{"id":"20bf38ae-fcb4-4b0e-bed7-81e31fc549a6","arxiv_id":"2506.22106","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Adapted total variation between laws of n-step processes satisfies ATV(μ,ν) ≤ sqrt(n) sqrt(2H(μ|ν)), and the constant sqrt(n) is tight.","lead":"The authors prove a Pinsker-type inequality for adapted total variation on n-step stochastic processes: the distance is at most sqrt(n) times the familiar entropy-based bound, and this sqrt(n) factor is shown to be tight. The result gives quantitative control of causal transport distances, which is useful in stochastic control, finance, and machine learning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the proof is sound and the only imported premise is the standard bicausal-coupling characterization.","rationale":"The reader's accepted verdict is appropriate. The proof is short but self-contained except for the standard bicausal-coupling characterization. I checked the two places where a subtle error could hide: the interchange of infimum and integral in Lemma 2.1, and the Jensen/chain-rule argument in Theorem 1.1. The former is justified by standard measurable selection and by the fact that off-diagonal kernel choices cannot lower the one-step cost below 2. The latter is a correct weighted Cauchy-Schwarz/Jensen estimate with the total mass of the auxiliary measure bounded by n and each minimum dominated by the corresponding μ-kernel. The tightness example is exact and gives ratio 2. I disagree with no step in the proof.","tokens_in":6767,"tokens_out":26259,"duration_ms":272158,"concrete_test":"Run an independent brute-force LP for a small finite-state instance (e.g., n=3, each X_k={0,1}, random transition kernels) computing ATV as 2 times the minimum over bicausal couplings of expected discrete-metric cost, and compare with the closed-form expression in Lemma 2.1. This checks both the external bicausal characterization and the diagonal-max optimality claim without relying on the cited [2, Props. 5.1/5.2].","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read in good faith and find no load-bearing flaw. The central inequality rests on Lemma 2.1, whose proof uses the standard bicausal-coupling characterization of Backhoff-Veraguas et al. [2, Props. 5.1/5.2] to interchange the infimum over the conditional kernels with the outer integral, and then solves the resulting one-step transport problem with cost max(2ρ, A), A≤2, by concentrating as much mass as possible on the diagonal. This is a standard measurable-selection argument. The subsequent Pinsker + Jensen + chain-rule step is algebraically sound: the auxiliary measure m has total mass at most n, each minimum μ∧ν is dominated by μ, and the chain rule gives 2H(μ|ν). The tightness computation in Corollary 2.3 is also correct. The induction display in Lemma 2.1 contains a small notational slip (ν_{x1:2} where ν_{y1:2} is expected for off-diagonal pairs), but it is harmless because off-diagonal costs are capped by 2 and on-diagonal the expressions coincide.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This short note proves an adapted Pinsker inequality: for probability measures μ and ν on a product Polish space X1 × ... × Xn, the adapted total variation satisfies ATV(μ, ν) ≤ √n √(2H(μ|ν)). The proof decomposes ATV recursively into a sum of conditional total variation terms (Lemma 2.1), applies classical Pinsker termwise, and then uses Jensen's inequality together with the entropy chain rule to obtain the √n factor. A product Bernoulli example (Corollary 2.3) shows that the constant √n is asymptotically sharp as the marginals approach each other.","tokens_in":6968,"tokens_out":12595,"duration_ms":104727,"significance":"The result gives a clean, dimension-dependent Pinsker inequality for adapted transport, complementing the known equivalence ATV ≤ (2n−1)TV and improving the dependence on n. The proof is transparent, parameter-free, and the tightness example is explicit. The recursive decomposition lemma for ATV is independently useful. The main limitation is that this is a short observation rather than a deep structural theorem, and the proof imports the standard bicausal-coupling characterization from [2] without proof.","major_comments":[],"minor_comments":[{"comment":"In the displayed induction hypothesis, the differential 'dπ_{x1,y1}(x1,y1)' should be 'dπ_{x1,y1}(x2,y2)', and the terms 'ν_{x1:2}' and 'µ_{x1}∧ν_{x1}' in the off-diagonal part should read 'ν_{y1:2}' and 'µ_{x1}∧ν_{y1}'; although these coincide on the diagonal and the intended formula is clear, the current notation makes the induction step unnecessarily confusing.","section":"Lemma 2.1 proof, general induction"},{"comment":"In the definition of the auxiliary measure m, the atom should be written 'δ0' rather than 'δt'.","section":"Theorem 1.1 proof"},{"comment":"The entropy expression 'H(µ, |ν)' contains a misplaced comma and bar; it should be 'H(µ|ν)'.","section":"Corollary 2.3"},{"comment":"The displayed expansion of '(1−ε)^{2n}' has a typographical garble '42n(2n−1)/2'; the intended coefficient '4·(2n)(2n−1)/2' is correct.","section":"Corollary 2.3"},{"comment":"Reference [2] spells 'Proposition' as 'Propositon'.","section":"References"},{"comment":"The notation '({Ω = {0,1}, 2Ω})' appears with a mismatched brace; this is a minor formatting issue.","section":"Lemma 2.2"}],"recommendation":"minor_revision","confidential_remarks":"The paper is correct and the central claim is sound; the errors are typographical. The proof borrows the bicausal-coupling characterization from [2] as an external input, which is standard and acceptable. I see no reason to delay publication beyond a quick revision of the notation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know that this short note delivers exactly what it claims: ATV(μ,ν) ≤ √n √(2H(μ|ν)) for laws of n-step processes, with a tightness example showing the √n constant can't be improved in general. Lemma 2.1 is the real content—a recursive decomposition of ATV into a sum of conditional total variations weighted by the minima of the marginals. The proof of that lemma uses the standard bicausal-coupling characterization from Backhoff-Veraguas et al. and a measurable selection argument; both are imported from a cited source, but the import is legitimate and the induction is clean.\n\nThe main inequality then follows by applying classical Pinsker to each term, Jensen to a measure of total mass at most n, and the entropy chain rule. That's a natural path, and the authors are honest that the decomposition was independently observed by Acciaio–Hou–Pammer. The tightness construction—product Bernoulli measures with bias ε—is correct and even walks through the limit carefully.\n\nSoft spots are minor. The induction display in Lemma 2.1 has a notational slip (ν_{x1:2} where ν_{y1:2} is meant off-diagonal), but it's harmless because the off-diagonal cost is capped at 2. The general induction is written a bit tersely; the measure m in the Jensen step is spelled out only abstractly, though the n=2 case makes the pattern obvious. A careful reader will fill in the details without trouble. The paper doesn't discuss possible improvements or the non-tightness for non-product measures; that's fine for a note.\n\nCitations look appropriate: the key prior work on ATV and adapted Wasserstein bounds is credited, and the chain-rule tensorization point is tied to known references. I don't see a load-bearing flaw or circularity.\n\nWho's this for? Anyone working in causal optimal transport, multistage stochastic optimization, or robust finance who needs a quick bound on ATV from entropy. It's a solid short note, not a paradigm shift. I'd send it to a good probabilist for a quick review, and expect it to be accepted with minor revisions.\n\nRecommendation: engage with it. It's worth a serious referee.","headline":"A clean, tight adapted Pinsker inequality with a sound proof; worth refereeing as a short note.","tokens_in":7458,"tokens_out":1849,"would_cite":true,"duration_ms":18913,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"An adapted version of Pinsker's inequality holds: $ATV(\\mu,\\nu)\\le \\sqrt{n}\\sqrt{2H(\\mu|\\nu)}$, and the constant $\\sqrt{n}$ is sharp.","keywords":["Pinsker inequality","adapted total variation","relative entropy","bicausal coupling","causal optimal transport","adapted Wasserstein distance","tightness","stochastic processes"],"falsifier":"Compute $ATV(\\mu,\\nu)$ and $H(\\mu|\\nu)$ for any pair of $n$-step laws: a single pair with $H(\\mu|\\nu)<\\infty$ and $ATV(\\mu,\\nu)>\\sqrt{n}\\sqrt{2H(\\mu|\\nu)}$ would refute the theorem. For the sharpness claim, evaluate the product Bernoulli family of Corollary 2.3 with fixed $n$ and $0<\\varepsilon\\ll 1$: here $ATV(\\mu,\\nu)=2-2(1-\\varepsilon)^n$ and $H(\\mu|\\nu)=n(2\\varepsilon^2+o(\\varepsilon^2))$, so the normalized ratio $ATV^2/(nH)$ tends to $2$; if a numerical evaluation of this family ever produced a ratio above $2$, the bound as stated would be false.","tokens_in":6601,"feed_emoji":"🎲","tokens_out":10648,"duration_ms":107747,"temperature":0.7,"pith_summary":"This note establishes an adapted Pinsker inequality: for probability measures $\\mu,\\nu$ on a product space $X_1\\times\\cdots\\times X_n$, the adapted total variation distance satisfies $ATV(\\mu,\\nu)\\le \\sqrt{n}\\sqrt{2H(\\mu|\\nu)}$. The extra factor $\\sqrt{n}$ is the price of requiring the comparison to respect the causal filtration structure of the processes, and it is shown to be unavoidable: for $n$-fold products of slightly biased Bernoulli measures the ratio $ATV(\\mu,\\nu)^2/(nH(\\mu|\\nu))$ tends to $2$ as the bias goes to zero, so no smaller constant works in general. The statement matters because adapted Wasserstein distances are the natural way to compare laws of stochastic processes in applications from stochastic control to distributionally robust optimization and machine learning, where transports must not use future information.","feed_headline":"Pinsker bound survives adaptation, with a √n cost","feed_subtitle":"Relative entropy controls causal disagreement between two n-step processes, and √n is best possible.","key_machinery":"The load-bearing object is Lemma 2.1, a recursive decomposition of $ATV$. It says that $ATV$ equals the total variation of the first marginal plus the integral, against the pointwise minimum of the two marginals, of the total variation between the successive conditional kernels, continued stage by stage; an optimal bicausal coupling is exactly one whose kernels place maximal mass on the diagonal at every layer. The decomposition turns a single constrained infimum over bicausal couplings into a sequence of ordinary transport problems, using the characterization that a coupling is bicausal precisely when its successive disintegrations are themselves ordinary couplings of the corresponding conditional laws, chosen measurably. Once the decomposition is available, the proof needs only the classical Pinsker inequality, Jensen's inequality, and the chain rule of relative entropy.","core_discovery":"The central claim is that causal, filtration-respecting comparisons of process laws are quantitatively controlled by relative entropy. Theorem 1.1 asserts $ATV(\\mu,\\nu)\\le \\sqrt{n}\\sqrt{2H(\\mu|\\nu)}$ for all probabilities on $X_1\\times\\cdots\\times X_n$, where $ATV$ is the adapted total variation distance obtained by restricting couplings to bicausal transport plans. This is the direct generalization of the classical bound $TV(\\mu,\\nu)\\le \\sqrt{2H(\\mu|\\nu)}$, recovered when $n=1$. The proof decomposes $ATV$ into the total variation of the first marginals plus the expected total variation of the successive conditional kernels, applies the ordinary Pinsker bound to each layer, and reassembles the layers through Jensen's inequality and the chain rule of relative entropy. Corollary 2.3 proves the bound is tight: in the product Bernoulli example, $ATV(\\mu,\\nu)^2/(nH(\\mu|\\nu))\\to 2$ as the bias tends to zero, so the constant $\\sqrt{n}$ cannot be improved.","pith_inferences":["The recursive decomposition suggests a template: any stagewise cost built from a bounded metric by taking the maximum over coordinates should admit an entropy bound with the same $\\sqrt{n}$ horizon factor; the discrete metric here is the simplest instance.","Sharpness among product measures indicates that the $\\sqrt{n}$ factor is a genuine cost of the horizon, not of path dependence or long memory; even independent repeated experiments pay the full factor in the worst case.","A natural testable extension is whether the constant can be improved when the conditional kernels are Lipschitz or otherwise regular, since the paper's examples are purely atomic and do not exercise regularity.","The discrete-time proof is built on summing $n$ layers, so a continuous-time analogue would need a different argument; the paper's own mention of continuous-time transport inequalities suggests such an analogue is not immediate."],"forward_implications":["If two $n$-step process laws have relative entropy at most $\\delta$, their adapted total variation is at most $\\sqrt{n}\\sqrt{2\\delta}$; in particular, entropy convergence of process laws forces causal transport convergence at rate $\\sqrt{n\\delta}$.","The bound is tight even among product measures: flat product Bernoulli laws already drive the ratio $ATV^2/(nH)$ to its maximal value $2$, so the $\\sqrt{n}$ constant cannot be improved by restricting to exchangeable or memoryless processes.","Since $ATV\\ge TV$, the inequality subsumes the classical Pinsker bound and gives a quantitative link between KL divergence and process-level causal distances, which is useful whenever a model is trained by minimizing relative entropy against a target process law.","Combined with the earlier linear comparison $ATV\\le (2n-1)TV$, the new inequality gives a dimension-dependent bound $ATV\\le \\sqrt{n}\\sqrt{2H}$, which is sharper than the linear comparison precisely when entropy is small relative to TV."],"supporting_citations":[{"why":"Supplies the characterization of bicausal couplings via successive conditional couplings and the measurable-selection step on which Lemma 2.1 rests.","marker":"[2]"},{"why":"Introduces the adapted total variation distance and proves the earlier linear-in-$n$ comparison with total variation that frames the new bound.","marker":"[7]"},{"why":"Provides the relative entropy chain rule and the tensorization property used to reassemble the per-layer Pinsker estimates.","marker":"[12]"}],"fun_headline_variants":["Adapted total variation obeys Pinsker, up to √n","Causal Pinsker: ATV ≤ √n √(2H)","Pinsker for processes: √n factor is sharp","ATV meets Pinsker, cost scales as √n","Adapted Pinsker bound: √n tight"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the imported structural fact that any transport plan respecting the time order of both processes can be assembled step-by-step from ordinary couplings of their successive conditional laws, with the kernels chosen measurably; if that fact failed, the recursive decomposition of $ATV$ that carries the whole proof would not be available.","fun_headline_variants_meta":{"raw":{"variants":["Adapted total variation obeys Pinsker, up to √n","Causal Pinsker: ATV ≤ √n √(2H)","Pinsker for processes: √n factor is sharp","ATV meets Pinsker, cost scales as √n","Adapted Pinsker bound: √n tight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1462,"prompt_tokens":923,"completion_tokens":539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":539,"tokens_out":539,"duration_ms":5041,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:11:07.772114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $ATV(\\mu,\\nu)$ and $H(\\mu|\\nu)$ for any pair of $n$-step laws: a single pair with $H(\\mu|\\nu)<\\infty$ and $ATV(\\mu,\\nu)>\\sqrt{n}\\sqrt{2H(\\mu|\\nu)}$ would refute the theorem. For the sharpness claim, evaluate the product Bernoulli family of Corollary 2.3 with fixed $n$ and $0<\\varepsilon\\ll 1$: here $ATV(\\mu,\\nu)=2-2(1-\\varepsilon)^n$ and $H(\\mu|\\nu)=n(2\\varepsilon^2+o(\\varepsilon^2))$, so the normalized ratio $ATV^2/(nH)$ tends to $2$; if a numerical evaluation of this family ever produced a ratio above $2$, the bound as stated would be false.","supporting_citations":[{"cited_title":"Backhoff-Veraguas, M","cited_arxiv_id":null,"evidence_quote":"Supplies the characterization of bicausal couplings via successive conditional couplings and the measurable-selection step on which Lemma 2.1 rests."},{"cited_title":"Eckstein and G","cited_arxiv_id":null,"evidence_quote":"Introduces the adapted total variation distance and proves the earlier linear-in-$n$ comparison with total variation that frames the new bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the relative entropy chain rule and the tensorization property used to reassemble the per-layer Pinsker estimates."}],"review_version":1}