Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Seeing the Forest for the Trees: The Gaussian Process Limit of BART

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read With infinitely many trees, BART is a Gaussian process, and its kernel's function space is a tensor-product Sobolev space—so random tree features learn at rates that barely depend on dimension.

desk verdict The GP limit theorem for symmetric BART is a real contribution, but the bridge to the BART people actually use is empirical, not formal. read the letter →

arxiv 2607.28844 v1 pith:DGUGS75U submitted 2026-07-30 math.ST stat.MEstat.MLstat.TH

Cory McCartan , Melody Huang This is my paper · ORCID
classification math.STstat.MEstat.MLstat.TH MSC 62G0860F1762G2046E2260G15
keywords BARTGaussianprocessrandomtreefeaturesreproducingkernelHilbertspacetensor-productSobolevminimaxrateridgeregressionmultiparameterweakconvergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

BART models an unknown regression function as a sum of many small decision trees. This paper asks what happens as the number of trees grows without bound, and proves that a symmetric version of the BART prior converges to a Gaussian process with an explicit covariance kernel k_BART. The kernel's reproducing kernel Hilbert space is norm-equivalent to the tensor-product Sobolev space S^1([0,1]^p), the set of functions built from products of one-dimensional smooth pieces; regression in this space has a minimax error that depends on dimension only through a logarithmic factor. The paper further shows that simply drawing trees from the prior and doing ridge regression on their leaf indicators—random tree features, with no structural learning—achieves this same n^{-1/3} log(n)^{(p-1)/3} rate, which explains the empirical success of BART and offers a cheaper computational route. An ablation study argues that, once many trees are used, Bayesian averaging and tree-structure learning are not what drives performance.

What carries the argument

The load-bearing object is the BART kernel k_BART(x,x') = sigma_mu^2 E[(1 - d_BART(x,x'))^D], which is exactly the covariance of a single tree evaluated at two points—sigma_mu^2 times the probability that x and x' fall in the same leaf. The paper computes this probability in closed form for symmetric trees in terms of the metric d_BART(x,x') = sum_v f_V(v) |F_S(x_v) - F_S(x'_v)|, where F_S is the distribution function of the split threshold and f_V the distribution of the splitting variable. This kernel is what makes the link to tensor-product Sobolev spaces: the proof sandwiches k_BART between two multiples of the ell1-Laplacian product kernel, which has RKHS S^1([0,1]^p). The second key ob

What would settle it

Simulate the covariance of standard asymmetric BART functions in the large-T limit, or compute the same-leaf probability under a branching-process depth prior, and compare it with k_BART; if the two kernels are not Loewner-comparable within universal constants, or if the RKHS is not norm-equivalent to S^1([0,1]^p), the paper's explanation of BART's empirical performance would not carry over to the method actually used.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a limit theorem and its consequences: for symmetric BART functions g_T(x) = T^{-1/2} sum_j h_j(x), with trees drawn i.i.d. from a prior with a fixed depth distribution, Gaussian leaf values, and independent split variables and thresholds, g_T converges weakly on the multidimensional cadlag space D_p to GP(0, k_BART), where k_BART(x,x') = sigma_mu^2 E[(1 - d_BART(x,x'))^D] and d_BART is a covariate-weighted L1 distance derived from the split priors. The paper then proves that the RKHS of k_BART is norm-equivalent to S^1([0,1]^p), so functions in this space can be represented as additive combinations of p-way interactions of univariate smooth

Load-bearing premise

The theoretical limit and RKHS characterization are proved only for symmetric trees with i.i.d. split rules at a common depth, while the paper's extension to the asymmetric, variable-depth trees of standard BART rests on an ablation study rather than a theorem.

Editorial extensions

If this is right

  • As T goes to infinity, finite-dimensional distributions of g_T converge to a multivariate Gaussian with covariance k_BART, and the whole random function converges weakly on D_p; the infinite-tree BART model is literally a Gaussian process.
  • The RKHS is norm-equivalent to the tensor-product Sobolev space S^1([0,1]^p), so functions additively decompose into p-way interactions of univariate smooth components; this explains why BART's learning rates degrade only logarithmically with p.
  • Ridge regression on random tree features drawn from the prior, without any tree-structure learning, attains expected squared error of order n^{-1/3} log(n)^{(p-1)/3} for the stated lambda_n and T_n, matching the minimax rate for S^1; bounded-depth trees need only T_n at least on the order of n^{1/3} log(n)^{1+2(p-1)/3}.
  • The rate beats n^{-1/4} for every p, so random tree features qualify as nuisance-function estimators in double/debiased machine learning procedures.
  • The ablation experiments indicate that with enough trees, Bayesian averaging, tree-structure learning, and symmetric-versus-asymmetric tree shape contribute little to predictive performance, pointing to the additive combination of many small step functions as the key ingredient.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • All theorems concern symmetric trees at a common depth with i.i.d. split rules; the paper's extension of the GP limit to the variable-depth asymmetric trees of standard BART is supported only by an empirical ablation, so the strongest rigorous statement is about the symmetric variant.
  • The random-feature perspective suggests a template for embedding BART-like flexibility into any model with a linear predictor—hazard models, varying-coefficient models, fixed-effects specifications—without writing bespoke MCMC samplers; a natural test is whether the n^{-1/3} rate persists with a light-tailed but unbounded depth prior instead of a bounded-depth assumption.
  • Since the kernel is anisotropic and shaped by f_V and F_S, choosing these split priors to match the true covariate distribution (or adapting them to the data) could improve finite-sample performance; the paper leaves this tuning question open.
  • The uncertainty-quantification experiments show that restricting split locations to the observed covariate range makes posterior intervals from random features track full BART rather than the limiting GP, suggesting that finite-tree effects remain visible and may matter for calibration.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies the large-ensemble limit of Bayesian Additive Regression Trees. It defines a BART function g_T as a T^{-1/2}-scaled sum of T independent tree functions, and proves that for a specially defined 'symmetric tree' prior (every leaf at the same depth D, with D i.i.d. split rules), g_T converges weakly in the multiparameter cadlag space D_p to a Gaussian process GP(0,k_BART), where k_BART(x,x') = sigma_mu^2 E[(1-d_BART(x,x'))^D] (Theorem 3.1). It then shows that the RKHS of k_BART is norm-equivalent to the tensor-product Sobolev space S^1([0,1]^p) (Proposition 3.3), obtains eigenvalue decay j^{-2} log^{2(p-1)}(j) for the corresponding integral operator, and derives n^{-1/3} log^{(p-1)/3} learning rates for ridge regression on random tree features drawn from the symmetric prior (Theorem 4.1). A substantial empirical section compares random tree features with full BART, boosted trees, and random forests, and studies uncertainty quantification. The key caveat, acknowledged in the text only through an empirical ablation, is that the theoretical limit and RKHS characterization are for symmetric trees, whereas standard BART as implemented in the experiments uses an asymmetric branching-process prior.

Significance. If the results hold as stated, they give a clean and surprising explanation of BART's behavior: in the infinite-tree limit the ensemble is a Gaussian process whose RKHS is the space of functions with dominating mixed derivatives, yielding rates whose dimension dependence is only logarithmic. The derivations are parameter-free in the theorems, with no fitted constants entering the kernel or the rates; the random-feature ridge estimator is a concrete, computationally attractive object. The proof strategy is novel in adapting Bickel-Wichura tightness to tree ensembles and in sandwiching the BART kernel between tensor-product Laplacian kernels. These are real contributions. However, the central theorem is proved only for symmetric trees, and the paper's claim to explain BART 'as actually used' is therefore currently supported by an empirical ablation rather than by a theorem. The value of the paper is thus conditional: the symmetric-tree results and the random-feature rates stand on their own, but the connection to standard BART is not established.

major comments (4)
  1. [§2.2, §3.1–3.3, §4] The Gaussian-process limit and the RKHS equivalence are proved only for symmetric trees, in which every leaf is at depth D and the D split rules are i.i.d. Standard BART — including the dbarts implementation used in the paper's own experiments — uses an asymmetric branching-process prior with variable-depth leaves. This gap is load-bearing: for any symmetric tree with D>=1, P[x ~_h x'] = 0 when d_BART(x,x')=1, but for an asymmetric branching-process prior there is positive probability that the root is terminal and x,x' remain in the same leaf even at maximal distance. Hence the limiting covariance of asymmetric BART, if it exists, is not k_BART, and the Loewner sandwich of Proposition 3.3 need not transfer. The only bridge offered in Sections 2.3 and 5.1 is an empirical ablation, not a theorem. The paper should either extend the theory to the asymmetric prior or substantially reframe the
  2. [Appendix A, Lemma A.6] Lemma A.6 is the tightness engine behind Theorem 3.1, but its proof is largely a sketch. The text says the proof 'closely follows' Bickel-Wichura, refers to their base case (i), and in the inductive step says the bound for M''(g) follows 'by the inductive hypothesis (see Bickel and Wichura, 1971)' without displaying the needed conditions. Since the paper explicitly modifies the framework (non-vanishing lower boundary, the C_T class, and the 1/T term), the reader cannot verify that these modifications preserve the induction. Please give a complete, self-contained proof of Lemma A.6, including the base case and the inductive step.
  3. [Appendix B, Lemma B.8] Lemma B.8 relies on the statement that the eigenpairs of the exponential covariance kernel on a symmetric interval are known explicitly, with eigenvalues 2/(1+omega_j^2), citing Xiu (2010), Example 4.1. That exact eigenvalue formula is not valid for the kernel L(s,t)=exp(-|s-t|) on [0,1] used in the proof: the integral operator on a bounded interval has boundary-layer terms, and sinusoids with the stated eigenvalues are not exact eigenfunctions. The needed bound |∫_I e_j|^2 ≲ lambda_j can likely be recovered from Sturm-Liouville asymptotics for the exponential kernel, but as written the proof is incorrect. This affects F_infty(lambda) ≲ N(lambda) and therefore the bounded-depth statement of Theorem 4.1; it should be repaired with a correct spectral argument.
  4. [§5.2–5.3] The empirical random tree features are compared with full BART and with the GP limit, but the text does not specify whether the random tree features used in the simulations are drawn from the symmetric prior (the object of Theorem 4.1) or from the standard asymmetric BART prior. If the latter, the experiments are outside the theory; if the former, this should be stated explicitly and the comparison to full BART should be interpreted accordingly. This is a presentation issue connected to the main scope gap, but it is important for reproducibility.
minor comments (5)
  1. [Abstract/§1] The abstract and introduction state that 'BART converges to a Gaussian process' without the qualifier 'symmetric-tree'. Given that the main theorem is for symmetric trees, the caveat should appear in the abstract.
  2. [§4.1, Proposition 4.1] Proposition 4.1 gives a n^{-1/3} log^{p-1}(n) rate for kernel ridge regression, while Theorem 4.1 gives n^{-1/3} log^{(p-1)/3}(n). If the larger log power comes from the cited Bak-Lee bound, please say so explicitly; as written, the two log exponents appear inconsistent and may confuse readers about the minimax-optimal rate.
  3. [Appendix B, Lemma B.4] The proof of Lemma B.4 refers to 'an algebra error in the definition of their sigma-bar' and says it is corrected, but the correction is not displayed. Please state precisely what the correct expression is.
  4. [Appendix B, Theorem B.1] The notation '(omega-) probability' in Lemma B.4 is not defined. Either define it or rephrase to 'probability conditional on the random features'.
  5. [§2.1] The definition of symmetric tree says the order of the D decision rules is arbitrary; later, for the co-leaf probability, the split rules are drawn i.i.d. and the result is independent of order. This is fine, but the sentence 'the order ... is arbitrary' is a bit ambiguous and could be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the limiting kernel, RKHS, and learning rates are derived from the stated symmetric-tree prior without fitted parameters or load-bearing self-citation.

full rationale

The paper's main chain is self-contained. k_BART is computed from the symmetric BART prior (Prop 3.1: covariance equals sigma_mu^2 times co-leaf probability; Prop 3.2 evaluates that probability as E[(1-d_BART)^D]; Cor 3.1 is a closed-form special case). Theorem 3.1's GP convergence uses the CLT for finite-dimensional distributions (Lemma 3.1) and a tightness proof that adapts Bickel-Wichura via moment bounds (Lemmas A.2-A.6); no target GP is assumed. Proposition 3.3's Loewner sandwich c K_tensor <= k_BART <= C K_tensor is proved by explicit Fourier/operator bounds on the triangular kernels, not by assuming the Sobolev equivalence. Corollary 4.1's eigenvalue decay j^{-2} log^{2(p-1)}(j) follows from the sandwich plus the known spectrum of the tensor-product Laplacian kernel, and Theorem 4.1 obtains the n^{-1/3} log^{(p-1)/3} rate from a generalized Rudi-Rosasco bound with effective dimension N(lambda) as lambda^{-1/2} log^{p-1}(1/lambda); no fitted constants or data-dependent parameters enter. The only scope limitation, that the theorems concern symmetric trees while standard BART uses asymmetric branching-process trees, is explicitly acknowledged and bridged empirically (Sections 2.3, 5.1), but this is an external-validity gap, not a circular derivation. The Section 5.3 simulation draws the true function from GP(0,k_BART), but that is a designed illustration for UQ comparison, not a claim that derives the kernel from data; it does not feed back into Theorems 3.1-4.1. No self-citation is load-bearing, and no prediction is equivalent to a fitted input by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central derivation rests on the symmetric-tree prior, the source condition g0∈H_BART, standard regularity (bounded density, sub-exponential noise), and finite moment assumptions on tree depth. The free parameters are prior scaling and tuning choices that do not enter the asymptotic theorems. No new forces or entities are postulated; random tree features are a constructive approximation of the derived kernel, not an ad hoc invention.

free parameters (4)
  • leaf prior variance σ_μ^2 = not fitted in theory; in ablation experiments tuned with LOOCV (bottom panels of Figure 3), otherwise default
    Scales the kernel k_BART but does not affect the structural claims; any positive value yields the same RKHS and rates up to constant factors.
  • depth prior parameter r (Poisson D|D>0) = r=1.25 chosen to match the mean depth of the traditional BART prior (Section 3.1, Corollary 3.1)
    Determines the concrete kernel in Corollary 3.1 and the length scale in Figure 2; the theory allows any regular depth prior, so this choice is illustrative rather than fit to the data.
  • ridge penalty λ_n = theory sets λ_n ≍ n^{-2/3} log(n)^{2(p-1)/3}; in experiments tuned via LOOCV/GCV
    The asymptotic choice is derived to balance bias and variance in the proof, not fitted to data; the experimental tuning is standard practice and does not feed back into the theorems.
  • bounded depth d+ (in Theorem 4.1 second statement) = not specified; standard BART priors satisfy P(D>8)<10^{-5}
    A theoretical condition enabling the sharper T_n ≍ n^{1/3} rate; the authors argue it is not restrictive for standard priors.
assumptions (6)
  • domain assumption Symmetric tree structure: all leaves at depth D, split rules (V_k,S_k) i.i.d., V draws from f_V with full support, S from continuous f_S with positive density
    This is the core modeling simplification; the theorems and kernel derivation depend on it. It is stated in Section 2.2 and used throughout.
  • domain assumption g0 ∈ H_BART with ||g0||_{H_BART} ≤ R (source condition) for the random-feature rates
    The learning-rate results (Theorem 4.1) assume the true CEF lies in the RKHS of the limiting kernel; this is a standard but load-bearing smoothness condition, stated in Section 4.
  • domain assumption Covariate density ρ bounded above and below on [0,1]^p; sub-exponential noise (bounded conditional moments)
    Regularity conditions on the data-generating process, stated in Section 4, used for eigenvalue comparison (Lemma A.1) and for the Bernstein tail bounds.
  • domain assumption Depth distribution D has E[D^{2p}] < ∞ for Theorem 3.1, and E[D^p]<∞ with P(D≥p)>0 for Proposition 3.3
    Finite moments ensure the number of split rules is controlled; the positivity condition ensures the RKHS contains p-way interactions. Standard BART priors satisfy these.
  • domain assumption F_S (split CDF) is Lipschitz continuous with constant K_S
    Used in Lemma A.2 to bound separation probability by L1 distance, a key step for tightness and for the second-moment bounds.
  • standard math Operator Bernstein inequalities, negative association of balls-and-bins, spectral characterization of stationary kernels, Schur product theorem, explicit eigenfunctions of the exponential kernel, eigenvalue counting for tensor products
    Imported background results cited to Rudi-Rosasco (2017), Dubhashi-Ranjan (1996), Ritter et al. (1995), Xiu (2010), Zhang-Simon (2023), Kanagawa et al. (2018). These are standard tools; the paper does not prove them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seeing the Forest for the Trees: The Gaussian Process Limit of BART." pith.science (2026). https://pith.science/paper/DGUGS75U

@misc{pith2026260728844,
  author       = {Pith},
  title        = {Pith review of: Seeing the Forest for the Trees: The Gaussian Process Limit of BART},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGUGS75U}},
  note         = {Machine review of arXiv:2607.28844}
}
read the original abstract

Bayesian Additive Regression Trees (BART) have shown state-of-the-art performance in both prediction and causal inference problems. Previous theoretical work has attempted to explain BART's superior performance by establishing posterior contraction rates for standard BART models, but these rates depend strongly on the number of covariates. Here, we take a different approach and study the behavior of BART as the number of trees grows towards infinity. We show that in this regime, BART converges to a Gaussian process (GP) with a particular kernel. The kernel and its corresponding reproducing kernel Hilbert space (RKHS) have favorable inferential properties that help explain BART's excellent performance. We introduce *random tree features* as an approximation to this limiting GP, and establish minimax-optimal learning rates for ridge regression on these random features that depend only logarithmically on dimension. In addition to providing insight into the empirical success of BART, random tree features offer a computational benefit over traditional MCMC estimation. The random-features approximation also allows practitioners to easily incorporate BART into any model which has a linear predictor, expanding the applicability and flexibility of BART.

Figures

Figures reproduced from arXiv: 2607.28844 by the authors.

Figure 1
Figure 1. Three ablation studies on abalone data (𝑛 = 4 177, 𝑝 = 8). Each panel compares out-of-sample 𝑅 2 values of the posterior mean, for two different regression fits and different numbers of regression trees. The average difference in 𝑅 2 across 20 cross validation splits and multiple replicates is plotted along with 95% confidence intervals. (a) Ablating Bayes: 𝑅 2 for the full posterior mean versus the posterior mean c… view at source ↗
Figure 2
Figure 2. Visualization of 𝑘BART. The left panel shows 𝑘BART(·, 0) for different choices of the depth prior 𝑓𝐷. The colored lines correspond to 𝑓𝐷 that place all mass on a single depth, while the black dashed line corresponds to the Poisson 𝑓𝐷 in Corollary 3.1, with 𝑟 = 1.25 chosen to match the mean depth of the traditional BART prior. The right panel shows 𝑘BART(·, (0.5, 0.5)) on the unit square, using the same calibrated Po… view at source ↗
Figure 3
Figure 3. Full ablation study. Each panel compares out-of-sample 𝑅 2 values of the posterior mean, for two different regression fits and different numbers of regression trees, across 16 different datasets. The average difference in 𝑅 2 across 20 cross validation splits and multiple replicates is plotted along with 95% confidence intervals, with each dataset’s line colored according to the ratio of observations to covariates. … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Predictive performance of random BART features [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Comparison of uncertainty quantification [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 7 linked inside Pith

  1. [1]

    and Lee, W

    Bak, K.-Y. and Lee, W. (2025). Effect of dimensionality on convergence rates of kernel ridge regression estimator. Journal of Statistical Planning and Inference , 236:106228

    2025

  2. [2]

    M., and Teh, Y

    Balog, M., Lakshminarayanan, B., Ghahramani, Z., Roy, D. M., and Teh, Y. W. (2016). The mondrian kernel. arXiv preprint arXiv:1606.05241

    arXiv 2016

  3. [3]

    and Van der Laan , M

    Benkeser, D. and Van der Laan , M. (2016). The highly adaptive lasso estimator. In 2016 IEEE international conference on data science and advanced analytics (DSAA) , pages 689--696. IEEE

    2016

  4. [4]

    Bickel, P. J. and Wichura, M. J. (1971). Convergence criteria for multiparameter stochastic processes and some applications. The Annals of Mathematical Statistics , 42(5):1656--1670

    1971

  5. [5]

    and Guestrin, C

    Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages 785--794

    2016

  6. [6]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/debiased machine learning for treatment and structural parameters

    2018

  7. [7]

    Chipman, H., George, E., and McCulloch, R. (2007). Bayesian ensemble learning. Advances in neural information processing systems , 19

    2007

  8. [8]

    A., George, E

    Chipman, H. A., George, E. I., and McCulloch, R. E. (2010). Bart: Bayesian additive regression trees. The Annals of Applied Statistics , 4(1):266--298

    2010

Show all 41 references
  1. [9]

    Deshpande, S. K. (2025). flexbart: Flexible bayesian regression trees with categorical predictors. Journal of Computational and Graphical Statistics , 34(3):1117--1126

    2025

  2. [10]

    K., Bai, R., Balocchi, C., Starling, J

    Deshpande, S. K., Bai, R., Balocchi, C., Starling, J. E., and Weiss, J. (2026). Vcbart: Bayesian trees for varying coefficients. Bayesian Analysis , 21(1):281--308

    2026

  3. [11]

    Dorie, V. (2026). dbarts: Discrete Bayesian Additive Regression Trees Sampler . R package version 0.9-33

    2026

  4. [12]

    Dorie, V., Hill, J., Shalit, U., Scott, M., and Cervone, D. (2019). Automated versus do-it-yourself methods for causal inference: Lessons learned from a data analysis competition. Statistical Science , 34(1):43--68

    2019

  5. [13]

    Dubhashi, D. P. and Ranjan, D. (1996). Balls and bins: A study in negative dependence. BRICS Report Series , 3(25)

    1996

  6. [14]

    Hill, J., Linero, A., and Murray, J. (2020). Bayesian additive regression trees: A review and look forward. Annual Review of Statistics and Its Application , 7(1):251--278

    2020

  7. [15]

    and Rockova, V

    Jeong, S. and Rockova, V. (2023). The art of bart: Minimax optimality over nonhomogeneous smoothness in high dimension. Journal of Machine Learning Research , 24(337):1--65

    2023

  8. [16]

    Kanagawa, M., Hennig, P., Sejdinovic, D., and Sriperumbudur, B. K. (2018). Gaussian processes and kernel methods: A review on connections and equivalences. arXiv preprint arXiv:1807.02582

    2018 arXiv

  9. [17]

    Kern, J. (2024). Skorokhod topologies: What they are and why we should care. Mathematische Semesterberichte , 71(1):1--18

    2024

  10. [18]

    Lin, Y. (2000). Tensor product space anova models. The Annals of Statistics , 28(3):734--755

    2000

  11. [19]

    Linero, A. R. (2017). A review of tree-based bayesian methods. Communications for Statistical Applications and Methods , 24(6)

    2017

  12. [20]

    Linero, A. R. (2018). Bayesian regression trees for high-dimensional prediction and variable selection. Journal of the American Statistical Association , 113(522):626--636

    2018

  13. [21]

    R., Bose, S., and Murray, J

    Linero, A. R., Bose, S., and Murray, J. (2026). Bayesian additive distribution regression. arXiv preprint arXiv:2603.06462

    2026

  14. [22]

    Linero, A. R. and Yang, Y. (2018). Bayesian regression tree ensembles that adapt to smoothness and sparsity. Journal of the Royal Statistical Society Series B: Statistical Methodology , 80(5):1087--1110

    2018

  15. [23]

    Liu, F., Huang, X., Chen, Y., and Suykens, J. A. (2021). Random features for kernel approximation: A survey on algorithms, theory, and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(10):7128--7148

    2021

  16. [24]

    Murray, J. S. (2021). Log-linear bayesian additive regression trees for multinomial logistic and count regression models. Journal of the American Statistical Association , 116(534):756--769

    2021

  17. [25]

    Nash, W., Sellers, T., Talbot, S., Cawthorn, A., and Ford, W. (1994). Abalone . UCI Machine Learning Repository. DOI : https://doi.org/10.24432/C55C7W

    1994 doi

  18. [26]

    Neuhaus, G. (1971). On weak convergence of stochastic processes with multidimensional time parameter. The Annals of Mathematical Statistics , 42(4):1285--1295

    1971

  19. [27]

    Petrillo, G. (2024). On the gaussian process limit of bayesian additive regression trees. arXiv preprint arXiv:2410.20289

    2024 arXiv

  20. [28]

    and Recht, B

    Rahimi, A. and Recht, B. (2007). Random features for large-scale kernel machines. Advances in neural information processing systems , 20

    2007

  21. [29]

    W., and Wo \'z niakowski, H

    Ritter, K., Wasilkowski, G. W., and Wo \'z niakowski, H. (1995). Multivariate integration and approximation for random fields satisfying sacks-ylvisaker conditions. The Annals of Applied Probability , pages 518--540

    1995

  22. [30]

    and Saha, E

    Ro c kov \'a , V. and Saha, E. (2019). On theory for BART . In The 22nd International Conference on Artificial Intelligence and Statistics , pages 2839--2848. PMLR

    2019

  23. [31]

    and van der Pas , S

    Ro c kov \'a , V. and van der Pas , S. (2020). Posterior concentration for Bayesian regression trees and forests. The Annals of Statistics , 48(4):2108--2131

    2020

  24. [32]

    and Rosasco, L

    Rudi, A. and Rosasco, L. (2017). Generalization properties of learning with random features. Advances in neural information processing systems , 30

    2017

  25. [33]

    Schuler, A., Hagemeister, A., and van der Laan , M. (2024). Highly adaptive ridge. arXiv preprint arXiv:2410.02680

    2024 arXiv

  26. [34]

    Schuler, A., Li, Y., and van der Laan , M. (2022). Lassoed tree boosting. arXiv preprint arXiv:2205.10697

    2022 arXiv

  27. [35]

    A., Logan, B

    Sparapani, R. A., Logan, B. R., McCulloch, R. E., and Laud, P. W. (2016). Nonparametric survival analysis using bayesian additive regression trees (bart). Statistics in medicine , 35(16):2741--2753

    2016

  28. [36]

    Van der Laan , M. (2023). Higher order spline highly adaptive lasso estimators of functional parameters: Pointwise asymptotic normality and uniform convergence rates. arXiv preprint 2301.13354

    2023 arXiv

  29. [37]

    and Ro c kov \'a , V

    Van der Pas , S. and Ro c kov \'a , V. (2017). Bayesian dyadic trees and histograms for regression. Advances in Neural Information Processing Systems , 30

    2017

  30. [38]

    Wright, M. N. and Ziegler, A. (2017). ranger : A fast implementation of random forests for high dimensional data in C++ and R . Journal of Statistical Software , 77(1):1--17

    2017

  31. [39]

    Xiu, D. (2010). Numerical Methods for Stochastic Computations: A Spectral Method Approach . Princeton University Press

    2010

  32. [40]

    Yee, R., Ghosh, S., and Deshpande, S. K. (2024). Scalable piecewise smoothing with bart. arXiv preprint arXiv:2411.07984

    2024 arXiv

  33. [41]

    and Simon, N

    Zhang, T. and Simon, N. (2023). Regression in tensor product spaces by the method of sieves. Electronic journal of statistics , 17(2):3660

    2023

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.