Pith. sign in

REVIEW 1 major objections 5 minor 1 cited by

Classic kernel density estimates yield almost-sure weakly convergent predictives and, for Gaussians, a limiting density with credibility intervals.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 22:57 UTC pith:XZGPBZG3

load-bearing objection Solid independent theory for classical KDE predictives: a.s. weak convergence without c.i.d./a.c.i.d., plus Gaussian absolute continuity that justifies density credibility bands. the 1 major comments →

arxiv 2607.03927 v1 pith:XZGPBZG3 submitted 2026-07-04 stat.ME

Posterior uncertainty for kernel density estimates

Dennis Christensen , Torjus Svardal , Leiv R{\o}nneberg , Emil A. Stoltenberg This is my paper
classification stat.ME MSC 62G0762C1060F1560G09
keywords kernel density estimationpredictive Bayesian inferencemartingale posterioralmost sure weak convergenceasymptotic exchangeabilityGaussian kernelscredibility intervals
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper places ordinary kernel density estimation inside the predictive Bayesian framework. Starting from a finite sample and a KDE, one draws future observations from the updated kernel density and continues forever. The authors prove that the resulting sequence of predictive measures converges weakly almost surely to a random probability measure, so the simulated sequence is almost surely asymptotically exchangeable and a martingale posterior exists. They further show that these predictives are neither conditionally identically distributed nor almost c.i.d., giving a concrete counter-example that weak convergence can hold without those standard sufficient conditions. For the Gaussian kernel they strengthen the result: the limiting measure is almost surely absolutely continuous, hence possesses a density. That density supplies the missing object needed for posterior credibility bands around the KDE itself, which the authors compute on the Old Faithful and Galaxies data.

Core claim

Under standard moment conditions on the kernel, the sequence of random probability measures defined by classical KDEs (bandwidth decaying as n to a negative power) converges weakly almost surely to a random limiting measure; for Gaussian kernels that limit is almost surely absolutely continuous with respect to Lebesgue measure and therefore admits a density that can be used for credibility intervals on the density estimate.

What carries the argument

Almost-sure weak convergence of the predictive measures α_n via an almost-supermartingale argument on smooth test functions (Theorem 2.1), together with a uniform-integrability and relative-compactness argument that upgrades the Gaussian limit to absolute continuity (Theorem 2.3).

Load-bearing premise

Absolute continuity of the limiting measure—and therefore the existence of a limiting density usable for credibility intervals—is proved only when the kernel is Gaussian; for general kernels it remains a conjecture.

What would settle it

Construct a non-Gaussian kernel satisfying the paper’s moment conditions for which the sequence of predictive densities fails to be relatively compact in L1, or for which the limiting measure charges a set of Lebesgue measure zero with positive probability; either event would refute the conjectured absolute-continuity claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 5 minor

Summary. The paper places classical kernel density estimation inside the predictive Bayesian (martingale posterior) framework. Under standard moment conditions on the kernel and bandwidths h_n = c n^{-a} (a>0), the associated predictive measures α_n are shown to converge weakly almost surely to a random probability measure α (Theorem 2.1; Corollary 3.2 for the intrinsic/predictive-resampling version that re-uses the observed sample). Consequently the resampling sequence is a.s. asymptotically exchangeable. For Gaussian kernels the sequence is proved not to be a.c.i.d. (Proposition 2.2), furnishing a non-trivial example of a.s. weak convergence outside the usual c.i.d./quasi-martingale/a.c.i.d. criteria. Absolute continuity of the limiting measure with respect to Lebesgue measure is established for Gaussian kernels (Theorem 2.3), which legitimates pointwise credibility bands for the density itself. Estimators for moments of α are obtained via quasi-martingale arguments, and the procedure is illustrated on the Old Faithful and Galaxies data sets.

Significance. If the theorems hold, the work supplies a rigorous predictive-Bayes foundation for uncertainty quantification of classical KDEs, including density-level credibility intervals that do not require a Dirichlet-process or other fully Bayesian nonparametric prior. The explicit non-a.c.i.d. counter-example is of independent interest for the growing literature on predictive constructions. The absolute-continuity argument for Gaussians (L^{2} bound via almost-supermartingales, de-la-Vallée-Poussin, relative compactness in L^{1}) is self-contained and carefully checked; the real-data illustrations are transparent and reproducible in spirit. The paper openly flags the concurrent Hilbert preprint and the open status of total-variation convergence and non-Gaussian absolute continuity, which strengthens rather than weakens the contribution.

major comments (1)
  1. Theorem 2.3 (and its supporting Lemma A.2) claims absolute continuity for the Gaussian case under the general bandwidth regime (2.2), i.e., any a>0. The almost-supermartingale argument for the quadratic functional Q_n, however, requires ∑ n^{a-2}<∞ and therefore a<1 (explicitly noted in the proof of Lemma A.2). Standard rates satisfy a=1/5, but the theorem statement should be restricted to a∈(0,1) or the argument extended; otherwise the absolute-continuity claim that underpins the credibility intervals is not fully justified for the full range of a advertised in (2.2).
minor comments (5)
  1. Section 2.2 and the paragraph after Theorem 2.3 correctly flag that absolute continuity is proved only for Gaussians and that total-variation convergence remains open. A one-sentence forward reference in the abstract or introduction would help readers who primarily want the credibility-interval application.
  2. In the real-data experiments (Section 4) the bandwidth is written h_n=n^{-1/5}; for the intrinsic scheme of (3.1) it should be h_n∝(n_obs+n)^{-1/5}. The numerical difference is negligible for large N, but the notation should be consistent with the theory of Section 3.
  3. Appendix D sketches the posterior mean as an infinite product of linear operators. The calculation is suggestive but incomplete; either expand it into a short proposition or move it to a remark so that it does not appear as an unfinished claim.
  4. A few typographical slips: “predicitive” (Section 3), “sequece” (Section 3), and the double use of Y_i for both observed and simulated data (already flagged by the authors) could be cleaned for readability.
  5. Figure 2 (DPMM comparison) shows uncertainty bands that are narrowest where data are scarcest; a brief sentence explaining why this is expected under a DPMM (or citing the relevant literature) would help non-specialist readers.

Circularity Check

0 steps flagged

No significant circularity; a.s. weak convergence, non-a.c.i.d. counter-example and Gaussian absolute continuity are derived from first principles via Taylor expansions, Robbins–Siegmund and L1 compactness, with only a non-load-bearing concurrent citation.

full rationale

The paper’s central results (Theorem 2.1 / Corollary 3.2 on a.s. weak convergence of classical KDE predictives, Proposition 2.2 that the sequence is not a.c.i.d., and Theorem 2.3 that the Gaussian limit is a.s. absolutely continuous) rest on transparent, self-contained arguments that do not reduce to their own inputs by construction. Theorem 2.1 expands smooth test functions, obtains an almost-supermartingale remainder controlled by the bandwidth sum, and invokes the Robbins–Siegmund theorem plus the C^∞_b equivalence of Lemma B.3; none of these steps presuppose the claimed limit. Absolute continuity (Theorem 2.3) is obtained from an L^{2} bound on the densities (Lemma A.2, again via almost-supermartingales), de-la-Vallée-Poussin uniform absolute continuity and relative compactness in L^{1}, followed by uniqueness of weak limits; the argument never assumes the density exists. The non-a.c.i.d. calculation (Proposition 2.2) is a direct lower bound showing the one-step difference is order 1/n and therefore not summable. Moment estimators and uniform-integrability lemmas follow similarly from explicit recursions and quasi-martingale convergence. The only self-reference is the open citation of the concurrent Hilbert (2026) preprint, which shares the same resampling scheme but is not used as a premise for any proof; the authors supply an independent argument. Real-data illustrations tune a bandwidth constant post-hoc for visual comparison only and do not enter the theorems. No fitted parameter is renamed a prediction, no uniqueness theorem is imported from overlapping authors, and no known result is merely re-labelled. Score 1 reflects the single non-load-bearing concurrent citation; the derivation chain itself is free of circular reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The load-bearing mathematical claims rest on standard kernel-moment conditions, the bandwidth decay rate, and the Gaussian kernel for absolute continuity. Free parameters appear only in the numerical illustrations (bandwidth constant c and simulation horizon). No new physical entities are postulated; the ‘martingale posterior’ is an existing concept applied to KDEs.

free parameters (2)
  • bandwidth constant c in h_n = c n^{-a}
    In the real-data section the authors set a=1/5 and then empirically tune c≈1.5 so that the ordinary KDE matches the posterior mean; the theoretical theorems hold for any fixed c>0, but the displayed credibility bands depend on this choice.
  • resampling horizon N and number of paths B
    N=100000, B=1000 chosen for the illustrations; convergence diagnostics are shown but the values themselves are free computational parameters.
axioms (4)
  • domain assumption Kernel satisfies ∫z K(z) dz = 0 and ∫z² K(z) dz < ∞ (and higher absolute moments when needed for uniform integrability).
    Invoked throughout Theorems 2.1, 2.3, Lemmas 2.4–2.9 and their intrinsic counterparts; standard for second-order kernels.
  • domain assumption Bandwidth schedule h_n = c n^{-a} with a>0, c>0.
    Used for summability of the almost-supermartingale remainder terms (proof of Theorem 2.1).
  • standard math Robbins–Siegmund nonnegative almost-supermartingale convergence theorem and Berti et al. (2006) criteria for a.s. weak convergence of random measures.
    Core technical engines of the weak-convergence proof.
  • standard math Ionescu–Tulcea theorem guaranteeing existence of the process with the given conditionals.
    Cited to ensure the predictive resampling sequence is well-defined.

pith-pipeline@v1.1.0-grok45 · 28311 in / 2750 out tokens · 24302 ms · 2026-07-11T22:57:44.433326+00:00 · methodology

0 comments
read the original abstract

Recent work in predictive Bayesian inference has enabled novel Bayesian interpretations of many well-known stochastic one-step-ahead predictive algorithms. In this paper, we study classic kernel density estimation in the predictive Bayesian framework. We prove that their predictive measures converge weakly almost surely $\unicode{x2013}$ meaning that their associated predictive resampling sequences are almost surely asymptotically exchangeable $\unicode{x2013}$ and we provide estimators for moments of the limiting random probability measure. We also show that the resampling sequences do not satisfy standard assumptions like being conditionally identically distributed (c.i.d.) or almost c.i.d. (a.c.i.d.), thus providing a non-trivial example of a predictive sequence which is not a.c.i.d. but nevertheless converges weakly almost surely. For Gaussian kernels, we show that the limiting directing measure is almost surely absolutely continuous with respect to the Lebesgue measure, meaning it emits a probability density. This enables us to derive credibility intervals for kernel density estimates, which we illustrate on two real datasets.

Figures

Figures reproduced from arXiv: 2607.03927 by Dennis Christensen, Emil A. Stoltenberg, Leiv R{\o}nneberg, Torjus Svardal.

Figure 1
Figure 1. Figure 1: (a) Martingale posterior of KDE applied to the ’Old Faithful’ dataset of [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A DPMM applied to the old faithful dataset, annotations the same as in Figure [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Martingale posterior of the KDE fit to the Galaxies dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Some cautionary tales about Bayesian predictive inference

    math.ST 2026-07 conditional novelty 5.0

    Bayesian predictive rules that are asymptotically exchangeable can behave very differently from their exchangeable limit — even failing the law of large numbers the limit satisfies.

Reference graph

Works this paper leans on

64 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    1985 , publisher=

    Exchangeability and related topics , author=. 1985 , publisher=

    1985

  2. [2]

    The Annals of Probability , pages=

    On mixing and stability of limit theorems , author=. The Annals of Probability , pages=

  3. [3]

    Mixtures of

    Antoniak, Charles E , journal=. Mixtures of

  4. [4]

    , title =

    Billingsley, P. , title =

  5. [5]

    Bogachev, Vladimir Igorevich , year=. Measure

  6. [6]

    Statistical Science , volume=

    Martingale Posterior Distributions for Time-Series Models , author=. Statistical Science , volume=

  7. [7]

    The total variation distance between high-dimensional

    Devroye, Luc and Mehrabian, Abbas and Reddad, Tommy , journal=. The total variation distance between high-dimensional

  8. [8]

    Stochastics and Stochastic Reports , volume=

    Almost sure weak convergence of random probability measures , author=. Stochastics and Stochastic Reports , volume=

  9. [9]

    The Annals of Probability , volume =

    Exchangeable sequences driven by an absolutely continuous random measure , author=. The Annals of Probability , volume =

  10. [10]

    Journal of Applied Probability , volume=

    A central limit theorem and its applications to multicolor randomly reinforced urns , author=. Journal of Applied Probability , volume=

  11. [11]

    Stochastic

    Cohen, Samuel N and Elliott, Robert James , year=. Stochastic

  12. [12]

    Foundations of

    Kallenberg, Olav , year=. Foundations of

  13. [13]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Martingale posterior distributions , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

  14. [14]

    Martingale

    Hall, Peter and Heyde, Christopher C , year=. Martingale

  15. [15]

    Scandinavian Journal of Statistics , pages=

    Central limit theorems for martingales with discrete or continuous time , author=. Scandinavian Journal of Statistics , pages=

  16. [16]

    Ferguson, Thomas S , year=. A

  17. [17]

    Fundamentals of

    Ghosal, Subhashis and van der Vaart, Aad W , year=. Fundamentals of

  18. [18]

    arXiv preprint arXiv:2507.19169 , year=

    Weak convergence of predictive distributions , author=. arXiv preprint arXiv:2507.19169 , year=

    Pith/arXiv arXiv

  19. [19]

    A detailed treatment of

    Miller, Jeffrey W , journal=. A detailed treatment of

  20. [20]

    2014 , journal=

    Steven Lalley , title =. 2014 , journal=

    2014

  21. [21]

    H. Stable. 2015 , publisher=

    2015

  22. [22]

    arXiv preprint arXiv:2605.14008 , year=

    Predictive Inference via Kernel Density Estimates , author=. arXiv preprint arXiv:2605.14008 , year=

    Pith/arXiv arXiv

  23. [23]

    Convergence of

    Pollard, David , year=. Convergence of

  24. [24]

    Second Edition , author=

    Probability. Second Edition , author=. 1996 , publisher=

    1996

  25. [25]

    Le calcul des probabilit

    Application of the theory of martingales , author=. Le calcul des probabilit

  26. [26]

    Jacod, Jean and Shiryaev, Albert N , year=. Limit

  27. [27]

    Probability

    Jacod, Jean and Protter, Philip , year=. Probability

  28. [28]

    A. W. van der Vaart , year=. Asymptotic

  29. [29]

    Asymptotics in

    Le Cam, Lucien and Yang, Grace Lo , year=. Asymptotics in

  30. [30]

    A revisit to

    Babu, G Jogesh and Li, Bing , journal=. A revisit to

  31. [31]

    1998 , publisher=

    Markov Chains , author=. 1998 , publisher=

    1998

  32. [32]

    Introduction to

    Ross, Sheldon M , year=. Introduction to

  33. [33]

    Dudley, Richard M , year=. Real

  34. [34]

    Exchangeability, prediction and predictive modeling in

    Fortini, Sandra and Petrone, Sonia , journal=. Exchangeability, prediction and predictive modeling in

  35. [35]

    Philosophical Transactions of the Royal Society A , volume=

    Prediction-based uncertainty quantification for exchangeable sequences , author=. Philosophical Transactions of the Royal Society A , volume=

  36. [36]

    Philosophical Transactions of the Royal Society A , volume=

    Statistical inference with exchangeability and martingales , author=. Philosophical Transactions of the Royal Society A , volume=

  37. [37]

    The Annals of Probability , volume =

    Limit theorems for a class of identically distributed random variables , author=. The Annals of Probability , volume =

  38. [38]

    A probabilistic view on predictive constructions for

    Berti, Patrizia and Dreassi, Emanuela and Leisen, Fabrizio and Pratelli, Luca and Rigo, Pietro , journal=. A probabilistic view on predictive constructions for

  39. [39]

    Kernel based

    Berti, Patrizia and Dreassi, Emanuela and Leisen, Fabrizio and Pratelli, Luca and Rigo, Pietro , journal=. Kernel based

  40. [40]

    Bogachev, Vladimir I , year=. Measure

  41. [41]

    arXiv preprint arXiv:2507.21874 , year=

    Bayesian predictive inference beyond martingales , author=. arXiv preprint arXiv:2507.21874 , year=

    arXiv

  42. [42]

    Theory of

    Schervish, Mark J , year=. Theory of

  43. [43]

    Probability with

    Williams, David , year=. Probability with

  44. [44]

    Probability:

    Durrett, Rick , volume=. Probability:. 2019 , publisher=

    2019

  45. [45]

    An almost sure conditional convergence result and an application to a generalized

    Crimaldi, Irene , booktitle=. An almost sure conditional convergence result and an application to a generalized

  46. [46]

    Econometrica , volume=

    Inference for continuous semimartingales observed at high frequency , author=. Econometrica , volume=

  47. [47]

    arXiv preprint arXiv:2403.16828 , year=

    Asymptotics of predictive distributions driven by sample means and variances , author=. arXiv preprint arXiv:2403.16828 , year=

    Pith/arXiv arXiv

  48. [48]

    van der Vaart, Aad W and Wellner, Jon A , year=. Weak

  49. [49]

    The Annals of Statistics , pages=

    Nonparametric density estimation with a parametric start , author=. The Annals of Statistics , pages=

  50. [50]

    The Annals of Statistics , volume =

    Smoothing counting process intensities by means of kernel functions , author=. The Annals of Statistics , volume =

  51. [51]

    Scandinavian Journal of Statistics , volume=

    Martingale posterior distributions for cumulative hazard functions , author=. Scandinavian Journal of Statistics , volume=

  52. [52]

    2009 , publisher=

    Introduction to Nonparametric Estimation , author=. 2009 , publisher=

    2009

  53. [53]

    Moment Martingale Posteriors for Semiparametric Predictive

    Yung, Yiu Yin and Lee, Stephen and Fong, Edwin , journal=. Moment Martingale Posteriors for Semiparametric Predictive

  54. [54]

    Optimizing Methods in Statistics , pages=

    A convergence theorem for non negative almost supermartingales and some applications , author=. Optimizing Methods in Statistics , pages=. 1971 , publisher=

    1971

  55. [55]

    2026 , journal =

    Uncertainty Quantification for Prior-Fitted Networks using Martingale Posteriors , author=. 2026 , journal =

    2026

  56. [56]

    arXiv preprint arXiv:2510.25154 , year=

    TabMGP: Martingale Posterior with TabPFN , author=. arXiv preprint arXiv:2510.25154 , year=

    Pith/arXiv arXiv

  57. [57]

    A principled framework for uncertainty decomposition in

    Fortini, Sandra and Ng, Kenyon and Petrone, Sonia and Rousseau, Judith and Wei, Susan , journal=. A principled framework for uncertainty decomposition in

  58. [58]

    The Annals of Applied Probability , pages=

    A functional central limit theorem for branching random walks, almost sure weak convergence and applications to random trees , author=. The Annals of Applied Probability , pages=

  59. [59]

    Confidence,

    Schweder, Tore and Hjort, Nils Lid , year=. Confidence,

  60. [60]

    2023 , note =

    dirichletprocess: Build Dirichlet process objects for Bayesian modelling , author =. 2023 , note =

    2023

  61. [61]

    and Bowman, A

    Azzalini, A. and Bowman, A. W. , year =. A look at some data on the. doi:10.2307/2347385 , journal =

    work page doi:10.2307/2347385

  62. [62]

    Journal of the American Statistical Association , volume=

    Density estimation with confidence sets exemplified by superclusters and voids in the galaxies , author=. Journal of the American Statistical Association , volume=. 1990 , publisher=

    1990

  63. [63]

    Communications in Statistics - Simulation and Computation , volume=

    A table of normal integrals , author=. Communications in Statistics - Simulation and Computation , volume=. doi:10.1080/03610918008812164 , year=

    work page doi:10.1080/03610918008812164

  64. [64]

    Transactions of the American Mathematical Society , volume=

    Uniformity in linear spaces , author=. Transactions of the American Mathematical Society , volume=. 1938 , doi=

    1938