{"id":"dcdc4d7e-a268-4db9-b8ec-343481deae60","arxiv_id":"2411.18555","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Mutual absolute continuity of probability measures on a filtered space holds if and only if a certain martingale limit M equals 1 almost surely, and then the square roots of the Radon-Nikodym derivatives converge in L^2.","lead":"The paper proves a new characterization of mutual absolute continuity of probability measures on a filtered space, using a martingale limit M that compares the tails of the measures. It applies to families of random variables and stochastic processes, offering a new theoretical tool for equivalence of measures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.2.5's proof relies on an unjustified equality relating a capped limit to M_n; as written this breaks the 'only if' direction, though a valid inequality likely repairs it.","rationale":"I read the paper in good faith and believe the central theorem is correct: it is consistent with the classical Kakutani criterion for product measures and with the Kabanov-Liptser-Shiryaev absolute-continuity criteria, and I found no counterexample to the characterization itself. The main issue is not the theorem's truth but the rigor of the proof of the 'only if' direction. The reader's stated weakest assumption was the cap-stability of the generator union F_n, which is standard and is not the most fragile point; the genuinely load-bearing problem is the unjustified equality in the proof of Theorem 4.2.5. That equality appears to be false in general, but the proof only needs the corresponding inequality, and the inequality follows from Fatou plus M_m ≤ 1. The missing limiting argument in Theorem 4.3.1 (passing from approximate orthogonality to exact orthogonality via a Borel-Cantelli type construction), the mis-stated conclusion of Corollary 4.3.2, the reversed monotonicity wording in Lemma 4.4.2, and the notation errors in Corollary 5.2 are all real but repairable presentation or proof-completion issues. Because the suggested fix is routine and I know of no reason to doubt the theorem, the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":10460,"tokens_out":32468,"duration_ms":294399,"concrete_test":"Re-derive equation (2) in Theorem 4.2.5 using only the valid inequality E(sqrt(varphi_{m,...}) ∧ 1 | F_m) ≤ M_m, together with M_m → M in L^1 for D ∈ F_n. Check that the resulting chain E(1_D liminf_m sqrt(varphi_{m,...}) ∧ 1) ≤ liminf_m E(1_D M_m) = E(1_D M) holds for every D ∈ F_n. If it does, the proof is repairable and the theorem stands, pending the same fix in the published version; if it does not, the 'only if' direction has a genuine gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The 'only if' direction of Theorem 5.1 depends on Theorem 4.2.5, which asserts that M=1 almost surely on {N>0}. In that proof the displayed chain after equation (2) contains the step\n\nE(1_D liminf_m sqrt(varphi_{m,...}) ∧ 1) ≤ liminf_m E(1_D sqrt(varphi_{m,...}) ∧ 1)\n= liminf_m E(1_D E(sqrt(varphi_{m,...}) ∧ 1 | F_m))\n= lim_m E(1_D M_m).\n\nThe final equality asserts that E(sqrt(varphi_{m,...}) ∧ 1 | F_m) has the same asymptotic behavior as M_m = lim_k E(sqrt(varphi_{m,...,k}) | F_m). This is not justified: Fatou gives only E(sqrt(varphi_{m,...}) | F_m) ≤ M_m, and capping with ∧1 can only make the left side smaller. In general the two quantities are not equal; for product measures one can have E(sqrt(varphi_{m,...}) ∧ 1) < M_m when the infinite product occasionally exceeds 1. Because this step is what forces M=1 on {N>0}, and Corollary 4.3.2 then uses the contrapositive to prove that failure of M=1 implies non-equivalence, the argument is load-bearing. The proof can probably be patched by replacing the unjustified equality with the valid inequality E(1_D sqrt(varphi_{m,...}) ∧ 1) ≤ E(1_D M_m), which still yields E(1_C) ≤ E(1_C M) and hence M=1 on C. But as written, the central 'only if' direction is not fully proved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies mutual absolute continuity (P ~ P') of two probability measures on a filtered space (Omega, F, (F_n)) with F = sigma(union_n F_n), under the standing assumption that P|F_n ~ P'|F_n for every n. Writing Phi_n = dP'|F_n/dP|F_n and phi_n = Phi_{n+1}/Phi_n, the author defines M_{n,k} = E( product_{i=n}^k sqrt(phi_i) | F_n ), proves that the double limit M = lim_n lim_k M_{n,k} exists (P+P')-almost everywhere, and establishes the main characterization (Theorem 5.1): P ~ P' if and only if M = 1 (P+P')-almost surely, in which case sqrt(Phi_n) -> sqrt(dP'/dP) in L^2(P). The proof is organized in four stages: convergence of M and the companion martingale N (Section 4.1); the relations between N and M, including M = 1 on {N > 0} (Section 4.2); orthogonality on {N = 0} and {N' = 0} (Section 4.3); and L^2/L^1 convergence of densities under M = 1 (Section 4.4). A product-space corollary (Corollary 5.2) gives applications to laws of families of random variables and stochastic processes.","tokens_in":2350,"tokens_out":5920,"duration_ms":357960,"significance":"If the main theorem is correct, this is a genuine addition to the classical criteria of Kakutani and Kabanov-Lipcer-Sirjaev: the M = 1 criterion is parameter-free, derived purely from the Radon-Nikodym densities, and comes with an L^2-convergence statement for square-root likelihood ratios that strengthens the existing absolute-continuity criteria. The construction of M is natural (a decreasing conditional-expectation limit in k followed by a bounded-submartingale limit in n), and no ad-hoc entities or fitted parameters enter. The proof is self-contained modulo standard martingale theorems and the generator approximation theorem, and the author is honest about the scope of the result, explicitly noting that unlike the KLS criterion, M is not a limit of a predictable sequence and is therefore 'of a more theoretical nature'. The applications to product spaces are natural and correctly reduce to Theorem 5.1 via Lemma 2.8. The contribution is moderate in scope - a new equivalence rather than a new phenomenon - but it is solid and useful, provided the proof gaps identified in the major comments are closed.","major_comments":[{"comment":"The stress-test concern is confirmed: the displayed chain (2) contains the step liminf_n E(1_D E(sqrt(phi_{n,...}) wedge 1 | F_n)) = lim_n E(1_D M_n), which is not justified by the stated tools and is false in general. Fatou's lemma for conditional expectations gives only E(sqrt(phi_{n,...}) | F_n) <= liminf_k E(sqrt(phi_{n,...,k}) | F_n) = M_n, and the cap wedge 1 can only decrease the left-hand side; the two quantities differ whenever the infinite product sqrt(phi_{n,...}) exceeds 1 on a set of positive conditional probability, which already occurs in the Kakutani product-measure case. Because this step is what yields E(1_C) <= E(1_C M) on C = {N > 0}, and hence M = 1 on C, the 'only if' direction of Theorem 5.1 is not fully proved as written. The repair is local: replacing the equality with the valid inequality E(1_D E(sqrt(phi_{n,...}) wedge 1 | F_n)) <= E(1_D M_n), together with dominated convergence for E(1_D M_n) -> E(1_D M), gives the same conclusion; the authors should make this change and justify each limit in the chain.","section":"Section 4.2, Theorem 4.2.5, Eq. (2)"},{"comment":"The statement of Corollary 4.3.2 is the reverse of the implication its proof establishes. As printed, 'Suppose M = 1 does not hold (P+P')-a.e. Then P ~ P'' contradicts the argument, which proves the contrapositive: if M < 1 on a set C with P(C) > 0, then Theorem 4.2.5 gives N = 0 a.e. on C, so P({N = 0}) > 0, and Theorem 4.3.1 yields P|C orthogonal to P'|C, which excludes P ~ P'. The final sentence of the proof, 'Since P(C) != 0, this implies P|C ~ P'|C', is likewise inconsistent with the preceding orthogonality conclusion and should instead note that orthogonality on a set of positive P-measure contradicts mutual absolute continuity. The corollary should be restated as: if M = 1 does not hold (P+P')-a.e., then P and P' are not mutually absolutely continuous; with that restatement, the citation of Corollary 4.3.2 as the '=>' direction in Theorem 5.1 becomes correct.","section":"Section 4.3, Corollary 4.3.2"}],"minor_comments":[{"comment":"In the introduction and in the statement of Theorem 5.1, 'F_1 = empty' should read 'F_1 = {empty, Omega}', and 'F = sigma(union_{n in N} F_n)' has a typographical omission of the subscript n in the published text.","section":"Section 1 and Theorem 5.1"},{"comment":"The sentence 'M_{n,k} is F_k-measurable' should read 'F_n-measurable'; the conditional expectation with respect to F_n is F_n-measurable.","section":"Section 4.1, after Definition 4.1.1"},{"comment":"The simultaneous choice of n and C' in F_n needs explicit justification: Theorem 2.6 applies with the generator union_m F_m (which is cap-stable and complement-closed because each F_m is a sigma-algebra), and applying it to the measure P + P' yields C' in union_m F_m with both P(C' Delta C) and P'(C' Delta C) small, after which n must be enlarged to cover the index of C', which is harmless because int_C N_n <= epsilon for all sufficiently large n. The parenthetical '(uniform integrability)' does not by itself give the third condition; that condition follows from Cauchy-Schwarz and the uniform L^2 bound on N_m once (P+P')(C' Delta C) < epsilon^2. Finally, the passage from the epsilon-bounds to P|C orthogonal to P'|C (e.g., a Borel-Cantelli argument along epsilon_m decreasing to 0) is omitted and should be stated.","section":"Section 4.3, proof of Theorem 4.3.1"},{"comment":"The displayed formula in the proof has an unbalanced parenthesis ('E(phi_{n,...,k})) = 1'), and the inequality E(M_n^2) <= lim_k E(phi_{n,...,k}) is a consequence of Fatou's lemma together with Jensen's inequality (M_{n,k}^2 <= E(phi_{n,...,k} | F_n)), not of dominated convergence by itself; the proof should be rewritten for clarity.","section":"Section 4.1, Lemma 4.1.6"},{"comment":"The expression 'lim inf_n sqrt(phi_{n,...}) wedge 1' is ambiguous; the Fatou step requires the cap to be inside the liminf, that is, liminf_n (sqrt(phi_{n,...}) wedge 1).","section":"Section 4.2, Eq. (2)"},{"comment":"The assertion that phi_{1,...,n} converges (P+P')-a.e. cites Lemma 4.2.2, which is stated for P only; the P' half follows by symmetry of the setup (M' = M) and should be said explicitly.","section":"Section 4.4, Lemma 4.4.2"},{"comment":"The symbol phi_{J_i} in the product product_{i=n}^k sqrt(phi_{J_i}) is not defined by the preceding definition phi_{J,K} = Phi_J/Phi_K; it should be the ratio Phi_{J_i}/Phi_{J_{i-1}} (with Phi_empty = 1), and the reference measure for the asserted L^2 convergence of (Phi_{J_n})^{1/2} should be specified.","section":"Section 5, Corollary 5.2"},{"comment":"The name 'Kabunov' is used inconsistently alongside 'Kabanov' (see also the bibliography), and the related-work discussion would benefit from a concrete comparison of the new M-criterion with the collapsed KLS criterion of Corollary 3.2, going beyond the qualitative remark that M is not a predictable limit.","section":"Section 3 Related Work"}],"recommendation":"major_revision","confidential_remarks":"Both major issues are localized in the 'only if' direction of the main theorem and admit straightforward repairs, so I do not see grounds for rejection; a careful revision should be able to close them. The contribution is solid but incremental: the M-criterion is a new equivalent formulation with a nice L^2-convergence companion, and the author appropriately characterizes its scope as 'of a more theoretical nature' than the predictable KLS criterion. The editor may also want the related-work section to spell out the precise logical relationship between the new criterion and Corollary 3.2, and to fix the 'Kabunov'/'Kabanov' spelling inconsistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely new criterion for when two measures on a filtered space are mutually absolutely continuous, and the core idea is right. The written proof has one real gap and a few typos, but the gap closes easily. Worth sending to review.\n\nWhat's new: the martingale limit M, defined as lim_n lim_k E(sqrt(phi_n...phi_k) | F_n), is not just a repackaging of the Kabanov-Lipcer-Shiryaev sum-of-log criterion. It stays closer to Kakutani and gives an L^2 convergence statement for sqrt(Phi_n) that I don't think appears elsewhere. The proof is self-contained and builds on standard martingale convergence and Fatou arguments. The generator approximation step in Theorem 4.3.1 is a nice touch.\n\nSoft spots: the proof of Theorem 4.2.5 uses an equality that isn't justified. After defining phi_{n,...} = liminf_k phi_{n,...,k}, the chain has E(1_D sqrt(phi_{n,...}) ∧ 1) = ... = E(1_D M_n). Fatou only gives a ≤ for the conditional expectation of the limit, and the cap makes it worse. This equality is load-bearing for the 'only if' direction. Fortunately, replacing it with ≤ still gives E(1_C) ≤ E(1_C M) on {N>0}, so the gap is patchable. Corollary 4.3.2 states 'P ~ P'' when the proof actually shows non-equivalence; the intended statement is clear. There are also typos: 'Kabunov' for Kabanov, and Theorem 5.1 says F_1 = empty set instead of the trivial sigma-algebra. None of these are deep.\n\nThe central theorem, Theorem 5.1, is likely correct as stated, and the L^2 convergence claim follows from Lemma 4.4.2. The paper is for probability theorists who work on absolute continuity, equivalence of measures, and martingale limits. I'd cite it once the fixes are in.\n\nRecommendation: yes, send to peer review. The result is novel and the reasoning is sound modulo a fixable gap.","headline":"New and plausible characterization of mutual absolute continuity; the proof has a fillable gap in the key theorem and a mis-stated corollary, but the result deserves a referee.","tokens_in":11352,"tokens_out":3433,"would_cite":true,"duration_ms":28489,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G","60B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Mutual absolute continuity on a filtered space is characterized by a martingale limit $M$ being equal to 1 almost surely.","keywords":["mutual absolute continuity","filtered space","Radon-Nikodym derivative","martingale limit","product probability measures","stochastic processes","absolute continuity of measures"],"falsifier":"Compute $M$ for a pair of singular product measures, for example two independent Bernoulli laws with different success probabilities; the theorem predicts $M<1$ on a set of positive measure, and if the computation instead gives $M=1$ almost surely, the characterization is false.","tokens_in":10291,"feed_emoji":"🎲","tokens_out":6816,"duration_ms":56632,"temperature":0.7,"pith_summary":"This paper gives a new criterion for deciding whether two probability measures on the same filtered space are mutually absolutely continuous, meaning they assign zero probability to exactly the same events. The criterion is a single martingale limit $M$ built from the square roots of the successive Radon-Nikodym density ratios between the two measures. The paper proves that the two measures are mutually absolutely continuous if and only if $M=1$ almost surely under both measures, and that in this case the square-root densities converge in $L^2$ to the square root of the final density. Because laws of random variables and stochastic processes live on product spaces, the result immediately gives a characterization of equivalence of such laws in terms of finite-dimensional marginals.","feed_headline":"A single martingale limit decides when two probability laws are equivalent","feed_subtitle":"The limit M compares the tails of the two measures on the filtration; M=1 is exactly mutual absolute continuity","key_machinery":"The central object is the martingale limit $M=\\lim_n \\lim_k \\mathbb{E}(\\prod_{i=n}^k \\sqrt{\\varphi_i}\\,|\\,\\mathcal{F}_n)$, where $\\varphi_n=\\Phi_{n+1}/\\Phi_n$ is the density ratio between consecutive filtration levels. Morally, $M$ measures the similarity of the two measures at infinity, after removing the information already seen at time $n$. The proof also uses the companion process $N_n=\\sqrt{\\varphi_1\\cdots\\varphi_{n-1}}\\,M_n$, which is a uniform $L^2$-martingale whose limit $N$ records whether the product of all densities converges to a positive limit. The argument splits the sample space according to whether $N$ and its counterpart $N'$ vanish: on $\\{N>0\\}$ the paper forces $M=1$, while on $\\{N=0\\}$ it shows the two measures are orthogonal. The generator approximation theorem is the tool that lets the orthogonality proof approximate the set $\\{N=0\\}$ by filtration sets.","core_discovery":"Let $\\mathbb{P}$ and $\\mathbb{P}'$ be probability measures on a filtered space $(\\Omega,\\mathcal{F},(\\mathcal{F}_n)_{n\\in\\mathbb{N}})$ with $\\mathcal{F}_1=\\{\\emptyset,\\Omega\\}$ and $\\mathcal{F}=\\sigma(\\bigcup_n \\mathcal{F}_n)$, and assume $\\mathbb{P}|_{\\mathcal{F}_n}\\sim\\mathbb{P}'|_{\\mathcal{F}_n}$ for every $n$. Write $\\Phi_n=d\\mathbb{P}'|_{\\mathcal{F}_n}/d\\mathbb{P}|_{\\mathcal{F}_n}$ and $\\varphi_n=\\Phi_{n+1}/\\Phi_n$. The paper's central claim is that the doubly indexed conditional expectation $M_{n,k}=\\mathbb{E}(\\prod_{i=n}^k \\sqrt{\\varphi_i}\\,|\\,\\mathcal{F}_n)$ has limits $M_n=\\lim_k M_{n,k}$ and $M=\\lim_n M_n$ almost surely with respect to $\\mathbb{P}+\\mathbb{P}'$, and that $\\mathbb{P}\\sim\\mathbb{P}'$ holds exactly when $M=1$ almost surely for both measures. In that case $(\\Phi_n)^{1/2}$ converges in $L^2(\\mathbb{P})$ to $(d\\mathbb{P}'/d\\mathbb{P})^{1/2}$. The 'only if' direction is obtained by showing that wherever $M<1$ the two measures are orthogonal, and the 'if' direction by using $M=1$ to prove that $\\sqrt{\\Phi_n}$ is a Cauchy sequence in $L^2(\\mathbb{P})$.","pith_inferences":["An extension the authors do not spell out: because $M$ is built only from conditional expectations of square-root density ratios, it can in principle be estimated pathwise from simulations of the two measures along a filtration, making the criterion computationally testable despite being non-predictable.","The condition $M=1$ is a tail condition independent of $\\mathbb{P}|_{\\mathcal{F}_n}$ and $\\mathbb{P}'|_{\\mathcal{F}_n}$ for every fixed $n$; one could use the gap $1-M$ as a quantitative measure of distance from equivalence, although the paper does not pursue that.","Because the 'only if' direction relies on approximating arbitrary measurable sets by filtration sets, one could replace the filtration by any family closed under finite intersections and complements that generates the final $\\sigma$-algebra; checking whether the characterization survives that replacement would map the theorem's structural boundary."],"forward_implications":["For any pair of measures on a filtered space whose finite-dimensional restrictions are equivalent, mutual absolute continuity of the full measures is decided by the single condition $M=1$ $(\\mathbb{P}+\\mathbb{P}')$-almost surely.","When the condition holds, the square roots of the Radon-Nikodym densities $\\Phi_n^{1/2}$ converge in $L^2(\\mathbb{P})$ to $(d\\mathbb{P}'/d\\mathbb{P})^{1/2}$, giving an $L^2$ approximation of the final density by its filtration approximations.","For laws of families of random variables or stochastic processes on a product space, equivalence is characterized by applying the criterion along every increasing sequence of finite coordinate sets, reducing the infinite-dimensional equivalence question to a family of finite-dimensional calculations.","If $M$ fails to equal 1, the measures split orthogonally on the set where the martingale limit $N$ is zero, so the equivalence-versus-singularity dichotomy is governed by the same tail object."],"supporting_citations":[{"why":"Supplies the generator-approximation and martingale-convergence results used in the proof of the main theorem.","marker":"[1]"},{"why":"Provides the earlier absolute-continuity criterion on filtered spaces that the new $M=1$ characterization is compared against.","marker":"[2]"},{"why":"Supplies the related absolute-continuity result referenced in the comparison of existing criteria.","marker":"[3]"},{"why":"Gives the square-root-density method and the product-measure equivalence criterion that the new martingale limit generalizes.","marker":"[4]"}],"fun_headline_variants":["M=1: The New Criterion for Equivalent Probability Laws","One Martingale Limit to Rule Them All: Measure Equivalence","Measure Equivalence: A Single Martingale Limit Tells All","Martingale Limit M=1: The Decisive Test for Equivalent Measures","When Probability Measures Agree: A Martingale Limit Decides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The characterization assumes the two measures agree on every finite filtration level and that the filtration's union is a family closed under finite intersections and complements that generates the final $\\sigma$-algebra; if either fails, the density ratios or the orthogonality argument can break down.","fun_headline_variants_meta":{"raw":{"variants":["M=1: The New Criterion for Equivalent Probability Laws","One Martingale Limit to Rule Them All: Measure Equivalence","Measure Equivalence: A Single Martingale Limit Tells All","Martingale Limit M=1: The Decisive Test for Equivalent Measures","When Probability Measures Agree: A Martingale Limit Decides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000817,"raw_usage":{"total_tokens":3581,"prompt_tokens":950,"completion_tokens":2631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2542}},"tokens_in":566,"tokens_out":2631,"duration_ms":15859,"temperature":1.0,"reasoning_tokens":2542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:05:39.506490+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $M$ for a pair of singular product measures, for example two independent Bernoulli laws with different success probabilities; the theorem predicts $M<1$ on a set of positive measure, and if the computation instead gives $M=1$ almost surely, the characterization is false.","supporting_citations":[{"cited_title":"Durrett, Probability: theory and examples","cited_arxiv_id":null,"evidence_quote":"Supplies the generator-approximation and martingale-convergence results used in the proof of the main theorem."},{"cited_title":"On the question of absolute continuity and singularity of probability measures,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier absolute-continuity criterion on filtered spaces that the new $M=1$ characterization is compared against."},{"cited_title":"On absolute continuity and singularity of probability measures,","cited_arxiv_id":null,"evidence_quote":"Supplies the related absolute-continuity result referenced in the comparison of existing criteria."},{"cited_title":"On equivalence of infinite product measures,","cited_arxiv_id":null,"evidence_quote":"Gives the square-root-density method and the product-measure equivalence criterion that the new martingale limit generalizes."}],"review_version":1}