{"id":"cc7a4fd2-7193-400c-a38d-9433f228372b","arxiv_id":"2505.03862","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper is a survey of the authors' categorical and geometric framework for statistical learning, with no substantially new results.","lead":"This paper summarizes the authors' recent and forthcoming work on using the category of probabilistic morphisms for supervised learning, along with geometric kernel methods on matrix manifolds. It is a survey and book preview rather than a new research result.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Example 2.9 overclaims: the 0-1 loss is not correct under Definition 2.8; for a Bernoulli label with P(Y=1)=0.6, the regular conditional 0.6 is not an argmin of 0-1 risk, the constant classifier 1 is.","rationale":"The reader's weakest_assumption concerned condition (L) in Proposition 2.18. That is a legitimate limitation of the learnability theorem, but the authors supply a concrete smooth-density example satisfying it, so I do not treat it as a correctness defect. The more pressing issue is the false assertion in Example 2.9 that the 0-1 loss is 'correct' under Definition 2.8. The paper's own definition is unambiguous: the argmin over the extended hypothesis class must equal the set of regular conditional probability measures. The 0-1 loss fails this even in the simplest Bernoulli case, since its minimizer is the Bayes classifier (the mode), not the conditional distribution. This is a concrete, checkable mathematical error in a section whose purpose is to demonstrate that the categorical framework yields correct loss functions. It does not doom Theorem 2.6 or the kernel mean embedding loss, but it does mean the paper currently contains a false claim about a standard loss, and the claim should be removed or replaced with a qualified statement. A conditional verdict acknowledges both the survey's useful restatement of known results and the need for a specific correction.","tokens_in":31860,"tokens_out":17635,"duration_ms":178340,"concrete_test":"Take X={x0}, Y={0,1}, μ=0.6·δ_{(x0,1)}+0.4·δ_{(x0,0)}. Compute the expected 0-1 risk for h∈Meas(X,Y): R(h)=0.6 if h(x0)=0 and R(h)=0.4 if h(x0)=1. The unique 0-1 risk minimizer is the constant classifier 1, which is not the regular conditional μ_{Y|X}(x0)=0.6. This directly violates Definition 2.8, which requires the argmin over the extended class ~H to be exactly the set of regular conditionals. To settle any ambiguity, ask the authors to specify the extension ~R to randomized classifiers under which the conditional is the argmin; any such extension cannot agree with the instantaneous 0-1 loss on hard classifiers.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Definition 2.8 defines a loss function R on H to be correct if there exists a larger hypothesis class ~H containing the regular conditionals such that the argmin of ~R over ~H is exactly the set of regular conditional probability measures for every μ in the model. Example 2.9 ends by asserting that the instantaneous 0-1 loss L_{0,1}(x,y,h)=d_{0-1}(y,h(x)), with h∈Meas(X,Y), generates a correct loss function. This is false for the paper's own definition. Let X={x0} and Y={0,1}, with μ=0.6·δ_{(x0,1)}+0.4·δ_{(x0,0)}. The regular conditional is μ_{Y|X}(x0)=0.6, i.e. the probability measure putting mass 0.6 on 1. For a hard classifier h∈Meas(X,Y), the expected 0-1 risk is 0.6 if h(x0)=0 and 0.4 if h(x0)=1, so the unique minimizer is the constant classifier h=1, not the conditional probability 0.6. If one instead extends the loss to randomized classifiers h∈Meas(X,P(Y)) by averaging over h(x0), the minimizer is still the mode 1, never the conditional 0.6, because the minimum is achieved by a deterministic choice. Thus no extension of the instantaneous 0-1 loss can have the regular conditional as its argmin unless the loss is changed to depend on the full conditional distribution, which the stated instantaneous loss does not. This contradiction invalidates the closing claim of Example 2.9. The earlier part of Example 2.9 and the kernel mean embedding example 2.11 are not affected, but the paper should correct or remove the 0-1 statement.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents the category of probabilistic morphisms as a framework for supervised learning, reviews geometric kernel methods for positive definite matrices and operators, and surveys two manifold-learning results. Section 2 recalls regular conditional probability measures, proves a characterization via the graph probabilistic morphism (Theorem 2.6), defines generative models and correct loss functions, gives quadratic and kernel mean embedding examples, and states a learnability theorem for overparameterized models (Proposition 2.18) using condition (L) and Vapnik-Stefanyuk regularization. Section 3 summarizes positive definite kernels on SPD matrices and Log-Hilbert-Schmidt operators, and Section 4 reviews the Belkin-Niyogi Laplacian eigenmap justification and Fefferman et al. manifold reconstruction.","tokens_in":32271,"tokens_out":8055,"duration_ms":79068,"significance":"If the claims held, the categorical formalism would provide a unified way to discuss generative models, correct losses, and learnability, and the geometric kernel sections give a useful compendium of known constructions with explicit Gram-matrix formulas. The paper is largely expository: Theorem 2.6 is clean and correct, but the main learnability theorem and its supporting Lemma 2.21 are quoted from the authors' prior work [37], and Sections 3 and 4 mostly restate known results. The false 0-1 loss claim in Example 2.9 undercuts one of the two concrete examples of correct loss functions. There are no new experiments and no fully self-contained proofs for the central theorem. The paper is best classified as a survey/preview of the forthcoming book [40], and its final claim of having demonstrated usefulness is stronger than what is established in this manuscript.","major_comments":[{"comment":"The assertion that the instantaneous 0-1 loss L_{0,1}(x,y,h)=d_{0-1}(y,h(x)) generates a correct loss function is false under Definition 2.8. For X={x0}, Y={0,1}, and μ=0.6δ_{(x0,1)}+0.4δ_{(x0,0)}, the regular conditional probability measure is μ_{Y|X}(x0)=0.6. The 0-1 risk of a hard classifier h∈Meas(X,Y) equals 0.6 if h(x0)=0 and 0.4 if h(x0)=1, so the unique minimizer is the constant classifier h≡1, not the conditional probability 0.6. Extending the loss to randomized classifiers h∈Meas(X,P(Y)) does not help, because the expected 0-1 risk is minimized at a deterministic mode. Hence no extension of this instantaneous loss satisfies the argmin condition in Definition 2.8. The quadratic-loss part of Example 2.9 and the kernel mean embedding example 2.11 are not affected, but the 0-1 statement must be corrected or removed.","section":"Example 2.9, last sentence"},{"comment":"The paper's central learnability theorem is not proved in this manuscript: Lemma 2.21 is deferred to [37, Proposition 6.1], Proposition 2.22 to [37, Theorem 6.5], and Proposition 2.18 is stated as [37, Corollary 6.3]. The outline in Section 2.4.2 does not supply the compactness and rate arguments needed to pass from Proposition 2.22 to uniform consistency under condition (L), nor does it specify how C and Γ are chosen. In addition, condition (L) defines L(μ_Y|X) as a function of μ, but regular conditional measures are only unique μ_X-a.e.; since P_Lip assumes full support, this is still an ambiguity unless a canonical representative is chosen. The authors should either state explicitly that this is a survey with proofs in [37], or provide the missing arguments. As written, the final sentence of Section 5 overstates what is demonstrated here.","section":"Proposition 2.18 and condition (L)"},{"comment":"The proposed concrete example of a family satisfying condition (L) is incomplete. The text says P_X×Y consists of all μ_f=f dxdy for which there exist c1,c0>0 with f∈C^1(X×Y), L(f)≤c1, and c1≥f≥c0. If c0,c1 are allowed to depend on f, the family is not compact in the weak*-topology and the Lipschitz constants L(μ_Y|X) are not uniformly bounded in a fixed interval [a,b] as required by condition (L). If c0,c1 are meant to be fixed uniform constants, the remark should say so and provide the verification. Without this, the example does not demonstrate that condition (L) is satisfiable in a nontrivial setting.","section":"Remark 2.19"}],"minor_comments":[{"comment":"The abstract contains a broken phrase 'as well a s'; Section 2.1's first bullet says the spaces S(Y), M(Y), P(Y) are on X, which should be on Y.","section":"Abstract and Section 2.1"},{"comment":"In the definition of instantaneous loss, 'for any h∈R' should be 'for any h∈H', and the sentence about E_μ switches between f and h; please correct the variable names.","section":"Definition 2.8"},{"comment":"There is a typo 'P(X×Y 9)' in the first paragraph, and the sentence 'any minimizer R_L^μ to Meas(X,[0,1])' should read 'any minimizer of R_L^μ over Meas(X,[0,1])'.","section":"Example 2.9"},{"comment":"The heading 'Leanability' should be 'Learnability'.","section":"Section 2.4 heading"},{"comment":"The phrase 'a generative model ofs supervised learning' contains a typo 'ofs'.","section":"Definition 2.16"},{"comment":"Remark 2.12(1) ends with the incomplete sentence 'They proved the following beautiful result on estimating probab.'; the proposition should be introduced by a complete sentence.","section":"Remark 2.12 and Proposition 2.13"},{"comment":"The term ||Γ_f||_{K3,K1} in equation (2.24) uses an undefined kernel or metric K3; please define it or remove it if it is a typo.","section":"Lemma 2.21"},{"comment":"The domain of K1 is written as 'X×Y ) × (X×Y)', which contains a misplaced parenthesis; it should be (X×Y)×(X×Y).","section":"Proposition 2.22"},{"comment":"The word 'languaguage' is a typo for 'language', and the spelling of 'Vapnik-Stefanyuk' is inconsistent with 'Stefanyuk' used in Proposition 2.20 and the references.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is largely a survey of the authors' own prior work and forthcoming book. The principal learnability theorem is quoted from [37] with only an outline, and the false 0-1 loss example is a genuine error that must be fixed. The paper may be suitable for the journal if survey-style submissions are within scope, but the authors should clearly mark which results are new to this paper and which are restated from [37] or [40]. The citation pattern is heavily self-referential, so the editor may want to check that the contribution is sufficiently novel for a research paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a survey/book preview, not a new-results paper. It restates the categorical characterization of regular conditional probabilities (Theorem 2.6), the learnability result (Proposition 2.18), the geometric kernel theorems, and the Belkin–Niyogi result from earlier work by the authors and others. Read as an expository chapter, it is mostly competent and useful. Read as a demonstration that the framework works, it overreaches, and there is a concrete error in Example 2.9.\n\nWhat is genuinely useful: the graph characterization (\\Gamma_T)_* \\mu_X = \\mu is a clean way to think about regular conditionals, and it leads naturally to the kernel mean embedding loss in (2.11). The geometric kernel section collects Schoenberg-type obstructions, positive definiteness of Log-Euclidean and Log-Hilbert-Schmidt kernels, and the closed-form Gram matrix expression for distances between RKHS covariance operators. That part is practical and well organized. The manifold learning discussion is a fair survey of Belkin–Niyogi and Fefferman et al.\n\nSoft spots, in order of seriousness. First, there are no new theorems; every central claim is attributed to [37], [45], [55,56], or [9], and the proofs of Proposition 2.18 and Lemma 2.21 are only outlined, with details deferred. Second, condition (L) — compactness of the model and uniform boundedness of the Lipschitz constants of conditional measures — is doing real work, and the paper does not discuss how often it holds in realistic high-dimensional settings. That limits the practical punch of the learnability theorem. Third, the stress-test note is right: the last sentence of Example 2.9 is false. For X={x0}, Y={0,1}, and \\mu with P(Y=1)=0.6, the regular conditional is 0.6, but the constant classifier h\\equiv 1 has 0-1 risk 0.4, strictly less than any randomized classifier at 0.6. Under Definition 2.8, no extension of the instantaneous 0-1 loss can make the conditional the unique argmin. That sentence should be removed or corrected.\n\nWho it is for: readers who want an accessible map of this categorical framework and the Log-Hilbert-Schmidt kernel toolbox. I would not cite it for a new result, but it could serve as a useful survey. If submitted as a survey/book chapter, it deserves a serious referee; if submitted as a research paper claiming new results, the lack of novelty and the Example 2.9 error would be a problem. I would ask for revision: fix Example 2.9, state openly that all theorems are prior results with proofs elsewhere, and add a paragraph on when condition (L) is plausibly satisfied.","headline":"A competent survey of the authors' categorical framework and geometric kernels, not a new-results paper; Example 2.9 overclaims the 0-1 loss and the learnability theorem rests on deferred self-citations.","tokens_in":32788,"tokens_out":5002,"would_cite":false,"duration_ms":48259,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","46E22","60B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A map $T : X \\to P(Y)$ is the correct conditional model for a joint distribution exactly when its graph sends the input marginal to the full joint; on this identity the paper builds correct loss functions and learnability.","keywords":["probabilistic morphisms","Markov kernels","regular conditional probability measures","supervised learning","empirical risk minimization","kernel mean embeddings","Log-Hilbert-Schmidt metric","manifold learning"],"falsifier":"Take $X = [0,1]$, $Y = \\{0,1\\}$, and a joint distribution whose conditional mean is a step function with a jump, so the conditional measure is not Lipschitz at the jump, and run the $(C,\\Gamma)$-regularized ERM algorithm from Proposition 2.18 with the kernel-mean-embedding loss; if uniform consistency still holds, condition (L) is stronger than necessary, and if it fails, the compactness-and-Lipschitz assumption is doing real work. For Theorem 2.6 itself the claim follows directly from the definition of a regular conditional probability measure, so the empirically checkable part is whether the constructed losses are actually minimized only at correct conditionals on finite samples, which can be tested directly by estimating both sides of the graph equation.","tokens_in":31655,"feed_emoji":"🎯","tokens_out":23399,"duration_ms":183557,"temperature":0.7,"pith_summary":"This paper argues that the category of probabilistic morphisms — Markov kernels treated as arrows between measurable spaces — gives supervised learning a unified formal setting. Its central technical claim is Theorem 2.6: a measurable map $T : X \\to P(Y)$ is a regular conditional probability measure for the joint distribution $\\mu$ exactly when pushing the input marginal $\\mu_X$ through the graph of $T$ reproduces $\\mu$, i.e. $(\\Gamma_T)_*\\mu_X = \\mu$. From this equivalence the paper derives correct loss functions, in particular a kernel-mean-embedding loss whose minimizers are precisely the correct conditional models, and proves that under a compactness and uniform-Lipschitz condition there exists a uniformly consistent regularized empirical-risk-minimization algorithm for an overparameterized learning model (Proposition 2.18). The same language frames classical regression, since with zero-mean noise the regression function is the conditional mean of the labels. The payoff, if the framework is right, is that correct loss functions and learnability follow from one structural identity instead of being assumed problem by problem.","feed_headline":"One equation tells when a learning model is correct","feed_subtitle":"The graph identity that defines conditional models directly yields correct loss functions and learnability guarantees.","key_machinery":"The machinery carrying the argument is the graph of a probabilistic morphism. Given a measurable map $T : X \\to P(Y)$, its graph $\\Gamma_T$ joins the identity on $X$ with $T$ to form a probabilistic morphism $\\Gamma_T : X \\rightsquigarrow X \\times Y$, and the identity doing the work is $(\\Gamma_T)_*\\mu_X = \\mu$: Theorem 2.6 shows this equality holds exactly when $T$ is a regular conditional probability measure for $\\mu$, turning the search for the correct stochastic model of labels given inputs into the search for solutions of one measure-valued equation. The learnability proof is carried by the regularizer $W(f) = \\|f\\|_M + L(f) + \\|\\Gamma_f\\|_{\\widetilde{K}_3,\\widetilde{K}_1}$, whose sublevel sets are compact (Lemma 2.21), together with a classical consistency estimate for stochastic ill-posed problems (Proposition 2.20) that bounds the probability a regularized minimizer is far from the true conditional map by the probability that the empirical data are far from the true joint distribution; condition (L) makes that bound vanish uniformly as the sample size grows. For the geometric half, the parallel identity is the classical characterization that $\\exp(-\\gamma d^2)$ is positive definite for every $\\gamma > 0$ exactly when the metric space embeds isometrically into a Hilbert space, which rules out squared-distance Gaussian kernels on curved Riemannian manifolds and motivates the Log-Euclidean and Log-Hilbert–Schmidt constructions with their induced vector-space structures.","core_discovery":"On its own terms, the paper's central discovery is that regular conditional probability measures are captured by a single graph equation. For $\\mu \\in P(X \\times Y)$, a measurable map $T : X \\to P(Y)$ is a regular conditional probability measure with respect to the projection onto $X$ if and only if $(\\Gamma_T)_*\\mu_X = \\mu$: the graph $\\Gamma_T$, the join of the identity on $X$ with $T$, pushes the input marginal forward to the full joint measure (Theorem 2.6). Two conditional models for the same $\\mu$ agree $\\mu_X$-almost everywhere, and any map equal $\\mu_X$-almost everywhere to a regular conditional is again one. Because the condition is an equality of measures, it converts into losses: the quadratic loss on $\\{0,1\\}$ labels is correct because its minimizer is the conditional mean (Example 2.9), and the kernel-mean-embedding loss $R_{K_1}(h,\\mu) = \\|M_{K_1}((\\Gamma_h)_*\\mu_X) - M_{K_1}(\\mu)\\|_{\\widetilde{K}_1}$ is correct whenever the kernel mean embedding is injective (Example 2.11). For the overparameterized model with hypothesis space $C_{\\mathrm{Lip}}(X, P(Y)_{\\widetilde{K}_2})$, the loss $R_{K_1}$, and a family of joint distributions satisfying condition (L) — weak*-compactness plus uniformly bounded Lipschitz constants of the conditional measures — Proposition 2.18 asserts the existence of a uniformly consistent $(C,\\Gamma)$-regularized ERM algorithm, obtained by combining the graph equation in variational form with a classical consistency estimate for stochastic ill-posed problems and the compact sublevel sets of a Lipschitz-type regularizer (Lemma 2.21).","pith_inferences":["The graph equation suggests a model-criticism tool the paper does not develop: for any candidate $h$, the sample estimate of $\\|M_{K_1}((\\Gamma_h)_*\\mu_X) - M_{K_1}(\\mu)\\|$ is a finite-sample diagnostic that is zero in expectation only at a correct conditional model, so it could be used to test generative models outside the paper's Lipschitz setting.","Condition (L) is likely the assumption that fails first in high-dimensional practice: uniformity over all $\\mu \\in P_{X\\times Y}$ requires one fixed interval $[a,b]$ to contain the Lipschitz constants of every possible conditional distribution, and sharp decision boundaries would exceed any fixed bound; a local or scale-dependent Lipschitz assumption would be a natural weakening the paper does not","The Log-Euclidean and Log-Hilbert–Schmidt constructions encode a transferable recipe: whenever a set of non-Euclidean objects carries a commutative group operation making it a vector space, every inner-product kernel from Euclidean theory can be carried over to that set; the same recipe could apply to other cones of positive operators or matrix groups.","Since a posterior distribution is itself a regular conditional probability measure, the same graph equation would characterize Bayesian posteriors given data; the paper cites this connection only in passing, so formalizing Bayesian consistency through the graph identity is a natural next step."],"forward_implications":["Correct loss functions can be built directly from the graph equation rather than from an instantaneous loss: the kernel-mean-embedding loss $R_{K_1}(h,\\mu) = \\|M_{K_1}((\\Gamma_h)_*\\mu_X) - M_{K_1}(\\mu)\\|_{\\widetilde{K}_1}$ is correct whenever the embedding is injective, and it is empirically definable, so the framework yields losses that the instantaneous-loss route does not naturally produce.","Overparameterized supervised learning can be learnable: under condition (L), the model $(X, Y, C_{\\mathrm{Lip}}(X, P(Y)_{\\widetilde{K}_2}), R_{K_1}, P_{X\\times Y})$ admits a uniformly consistent $(C,\\Gamma)$-regularized ERM algorithm, and the same holds for discriminative subspaces $H \\subset C_{\\mathrm{Lip}}(X,Y)$ (Proposition 2.18).","Classical regression falls out of the framework: when the noise has zero mean, the regression function is the conditional mean $r_\\mu(x) = \\int_Y y\\, d\\mu_{Y|X}(y|x)$, so the categorical formulation subsumes standard supervised learning as a special case (Example 2.7).","On a geodesically complete Riemannian manifold, the kernel $\\exp(-\\gamma d^2)$ is positive definite for all $\\gamma > 0$ if and only if the manifold is isometric to Euclidean space; consequently the affine-invariant and Bures–Wasserstein distances on positive definite matrices cannot yield such kernels, while the Log-Euclidean and Log-Hilbert–Schmidt kernels are unconditionally positive definite (","Riemannian geometry can be learned from point clouds: a point-cloud Laplacian converges in probability to the Laplace–Beltrami operator at a suitable bandwidth scaling (Theorem 4.1), and a Riemannian manifold can be reconstructed from noisy intrinsic distances between sample points."],"supporting_citations":[{"why":"The cited paper supplies the central results of Section 2: the graph characterization of regular conditional measures (Theorem 2.6), the regularizer with compact sublevel sets (Lemma 2.21), and the learnable overparameterized model (Propositions 2.18 and 2.22).","marker":"[37]"},{"why":"The cited monograph is credited with the category of probabilistic morphisms and the faithful functor to Banach spaces used to push measures along kernels.","marker":"[16]"},{"why":"The cited 1962 notes are the independent origin of the category of probabilistic mappings, anchoring the framework historically.","marker":"[36]"},{"why":"The cited monograph supplies the classical ERM consistency framework and the Vapnik–Stefanyuk method for stochastic ill-posed problems that Proposition 2.18 extends.","marker":"[64]"},{"why":"The cited paper introduces the perturbation method for solving stochastic ill-posed problems, the template for the regularized ERM construction.","marker":"[68]"},{"why":"The cited theorem is the consistency estimate quoted as Proposition 2.20, bounding the probability that regularized minimizers leave the neighborhood of the true solution.","marker":"[58]"},{"why":"The cited work provides the finite-sample kernel mean embedding bound (Proposition 2.13) used to control empirical terms under condition (L).","marker":"[42]"},{"why":"The cited work proves the kernel mean embeddings are injective and their induced metrics generate the weak*-topology, assumptions the loss functions and condition (L) rely on.","marker":"[57]"},{"why":"The cited theorem on metric spaces and positive definite functions underlies the impossibility of squared-distance Gaussian kernels on non-Euclidean manifolds (Theorem 3.3).","marker":"[53]"},{"why":"The cited paper introduces the Log-Hilbert-Schmidt metric and its kernel constructions, which Section 3.3 generalizes.","marker":"[45]"}],"fun_headline_variants":["Graph equation defines correct losses and learnability","One identity yields loss correctness and learnable models","Single graph test for conditional model correctness","Graph condition ensures learnable loss functions","A graph equality enforces accurate learning guarantees"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise of the learnability claim is condition (L): all possible joint distributions must lie in a weak*-compact family whose conditional label distributions all have Lipschitz constants within one fixed finite interval $[a,b]$; if real data can contain arbitrarily sharp or rough conditional distributions, the uniform consistency guarantee of Proposition 2.18 does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Graph equation defines correct losses and learnability","One identity yields loss correctness and learnable models","Single graph test for conditional model correctness","Graph condition ensures learnable loss functions","A graph equality enforces accurate learning guarantees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2270,"prompt_tokens":962,"completion_tokens":1308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1243}},"tokens_in":578,"tokens_out":1308,"duration_ms":9710,"temperature":1.0,"reasoning_tokens":1243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:45:22.885907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $X = [0,1]$, $Y = \\{0,1\\}$, and a joint distribution whose conditional mean is a step function with a jump, so the conditional measure is not Lipschitz at the jump, and run the $(C,\\Gamma)$-regularized ERM algorithm from Proposition 2.18 with the kernel-mean-embedding loss; if uniform consistency still holds, condition (L) is stronger than necessary, and if it fails, the compactness-and-Lipschitz assumption is doing real work. For Theorem 2.6 itself the claim follows directly from the definition of a regular conditional probability measure, so the empirically checkable part is whether the constructed losses are actually minimized only at correct conditionals on finite samples, which can be tested directly by estimating both sides of the graph equation.","supporting_citations":[{"cited_title":"Supervised learning with probabilistic morphisms and kernel mean embeddings","cited_arxiv_id":"2305.06348","evidence_quote":"The cited paper supplies the central results of Section 2: the graph characterization of regular conditional measures (Theorem 2.6), the regularizer with compact sublevel sets (Lemma 2.21), and the learnable overparameterized model (Propositions 2.18 and 2.22)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The cited monograph is credited with the category of probabilistic morphisms and the faithful functor to Banach spaces used to push measures along kernels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The cited 1962 notes are the independent origin of the category of probabilistic mappings, anchoring the framework historically."},{"cited_title":"Vapnik, Statistical Learning Theory","cited_arxiv_id":null,"evidence_quote":"The cited monograph supplies the classical ERM consistency framework and the Vapnik–Stefanyuk method for stochastic ill-posed problems that Proposition 2.18 extends."},{"cited_title":"Vapnik and A","cited_arxiv_id":null,"evidence_quote":"The cited paper introduces the perturbation method for solving stochastic ill-posed problems, the template for the regularized ERM construction."},{"cited_title":"disorder","cited_arxiv_id":null,"evidence_quote":"The cited theorem is the consistency estimate quoted as Proposition 2.20, bounding the probability that regularized minimizers leave the neighborhood of the true solution."},{"cited_title":"Lopez-Paz, K","cited_arxiv_id":null,"evidence_quote":"The cited work provides the finite-sample kernel mean embedding bound (Proposition 2.13) used to control empirical terms under condition (L)."},{"cited_title":"Sriperumbudur, On the optimal estimation of probabi lity measures in weak and strong topologies","cited_arxiv_id":null,"evidence_quote":"The cited work proves the kernel mean embeddings are injective and their induced metrics generate the weak*-topology, assumptions the loss functions and condition (L) rely on."},{"cited_title":"Schoenberg","cited_arxiv_id":null,"evidence_quote":"The cited theorem on metric spaces and positive definite functions underlies the impossibility of squared-distance Gaussian kernels on non-Euclidean manifolds (Theorem 3.3)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The cited paper introduces the Log-Hilbert-Schmidt metric and its kernel constructions, which Section 3.3 generalizes."}],"review_version":1}