, the paper applies the hypergraph regularity lemma to each $P^e$, intersects the resulting regular partitions to form 'cells', 'faces' and 'bases', and then uses quasi-randomness of the $P^e$ on faces to show, by successive slicing over an ordering of the $d$-subsets, that the intersection-measure function is approximated by a product of face densities, uniformly in most points. For Theorem 3.9, the machinery is a direct partition construction on pairs of coordinates using the monotone order inherited from the half-simplex embedding: it isolat","core_discovery":"On the paper's own terms, the central discovery is Theorem 2.1: for any $1\\le d<k$, $\\varepsilon>0$ and $F:\\mathbb{N}\\to(0,1]$, there is $N(d,k,\\varepsilon,F)$ such that every function $f(x_1,\\dots,x_k)=\\mu\\left(\\bigcap_{e\\in\\binom{[k]}{d}} P^e_{x_e}\\right)$ admits partitions $\\bigsqcup_{i\\le b_e} S_{e,i}$ of each $\\prod_{i\\in e} X_i$ with $b_e\\le N$, exceptional part of density $<\\varepsilon$, and on every cylinder intersection $\\bigcap_e S_{e,j_e}$ with each $j_e>0$, $f$ is contained in an interval of length $<\\varepsilon$ except on a set of density $<F(\\max_e b_e)$. Corollary 2.2 extends the same conclusion to $f(\\vec x)=\\int h\\left((f^e_{\\vec x_e}(z))_{e}\\right)d\\mu(z)$ for continuous $h","pith_inferences":["Inference: If the infinite-$\\Omega$ step can be supplied, the same strong regularity should hold uniformly for definable families in arbitrary probability algebras, giving a model-theoretic 'perfect regularity' companion for higher arities analogous to stable graph regularity.","Inference: The partition built in Theorem 3.9 is constructive and does not pass through a combinatorial characterization of strong 2-stability; this suggests that the right 'forbidden configuration' picture for strong $k$-stability may involve ordered or continuous configurations rather than finite hypergraphs.","Inference: The $R$/$sR$ rectangle decomposition used in the proof may yield explicit bounds on the number of parts that are far smaller than iterated hypergraph regularity; this quantitative question is not addressed in the paper.","Inference: Because the parity example in Remark 3.5 shows that $NFOP_k$ functions can lack slice-wise NIP, the hierarchy of higher-arity tameness notions branches; one testable extension is whether the half-simplex/$GS_3$ common class is exactly the class of functions with perfect 2-stable regularity, which the paper leaves open."],"forward_implications":["Every function of the form $f(x_1,\\dots,x_k)=\\mu\\left(\\bigcap_e P^e_{x_e}\\right)$ satisfies strong $d$-stable regularity: partitions of bounded size, small exceptional part, and near-constancy up to a density-$F(b)$ error.","The same conclusion holds for integrals $\\int h\\left((f^e_{\\vec x_e}(z))_e\\right)d\\mu(z)$ with $h$ continuous, so the class of well-behaved functions is closed under continuous combinations of smaller-arity measurable functions.","Every finite 3-hypergraph that embeds into both $GS_3$ and the half-simplex has perfect 2-stable regularity, hence strong 2-stable regularity.","Strong 2-stability is not characterized by a single excluded finite induced subhypergraph, partially addressing a conjecture in the literature on ternary stability.","The functions considered in Theorem 2.1 are $NFOP_k$ and $NOP_2$ (for $k=3,d=2$) but are not slice-wise NIP, so they occupy an intermediate position in the higher-arity tameness hierarchy."],"supporting_citations":[{"why":"Provides the counting lemma for regular $k$-uniform hypergraphs used in the quasi-randomness estimates of Theorem 2.1.","marker":"[NRS06]"},{"why":"Supplies the hypergraph regularity lemma for $k$-uniform hypergraphs applied to each auxiliary hypergraph $P^e$.","marker":"[RS04]"},{"why":"Gives a variant of hypergraph regularity and counting that underpins the successive slicing argument.","marker":"[Gow07]"},{"why":"Provides the hypergraph regularity/counting framework used in refining the partitions.","marker":"[Tao06]"},{"why":"Defines the half-simplex, $GS_p$, and binary $\\mathrm{disc}_{2,3}$ error, and proves these hypergraphs fail strong 2-stability; it supplies the target notion for Theorem 3.9.","marker":"[TW21b]"},{"why":"Establishes the non-embedding of the half-simplex into $GS_3$ and related structural facts needed to show no single excluded hypergraph works.","marker":"[TW21a]"},{"why":"Proves that functions of the form $\\mu(P_x\\cap Q_y)$ are stable, the binary base case that Theorem 2.1 generalizes to higher arity.","marker":"[Hru12]"},{"why":"Supplies the stable graph regularity lemma whose strong form motivates the density-$F(b)$ strengthening in Theorem 2.1.","marker":"[MS14]"},{"why":"Provides the graded probability space setup and the Aldous-Hoover-Kallenberg style presentation that frame the higher-arity setting.","marker":"[CT20]"},{"why":"Distinguishes perfect, strong, and ordinary regularity and gives the example showing perfect regularity is strictly stronger.","marker":"[CT24b]"}],"fun_headline_variants":["Hypergraph regularity holds for higher-arity averages","Strong regularity for averages of set intersections","Higher-arity stability without excluded hypergraphs","Averages of intersections satisfy robust hypergraph regularity","No excluded hypergraphs needed for strong ternary stability"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The proof of Theorem 2.1 applies hypergraph regularity to a $(d+1)$-uniform hypergraph whose last vertex set is an arbitrary probability space $\\Omega$, which may be infinite, while the cited regularity lemmas are stated for finite hypergraphs; the paper does not prove the needed infinite-vertex version with the stated error bounds.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph regularity holds for higher-arity averages","Strong regularity for averages of set intersections","Higher-arity stability without excluded hypergraphs","Averages of intersections satisfy robust hypergraph regularity","No excluded hypergraphs needed for strong ternary stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1153,"prompt_tokens":766,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":319}},"tokens_in":510,"tokens_out":387,"duration_ms":4475,"temperature":1.0,"reasoning_tokens":319,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:06:47.353523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct refutation would exhibit $k=3,d=2$, some $\\varepsilon>0$ and $F:\\mathbb{N}\\to(0,1]$, and families $P_{x,y},Q_{x,z},R_{y,z}$ of measurable sets such that for every $N$ there is some finite $X,Y,Z$ and $\\Omega$ with no partitions of the three pair-products into $\\le N$ parts having exceptional density $<\\varepsilon$, and cylinder intersections on which $f=\\mu(P\\cap Q\\cap R)$ varies by more than $\\varepsilon$ outside any set of density $<F(b)$. For Theorem 3.9, a refutation would be one sequence of finite 3-partite 3-hypergraphs, each an induced subhypergraph of both $GS_3$ and the half-","supporting_citations":[],"review_version":1}