{"id":"717a57e6-3488-49ed-bed3-0fde6708686d","arxiv_id":"2608.08247","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Functional distributional random forests estimate covariate-dependent distributions over curves using an MMD-based split criterion, with consistency theory and proper scoring rules for evaluation.","lead":"The paper introduces functional distributional random forests, a method that estimates the full conditional distribution of a functional outcome (a curve) given covariates, rather than just its mean. This makes it possible to ask how covariates change the shape, timing, variability, or tail behavior of entire trajectories, and it is tested on simulations and NHANES activity data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The implemented random-feature MMD splitter is not covered by the consistency theorems, so the central plug-in claim is not yet established for the estimator actually used.","rationale":"The reader's weakest-assumption analysis identifies the same core issue: the consistency theorems apply to an idealized infinite-forest estimator using the exact MMD split rule, while the implemented Algorithm 2 uses a random-feature approximation with finite B and a discretized Brownian motion. I agree that this is the most load-bearing concern. I reviewed the supplementary proofs of Theorems 2–3; the subsequence-and-identification argument is internally sound given Theorem 1, and the tightness lemmas are reasonable. The port of Čevid et al.'s Theorem 2 to separable Banach spaces is plausible but not independently verified; however, the more direct and concrete gap is the mismatch between the theoretical split criterion and the algorithm actually run. Assumption 5 (uniform tightness) is nontrivial but is a standard regularity condition with explicit sufficient conditions supplied in Remark 1, so it is less concerning than the algorithm gap. The paper does not prove that the random-feature approximation preserves the consistency guarantees, nor does it provide any rate or condition on B and R_j. Because the central claim is that the fitted forest yields consistent plug-in inference for arbitrary functionals, this gap prevents the theory from covering the estimator used in the empirical sections. This is not a fatal flaw—an extended theorem with B,R_j growing at suitable rates could close it—so the reader's CONDITIONAL verdict remains appropriate. I did not find an independent, more severe objection: the proofs of Theorem 2 and 3 are coherent under their stated assumptions, and the empirical comparisons are extensive. The recommendation is therefore to keep the CONDITIONAL verdict, requiring the authors to either prove consistency for the approximate split criterion or explicitly restrict the theoretical claims and provide a separate justification for the implemented procedure.","tokens_in":34433,"tokens_out":7871,"duration_ms":80901,"concrete_test":"On a small simulation (e.g., N=200, grid of 20 points, one DGP from Section 4), compute the fDRF using the exact MMD split criterion (feasible at this size) and using the Algorithm 2 random-feature criterion with B=20, B=200, and B=2000. Compare the resulting conditional-law estimates bF_x via the energy score and via the plug-in estimates of a target functional (e.g., conditional exceedance probability). If the B=200 result differs materially from the exact-MMD result, or if the difference does not shrink as B grows, then the implemented estimator is not shown to inherit the consistency theorems. Analytically, attempt to bound |D̂_MMD^RF(U,V) - D_MMD(U,V)| in terms of B, R, and sup norms of U,V; if such a bound holds uniformly over candidate splits with B,R→∞, the proofs of Theorems 1–4 can be revisited with the approximate criterion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The consistency results in Section 5 are stated for an 'ideal' fDRF that uses M→∞ trees and the exact MMD split criterion of Eq. (1), with D_MMD defined in Eq. (2). Theorem 1 is ported from Čevid et al. for this exact-criterion forest, and Theorems 2–3 and the supplementary proofs (Lemmas 1–4, Theorem 4) build directly on Theorem 1. However, Algorithm 2 in the Supplementary Material explicitly replaces D_MMD in Eq. (1) by a random-feature approximation based on Eq. (4): B random features, a finite-grid discretization of the Itô integral, and fresh randomness at each split. No theorem, uniform bound, or rate is given linking this approximate criterion to the exact MMD criterion, and no condition on B or R_j is stated that would let the approximation error vanish as N→∞. The estimator whose performance is reported in Section 4 (simulations) and Section 6 (NHANES) is therefore not the estimator shown to be consistent. This gap is load-bearing because the paper's central claim is that one fitted forest supports plug-in inference on arbitrary functionals of the conditional law; that claim requires consistency of the fitted forest, not of a hypothetical exact-MMD forest. The discretization compounds the issue: the split criterion depends only on function evaluations at a finite grid, so for a fixed R_j it is not the MMD of the L2 Gaussian kernel acting on full curves. The paper's own Discussion (Section 7) acknowledges that extension to discretely-sampled approximations is future work, which applies to the consistency theory as much as to the application.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces the functional distributional random forest (fDRF), a random-forest method that estimates the conditional distribution of a functional response Y given covariates X by a weighted empirical distribution over the observed curves. Splits are chosen to maximize an empirical maximum mean discrepancy (MMD) between child-node response distributions, with kernels defined on function spaces such as L2 and Sobolev spaces. The paper establishes consistency of an idealized infinite-forest version of the estimator (Theorems 2 and 3), proposes proper scoring rules for evaluating conditional distribution estimators, reports simulation comparisons against K-nearest neighbors and functional linear models, and applies the method to NHANES accelerometer data to study how age and income affect activity profiles. The theoretical development explicitly addresses the fact that MMD convergence does not automatically imply weak convergence on infinite-dimensional spaces, using uniform tightness assumptions to upgrade the consistency statement.","tokens_in":34784,"tokens_out":15558,"duration_ms":139070,"significance":"The contribution is potentially significant: it extends distributional random forests from Euclidean responses to functional responses, provides a practical plug-in framework for inference on distributional features of curves, and gives a theoretically grounded treatment of weak convergence of the estimated conditional laws. The paper's handling of the non-local-compactness issue is a real technical contribution, and the proposed Sobolev-transform kernels and proper scoring rules are useful tools. The supplementary material includes detailed proofs, implementation algorithms, and NHANES analysis code, which supports reproducibility. The simulation study is reasonably extensive and the NHANES application illustrates genuinely distributional findings beyond conditional means. However, the consistency theorems apply to an idealized estimator using exact MMD and a fixed kernel, while the implemented algorithm uses a random-feature approximation and data-dependent kernel tuning; this gap must be addressed before the central plug-in claim can be regarded as fully established.","major_comments":[{"comment":"The consistency results in Section 5 are stated for an ideal infinite forest that uses the exact MMD split criterion (1) with D_MMD from (2), but Algorithm 2 of the Supplementary Material replaces D_MMD by the random-feature approximation in (4), with B random features, a finite-grid discretization of the Itô integral, and fresh randomness at each split. No theorem, uniform bound, or rate is provided connecting the approximate criterion to the exact one, and no condition on B or R_j is stated under which the approximation error vanishes as N→∞. Consequently, the estimator whose performance is reported in Sections 4 and 6 is not the estimator shown to be consistent in Theorems 2 and 3. This gap is load-bearing because the paper's central plug-in claim requires consistency of the fitted forest, not of a hypothetical exact-MMD forest. The Discussion (Section 7) acknowledges that discretely-sampled approximations are future work, which confirms the scope of the current theory.","section":"Section 5 / Supplementary Algorithm 2"},{"comment":"The theoretical results treat the kernel κ as fixed, whereas the implementation sets the length scale ℓ by the median-distance heuristic on the data and sets λ and the Sobolev normalizers (a,b) by data-dependent rules. Because Assumption 4 and the Lipschitz condition on x↦μ(F_x) in Assumption 2 are stated for a fixed kernel, the ported Theorem 1 and the proof of Theorem 2 do not directly cover the estimator actually fitted with a data-selected kernel. The paper should either state conditions under which the data-dependent kernel still satisfies the assumptions (e.g., a rate condition on ℓ̂), or clarify that the theory is for a pre-specified kernel and treat the heuristic as a practical implementation choice.","section":"Section 2.4 and Section 5 (Assumptions 2 and 4)"}],"minor_comments":[{"comment":"The abstract claims the method 'supports inference on arbitrary functionals of the conditional distribution,' but Theorems 2 and 3 only establish weak convergence, which guarantees convergence for continuous bounded functionals (or continuous functionals on C[0,1] under Assumption 6). Recommend replacing 'arbitrary' with 'a broad class of' or 'continuous.'","section":"Abstract"},{"comment":"The sentence 'We see this in Section 2.4, where it is shown that the L2-kernel is outperformed when functions differ by smoothness' appears to refer to the simulation results, which are in Section 4, not Section 2.4.","section":"Section 2.4"},{"comment":"There are several typos in this section: 'a broad a broad class', 'kernels htat', and 'obtaiend' should be corrected.","section":"Supplementary Material C.1"},{"comment":"The notation q_b = min{max[Poisson(m_try),1],P} is unclear; please define m_try and explain whether Poisson(m_try) denotes a Poisson random variable.","section":"Supplementary Algorithm 2"},{"comment":"In the displayed chain of equalities in the proof of Lemma 2, the third line writes w_i where w_{m,i} appears in the neighboring lines; the intended conditioning argument is clear, but the notation should be made consistent.","section":"Supplementary Lemma 2 proof"},{"comment":"The Discussion lists irregular and partially observed sampling as future work, but the NHANES application itself uses discretely sampled activity profiles; a sentence in Section 6 acknowledging that the theory is applied informally to discretely observed curves would help.","section":"Section 6 (NHANES)"}],"recommendation":"major_revision","confidential_remarks":"The main gap between theory and implementation is real but fixable; I would be willing to see a revision that either proves an approximation-error bound for the random-feature split criterion or explicitly restricts the theoretical claims to the exact-MMD estimator and presents the random-feature version as a computational approximation with empirical support. The paper's scope is appropriate for a statistics journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time, and worth refereeing. The core idea is natural: split on MMD between child-node empirical distributions of curves, then average leaf weights and plug in. The genuinely new part is making this work for functional responses in infinite-dimensional Banach spaces, where MMD convergence does not automatically give weak convergence. The tightness-plus-MMD argument in the supplement, upgrading MMD consistency to weak-in-probability in P(Y) and then P(C[0,1]), is a real contribution. The Sobolev transform kernel and the proper scoring rules for functional distributions are useful, and the simulations are honestly run: fDRF beats FLM and KNN precisely when the signal is in dispersion or roughness rather than the mean. The NHANES analysis shows what plug-in inference buys you.\n\nThe soft spot is load-bearing, and it is exactly the one in the stress-test note. Theorems 1–3 are for an ideal infinite forest using the exact MMD split criterion. Algorithm 2 replaces that criterion with a random-feature approximation: B random features, a discretized Brownian motion sampled fresh per split. The paper gives no theorem, uniform bound, or rate condition linking the approximate criterion to the exact one. So the estimator whose performance is actually reported is not the estimator shown to be consistent. The central plug-in claim depends on consistency of the fitted forest, not of a hypothetical exact-MMD forest. The Discussion even says extending to discretely-sampled approximations is future work, which is precisely this gap. This is not fatal; the random-feature approximation is standard and may well preserve consistency. But it needs to be said plainly, and a bound on B or a vanishing-error condition would close the gap.\n\nTwo smaller issues. The little-bag confidence intervals are honestly discussed as overestimating variance, and they work empirically, but the asymptotic justification is only sketched. And there is no public code, despite a claim of code in the supplementary; for a methods paper that is a reproducibility gap.\n\nMy bottom line: the central theoretical claim is plausible and the empirical work is persuasive. The theory-implementation gap is real but addressable. Send it to a serious referee with instructions to focus on the random-feature approximation and the uniform-tightness assumptions.","headline":"A genuine and useful extension of distributional random forests to functional responses, held back by a real gap between the consistency theory (exact MMD, infinite forest) and the implemented estimator (random-feature MMD, finite forest).","tokens_in":35298,"tokens_out":1990,"would_cite":true,"duration_ms":20290,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62G20","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Functional distributional random forests consistently estimate the conditional law of a curve-valued response, so a single fitted forest supports plug-in inference on any functional of that law.","keywords":["functional data analysis","random forests","conditional distribution estimation","maximum mean discrepancy","kernel methods","proper scoring rules","functional response","MMD split criterion"],"falsifier":"Fit the implemented fDRF on data from a Gaussian process whose length scale varies with a one-dimensional covariate, at sample sizes 500 and 4000, and estimate the conditional distribution of $\\sup_t Y(t)$ at a fixed covariate value. If the Wasserstein or Kolmogorov distance between the estimated and true laws of this supremum functional does not shrink as $N$ grows, the random-feature approximation of the MMD split criterion is not preserving the paper's weak-consistency guarantee.","tokens_in":34222,"feed_emoji":"🌲","tokens_out":9578,"duration_ms":83757,"temperature":0.7,"pith_summary":"The paper introduces the functional distributional random forest (fDRF), which estimates the conditional law $F_x$ of a curve-valued response $Y$ given covariates $X$ as a weighted empirical distribution over the observed functions. The paper's central claim is that this estimator is consistent: under assumptions on the tree-growing process, the kernel, and the data, the estimated law $\\hat{F}_x$ converges weakly-in-probability to $F_x$ as an element of $\\mathcal{P}(\\mathcal{Y})$, and with an additional tightness condition also in $\\mathcal{P}(C([0,1]))$. This matters because a single fitted forest can then answer many scientific questions at once—conditional medians, quantiles, peak times, exceedance probabilities, and predictive dispersion—by plugging the estimated law into any functional of interest. The paper also proposes proper scoring rules for comparing functional distribution estimators and shows in simulations and NHANES accelerometer data that distributional changes missed by functional linear models are recovered.","feed_headline":"One forest learns the whole distribution of curve data","feed_subtitle":"A single fit yields conditional medians, quantiles, peak times, and dispersion for functional responses.","key_machinery":"The central machinery is the MMD split criterion together with its random-feature approximation. For a positive definite kernel $\\kappa$ on the function space, each candidate split is scored by the size-weighted empirical MMD between the response samples in the two child nodes, $\\frac{n_{bL} n_{bR}}{n_b^2} D_{\\mathrm{MMD}(\\kappa)}$, which is CART applied to the RKHS feature map $\\varphi(Y_i)$. To make this computable, the paper writes a Gaussian kernel on $L^2([0,1])$ as an expectation over cosine features of an isonormal Gaussian process $W(y)=\\int_0^1 y(t)\\,dB(t)$, then approximates it with $B$ random features that include discretized Brownian motion and are sampled afresh for each split; the same construction extends to other radial kernels through mixing variables. The theoretical guarantee is proved for the ideal infinite-forest estimator whose weights $w_i(x)$ are conditional expectations over tree randomness, under honesty, $\\alpha$-regularity, subsampling, and uniform tightness assumptions.","core_discovery":"The central claim, stated on the paper's own terms, is that the conditional law $F_x = \\mathcal{L}(Y \\mid X = x)$ of a functional response can be estimated nonparametrically by a random forest whose splits maximize a maximum mean discrepancy (MMD) between the empirical distributions in the child nodes, with the estimated law $\\hat{F}_x = \\sum_{i=1}^N w_i(x)\\delta_{Y_i}$. Theorem 2 proves $\\hat{F}_x \\to F_x$ weakly-in-probability as probability measures on $\\mathcal{Y}$ under Assumptions 1–5, and Theorem 3 upgrades this to $\\mathcal{P}(C([0,1]))$ under Assumption 6. The paper is careful that MMD convergence alone does not imply weak convergence in infinite-dimensional spaces, so the proof adds uniform tightness and uses Prohorov's theorem and a subsequence argument. Once consistency holds, the same $\\hat{F}_x$ can be used for plug-in inference on functionals that are discontinuous in $L^2([0,1])$, such as point evaluations, suprema, peak times, and exceedance sets.","pith_inferences":["Editorial extension: the consistency theorems cover an idealized infinite forest with the exact MMD, while the implemented algorithm uses $B$ random features with discretized Brownian motion resampled for each split; a uniform error bound for this approximation as $B,R\\to\\infty$ would be the natural step to make the applied estimator match the proven one.","Editorial extension: because the estimator is a weighted empirical law, the fDRF could be coupled with smoothed or calibrated weights to serve as a common-support device for causal or mediation estimators, a use the paper lists as future work.","Editorial extension: the same MMD-split-plus-tightness argument should extend to responses in other infinite-dimensional or non-Euclidean metric spaces—direction-valued, graph-valued, or manifold-valued data—whenever a characteristic kernel and uniform tightness are available.","Editorial extension: the little-bags intervals intentionally overstate variability because the within-bag correction is omitted; a testable refinement is to replace the normal quantile with a calibrated bootstrap quantile, especially for nonregular functionals like argmax peak times."],"forward_implications":["One fitted fDRF yields plug-in estimates for any functional $\\tau(F_x)$—conditional means, quantiles, exceedance probabilities, peak times, and predictive dispersion—without refitting.","Because $\\hat{F}_x$ is supported on the observed curves, predictive draws preserve realistic geometry and automatically respect constraints such as nonnegativity or density normalization.","Under Assumptions 1–5 the estimated law converges weakly-in-probability in $\\mathcal{P}(\\mathcal{Y})$; adding Assumption 6 extends this to $\\mathcal{P}(C([0,1]))$, justifying inference on point evaluations and suprema.","Energy and kernel scoring rules for functional responses allow comparison of full conditional-distribution estimators; Sobolev-based scores are more sensitive to roughness differences than $L^2$-based scores.","Simulations show the fDRF outperforms K-nearest neighbors and functional linear model baselines on heteroskedastic, mixture, warped, and variable-roughness Gaussian processes, with the weighted Sobolev kernel the strongest default."],"supporting_citations":[{"why":"Supplies the distributional random forest whose MMD split criterion and consistency theory the fDRF extends to functional responses; Theorem 1 ports its Theorem 2.","marker":"´Cevid et al. (2022)"},{"why":"Provides the kernel two-sample test for functional data and the transform-kernel construction, including characteristicness of Sobolev-type kernels used by the fDRF.","marker":"Wynne and Duncan (2022)"},{"why":"Defines the maximum mean discrepancy and its empirical estimator used as the split criterion and in scoring rules.","marker":"Gretton et al. (2012)"},{"why":"Supplies the honesty, subsampling, and alpha-regularity assumptions on tree growth and the asymptotic normality result used for confidence intervals.","marker":"Wager and Athey (2018)"},{"why":"Provides the random Fourier feature representation that makes the MMD split criterion computationally feasible, extended here to Brownian features.","marker":"Rahimi and Recht (2007)"},{"why":"Gives the weak-convergence, tightness, Prohorov, and subsequence tools used to upgrade MMD convergence to weak convergence of probability measures.","marker":"Billingsley (1999)"},{"why":"Provides the theory of proper scoring rules that underlies the energy and kernel scores used to evaluate conditional distribution estimators.","marker":"Gneiting and Raftery (2007)"},{"why":"Supplies the little-bags bootstrap used for pointwise confidence intervals for plug-in functionals of the estimated law.","marker":"Näf et al. (2023)"},{"why":"Supplies generalized random forest theory used to justify inference on smooth functions of conditional moments.","marker":"Athey et al. (2019)"}],"fun_headline_variants":["Forest predicts entire conditional law of curves","Random forest learns full distribution of functional responses","MMD-forest estimates entire curve distribution","One forest yields medians, quantiles, and peak times for curves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is uniform tightness of the conditional laws (Assumption 5), which is what upgrades MMD convergence to genuine weak convergence in infinite-dimensional function spaces; a second fragile point is that the theorem treats an idealized infinite forest with the exact MMD criterion, while the implemented algorithm uses a random-feature approximation whose error is not analyzed.","fun_headline_variants_meta":{"raw":{"variants":["Forest predicts entire conditional law of curves","Random forest learns full distribution of functional responses","MMD-forest estimates entire curve distribution","One forest yields medians, quantiles, and peak times for curves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000795,"raw_usage":{"total_tokens":3512,"prompt_tokens":969,"completion_tokens":2543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2495}},"tokens_in":585,"tokens_out":2543,"duration_ms":19058,"temperature":1.0,"reasoning_tokens":2495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:13:59.458849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the implemented fDRF on data from a Gaussian process whose length scale varies with a one-dimensional covariate, at sample sizes 500 and 4000, and estimate the conditional distribution of $\\sup_t Y(t)$ at a fixed covariate value. If the Wasserstein or Kolmogorov distance between the estimated and true laws of this supremum functional does not shrink as $N$ grows, the random-feature approximation of the MMD split criterion is not preserving the paper's weak-consistency guarantee.","supporting_citations":[],"review_version":1}