Weak convergence of predictive distributions

Fabrizio Leisen; Luca Pratelli; Pietro Rigo

Weak convergence of predictive distributions

T0 review · 0 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Predictive distributions converge in probability exactly when three separate conditions are met.

desk verdict A clean, correct characterization of when predictive distributions converge in probability, with sharp counterexamples; worth a serious referee. read the letter →

arxiv 2507.19169 v1 pith:HRZSLEVE submitted 2025-07-25 math.PR math.STstat.MEstat.TH

Fabrizio Leisen , Luca Pratelli , Pietro Rigo This is my paper · ORCID

classification math.PRmath.STstat.MEstat.TH MSC 60B1060G5760G0960F99

keywords predictivedistributionsstableconvergencerandomprobabilitymeasureconditionallyidenticallydistributedasymptoticexchangeabilityweaksecond-momentcondition

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

This paper asks when the sequence of predictive distributions $\alpha_n(f)=E[f(X_{n+1})\mid X_1,\dots,X_n]$ settles down: when, for every bounded Borel function $f$, these conditional expectations converge in probability. The main result, Theorem 3, is a complete characterization: for a standard Borel state space, this happens exactly when the marginal laws converge setwise, the variables $X_n$ converge stably to a random probability measure $\alpha$, and the second moments of the predictive means converge to $E[\alpha(f)^2]$. The characterization matters because predictive convergence makes empirical measures consistent estimates of predictive distributions, and it implies a weak law of large numbers and asymptotic exchangeability. The paper also shows which weak versions of conditional identity in distribution still deliver such convergence, and which do not.

What carries the argument

The central object is the predictive distribution $\alpha_n(B)=P(X_{n+1}\in B\mid F_n)$, a random probability measure. The load-bearing identity is Theorem 3's equivalence: condition (1) holds if and only if condition (4) holds, $X_n$ converges stably to a random probability measure $\alpha$, and condition (5) holds. Stable convergence is the requirement that $P(X_n\in\cdot\mid H)$ converge weakly to $E[\alpha(\cdot)\mid H]$ for every event $H$ of positive probability; it is stronger than convergence in distribution but weaker than almost-sure convergence of the laws. Condition (5) is the second-moment match that closes the argument: it lets the proof show $E[(\alpha_n(f)-\alpha(f))^2]\to 0$, turning conditional convergence into genuine convergence in probability.

What would settle it

Take a measurable space that is not standard Borel and construct a sequence satisfying condition (4), stable convergence to some random probability measure $\alpha$, and $E[\alpha(f)^2]=\lim_n E[\alpha_n(f)^2]$ for continuous bounded $f$, but with $\alpha_n(f)$ failing to converge in probability for some bounded Borel $f$. Such an example would show the standard-Borel hypothesis is necessary; the paper's Theorem 3 predicts none exists when $S$ is standard Borel.

Watch

Extended reading notes

Core claim

The central discovery is that convergence in probability of all predictive distributions is not a loose or accidental property; it is equivalent to three conditions that can be checked separately. Setwise convergence of the marginal laws $P(X_n\in\cdot)$ must hold; the sequence $X_n$ must converge stably to a random probability measure $\alpha$; and the second-moment matching condition $E[\alpha(f)^2]=\lim_n E[\alpha_n(f)^2]$ must hold for every continuous bounded $f$. Under those conditions, and only then, $\alpha_n(f)$ converges in probability to $\alpha(f)$ for every bounded Borel $f$. The proof passes from continuous functions to all bounded Borel functions by a monotone class argument, and it uses the standard-Borel assumption to realize the limiting object as a single random probability measure.

Load-bearing premise

The characterization assumes the state space $S$ is standard Borel, i.e., a Borel subset of a Polish space; this is what lets the paper pass from convergence on continuous functions to a single random probability measure $\alpha$, and if that assumption fails the equivalence in Theorem 3 can break down.

Editorial extensions

If this is right

Under condition (1), there is a single random probability measure $\alpha$ such that $\alpha_n(f)\xrightarrow{P}\alpha(f)$ for all bounded Borel $f$.
Under condition (1), the empirical measure $\mu_n$ also satisfies $\mu_n(f)\xrightarrow{P}\alpha(f)$, so empirical distributions are consistent estimates of predictive distributions.
Under condition (1), $X$ is asymptotically exchangeable: the shifted block $(X_{n+1},X_{n+2},\dots)$ converges in distribution to an exchangeable sequence.
Conditionally identically distributed sequences imply almost-sure convergence of predictive distributions, hence condition (1); the paper shows that quasi-martingale predictive distributions and non-adapted c.i.d. sequences still give sufficient conditions, while higher-order c.i.d. does not.
The second-moment condition (5) is automatic when condition (1) holds, so in that direction it suffices to check stable convergence and setwise marginal convergence.

Reading between the lines

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

Because condition (5) is a variance-type matching condition, a practical diagnostic suggests itself: in simulation output, monitor the squared predictive expectations $E[\alpha_n(f)^2]$ against $E[\alpha(f)]^2$; a persistent gap would signal failure of condition (1). This diagnostic is not proposed in the paper.
The theorem suggests that almost-sure convergence of predictive distributions may be unnecessarily strong for many Bayesian consistency arguments; convergence in probability plus the second-moment match could serve as the right standing hypothesis for posterior predictive asymptotics.
The counterexamples in Section 4 can be read as a construction kit: because the three conditions are independent, one can build sequences satisfying any two of them while failing condition (1), which may help in designing statistical models with a controlled degree of predictive instability.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

This is a solid paper from people who clearly know this territory. The core result, Theorem 3, says condition (1) (predictive distributions converging in probability for every bounded Borel function) is equivalent to setwise marginal convergence, stable convergence of X_n to a random measure α, and a second-moment matching condition E[α(f)^2] = lim E[α_n(f)^2]. That's a genuinely useful equivalence, and it's not in the earlier literature. Theorem 5 gives a version that avoids explicitly identifying α, which is practical for applications. The counterexamples are also well chosen: Example 11 shows (1) does not imply a.s. convergence even for pairwise independent sequences, and Example 15 shows that even stable convergence plus asymptotic exchangeability plus the obvious necessary condition (15) do not force (1). Those examples delimit the theory honestly.

The proofs are generally rigorous. I have small quibbles rather than real objections. The proof of Theorem 5 contains a terse limit interchange: the equality between the Cesàro conditional expectations and E[μ_n(f)|H] requires conditioning the a.s. convergence in Remark (iii) on H, and the step from stable convergence to E[α(f)|H] = lim E[α_n(f)|H] is compressed. But these are standard moves and they work. The paper leans on the authors' own earlier papers, especially [3] and [4], for the existence of the limiting random measure under standard Borel assumptions. That is not circular — those theorems are published and independently checkable — but it does mean a reader needs to trust that prior work. The standard Borel assumption is explicit and is used essentially; within that scope the results are fine.

Who benefits? Anyone working on Bayesian predictive inference, species sampling, or empirical processes who wants to know when the predictive distribution stabilizes in probability rather than almost surely. The paper doesn't open a new subfield, but it sharpens the boundary conditions for a commonly used property and provides useful vocabulary (the weak c.i.d. notions, the second-moment condition).

I'd send this to a serious referee. It's not a breakthrough, but it is correct, careful, and a real advance over the cited literature. If I were working on predictive limits, I'd cite it.

Referee Report

0 major / 6 minor

Summary. The paper studies the condition that the predictive distributions α_n(·)=P(X_{n+1}∈·|F_n) converge in probability for every bounded Borel function f, for a sequence X_n with values in a standard Borel space. The main result, Theorem 3, characterizes this condition as the conjunction of setwise convergence of the marginal laws (condition (4)), stable convergence of X_n to a random probability measure α, and the second-moment matching condition E[α(f)^2]=lim_n E[α_n(f)^2] (condition (5)). The paper also shows that condition (1) implies a weak law of large numbers and asymptotic exchangeability (Theorem 2), gives a version of the characterization that avoids explicit mention of α (Theorem 5), and proves a special case under a degenerate empirical-limit assumption (Theorem 4). Three weakenings of conditional identity in distribution are examined: higher-order c.i.d., quasi-martingale predictive distributions, and non-adapted c.i.d. with respect to a coarser filtration. The final section contains five counterexamples delineating the boundaries of the implications, including examples showing that condition (1) does not imply almost-sure convergence, that asymptotic exchangeability does not imply condition (3), and that stable convergence plus asymptotic exchangeability and condition (15) do not suffice for condition (1).

Significance. If correct, Theorem 3 gives a clean and useful characterization that reduces a seemingly non-linear convergence-in-probability condition on conditional laws to stable convergence plus a second-moment condition. This should be of interest to researchers in Bayesian asymptotics, species sampling, and predictive inference. The paper is careful in marking which implications are strict and supplies constructive counterexamples (Examples 11–15) that check out. The standard Borel assumption is explicit and used only through the cited compactness result [4, Cor. 2.4]; no hidden parameters or ad hoc assumptions enter. The main proofs are rigorous, though a few limit-interchange and extension steps are compressed.

minor comments (6)

[§2, Theorem 3 proof] The step labeled 'arguing as above' in the converse direction, which extends the equality E[α(f)|H] = lim_n E[α_n(f)|H] from H∈∪_k F_k to all H∈A_+, is the most delicate point in the paper; please expand it into an explicit argument or state the lemma being used, because this equality is what connects stable convergence of X_n to convergence of the predictive distributions.
[§2, Theorem 5 proof] The double-limit interchange lim_k lim_n E[U_k α_n(f)] = lim_n lim_k E[U_k α_n(f)] should be justified: it is valid because U_k→α(f) uniformly and α_n(f) is uniformly bounded, so the inner convergence is uniform in n, but this justification is currently omitted.
[§4, Example 15] The statement that 'X_n converges in total variation' is stronger than what the stated assumptions obviously imply; the subsequent approximation argument only requires weak convergence together with a compactness/Lusin approximation, so please weaken the claim accordingly or add a proof.
[§2, before Theorem 1] The notation for the tail σ-field and the union of the initial filtrations is corrupted in the typesetting: 'T =T nσ(X_n,X_{n+1},...)' should read 'T = ∩_n σ(X_n,X_{n+1},...)', and 'if H∈S kFk' should read 'if H∈∪_k F_k'.
[§2, Theorem 4 statement] In condition (9), the right-hand side is ambiguous: it should be read as (lim_n E[f(X_n)])^2, as used in the proof immediately below, not as lim_n (E[f(X_n)])^2.
[§2, Theorem 5 statement] Condition (10) uses conditional expectation given an event H, as in E[f(X_n)|H], while similar notation is used elsewhere for conditional expectations given σ-fields; the paper is consistent, but the convention would be clearer if stated explicitly when the condition is introduced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the characterization is proven from stated conditions.

full rationale

The paper's central claim, Theorem 3, is an equivalence between condition (1) and the conjunction of condition (4), stable convergence of X_n to a random probability measure α, and the second-moment condition (5). The forward direction derives α from convergence in probability on continuous bounded functions via the cited Corollary 2.4 of [4], then extends to all bounded Borel functions through a monotone-class argument and the Vitali–Hahn–Saks theorem. The converse direction assumes stable convergence and condition (5), identifies conditional expectations using equation (6) plus a standard passage from finite unions of F_k to A_+, and then uses condition (5) to force L2 convergence of α_n(f) to α(f). At no point is condition (1) assumed in the converse, nor is condition (5) derived from (1) except in the forward direction where it is a consequence of bounded convergence. The self-citations [3], [4], and [5] provide external, published support: [4, Cor. 2.4] is a general existence theorem for limiting random probability measures from convergence in probability on continuous functions; it does not assume the paper's target result. [3] supplies the known sufficiency of c.i.d. for a.s. convergence, which is used only as a sufficient condition, and [5] is used only in an example. There are no fitted parameters relabeled as predictions and no definitional equivalence between the hypotheses and the conclusion. The standard Borel assumption is explicit and is used in a standard way. Overall, the derivation chain is self-contained relative to its cited external theorems and no circular step was found.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no fitted parameters and no new postulated entities. All inputs are standard measure-theoretic tools plus two published theorems from the same research group; those theorems are externally published, not derived inside this paper, so their use raises a minor self-citation burden but not circularity.

assumptions (5)

standard math Monotone class theorem for bounded functions
Used in Theorem 1 to extend the limit identification from C_b(S) to M_b(S) after proving closure under bounded monotone pointwise limits.
standard math Vitali-Hahn-Saks theorem
Used in Theorems 1 and 4 to conclude that the setwise limit λ(B)=lim P(X_n∈B) is a probability measure.
standard math Quasi-martingale convergence theorem
Used in Section 3.2 to conclude α_n(f) converges almost surely under the quasi-martingale condition (11).
standard math Corollary 2.4 of Berti, Pratelli and Rigo (2006)
Load-bearing external theorem: under convergence in probability of α_n(f) for every f in C_b(S), there exists a random probability measure α with α_n(f) converging in probability to α(f) for all f in C_b(S). Used in Theorems 1, 2, 3, and 5.
domain assumption Standard Borel assumption on the state space S
Assumed throughout; guarantees regular conditional distributions and is used in Theorem 1, the characterization theorems, and the examples.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Weak convergence of predictive distributions." pith.science (2026). https://pith.science/paper/HRZSLEVE

@misc{pith2026250719169,
  author       = {Pith},
  title        = {Pith review of: Weak convergence of predictive distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRZSLEVE}},
  note         = {Machine review of arXiv:2507.19169}
}

abstract

Let $(X_n)$ be a sequence of random variables with values in a standard Borel space $S$. We investigate the condition \begin{gather}\label{x56w1q} E\bigl\{f(X_{n+1})\mid X_1,\ldots,X_n\bigr\}\,\quad\text{converges in probability,}\tag{*} \\\text{as }n\rightarrow\infty,\text{ for each bounded Borel function }f:S\rightarrow\mathbb{R}.\notag \end{gather} Some consequences of \eqref{x56w1q} are highlighted and various sufficient conditions for it are obtained. In particular, \eqref{x56w1q} is characterized in terms of stable convergence. Since \eqref{x56w1q} holds whenever $(X_n)$ is conditionally identically distributed, three weak versions of the latter condition are investigated as well. For each of such versions, our main goal is proving (or disproving) that \eqref{x56w1q} holds. Several counterexamples are given.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Posterior uncertainty for kernel density estimates
stat.ME 2026-07 accept novelty 6.5 of 10

Kernel density estimate predictives converge weakly almost surely (hence are asymptotically exchangeable) even though they are neither c.i.d. nor a.c.i.d., and for Gaussian kernels the limit is a.s. absolutely continu...

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages · cited by 1 Pith paper

[3]

(2004) Limit theorems for a class of identically distributed random variables, Ann

Berti P., Pratelli L., Rigo P. (2004) Limit theorems for a class of identically distributed random variables, Ann. Probab., 32, 2029-2052

work page 2004
[4]

(2006) Almost sure weak convergence of random probability measures, Stochastics, 78, 91-97

Berti P., Pratelli L., Rigo P. (2006) Almost sure weak convergence of random probability measures, Stochastics, 78, 91-97

work page 2006
[1]

(1985) Exchangeability and related topics, Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math

Aldous D.J. (1985) Exchangeability and related topics, Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math. 1117, Springer, Berlin

work page 1985
[2]

(2010) Conditionally identically distributed species sam- pling sequences, Adv

Bassetti F., Crimaldi I., Leisen F. (2010) Conditionally identically distributed species sam- pling sequences, Adv. Appl. Probab., 42, 433-459

work page 2010
[5]

(2011) A central limit theorem and its applications to multicolor randomly reinforced urns, J

Berti P., Crimaldi I., Pratelli L., Rigo P. (2011) A central limit theorem and its applications to multicolor randomly reinforced urns, J. Appl. Probab., 48, 527-546

work page 2011
[6]

(2012) Limit theorems for empirical processes based on dependent data, Electr

Berti P., Pratelli L., Rigo P. (2012) Limit theorems for empirical processes based on dependent data, Electr. J. Probab., 17, 1-18

work page 2012
[7]

(2013) Exchangeable sequences driven by an absolutely contin- uous random measure, Ann

Berti P., Pratelli L., Rigo P. (2013) Exchangeable sequences driven by an absolutely contin- uous random measure, Ann. Probab, 41, 2090-2102

work page 2013
[8]

(2021) A class of models for Bayesian predictive inference, Bernoulli, 27, 702-726

Berti P., Dreassi E., Pratelli L., Rigo P. (2021) A class of models for Bayesian predictive inference, Bernoulli, 27, 702-726

work page 2021

Show all 22 references

[9]

(2025) Bayesian analysis with conditionally identically distributed sequences, Electr

Bissiri P.G., Walker S.G. (2025) Bayesian analysis with conditionally identically distributed sequences, Electr. J. Statistics , 19, 1609-1632

2025
[10]

(1962) Merging of opinions with increasing information, Ann

Blackwell D., Dubins L.E. (1962) Merging of opinions with increasing information, Ann. Math. Statist., 33, 882-886

1962
[11]

(2019) A Bayesian nonparametric spiked process prior for dynamic model selection, Bayesian Anal., 14, 553-572

Cassese A., Zhu W., Guindani M., Vannucci M. (2019) A Bayesian nonparametric spiked process prior for dynamic model selection, Bayesian Anal., 14, 553-572

2019
[12]

(1999) Prequential probability: principles and properties, Bernoulli, 5, 125-162

Dawid A.P., Vovk V.G. (1999) Prequential probability: principles and properties, Bernoulli, 5, 125-162

1999
[13]

(2004) Convergence of sequences of pairwise independent random variables, Proc

Etemadi N., Lenzhen A. (2004) Convergence of sequences of pairwise independent random variables, Proc. Amer. Math. Soc. , 132, 1201-1202

2004
[14]

(2023) Martingale posterior distributions (with discussion), J

Fong E., Holmes C., Walker S.G. (2023) Martingale posterior distributions (with discussion), J. Royal Stat. Soc. B, 85, 1357-1391

2023
[15]

(2024) Asymptotics for parametric martingale posteriors, arXiv:2410.17692v1

Fong E., Yiu A. (2024) Asymptotics for parametric martingale posteriors, arXiv:2410.17692v1

2024
[16]

(2018) On a notion of partially conditionally identically distributed sequences, Stoch

Fortini S., Petrone S., Sporysheva P. (2018) On a notion of partially conditionally identically distributed sequences, Stoch. Proc. Appl., 128, 819-846

2018
[17]

(2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

Hahn P.R., Martin R., Walker S.G. (2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

2018
[18]

(1988) Spreading and predictable sampling in exchangeable sequences and processes, Ann

Kallenberg O. (1988) Spreading and predictable sampling in exchangeable sequences and processes, Ann. Probab., 16, 508-534

1988
[19]

(2009) Asymptotics in response-adaptive designs generated by a two- color, randomly reinforced urn, Ann

May C., Flournoy N. (2009) Asymptotics in response-adaptive designs generated by a two- color, randomly reinforced urn, Ann. Statist., 37, 1058-1078

2009
[20]

(2025) Martingale posterior distributions for time series models,Statist

Moya B., Walker S.G. (2025) Martingale posterior distributions for time series models,Statist. Sci., 40, 68-80

2025
[21]

(2007) A survey of random processes with reinforcement, Prob

Pemantle R. (2007) A survey of random processes with reinforcement, Prob. Surveys, 4, 1-79

2007
[22]

P. Fortu- nati

Pitman J. (1996) Some developments of the Blackwell-MacQueen urn scheme, Statistics, Probability and Game Theory, IMS Lect. Notes Mon. Series , 30, 245-267. Fabrizio Leisen, Department of Mathematics, king’s College, Strand WC2R 2LS, Lon- don, Uk Email address: fabrizio.leisen...

1996

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [3]

(2004) Limit theorems for a class of identically distributed random variables, Ann

Berti P., Pratelli L., Rigo P. (2004) Limit theorems for a class of identically distributed random variables, Ann. Probab., 32, 2029-2052

work page 2004

[2] [4]

(2006) Almost sure weak convergence of random probability measures, Stochastics, 78, 91-97

Berti P., Pratelli L., Rigo P. (2006) Almost sure weak convergence of random probability measures, Stochastics, 78, 91-97

work page 2006

[3] [1]

(1985) Exchangeability and related topics, Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math

Aldous D.J. (1985) Exchangeability and related topics, Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math. 1117, Springer, Berlin

work page 1985

[4] [2]

(2010) Conditionally identically distributed species sam- pling sequences, Adv

Bassetti F., Crimaldi I., Leisen F. (2010) Conditionally identically distributed species sam- pling sequences, Adv. Appl. Probab., 42, 433-459

work page 2010

[5] [5]

(2011) A central limit theorem and its applications to multicolor randomly reinforced urns, J

Berti P., Crimaldi I., Pratelli L., Rigo P. (2011) A central limit theorem and its applications to multicolor randomly reinforced urns, J. Appl. Probab., 48, 527-546

work page 2011

[6] [6]

(2012) Limit theorems for empirical processes based on dependent data, Electr

Berti P., Pratelli L., Rigo P. (2012) Limit theorems for empirical processes based on dependent data, Electr. J. Probab., 17, 1-18

work page 2012

[7] [7]

(2013) Exchangeable sequences driven by an absolutely contin- uous random measure, Ann

Berti P., Pratelli L., Rigo P. (2013) Exchangeable sequences driven by an absolutely contin- uous random measure, Ann. Probab, 41, 2090-2102

work page 2013

[8] [8]

(2021) A class of models for Bayesian predictive inference, Bernoulli, 27, 702-726

Berti P., Dreassi E., Pratelli L., Rigo P. (2021) A class of models for Bayesian predictive inference, Bernoulli, 27, 702-726

work page 2021