{"id":"4431f3f5-1678-4917-8ac1-6c7a586da31f","arxiv_id":"2607.29429","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For both Steinhaus and Rademacher random multiplicative functions, almost surely |sum_{n≤x} f(n)| ≪ sqrt(x)(log log x)^{1/4+ε}, matching Harper's lower bound.","lead":"This paper proves that partial sums of Steinhaus and Rademacher random multiplicative functions almost surely stay below sqrt(x)(log log x)^{1/4+ε}. This matches the known lower bound and settles Harper's conjecture on the sharp logarithmic exponent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof imports Caich's deterministic smoothing inequality (Prop. 3.15, eq. 3.20) without proof; a failure or convention mismatch there would invalidate Proposition 5.4 and hence Theorem 1.","rationale":"The reader's weakest-assumption analysis identifies exactly the point I would stress: Proposition 3.15's (3.20) is imported from Caich's unpublished preprint and is not proved in this manuscript. The paper proves the auxiliary estimates (3.21) and the block-moment estimate Proposition 5.2, but the deterministic smoothing inequality is the hinge connecting the principal mean square to V_ell. Both the Steinhaus and Rademacher proofs rely on it (Prop. 5.4 and Prop. 7.3), so this single unproved input controls the theorem's conclusion. I agree with the reader that the right verdict is conditional rather than accept: the concern is specific and load-bearing, but it is not a demonstrated error, and the paper's own contributions (the fixed conditional r-th moment, the stopping argument, the Rademacher transfer) are argued with care. No change to the reader's verdict is needed; the recommended action is to require an independent verification of Caich's inequality, ideally by including a full proof or a machine-checked version in a revision.","tokens_in":28699,"tokens_out":6564,"duration_ms":65501,"concrete_test":"Independently re-derive (3.20) from the definitions in Sections 3.4-3.6 without invoking [5]. In particular, expand |Psi'_f(x_i/p,p)|^2 via the block decomposition and verify the smoothing step for the final block y_{j-1}<x_i<y_j, where terms with p>x_i are absent. As a local check, specialize to J=1 and compute both sides of (3.20) for a fixed small ell and an explicit Steinhaus realization on the primes in (y_0,y_1]; if any boundary term or convention mismatch appears, the inequality needs modification. A positive result restores the proof; a negative one leaves Theorem 1 unproved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 3.15's pointwise inequality (3.20) is the only bridge from the principal logarithmic mean square M_ell to the quadratic variation V_ell of the largest-prime martingale. The proof in the paper establishes only the auxiliary probability bounds (3.21); for (3.20) it says it is 'the deterministic estimate obtained from [5, Section 7.1]' and gives no derivation. This is load-bearing: Proposition 5.4 uses (3.20) to conclude V_ell(x_i;f)/x_i << T(ell)L^{-1/2+1/r}, and the final martingale step (Lemma 3.4) then controls M_f^{(1)} at the claimed exponent. If Caich's inequality has a hidden hypothesis, or if the strict smoothness convention Psi'_f used in V_ell and W_ell is not the convention under which Caich's equation (24) was derived, the quadratic-variation bound and hence Theorem 1 fail. The same inequality is transferred verbatim to the Rademacher case (Prop. 7.3), so the issue is not model-specific. This is an omitted proof of a central deterministic estimate, not a mere citation gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that for both a Steinhaus and a Rademacher random multiplicative function f, and for every ε>0, almost surely |M_f(x)| ≪_{ε,f} √x (log log x)^{1/4+ε}. Together with Harper's almost sure lower bound, this indeed determines the logarithmic exponent and would settle Harper's conjecture. The proof follows Caich's largest-prime-factor decomposition and smoothing architecture, but replaces Caich's conditional first-moment step by a conditional r-th moment estimate on a single prime block (Proposition 5.2), which removes a √(log log x) loss. The main new technical content is the supermartingale tilt (Lemma 5.1), the fixed conditional block moment (Proposition 5.2), the combined stopping argument (Proposition 5.3), and the transfer to quadratic variation (Proposition 5.4). The Rademacher case is treated in Section 7 by parallel arguments with a shifted low-moment estimate.","tokens_in":29038,"tokens_out":6225,"duration_ms":66040,"significance":"If the proof is complete, the result is a major contribution: it settles the almost sure upper bound at the conjectured exponent 1/4 for both random multiplicative function models, matching Harper's lower bound. The paper's main new idea, a fixed conditional r-th moment estimate for one prime block, is natural and appears to give exactly the saving needed to compensate the union over the J ≍ log log x prime blocks. The exposition is generally careful, with explicit constants and many estimates verified in detail. The largest caveat is that the deterministic smoothing inequality at the heart of the argument is imported without proof from an unpublished preprint; see the major comments.","major_comments":[{"comment":"The pointwise smoothing inequality (3.20) is the only bridge from the principal logarithmic mean squares M_ℓ to the quadratic variation V_ℓ of the largest-prime martingale. In the proof of Proposition 3.15 the authors state only that this inequality is 'the deterministic estimate obtained from [5, Section 7.1]' and give no derivation. This is load-bearing: Proposition 5.4 uses (3.20) to obtain V_ℓ(x_i;f)/x_i ≪ T(ℓ)L^{-1/2+1/r}, and the final martingale step (Lemma 3.4) then produces the claimed exponent. The same unproved inequality is transferred verbatim to the Rademacher case in Proposition 7.3. Because [5] is a preprint and the strict/weak smoothness conventions and final-prime-block range are precisely the points where a mismatch could occur, the paper should either prove (3.20) in an appendix or state it as a theorem with a complete proof, explicitly verifying the conventions used","section":"§3.6, Proposition 3.15, Eq. (3.20)"},{"comment":"The Rademacher analogue of the smoothing inequality is asserted by saying that 'the deterministic decomposition leading to (3.20) uses the largest-prime representation ... and deterministic inequalities for the resulting smooth sums,' and applying it to (7.2). This inherits the same gap as the Steinhaus case: the pointwise inequality (7.3) is not proved. In particular, the auxiliary estimates (3.21) are verified, but the principal inequality itself is not. Since the Rademacher theorem depends on (7.3) exactly as the Steinhaus theorem depends on (3.20), this is a second load-bearing unproved ingredient, not a model-specific technicality.","section":"§7.1, Proposition 7.3"}],"minor_comments":[{"comment":"The notation 'X_ℓ = exp(2ℓK)' is easy to misread. From the subsequent display 'log y_0 = 2^{ℓ^K - Kℓ^{K-1}}' it is clear that X_ℓ = exp(2^{ℓ^K}) and L=ℓ^K, but the superscript is not visible in the typeset formula. Please re-set these displays so that the exponent is unambiguous.","section":"§3.3"},{"comment":"The range in Lemma 3.13 is stated uniformly for y_0 ≤ v ≤ y_J(1+1/X). It may be worth adding a short comment that the estimate remains valid when the interval (v/(1+1/X), v] extends beyond y_J, since the final block is truncated at x_i in the application.","section":"§3.6, Lemma 3.13"},{"comment":"In the quotation of Harper's estimate, the variable q is reused both as the exponent and later as a prime; although the local convention is clear, the notation is overloaded. A different symbol for the exponent would improve readability.","section":"§4, Lemma 4.1"},{"comment":"The conditional Doob inequality is applied to the martingale q ↦ Ψ_f(z,q) along Q_j. The text does not explicitly justify that the index set Q_j, which contains y_{j-1} and primes, is well-ordered and that the endpoint y_{j-1} is included; this is true, but a brief comment would help the reader.","section":"§5, proof of Proposition 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly serious and the new conditional block-moment argument is promising. However, the referee report must emphasize that Proposition 3.15 is not proved in the paper and is the single most important technical estimate. If the authors can provide a full proof of (3.20), or otherwise make the dependence on Caich's preprint completely transparent and verifiable, the result should be publishable. The present form is not self-contained enough for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper proves the (log log x)^{1/4+epsilon} almost sure upper bound for both Steinhaus and Rademacher random multiplicative functions, matching Harper's lower bound and settling his conjecture. If correct, it replaces Caich's 3/4 exponent and extends Mastrostefano's 1/4 from the restricted largest-prime sum to the full sum. That is a real result, not an incremental step.\n\nThe genuinely new ingredient is Proposition 5.2, a fixed conditional r-th moment estimate for one prime block, and it deserves credit. It kills the sqrt(log log x) loss that a conditional first/second moment step would incur, and the argument is careful, with explicit constants. The paper is detailed, the Rademacher case is handled seriously (the shifted low-moment estimate is nontrivial), and I see no circularity or fitted parameters. Harper's lower bound is used only to state sharpness, not in the upper-bound proof.\n\nThe soft spot, which the stress-test note identifies correctly, is Proposition 3.15. The pointwise inequality (3.20) is the entire bridge from the logarithmic mean square M_ell to the quadratic variation V_ell of the largest-prime martingale, and it is not proved. The paper says it is the deterministic estimate from Caich's unpublished preprint [5, Section 7.1], then proves only the auxiliary probabilistic bounds (3.21). This is load-bearing: Proposition 5.4 and Theorem 1 depend on (3.20) holding for the final prime block and under the strict smoothness convention used here. If Caich's inequality has a hidden hypothesis or a convention mismatch, the theorem fails. Caich's preprint is public and the authors thank him, so it is not an unverifiable black box, but it is an omitted proof of a central deterministic claim. A referee will need to verify the derivation and the convention alignment carefully. I did not machine-check the arithmetic, and that is a second reason to keep the verdict conditional.\n\nThis is not a desk-reject. It is an important paper that deserves serious refereeing. The recommendation: send it to review, and direct the referee to certify Prop 3.15 against [5] or ask the authors to include a proof in a later version.","headline":"Settles Harper's conjecture conditionally, but the load-bearing smoothing inequality is imported from an unpublished preprint without proof.","tokens_in":29500,"tokens_out":2692,"would_cite":true,"duration_ms":29713,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11N64","11K65","60G42","60G44"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that, almost surely, partial sums of Steinhaus and Rademacher random multiplicative functions are bounded by √x(log log x)^{1/4+ε} for every ε>0, matching the lower bound and settling the conjecture.","keywords":["random multiplicative functions","almost sure bounds","martingales","hypercontractivity","partial sums","Steinhaus model","Rademacher model","law of the iterated logarithm"],"falsifier":"Verify the smoothing inequality (3.20) on the deterministic function f≡1 (or any fixed assignment of prime values) for large x and the final prime block. The theorem's proof would collapse if for some ε and unbounded x the quadratic variation V_ℓ(x;1)/x exceeded the claimed right-hand side by a factor growing with x.","tokens_in":28629,"feed_emoji":"🔢","tokens_out":8246,"duration_ms":79966,"temperature":0.7,"pith_summary":"This paper proves that, for either a Steinhaus or a Rademacher random multiplicative function, the partial sums up to x are almost surely bounded by √x times (log log x)^{1/4+ε} for every ε>0. Because a matching lower bound was previously known, this determines the sharp logarithmic exponent in both models and settles the conjecture on the size of large fluctuations. The proof decomposes the sum by the largest prime factor, reduces the problem to controlling the quadratic variation of a prime-indexed martingale, and obtains the required saving through a new conditional r-th moment estimate on a single prime block.","feed_headline":"Random multiplicative sums almost surely match the 1/4 exponent","feed_subtitle":"Both Steinhaus and Rademacher models have the same sharp almost sure upper bound, settling the conjecture.","key_machinery":"The largest-prime decomposition writes M_f(x) as a y0-smooth sum, a prime-indexed martingale M^(1)_f(x) = ∑_{y0<p≤y_J} f(p)Ψ'_f(x/p,p), and a repeated-prime term (absent in the Rademacher model). The proof then uses a deterministic smoothing inequality, quoted from an unpublished preprint, to bound the quadratic variation of this martingale by a principal logarithmic mean square M_ℓ(x_i,y_j;f) plus auxiliary errors. The new device is the conditional block moment estimate E[U_j^r | F_{y_{j-1}}] ≪_r I_{j-1}^r, where U_j is the maximum over smoothness cutoffs in the j-th prime block and I_j is a tilted normalized Euler-product integral; together with a supermartingale tilt and a stopped concent","core_discovery":"The central assertion is that the almost sure upper bound for partial sums of random multiplicative functions is sharp: |M_f(x)| ≪_{ε,f} √x (log log x)^{1/4+ε} almost surely, for both the Steinhaus model (independent unit-circle phases) and the Rademacher model (independent signs on squarefree integers). Combined with a previously established lower bound, this fixes the exponent 1/4. The proof achieves this by showing that the principal logarithmic mean square, which dominates the quadratic variation of the largest-prime martingale, is controlled by a supermartingale of tilted Euler-product integrals; a fixed conditional r-th moment bound on one prime block replaces a first-moment estimate s","pith_inferences":["The conditional block moment technique could yield analogous sharp bounds for weighted partial sums or for higher moments of the martingale, since the loss in earlier approaches came entirely from summing first moments over many blocks.","If the quoted smoothing inequality were proven from scratch, the theorem would no longer depend on an unpublished preprint; if it fails on the final prime block, the main theorem would not follow.","A similar largest-prime decomposition with a conditional r-th moment step might apply to almost sure bounds for other random multiplicative objects, such as character sums or random Euler products at the critical line."],"forward_implications":["The logarithmic exponent 1/4 in the almost sure upper bound is sharp in both the Steinhaus and Rademacher models.","A law-of-the-iterated-logarithm-type statement follows: limsup_{x→∞} log(1+|M_f(x)|/√x)/log log log x = 1/4.","The fixed-order form (Proposition 1.1) provides almost sure bounds with any exponent 1/4+1/(2r)+η, potentially useful for later quantitative applications.","The method reduces the problem to a single-block moment estimate, which may be transferable to other random multiplicative structures or to character sums."],"fun_headline_variants":["Sharp almost sure bound for random multiplicative partial sums","Random multiplicative sums match sharp 1/4 exponent almost surely","Harper's conjecture settled: random multiplicative sums hit 1/4 exponent","Almost sure upper bound matches lower for random multiplicative sums","Sharp log log exponent for random multiplicative sums"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof assumes a deterministic smoothing inequality, quoted from an unpublished preprint, that bounds the quadratic variation of the largest-prime martingale by the principal logarithmic mean square; if that inequality fails for the final prime block or under the strict smoothness convention, the quadratic variation estimate and hence the theorem would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Sharp almost sure bound for random multiplicative partial sums","Random multiplicative sums match sharp 1/4 exponent almost surely","Harper's conjecture settled: random multiplicative sums hit 1/4 exponent","Almost sure upper bound matches lower for random multiplicative sums","Sharp log log exponent for random multiplicative sums"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001366,"raw_usage":{"total_tokens":5307,"prompt_tokens":606,"completion_tokens":4701,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":350,"completion_tokens_details":{"reasoning_tokens":4621}},"tokens_in":350,"tokens_out":4701,"duration_ms":31863,"temperature":1.0,"reasoning_tokens":4621,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:17:39.489599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Verify the smoothing inequality (3.20) on the deterministic function f≡1 (or any fixed assignment of prime values) for large x and the final prime block. The theorem's proof would collapse if for some ε and unbounded x the quadratic variation V_ℓ(x;1)/x exceeded the claimed right-hand side by a factor growing with x.","supporting_citations":[],"review_version":1}