{"id":"a8fb47f3-a72e-42ef-add1-d878885b9783","arxiv_id":"2411.13959","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For pairs of measures, the Legendre transform of the bivariate Lq-spectrum can be completely irrelevant as an upper bound, and for a random correlated pair of binomial measures the bivariate multifractal formalism holds exactly when the two parameters lie on the same side of 1/2.","lead":"This paper shows that the natural extension of the Legendre spectrum to pairs of measures does not bound the bivariate multifractal spectrum, and it builds explicit measures where the two spectra have disjoint supports. It also works out the bivariate multifractal analysis of a simple random mixture of two binomial measures, giving the first exact examples where the bivariate formalism can fail.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5.3's simultaneous-neighbor claim is false, not just weak: for p1=0.1, p2=0.45, A={1,2,3,6}, w=001000 the µp1-maximizer is 000111 but νη(3I_w)=0.00209 > 3νη(I_{000111})=0.000817, so the proof of τµp1,νη=T^η in Theorem 1.3(1) has a genuine gap.","rationale":"Agreement: the reader's weakest_assumption is precisely the simultaneous dyadic-neighbour domination in Lemma 5.3, and my check confirms that this is the right place to look — but it is worse than a weak assumption, it is a false assertion. The structural reason is that νη is not homogeneous: a 1-bit at an A-generation weighs 1−p1, at an A^c-generation weighs 1−p2, and 0-bits weigh p1 or p2 correspondingly, so the word with more 1s can have smaller νη-mass when p1 ≠ p2. My explicit example (p1=0.1, p2=0.45, A={1,2,3,6}, w=001000) has unique µp1-maximizer 000111 but νη(3I_w) = 0.00209 > 3·0.00081675, contradicting the proof's key inequality. Since this inequality converts the exact undilated identity (Lemma 5.1, eτ = T^η) into the dilated statement τ = T^η, Theorem 1.3(1)-(2) is unproved as written, and the failure affects almost every realization. I stress the limits of my finding: I have not shown τ ≠ T^η. The bad words are rare (they require a specific carry pattern and specific A-membership), and their total contribution may still be subdominant; the proposed exact enumeration for j ≤ 22 would settle this. The other central results are not affected by this specific gap: Theorem 1.2 is proved by an independent construction whose disjointness of supports is robust, and Theorem 1.4 rests on a different Section 6 computation in which I did not locate a concrete flaw. The paper is valuable and the results are plausibly true; however, a knowingly false claim inside a central lemma is a real correctness problem that a referee should require to be fixed. Hence I recommend CONDITIONAL rather than ACCEPT: accept once Lemma 5.3 (or a replacement argument establishing τµp1,νη = T^η) is proved correctly.","tokens_in":38152,"tokens_out":53337,"duration_ms":422155,"concrete_test":"Computational check, exact enumeration: (1) Reproduce the counterexample: with p1=0.1, p2=0.45, A={1,2,3,6}, confirm that for w=001000 the triples of masses are as claimed and νη(3I_w) > 3νη(I_{000111}). (2) The decisive test for Theorem 1.3(1): fix one pseudorandom infinite A with density η (say η=1/2), and for j up to 22 enumerate S_j = Σ_{w∈Σ_j} µp1(3I_w)^{q1}νη(3I_w)^{q2} exactly (2^22 ≈ 4.2×10^6 terms), comparing −(1/j)log2 S_j with T^η(q1,q2) for (q1,q2) = (1,1), (2,−1), (−1,2). If the difference does not tend to 0 at rate o(1), then τµp1,νη ≠ T^η and Theorem 1.3(1) is false; if it does, the conclusion survives but Lemma 5.3's proof must be replaced by an argument controlling the carry-run exceptional words.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Lemma 5.3 (Section 5.1) asserts: when p1, p2 lie on the same side of 1/2, for every dyadic word w the same neighboring word e_w maximizes both µp1 and νη among {I_w^-, I_w, I_w^+}, because the µp1-maximizer is determined only by the counts of 0s and 1s and pi ≤ 1 − pi. On this the proof bases the estimate Σ_w µp1(3I_w)^{q1}νη(3I_w)^{q2} ≤ C Σ_w µp1(I_w)^{q1}νη(I_w)^{q2} = C·2^{−jT^η(q1,q2)}, giving τµp1,νη = T^η and hence Theorem 1.3(1)–(2). The assertion is false. Write νη(I_w) = µp1(I_w^A)µp2(I_w^{A^c}): a 1-bit contributes weight 1−p1 (resp. 1−p2) and a 0-bit weight p1 (resp. p2), depending on whether the generation lies in A. These ratios differ when p1 ≠ p2, so more 1s need not mean larger νη-mass. Explicit counterexample: take p1 = 0.1, p2 = 0.45, A = {1,2,3,6}, and w = 001000 of length 6. Then w^− = 000111, w^+ = 001001. The µp1-masses are 0.000729, 0.000009, 0.000081, so e_w = 000111 is the unique µp1-maximizer. The νη-masses are 0.00027225, 0.00018225, 0.00164025, so the νη-maximizer is w^+, and νη(3I_w) = 0.00209475 > 3νη(I_{e_w}) = 0.00081675. The inequality νη(3I_w) ≤ 3νη(I_{e_w}) used in the proof fails at this word. The A-pattern has positive probability, so the failure occurs for a.e. realization. Consequence: the map from w to the pair of individual maximizers is not shown to be bounded-to-one, so the comparison between the dilated sum (defining τµp1,νη) and the undilated sum (exactly 2^{−jT^η} by Lemma 5.1) is not justified. I have not proved the theorem false — a large-deviation argument over the exceptional long-run words might rescue it — but the proof of Theorem 1.3 as written is invalid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the bivariate multifractal formalism for pairs of probability measures. Theorem 1.2 constructs two measures for which the support of the bivariate multifractal spectrum and the support of the Legendre transform of the bivariate Lq-spectrum are disjoint, disproving the natural multivariate analogue of the univariate upper bound. Theorems 1.3 and 1.4 then analyze the pair (μp1, νη), where νη is a random binomial cascade that changes parameter at random generations. The paper claims that when p1, p2 lie on the same side of 1/2 the bivariate Lq-spectrum is T^η and the multifractal formalism holds with parallelogram support, while for p1 < 1/2 < p2 the support is a pentagon strictly larger than the Legendre pentagon and the formalism fails. The proofs use explicit step constructions, large-deviation estimates, and a dimension formula for intersections of level sets.","tokens_in":38774,"tokens_out":21108,"duration_ms":183977,"significance":"If the results are correct, Theorem 1.2 is a striking and valuable counterexample, and the random pair (μp1, νη) is a natural test case for bivariate multifractal behavior. The paper provides explicit constructions with quantitative estimates, and it reproduces the needed computation from [19] in Lemma 5.1 rather than merely citing it. However, the proof of Theorem 1.3 currently rests on false statements in Section 5, so the advertised same-side positive result is not established as written. The disjoint-support counterexample and the opposite-side analysis may still be sound, but the manuscript needs substantive repair before the central claims can be accepted.","major_comments":[{"comment":"The simultaneous-maximizer assertion in Lemma 5.3 is false, and the counterexample is not marginal. Take p1 = 0.1, p2 = 0.45, A = {1,2,3,6}, and w = 001000 of length 6. For μp1 the three neighboring masses are μp1(I_{w^-}) = 0.000729, μp1(I_w) = 0.000009, μp1(I_{w^+}) = 0.000081, so e_w = w^- is the unique μp1-maximizer. For νη the masses are νη(I_{w^-}) = 0.00027225, νη(I_w) = 0.00018225, νη(I_{w^+}) = 0.00164025, so the νη-maximizer is w^+, not w^-. Moreover νη(3I_w) = 0.00209475 > 3νη(I_{w^-}) = 0.00081675, so the inequality νη(3I_w) ≤ 3νη(I_{e_w}) used in the proof fails at this word. Since A has positive probability, such words occur almost surely at infinitely many generations. Consequently the comparison between the dilated sum defining τμp1,νη and the undilated sum 2^{-jT^η(q1,q2)} of Lemma 5.1 is not justified. The proof of τμp1,νη = T^η, and with it the verification of the multifractal formalism in Section 5.3, has a genuine gap. I have not been able to determine whether Theorem 1.3(1) itself is false; the provided proof is invalid.","section":"§5.1, Lemma 5.3"},{"comment":"The slope assertion in Lemma 5.5 is reversed. The proof states that for 0 < p1 < p2 < 1/2 the slope of G1,2 is δ1,2 > 1. But δ1,2 = (H_{2,max} - H_{2,min}) / (H_{1,max} - H_{1,min}), and for p < 1/2 the width H_{p,max} - H_{p,min} = log2((1-p)/p) decreases as p increases toward 1/2. Hence 0 < p1 < p2 < 1/2 gives δ1,2 < 1, not > 1. The later sentence 'when 0 < p2 < p1 < 1/2, then δ1,2 < 1' is also reversed: for p2 < p1 < 1/2 one has δ1,2 > 1. This is not a harmless typo: the four-case analysis and the exclusion argument for the lower endpoint a3 = a1 explicitly use the sign of δ1,2 - 1 to decide whether the minimum of H2 corresponds to the maximum or the minimum of α. With the correct sign, the case labels and the extremal computations must be interchanged. The parallelogram support may still be correct, but the proof as written does not establish it for both orderings of p1 and p2 covered by Theorem 1.3.","section":"§5.2, Lemma 5.5"}],"minor_comments":[{"comment":"The hypothesis '0 < p1 < 1/2 < p2 < 1/2' is contradictory; it should read '0 < p1 < 1/2 < p2 < 1'.","section":"Theorem 1.4"},{"comment":"The title in the text appears as 'ON THE MULTIV ARIATE MULTIFRACTAL FORMALISM', with an unwanted space; also Theorem 1.3's phrase 'both greater or both larger than 1/2' should be 'both smaller than 1/2 or both larger than 1/2'.","section":"Title and Theorem 1.3"},{"comment":"The proof assumes 0 < p1 < p2 < 1/2 'without loss of generality', but Theorem 5.4 also covers the opposite ordering; since the sign of δ1,2 - 1 changes with the ordering, the reduction should be stated explicitly.","section":"§5.2, Lemma 5.5"}],"recommendation":"major_revision","confidential_remarks":"The reader report's acceptance is too optimistic given Section 5. The false Lemma 5.3 and the reversed slope assertions in Lemma 5.5 are load-bearing for Theorem 1.3, which is one of the paper's central advertised results. I would not accept the paper until Section 5 is rewritten with a correct proof of τμp1,νη = T^η (or the theorem modified), and until the parallelogram support proof handles the correct sign of δ1,2 - 1. The rest of the paper, in particular Theorem 1.2 and the Section 6 analysis, appears substantially sound, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe stress-test note is right: Lemma 5.3 is not just missing detail, it rests on a false assertion. The reader's ACCEPT is too optimistic.\n\nThe specific problem is the simultaneous-maximizer claim. For p1, p2 on the same side of 1/2, the paper says that because pi ≤ 1−pi, the neighbor interval that maximizes µp1 also maximizes νη. That is false. Take p1=0.1, p2=0.45, A={1,2,3,6}, w=001000. Then w−=000111 maximizes µp1 (mass 0.000729 vs 0.000009 and 0.000081), but νη is maximized by w+=001001 (mass 0.00164025 vs 0.00027225 and 0.00018225). Consequently νη(3I_w)=0.00209475 > 3νη(I_{e_w})=0.00081675, contradicting the inequality used in Lemma 5.3. Since the A-pattern has positive probability, this is not a measure-zero pathology. The chain from Lemma 5.3 to τ=T^η and the formalism in Theorem 1.3 is broken as written.\n\nI have not shown the theorem itself is false—a large-deviation argument over atypical words might repair it—but it is not proved in this manuscript. Section 5 needs a substantive rewrite.\n\nThat said, the paper is not without merit. The explicit construction for Theorem 1.2—two measures whose bivariate multifractal and Legendre spectra have disjoint supports—is a clean and interesting counterexample. The random correlated binomial model is a natural object, and the Section 6 computations for p1<1/2