{"id":"aa01ec6b-8beb-495d-b042-6d62caf0d41f","arxiv_id":"2506.08325","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conformalized kernel-depth algorithm constructs prediction regions for regression in separable Hilbert spaces with marginal non-asymptotic guarantees and asymptotic conditional consistency.","lead":"The paper introduces conformal prediction regions for regression problems where both inputs and outputs live in abstract vector spaces, using kernel-based depth scores. It offers marginal coverage guarantees and asymptotic conditional consistency, and demonstrates the method on physical activity data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Conditional consistency rests on an unproved sup-norm consistency condition that the paper itself flags as delicate; a failed Assumption 3(2) would leave Theorem 4 without support.","rationale":"The reader's weakest-assumption analysis and my stress-test converge: the marginal conformal coverage is a standard split-conformal argument and the simulation/application sections are illustrative, but the conditional consistency theorem depends entirely on an unverified, delicate uniform consistency condition. The paper's own Remark 3 strengthens the concern by acknowledging that Donsker-type hypotheses for kernel mean embeddings may fail in infinite-dimensional spaces. I do not see a more load-bearing flaw in the central argument: the conformal calibration logic, while presented with nonstandard naming (Dtest used for quantile calibration), still gives exchangeability for fresh observations, and the asymptotic oracle-region argument is plausible once Assumption 3 is granted. Therefore the reader's CONDITIONAL verdict remains appropriate: the manuscript should be accepted only subject to supplying a proof or credible empirical validation of Assumption 3(2), or substantially weakening the associated claims. No change to the verdict is needed.","tokens_in":19566,"tokens_out":9531,"duration_ms":108063,"concrete_test":"Simulate the paper's own functional setting: X~U[0,1], Y(t)=beta0(t)+X*beta1(t)+epsilon(t) with epsilon a Gaussian process in L2[0,1]; estimate the conditional kernel mean embedding by kernel ridge regression exactly as in Section 4.3 for n = 200, 800, 3200; approximate sup_{x,y} |hat_mu(y;x) - mu(y;x)| using a fine grid over x and a large set of y draws from the conditional response distribution plus a broader function set. If the sup-norm error does not decrease toward zero as n grows, Assumption 3(2) is untenable in the paper's flagship scenario and Theorem 4 needs additional hypotheses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 3(2) is the load-bearing premise of Theorem 4, the advertised conditional consistency result. It requires uniform sup-norm consistency of the conditional kernel mean embedding estimator, ||hat_mu_{P(Y|X)} - mu_{P(Y|X)}||_infinity = o_p(1), over the response space (and, as used in the theorem, effectively over x as well). The paper provides no proof of this uniformity. Standard kernel-ridge conditional mean embedding guarantees hold in RKHS norm or L2(P), neither of which implies sup-norm convergence when Y is infinite-dimensional. Uniform control over all y requires Glivenko-Cantelli/Donsker-type conditions on the class {k_Y(y, .): y in Y}; Remark 3 explicitly concedes that Donsker hypotheses may fail for kernel mean embeddings in infinite-dimensional spaces. Since Theorem 4's integrated symmetric-difference statement is exactly the conditional-consistency claim made in the abstract and Section 1.3, a failure of Assumption 3(2) removes the basis for that headline theoretical result. Formally, the theorem is conditional on the assumption, but no evidence (proof, reference, or simulation) is supplied that the assumption holds in the paper's main functional-response setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes prediction and tolerance regions for regression with responses in separable Hilbert spaces, based on conditional kernel mean embeddings interpreted as depth measures. The main algorithms combine a plug-in estimate of the conditional depth distribution with conformal calibration to obtain marginal coverage guarantees, and a bootstrap variant for tolerance regions in probability. The methods are applied to Euclidean, functional, and distributional (2-Wasserstein) responses, with simulation studies and an NHANES accelerometer case study. The paper claims non-asymptotic marginal conformal guarantees, asymptotic conditional consistency (Theorem 4), and fast convergence rates, but the theoretical apparatus is incomplete and at least one central definition appears to invert the intended coverage level.","tokens_in":19857,"tokens_out":11865,"duration_ms":141860,"significance":"If the main claims were rigorously established, this would be a useful contribution: it extends distributional conformal prediction to infinite-dimensional responses, connects depth bands with kernel mean embeddings, and provides a practical implementation in a health application. The simulation comparison in the Euclidean conditional-coverage setting and the functional-data illustration are valuable and suggest the approach works in practice. However, the advertised asymptotic results are currently conditional on an unproved and delicate sup-norm consistency assumption, and the formal coverage statements are not correctly specified. The paper does not provide machine-checked proofs or code; the conformal argument is standard and likely salvageable, but the present manuscript does not yet support its headline claims.","major_comments":[{"comment":"As written, the population region has coverage α, not 1−α. Since g(x,Y) = P(D_k(Y;P_{Y|X}) ≤ D_k(Y;P_{Y|X}) | X) is the probability integral transform of the conditional depth, it is Uniform(0,1) under continuity, so P(Y ∈ {y : g(x,y) ≥ q_{1−α}(x)} | X=x) = 1 − q_{1−α}(x) = α when q_{1−α}(x)=1−α. The same inversion appears in Algorithm 1's threshold bD_{⌈n+1−α(n+1)⌉} and in Algorithm 3's use of the empirical (1−α) quantile of depth values. To obtain 1−α coverage the threshold should be the α-quantile of the depth distribution, or the inequality should be reversed. This needs to be corrected and propagated through Theorem 4 and the simulation and application sections.","section":"§2.3, Eq. (5); Algorithms 1 and 3"},{"comment":"Theorem 4 is the paper's advertised conditional-consistency result, but it is stated without proof and rests entirely on Assumption 3(2), which requires sup-norm consistency of the conditional kernel mean embedding estimator. Standard kernel-ridge guarantees for conditional mean embeddings are in RKHS norm or L2(P), neither of which implies sup-norm convergence for infinite-dimensional responses; such uniformity requires Glivenko–Cantelli/Donsker conditions on {k_Y(y,·): y∈Y}, and Remark 3 concedes these may fail for infinite-dimensional Y. The theorem also needs extra conditions, for example continuity and level-set stability of g, convergence of the estimated quantile, and uniformity over x. Without a proof or a reference establishing Assumption 3(2), the conditional-consistency claim in the abstract and Section 1.3 is unsupported.","section":"§2.3, Assumption 3 and Theorem 4"},{"comment":"Propositions 5 and 6 are stated without proofs. Proposition 6 is the only result supporting the heteroscedastic Algorithm 2's conditional consistency, yet it asserts sup-norm consistency and √n-weak convergence of \\widehat{g} under a GAMLSS fit, without specifying how the GAMLSS model is identified with the conditional CDF g(x,y) or how the Donsker condition is verified. The Donsker assumption also appears to conflict with Remark 3 for infinite-dimensional responses. Please provide complete proofs or precise references, and state the regularity conditions on the GAMLSS parameterization explicitly.","section":"§2.4, Propositions 5 and 6"},{"comment":"The advertised 'non-asymptotic guarantees of the type P(Y ∈ \\widehat{C}_α(X)) ≥ 1−α' are not formalized anywhere: there is no theorem stating the coverage property, the exact quantile index, or the tie-breaking rule, and the roles of Dcalibration and Dtest in Algorithm 2 are not justified by an exchangeability argument. The standard split-conformal argument is available once the fitted depth and \\widehat{g} are fixed, but it must be stated carefully because \\widehat{g} itself is estimated on Dcalibration and the quantile is computed on Dtest. Please add a formal proposition for the marginal conformal guarantee.","section":"§2.3, Algorithm 2 and Contribution 2"},{"comment":"The abstract and Section 1.3 promise 'faster convergence rates' and 'fast convergence rates' in certain homoscedastic settings, but the manuscript gives no theorem quantifying rates for \\widehat{C}_α; Theorem 4 is only an o_p(1) consistency statement. Please either add a rate result with its assumptions or remove the rate claim from the contributions.","section":"Abstract and §1.3"}],"minor_comments":[{"comment":"Typos and leftover editing notes should be cleaned: 'popululation bases' in §1.1, 'continuty hyphotesis' in Assumption 3, 'hypthesis' in Remark 3, 'bellow' in §1.3, the duplicate 'Matabuena and Petersen (2023)Matabuena and Petersen (2023)' in §4.2, and the editorial note 'why not put more plots in each row? we are wasting a lot of space...' in the Figure 3 caption.","section":"Throughout"},{"comment":"The paper has duplicate equation numbers: Eq. (1) is used for the regression model and again for the GAMLSS additive predictor, and Eq. (2) is used for both the conditional-coverage target and the empirical-process display; please renumber them.","section":"§1 and §2.4"},{"comment":"Property (vi) says the depth is 'continuous in P' but does not specify a topology or metric on P; please make this formal.","section":"Definition 3"},{"comment":"Algorithm 1's threshold should be reconciled with Corollary 2: for an α-expectation tolerance region the coverage result suggests r_n = ⌈(n+1)α⌉, which is not what the displayed formula computes; this is related to Major Comment 1.","section":"Algorithm 1 and Corollary 2"},{"comment":"The population quantile condition ('By definition, under continuity hypothesis, q_{1−α}=1−α') assumes a continuous conditional depth distribution; the paper should address atoms and ties, which also matter for the conformal quantile.","section":"Assumption 3(3)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript appears to be an early draft with editorial artifacts, including a note left in the Figure 3 caption. The central idea is promising and the conformal part is largely salvageable, but the theoretical section needs a complete rewrite with proofs: Theorem 4 and Propositions 5–6 are currently stated without proof, and the coverage-level inversion in Eq. (5) and Algorithms 1–3 is a fundamental correctness issue. I would request a major revision with full proofs and a corrected coverage lemma before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible integration of known pieces—kernel mean embeddings as depth, split conformal calibration, GAMLSS for heteroscedasticity—into a prediction-region algorithm for functional and distributional responses. The marginal non-asymptotic coverage guarantee is standard split conformal and holds. The paper is worth engaging, but the headline theory overreaches.\n\nWhat is new: putting these pieces together as a unified algorithmic pipeline for responses in separable Hilbert spaces is genuinely useful. The application to accelerometer-derived distributional responses with a Wasserstein kernel is a nice demonstration. The simulation study includes a fair comparison in the scalar heteroscedastic setting, where the proposed method beats HPD-split and CQR on conditional-coverage error, and the functional simulation checks marginal coverage. The bootstrap tolerance region is a natural extension.\n\nSoft spots, in order of importance. First, Theorem 4 and Propositions 5-6 are stated without proof. Assumption 3(2) demands sup-norm consistency of the conditional mean embedding estimator, uniform over the response space. No proof is given, and Remark 3 concedes that Donsker-type conditions may fail in infinite-dimensional response spaces. So the asymptotic conditional consistency result, which is the advertised headline, is unsupported in the main functional setting. The theorem is formally conditional, but the assumption is not validated or even discussed for the paper's own examples. That is a load-bearing gap, not a cosmetic one.\n\nSecond, the paper is unfinished. There are typos (\"popululation\", \"bellow\"), an editorial aside in the figure caption, and the GAMLSS estimator section is sketchy; Proposition 6 assumes finite-dimensional X and Donsker conditions without proof. Third, the uniqueness claim—\"no other general predictive framework\"—is stronger than the cited literature supports; the paper itself cites Lugosi and Matabuena (2024) and Chernozhukov et al. (2021), so the novelty is the specific pipeline, not the absence of any competing approach.\n\nWhat is not a flaw: the self-citations are mostly to prior work that this paper builds on, and the conformal guarantee is external, not circular.\n\nBottom line: for a reader who wants a working method to produce prediction bands for curve-valued or distributional outcomes, this paper delivers. The theory needs to be brought in line with what is actually proved. I would send it to serious peer review, but the revision should include proofs or explicit relaxation of Theorem 4, a validation or discussion of Assumption 3(2), and a cleanup pass. It deserves referee time because the marginal guarantee defends the practical algorithm even if the asymptotic claims fail.","headline":"Useful synthesis for conformal UQ in Hilbert-space-valued regression, but the advertised conditional consistency result leans on an unproved sup-norm assumption the paper itself flags as delicate.","tokens_in":20347,"tokens_out":2439,"would_cite":true,"duration_ms":30158,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G15","62G20","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A model-free conformal recipe turns kernel depth bands into prediction regions for functional responses.","keywords":["conformal prediction","data depth","kernel mean embedding","prediction regions","functional data","separable Hilbert spaces","tolerance regions","uncertainty quantification"],"falsifier":"Construct a functional-response model with a kernel for which the conditional kernel mean embedding estimators fail the Donsker property that Remark 3 concedes may be absent, then check whether the integrated symmetric difference between the estimated and oracle $\\alpha$-level regions converges to zero; a concrete counterexample would settle the conditional-consistency claim.","tokens_in":19401,"feed_emoji":"📈","tokens_out":5565,"duration_ms":65899,"temperature":0.7,"pith_summary":"This paper aims to give a general recipe for prediction regions in regression when both predictor and response live in separable Hilbert spaces, such as functions or probability distributions, where standard ordering tools do not exist. It proposes ranking each candidate response by a conditional depth measure built from conditional kernel mean embeddings, then applying conformal calibration to the ranked values. The central claims are that the conformalized regions satisfy marginal non-asymptotic coverage $P(Y \\in \\hat C_\\alpha(X)) \\geq 1-\\alpha$, and that under uniform consistency of the embedding estimator the regions converge conditionally to the oracle level sets. A concrete motivation is personalized physical activity recommendations: given age, sex, and BMI, the method outputs a band of plausible activity-distribution profiles. If these claims hold, the framework fills a gap where few competing uncertainty-quantification methods exist for functional responses.","feed_headline":"Conformal depth bands for functional response prediction","feed_subtitle":"Kernel mean embeddings rank functional responses, with marginal coverage guarantee and conditional consistency.","key_machinery":"The central object is the conditional kernel mean embedding $\\mu_{Y|X}(x)=\\mathbb{E}[\\phi(Y)\\mid X=x]$, estimated by the ridge-regularized operator $\\hat C = \\Phi(K_X+n\\lambda I)^{-1}\\Psi^\\top$, which yields a depth value for each response through the kernel feature map. The paper identifies this embedding with an $h$-integrated depth band, so thresholding it gives nested level sets. The theoretical engine also includes the Li\\u2013Liu depth-spacing theorem, which makes the probability masses of depth spacings behave like uniform spacings and gives tolerance regions with Beta-distributed coverage. Conformal calibration on held-out splits supplies the quantile thresholds, and a bootstrap variant exploits the asymptotic Gaussianity of the estimators to produce tolerance regions in probability.","core_discovery":"The paper's central discovery is that kernel mean embeddings can serve as conditional depth measures, and that the resulting depth bands can be conformalized to produce prediction regions for responses in separable Hilbert spaces. For a candidate response $y$ at predictor $x$, the method evaluates the estimated conditional kernel mean embedding $\\hat\\mu_{Y|X}(x)$ and treats the inner product $D_k(y;\\hat P_{Y|X}) = \\langle \\phi(y), \\hat\\mu_{Y|X}(x)\\rangle$ as a ranking score. Conformal calibration converts these scores into regions of the form $\\hat C_\\alpha(x)=\\{y:\\hat g(x,y) \\geq \\hat q_{1-\\alpha}\\}$ that, by exchangeability, obey $P(Y \\in \\hat C_\\alpha(X)) \\geq 1-\\alpha$ regardless of model specification. Under Assumption 3, Theorem 4 establishes the stronger asymptotic statement that the integrated symmetric difference between estimated and oracle conditional regions is $o_P(1)$, and in homoscedastic settings the method needs only train/test splits. Heteroscedastic versions use GAMLSS to model how the depth distribution varies with $x$.","pith_inferences":["If the sup-norm consistency assumption highlighted in Remark 3 fails for infinite-dimensional responses, the conditional-consistency theorem no longer follows, but the marginal conformal guarantee would survive; practical users should then read the method as a marginal rather than conditional promise.","The method's stated speed advantage over density-level-set estimators comes from kernel smoothing: a natural next benchmark is comparing region volume or sharpness against CQR/HPD-style regions in a functional-to-functional setting.","Because conformal guarantees here rely on exchangeability, applying the algorithm to correlated wearable-device time series would require weighted or split-conformal adaptations; the paper lists dependent data as future work.","Using a 2-Wasserstein Gaussian kernel makes the response ranking interpretable in quantile space, so the same construction could directly produce clinical reference bands for biomarker distributions."],"forward_implications":["Prediction regions for functional or distributional responses come with a finite-sample marginal coverage guarantee $P(Y \\in \\hat C_\\alpha(X)) \\geq 1-\\alpha$ that does not depend on the correctness of any regression model.","In homoscedastic settings the procedure reduces to a two-split algorithm and still attains the same conformal guarantee.","Under Assumption 3, the estimated regions converge to the oracle conditional level sets in integrated symmetric difference, so conditional coverage approaches $1-\\alpha$ as sample size grows.","The bootstrap variant returns tolerance regions in probability, meaning the random region contains at least $1-\\alpha$ of the conditional mass with prescribed confidence $\\gamma$.","The depth-plus-conformal pattern extends beyond functional data to random objects in separable Hilbert spaces, including distributions equipped with Wasserstein-type kernels."],"supporting_citations":[{"why":"Supplies the depth-spacing theorem and the construction of nonparametric multivariate tolerance regions that the paper's depth-band machinery builds on.","marker":"Li and Liu (2008)"},{"why":"Provides the conformal prediction framework whose exchangeability argument yields the non-asymptotic marginal coverage guarantee.","marker":"Shafer and Vovk (2008)"},{"why":"Supplies the theory and estimation of kernel mean embeddings, the ranking measure at the heart of the proposed depth.","marker":"Muandet et al. (2017)"},{"why":"Provides the conditional kernel mean embedding estimator used in the prediction-region algorithms.","marker":"Park and Muandet (2020)"},{"why":"Supplies the GAMLSS framework that models the conditional distribution of depth values in the heteroscedastic variant.","marker":"Rigby and Stasinopoulos (2005)"},{"why":"Introduces the distributional conformal inference paradigm that the paper extends to separable Hilbert space responses.","marker":"Chernozhukov et al. (2021)"},{"why":"Establishes the mathematical connection between kernel mean embeddings and h-integrated depth bands that justifies using KME as a depth measure.","marker":"Wynne and Nagy (2025)"},{"why":"Serves as a state-of-the-art conformal quantile regression baseline compared in the conditional coverage simulations.","marker":"Romano et al. (2019)"},{"why":"Provides the earlier functional tolerance-band methodology whose bootstrap calibration approach is extended to conditional tolerance regions in probability.","marker":"Rathnayake and Choudhary (2016)"}],"fun_headline_variants":["Kernel embedding depth with conformal coverage","Model-free conformal prediction via kernel depth","Conformal depth bands for functional responses","Depth-based uncertainty sets with fast convergence","Kernel depth conformal regions for regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 3(2): the estimated conditional kernel mean embedding must converge uniformly over the whole response space in sup norm, $\\|\\hat\\mu_{Y|X}-\\mu_{Y|X}\\|_\\infty = o_P(1)$, and the paper does not prove this uniform convergence, which can fail for infinite-dimensional responses.","fun_headline_variants_meta":{"raw":{"variants":["Kernel embedding depth with conformal coverage","Model-free conformal prediction via kernel depth","Conformal depth bands for functional responses","Depth-based uncertainty sets with fast convergence","Kernel depth conformal regions for regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1466,"prompt_tokens":1001,"completion_tokens":465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":401}},"tokens_in":617,"tokens_out":465,"duration_ms":5686,"temperature":1.0,"reasoning_tokens":401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:13:33.298184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a functional-response model with a kernel for which the conditional kernel mean embedding estimators fail the Donsker property that Remark 3 concedes may be absent, then check whether the integrated symmetric difference between the estimated and oracle $\\alpha$-level regions converges to zero; a concrete counterexample would settle the conditional-consistency claim.","supporting_citations":[],"review_version":1}