{"id":"056b4c33-3f68-471d-89ea-ca36268723a4","arxiv_id":"1908.07021","paper_version":8,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Markov categories provide a synthetic, axiom-based framework in which conditional independence, sufficiency, completeness, and classical theorems such as Basu and Bahadur hold uniformly across many probability theories.","lead":"This paper develops Markov categories as a high-level algebraic language for probability and statistics, proving abstract versions of classical theorems like Basu's and Bahadur's theorems. A generalist should read it because it offers a unified framework that covers discrete, measurable, Gaussian, and stochastic-process probability at once.","discovery_kind":"paradigm_shift","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract conditional independence (Def. 12.1) is not verified to match the standard measure-theoretic notion in Stoch, weakening the uniform-treatment claim.","rationale":"The reader identified the unverified correspondence between the abstract conditional-independence notion and the standard measure-theoretic definition as the weakest assumption. I agree that this is the most load-bearing concern for the paper's broad claim of a uniform treatment of probability theory. The paper itself flags the issue in Remark 12.4, making it a self-admitted limitation rather than a speculative worry. My analysis adds that the concern is even more direct than the reader's phrasing suggests: Definition 12.1 in Stoch implies the existence of regular conditional distributions, which is a stronger requirement than the standard conditional-expectation formulation (12.2). Therefore, the semigraphoid properties and conditional-product results of Section 12 may not apply to the standard notion of conditional independence used by probabilists. I also note that the link to Basu's theorem is not the strongest consequence, since Basu relies on the input-conditional notion (Definition 12.12), which is less controversial. Nevertheless, the central claim that Markov categories provide a uniform treatment of measure-theoretic probability remains conditional on resolving the mismatch. The author's own open problem and the explicit scope limitation justify a conditional verdict rather than full acceptance; no internal inconsistency was found in the abstract proofs, so rejection is not warranted. The verdict should remain CONDITIONAL, and the authors should either establish the correspondence for standard Borel spaces (where regular conditionals exist) or clearly scope the claims to the abstract framework.","tokens_in":56640,"tokens_out":13813,"duration_ms":147862,"concrete_test":"Exhibit a concrete probability measure on a product of measurable spaces in Stoch for which the standard conditional independence (12.2) holds but the factorization of Definition 12.1 fails because no regular conditional distribution exists. For example, use the classical counterexample of a non-regular conditional distribution on [0,1] x [0,1] referenced in Example 11.3: define a joint distribution psi on X x W x Y such that x and y are conditionally independent given w in the conditional-expectation sense, but verify that no morphisms phi, f, g as in (12.1) can exist. If such an example is confirmed, Definition 12.1 is strictly stronger than the standard notion, and the uniform-treatment claim for Stoch is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Markov categories provide a uniform treatment of measure-theoretic probability depends on the abstract conditional-independence notion of Definition 12.1 matching the standard notion in Stoch. The paper's Remark 12.4 explicitly states that it does not see how to derive the standard conditional-expectation definition (12.2) within the framework. This is not a minor gap: Definition 12.1 requires a factorization of the joint distribution through a common marginal and two kernels f and g, which in Stoch forces the existence of regular conditional distributions. Standard conditional independence via conditional expectations (12.2) does not require regular conditionals, and is known to hold in cases where they do not exist, as the paper itself notes in Examples 11.3 and 11.18. Consequently, Definition 12.1 appears strictly stronger than the standard notion in Stoch, and the semigraphoid properties proven in Lemmas 12.5, 12.13, and 12.17 do not automatically transfer to the notion used in measure-theoretic probability. Since the abstract claims to provide various versions of conditional independence and a uniform treatment of measure-theoretic probability, the unverified correspondence leaves this central claim unsupported for the primary example Stoch. The reader links this concern to Basu's theorem, but Basu (Theorem 15.8) actually uses the input-conditional notion of Definition 12.12, which is less affected; the main impact is on the framework's claim to capture the standard conditional-independence concept itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Markov categories as a synthetic framework for probability and statistics, following Golubtsov and Cho–Jacobs. It gives purely categorical definitions of conditioning, several variants of conditional independence, almost sure equality, sufficient, complete and ancillary statistics, and minimal sufficient statistics, and proves abstract versions of the semigraphoid properties, the Fisher–Neyman factorization theorem, Basu's theorem, and Bahadur's theorem. The framework is illustrated with many examples: FinStoch, Stoch, BorelStoch, the Radon-monad Kleisli category, Gaussian probability, functor categories for stochastic processes, and hypergraph/cocommutative-comonoid constructions. The main advertised advantage is a uniform treatment of discrete, measure-theoretic, Gaussian, and process-level probability theories.","tokens_in":56916,"tokens_out":12015,"duration_ms":129965,"significance":"If the advertised dictionary with measure-theoretic probability is established, this would be a valuable unifying contribution to categorical probability and to the foundations of statistics. The paper is careful and honest: it states limitations explicitly (Remark 12.4, Example 10.4, the open measure-theoretic Fisher–Neyman instantiation) and provides many worked examples. The abstract proofs are clean string-diagram arguments, and the axiomatization is parameter-free in the sense that the theorems are derived from the stated categorical axioms. The current significance is tempered by the fact that the central example Stoch has not been shown to match the standard measure-theoretic notion of conditional independence, which is a key advertised application.","major_comments":[{"comment":"The paper leaves open whether Definition 12.1 agrees with the standard measure-theoretic conditional independence defined via conditional expectations, equation (12.2), in Stoch. In Stoch, Definition 12.1 requires a factorization through Markov kernels f and g, which by Remark 12.2 must be regular conditional distributions; this is strictly stronger than the conditional-expectation notion, which does not require regular conditionals, as the paper itself notes in Examples 11.3 and 11.18. Consequently, the semigraphoid properties proven in Lemmas 12.5, 12.13, 12.17, and 12.20 do not automatically transfer to the notion used in measure-theoretic probability. Since the abstract claims a uniform treatment of measure-theoretic probability, this gap weakens a central advertised contribution. The authors should either prove the correspondence (for BorelStoch, or via a Markov category as in Conjecture 11.10), or explicitly state that the framework currently captures a regular-conditional-based notion of conditional independence, which is stronger than the standard one in full Stoch.","section":"§12, Remark 12.4"},{"comment":"Theorem 14.5 characterizes sufficiency in strictly positive Markov categories via the splitting conditions αsp = p and sα =_{sp-a.s.} id_V, and the paper describes this as an abstract close relative of the Fisher–Neyman factorization theorem. However, it does not verify that the abstract sufficiency condition (14.1) coincides with the classical measure-theoretic definition of sufficiency in Stoch or BorelStoch, and on page 86 it explicitly leaves open whether the theorem specializes to the standard measure-theoretic Fisher–Neyman factorization theorem. Since Stoch is the primary example for measure-theoretic probability, the advertised instantiation of the theorem in that setting is not yet established. The authors should prove the equivalence at least for BorelStoch, where regular conditional distributions exist, or state more precisely which notion of sufficiency the theorem captures in general Stoch.","section":"§14, Theorem 14.5"}],"minor_comments":[{"comment":"The notation S ∩ T is abusive because S ∈ Σ_X and T ∈ Σ_Y are subsets of different spaces; the pullbacks of these sets to the product space should be used, or the conditional expectation should be written with the indicator functions explicitly.","section":"§12.4, Eq. (12.2)"},{"comment":"The definition of p-a.s. equality via equality after composing with copy is clear to category theorists but may be opaque to statisticians; a short concrete instantiation (e.g., in FinStoch, as in Example 13.2) immediately following the definition would help.","section":"§13.1, Definition 13.1"},{"comment":"Since statistics are defined as deterministic morphisms, and in Stoch deterministic morphisms are {0,1}-valued kernels that need not coincide with measurable maps unless the σ-algebras are countably generated and separating, the reader should be reminded that the clean identification of statistics with measurable functions holds in BorelStoch, which is the intended setting for the statistical instantiations.","section":"§14, Definition 14.2 and §10, Example 10.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is an honest and substantial contribution to categorical probability. The main issue is that the authors' strongest advertised claim—uniform treatment of measure-theoretic probability—is currently not supported for the central example Stoch, because the synthetic conditional independence notion has not been matched to the standard one. This is acknowledged in the paper, but the abstract and introduction still assert the uniform treatment. A revision that either closes this gap for BorelStoch or carefully qualifies the claims would largely resolve my concern. The proofs otherwise appear internally coherent, and the explicit discussion of limitations is commendable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper up front. First, it is the real follow-up to Cho–Jacobs and Golubtsov: it introduces conditional independence without requiring conditionals, proves semigraphoid properties in that setting, builds conditional products as gleaves, relativizes almost-sure equality, and proves abstract versions of Fisher–Neyman, Basu, and Bahadur. The string-diagram proofs are careful, and the strictification theorem and the 2-categorical clean-up are genuinely useful. Second, the paper is honest about a real gap: the abstract conditional-independence notion of Definition 12.1 is not known to coincide with the standard measure-theoretic definition in terms of conditional expectations. Remark 12.4 says this outright. The concern is substantive, because Definition 12.1 forces regular conditional distributions, which need not exist in Stoch. So the abstract notion is strictly stronger, and the 'uniform treatment of measure-theoretic probability' claim is scoped down until the correspondence is settled.\n\nWhat is actually new and good: the input-conditional notion (Def 12.12) used for Basu's theorem is less affected by the gap; the abstract Basu and Bahadur are real theorems, not just re-statements. The completeness notion in Stoch is shown to match bounded completeness (Example 15.4), which gives some confidence that the definitions are tracking the classical ones. The citation practice is fine: the paper credits Golubtsov and Cho–Jacobs appropriately and doesn't over-claim novelty.\n\nThe soft spots are the ones the paper itself flags. The Stoch instantiation of Definition 12.1 is the biggest. Relatedly, the general measure-theoretic Fisher–Neyman factorization is left open; the FinStoch case is worked out, but the general case is not. These are scope limitations, not internal contradictions. The abstract theorems hold; the question is how far they reach into classical measure theory. The paper could be more careful in the abstract and introduction: 'uniform treatment' should be qualified, since the most commonly used conditional independence in Stoch is not yet inside the framework.\n\nWho should read this: categorical probabilists, and statisticians with a taste for foundations. It deserves a serious referee: a careful reader should check the diagrammatic proofs and press on the Stoch instantiation, but the paper is important enough to warrant that time.\n\nRecommendation: send it out. The right outcome is probably a revision where the authors either prove the correspondence with the conditional-expectation definition (possibly via a Markov category in the direction of Conjecture 11.10) or explicitly scope the 'uniform treatment' claim. Either way, the paper is a real contribution.","headline":"Real contribution to categorical probability with an honest, substantive gap: the abstract conditional independence notion isn't yet matched to the standard measure-theoretic one in Stoch.","tokens_in":57428,"tokens_out":3631,"would_cite":true,"duration_ms":33931,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60A05","62A01","62B05","18D10","68Q55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sufficiency becomes a pure factorization condition in any Markov category, and the classical statistical theorems follow.","keywords":["Markov categories","conditional independence","sufficient statistics","Fisher–Neyman factorization","Basu's theorem","synthetic probability","semigraphoid properties","almost sure equality"],"falsifier":"In Stoch, take a joint distribution $\\psi$ on $X\\times W\\times Y$ and test whether the factorization condition of Definition 12.1 is equivalent to the standard conditional-expectation equality $\\mathbb{E}[1_S 1_T \\mid \\Sigma_W] = \\mathbb{E}[1_S \\mid \\Sigma_W] \\mathbb{E}[1_T \\mid \\Sigma_W]$; a distribution satisfying exactly one of the two would settle whether the synthetic and measure-theoretic notions coincide.","tokens_in":1590,"feed_emoji":"🎲","tokens_out":1508,"duration_ms":149751,"temperature":0.7,"pith_summary":"This paper argues that the core concepts of probability and statistics—Markov kernels, conditioning, conditional independence, almost sure equality, sufficiency, and completeness—can be developed once, abstractly, in any Markov category, meaning any system of noisy maps equipped with operations for copying and discarding values. The payoff is a single set of proofs that specialize to discrete probability, measure-theoretic probability, Gaussian probability, and stochastic processes. On the statistics side, it proves an abstract characterization of sufficient statistics and derives abstract versions of the Fisher–Neyman, Basu, and Bahadur theorems. A reader should care because the result suggests that measure theory is one concrete model of probability rather than its mandatory foundation.","feed_headline":"Sufficiency becomes one factorization condition","feed_subtitle":"A single axiom system recovers Fisher–Neyman, Basu, and Bahadur across all major probability styles.","key_machinery":"The central object is a Markov category: a symmetric monoidal category whose morphisms are interpreted as noisy maps, with a copying operation $X \\to X \\otimes X$ and a discarding operation $X \\to I$ on every object. This structure suffices to define joint distributions, conditionals, conditional independence via string-diagram factorization, almost sure equality, statistics, and completeness. The load-bearing identity is the sufficiency equation (14.1), which under strict positivity becomes equivalent to $\\alpha s p = p$ together with $s\\alpha =_{sp\\text{-a.s.}} \\mathrm{id}_V$; strict positivity is the axiom that a $p$-almost surely deterministic composite forces a product structure, and it holds in all the paper's main examples.","core_discovery":"On the paper's own terms, the central discovery is Theorem 14.5: in a strictly positive Markov category, a deterministic statistic $s: X \\to V$ is sufficient for a statistical model $p: \\Theta \\to X$ exactly when there is $\\alpha: V \\to X$ with $\\alpha s p = p$ and $s\\alpha =_{sp\\text{-a.s.}} \\mathrm{id}_V$. Sufficiency therefore becomes a pure factorization and almost-sure splitting condition, with no densities and no integration. The paper also proves that completeness is a property of a single morphism rather than of a statistic-model pair, that a complete sufficient statistic is independent of every ancillary statistic (Basu), and that a complete sufficient statistic is minimal sufficient whenever a minimal sufficient statistic exists (Bahadur).","pith_inferences":["If the paper's string-diagram conditional independence were shown to agree with the standard conditional-expectation definition in the measure-theoretic category Stoch, the uniform treatment would extend to the full classical theory; the author leaves that correspondence open.","The factorization characterization suggests a proof-search strategy for new statistical results: conjecture a string-diagram identity, check it in a concrete Markov category such as standard Borel spaces, and then export it to every other instance.","Because Markov categories can be built from relations, matrices over semirings, and commutative comonoids, the same theorems may have analogues in non-probabilistic settings such as multivalued functions or fuzzy information transformers.","A natural next step, left as Problem 16.4 in the paper, is to determine whether every statistical model in a category with sufficiently nice colimits has a minimal sufficient statistic."],"forward_implications":["The same abstract proofs instantiate in finite discrete probability, measure-theoretic probability over general measurable spaces, Gaussian probability, and diagram categories representing stochastic processes, so the statistical theorems transfer across all of these.","In the category of measurable spaces and Markov kernels, the abstract notion of completeness specializes to bounded completeness, and the abstract Basu theorem becomes the standard Basu theorem.","Any Markov category with conditionals supports conditional products that satisfy the Dawid–Studený axioms, giving a compositional calculus for gluing joint distributions.","In a causal Markov category, Bayesian inversion is a symmetric monoidal dagger functor on the category of probability spaces, resolving an open problem about the structure of Bayesian inversion.","Sufficiency is compositional: if a statistic $s$ is sufficient for $p$ and a statistic $t$ is sufficient for the pushed-forward model $sp$, then the composite $ts$ is sufficient for $p$."],"supporting_citations":[{"why":"Supplies the affine CD-category axiomatization and prior conditional-independence definitions from which the present Markov-category treatment grows.","marker":"[15]"},{"why":"Provides the precursor information-transformer framework and the abstract idea of conditioning on which Markov categories build.","marker":"[55]"},{"why":"Introduces the Giry monad construction that realizes Stoch as a Kleisli category, grounding the measure-theoretic instantiation.","marker":"[49]"},{"why":"Proves that standard Borel stochastic kernels admit conditionals depending measurably on a parameter, used to instantiate the conditioning axioms.","marker":"[8]"},{"why":"Supplies the Dawid–Studený axioms for conditional products that the paper shows hold in any Markov category with conditionals.","marker":"[27]"},{"why":"Contains the original Basu theorem whose abstract version is proved as Theorem 15.8.","marker":"[5]"},{"why":"Provides the original definition of completeness for statistics, which the paper's abstract completeness specializes to bounded completeness in Stoch.","marker":"[82]"},{"why":"Gives the composition formulas for Gaussian conditionals that the Gauss Markov category reproduces.","marker":"[79]"}],"fun_headline_variants":["Sufficiency = factorization","One factorization condition unifies Fisher, Basu, Bahadur","Sufficiency without densities: a single split","All probability styles, one sufficiency rule","Sufficiency: just a factorization and almost-sure split"],"cache_read_input_tokens":59520,"weakest_assumption_plain":"The unified-treatment claim rests on the open assumption that the paper's string-diagram conditional independence coincides with the standard conditional-expectation definition in Stoch; if that correspondence fails, some abstract theorems would still hold but their translation to classical statistics would be incomplete.","fun_headline_variants_meta":{"raw":{"variants":["Sufficiency = factorization","One factorization condition unifies Fisher, Basu, Bahadur","Sufficiency without densities: a single split","All probability styles, one sufficiency rule","Sufficiency: just a factorization and almost-sure split"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001812,"raw_usage":{"total_tokens":7068,"prompt_tokens":820,"completion_tokens":6248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":6188}},"tokens_in":436,"tokens_out":6248,"duration_ms":44803,"temperature":1.0,"reasoning_tokens":6188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:28:42.589414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In Stoch, take a joint distribution $\\psi$ on $X\\times W\\times Y$ and test whether the factorization condition of Definition 12.1 is equivalent to the standard conditional-expectation equality $\\mathbb{E}[1_S 1_T \\mid \\Sigma_W] = \\mathbb{E}[1_S \\mid \\Sigma_W] \\mathbb{E}[1_T \\mid \\Sigma_W]$; a distribution satisfying exactly one of the two would settle whether the synthetic and measure-theoretic notions coincide.","supporting_citations":[{"cited_title":"Disintegration and Bayesian Inversion via String Diagrams","cited_arxiv_id":"1709.00322","evidence_quote":"Supplies the affine CD-category axiomatization and prior conditional-independence definitions from which the present Markov-category treatment grows."},{"cited_title":"Kantorovich problems and conditional measures depending on a parameter","cited_arxiv_id":"1904.03642","evidence_quote":"Proves that standard Borel stochastic kernels admit conditionals depending measurably on a parameter, used to instantiate the conditioning axioms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the original definition of completeness for statistics, which the paper's abstract completeness specializes to bounded completeness in Stoch."},{"cited_title":"Lauritzen and Frank Jensen","cited_arxiv_id":null,"evidence_quote":"Gives the composition formulas for Gaussian conditionals that the Gauss Markov category reproduces."}],"review_version":1}