{"id":"223a9d56-022e-4951-ba51-8ebc08678fc1","arxiv_id":"1908.06349","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Given any exchangeable array of Bernoulli processes directed by a random base measure, the authors construct exchangeable multisets whose conditional law is an i.i.d. negative binomial process with the same base measure.","lead":"This paper proves a black-box recipe: if you can generate exchangeable Bernoulli processes sharing a random base measure, you can generate exchangeable negative binomial processes over the same base measure, without ever building the measure. It gives exact finite algorithms, including a rejection-sampling version for any positive dispersion parameter, so Bayesian nonparametric models with infinite atom sets become simulable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The central theorems are mathematically sound. I re-examined the two places where a hidden assumption could break the argument: the support/union identity for ordinary components and the rejection-sampling extension to r>0. For integer r, \tilde Xn = sum_{j≤r} \tilde Yn,j follows because independent Poisson processes with diffuse intensities have pairwise disjoint supports a.s.; the paper states only the weaker intersection-empty condition, but the needed property is true and standard. For r>0, Algorithm 4.1 is a valid rejection sampler and the one-success input for ordinary atoms yields a Bernoulli(r/⌈r⌉) retention, so thinning produces the correct Poisson intensity. The only concrete error I found is the expected-iterations formula in Lemma 4.1, which is the reciprocal of the correct value; since the lemma's distributional claim and the theorem's proof do not use that mean, this is a typo rather than a correctness issue. The M0 condition excluding atoms of mass 1 is genuinely load-bearing, and it is stated. Overall, the reader's accept verdict with high confidence is appropriate; no change is needed.","tokens_in":10554,"tokens_out":40493,"duration_ms":406700,"concrete_test":"Conduct a Monte Carlo test of Algorithm 4.1 with r=0.5, p=0.5: draw 10^5 outputs and compare the empirical distribution to NB(0.5,0.5); also record mean iterations and check it equals (1−p)^{−0.5}≈1.414 rather than (1−p)^{0.5}, confirming the Lemma 4.1 expectation is a typo.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I reviewed Theorems 3.1 and 4.1 against the definitions and found no load-bearing flaw in the central claim. The construction is a standard conditional-i.i.d. argument: fixed atoms are negative binomial counts from independent Bernoulli sequences, and the ordinary component is a superposition/thinning of independent diffuse Poisson processes. The weakest premises identified by the reader—B ∈ M0 (atoms <1) and diffuse Poisson ordinary components—are exactly what make the construction and the support/union argument valid. Two minor issues do not threaten the theorems: Lemma 4.1's stated expected number of iterations, (1−p)^{⌈r⌉−r}, is the reciprocal of the correct value ((1−p)^{r−⌈r⌉}); and Theorem 4.1's proof says 'only one entry is one' from ∩ supp = ∅, which requires the (true but unstated) pairwise disjointness of independent diffuse Poisson processes. Both are repairable and do not affect the output law.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops black-box constructions for exchangeable sequences of random multisets. Given an exchangeable array of Bernoulli processes directed by a random base measure B in M0, the authors construct a sequence (X_n) that is conditionally i.i.d. negative binomial process with parameter r and base measure B. Theorem 3.1 covers positive integer r by a 'negative binomial urn scheme' in which counts are generated as successes before r failures from the input Bernoulli processes. Theorem 4.1 extends the construction to any r>0 by adding a rejection-sampling subroutine called a negative binomial factory. The algorithms never represent B explicitly, and are finitary when B has finite total mass. Applications to beta processes, hierarchies of beta processes, and generalized beta processes are discussed.","tokens_in":10726,"tokens_out":18659,"duration_ms":184409,"significance":"If the results are correct, the paper gives a genuinely useful reduction: any finitary construction of an exchangeable sequence of Bernoulli processes directed by a random base measure can be converted, in a black-box way, into a finitary construction of exchangeable negative binomial processes directed by the same base measure. This covers several popular Bayesian nonparametric models and avoids the inexact truncations of stick-breaking or inverse-Levy methods. The integer-r construction is clean and the Poisson-superposition/thinning reasoning is elegant. The r>0 extension is more delicate and, in its current form, needs an important clarification in the proof.","major_comments":[{"comment":"The proof of the ordinary component in Theorem 4.1 is incomplete and relies on an implicit convention about how the rejection-sampling iterations consume the input p-coin sequence. As written, Algorithm 4.1 says only 'Simulate W_k ~ NB(⌈r⌉,p) with p-coins'; it does not state that each proposal uses a fresh disjoint block of the infinite sequence. Under the natural reading that W_k is computed from the same infinite sequence, in the diffuse case W_k=1 for every k, so the loop would always terminate with output 1, and the sentence 'otherwise it outputs \\bar X_{n,j}=0 a.s.' would be false; the claimed Poisson thinning to intensity r\\tilde H_0 would fail. Please specify explicitly that each proposal consumes a fresh block of the input sequence, and add the missing argument: after a rejected W_1=1, the next block contains no success, so W_2=0 and is accepted with probability 1. This is the load-bearing step for Theorem 4.1.","section":"Section 4, Algorithm 4.1 and proof of Theorem 4.1, Eqs. (4.7)-(4.8)"}],"minor_comments":[{"comment":"The stated expected number of iterations is the reciprocal of the correct value. For the rejection sampler with proposal NB(⌈r⌉,p) and target NB(r,p), the constant is k=(1-p)^{r-⌈r⌉}, and the mean number of iterations is k. Accordingly, in the proof, R+1 has mean k, not 1/k. The output distribution in Lemma 4.1 is unaffected, but the expectation claim should be corrected.","section":"Section 4, Lemma 4.1"},{"comment":"The constants \\bar b_s are said to lie in (0,1], but the definition of M0 requires atoms to be strictly less than one. Since the constructions require p<1 for the negative binomial counts to be finite, the interval should be (0,1) (or the definition of M0 should be reconciled with the notation).","section":"Section 2.2, Eq. (2.2)"},{"comment":"The condition ∩_m supp(\\tilde Y_{n,m}) = ∅ is weaker than the claimed conclusion that only one entry of the sequence \\tilde Y_{n,1}{s}, \\tilde Y_{n,2}{s}, ... equals one; what is needed is the pairwise disjointness of the supports of independent diffuse Poisson processes. This is a standard fact, but it should be stated and cited rather than left implicit.","section":"Section 4, proof of Theorem 4.1"},{"comment":"The algorithm should define explicitly how the input p-coin sequence is consumed (fresh block per iteration, consumed sequentially). Also, Eq. (4.4) contains a typo: '( X)n∈N' should be '(X_n)_{n∈N}'.","section":"Section 4, Algorithm 4.1 and Eq. (4.4)"},{"comment":"The statement 'Theorem 3.1 holds with this construction' is informal; it should be replaced by an explicit statement of the conclusion, namely that conditioned on B the (X_n) are i.i.d. NBP(r,B).","section":"Section 4, Theorem 4.1 statement"}],"recommendation":"major_revision","confidential_remarks":"The integer-r construction (Theorem 3.1) is sound and is the main contribution. The r>0 extension is likely correct, but the proof of Theorem 4.1 currently leaves a load-bearing ambiguity about how Algorithm 4.1 consumes its input sequence, and the expected-iteration claim in Lemma 4.1 is wrong as stated. These are fixable in revision, and I would be comfortable accepting once they are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read Heaukulani and Roy (1908.06349). The headline: they give a construction that turns any exchangeable array of Bernoulli processes directed by a random base measure B (atoms <1) into an exchangeable sequence of negative binomial processes directed by the same B, for any r>0. The construction never represents B explicitly and is finitary when B has finite total mass. The integer-r case is the natural idea—take the union of supports of r Bernoulli processes and count successes before r failures—and the r>0 case adds a rejection sampler (a 'negative binomial factory') whose acceptance probability is independent of p. Both are proved correctly.\n\nWhat's genuinely new: the earlier Heaukulani-Roy construction was tied to beta process base measures; this one is black-box, so it immediately applies to stable beta, Gibbs-type beta, and hierarchies, whatever classes already have Bernoulli process constructions. That is a real methodological advance for Bayesian nonparametrics, even if the mathematical ingredients are standard.\n\nThe main theorems are sound. Theorem 3.1's conditioning argument is the right one: fixed atoms reduce to i.i.d. Bernoulli trials; the diffuse part is a sum/thinning of independent Poisson processes. Theorem 4.1's thinning computation is also correct: each diffuse atom is retained with probability r/⌈r⌉, giving intensity r B_tilde. I checked the Laplace functional and the support/union argument; no load-bearing flaw.\n\nSoft spots, in ascending order of importance:\n\n- Lemma 4.1 states the expected number of iterations is (1−p)^{⌈r⌉−r}. That is the acceptance probability, not the expected number of proposals; the correct mean is its reciprocal, (1−p)^{r−⌈r⌉}. The proof repeats the inversion when it says R+1 has mean 1/k. This is a performance claim, not a distributional one, so Theorem 4.1's output law is unaffected. Still, it should be corrected.\n\n- Theorem 4.1's proof says for s in the support of the union, 'only one entry is one' from the empty intersection of supports. That conclusion needs the pairwise disjointness of independent diffuse Poisson processes, which is true but unstated. Minor.\n\n- Theorem 3.2 is asserted as a straightforward consequence; there is no proof. For a math.PR paper, that is acceptable but a bit thin.\n\n- The text contains a stray '/suppress' before Latuszyński's name in two places (and in the acknowledgements). Presumably a LaTeX artifact; should be cleaned.\n\nNone of these threaten the central results. The paper has no code or experiments, but it is a construction paper; reproducibility is at the methods level, and the proofs are checkable.\n\nWho is this for? Bayesian nonparametrics people who want exact simulation of negative binomial processes with non-beta base measures, and probabilists interested in exchangeable constructions. It deserves a serious referee.\n\nMy recommendation: send it to review. The main theorems are correct and the contribution is modest but real. The referee should ask for the Lemma 4.1 fix and a one-line clarification on the disjointness.","headline":"A clean black-box reduction from negative binomial to Bernoulli processes; the main theorems hold up, with two small blemishes that should be fixed but do not change the conclusions.","tokens_in":11204,"tokens_out":8470,"would_cite":true,"duration_ms":71629,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","60G09","60G57"],"pacs":[],"model":"deepseek-v4-flash","headline":"Exchangeable Bernoulli processes directed by an unknown random base measure can be converted, black-box, into an exchangeable sequence of negative binomial processes with the same directing measure, for any r>0, without ever constructing…","keywords":["exchangeable sequences","negative binomial process","Bernoulli process","random multisets","urn scheme","completely random measures","Indian buffet process","Bayesian nonparametrics"],"falsifier":"Take B with a single atom of mass p in (0,1) and no diffuse part, simulate the integer-r urn scheme from an exchangeable Bernoulli array directed by B, and compare the empirical distribution of X_1 at that atom to NB(r,p): a mismatch at any p would disprove Theorem 3.1. A sharper check is to add a diffuse component and test whether the ordinary part of X_1 is Poisson with intensity r times the diffuse part; if the Bernoulli inputs are modified so their ordinary components ever coincide, that Poisson character should fail, exposing the reliance on disjoint supports.","tokens_in":10391,"feed_emoji":"🎲","tokens_out":9457,"duration_ms":82285,"temperature":0.7,"pith_summary":"The paper proves a black-box construction: from any exchangeable array of Bernoulli processes (random point processes whose atoms carry mass 0 or 1) directed by a random base measure B whose atoms all have mass less than one, it produces an exchangeable sequence of point processes that, conditioned on B, are independent negative binomial processes NBP(r,B). The construction never represents B explicitly, and if B has finite total mass, each output multiset is almost surely determined by finitely many atoms from a prefix of the input array, so exact simulation is possible even for base measures with countably infinite support. This gives a uniform way to sample random multisets in Bayesian nonparametric models, where previous exact constructions were tailored to specific base measures such as the beta process. The method composes with known constructions of exchangeable Bernoulli processes for beta processes, generalized beta processes, and hierarchies, yielding exact negative binomial process samplers for all of them.","feed_headline":"Bernoulli arrays yield exact negative binomial multisets","feed_subtitle":"A black-box urn scheme converts exchangeable Bernoulli inputs into NBP outputs without building the base measure.","key_machinery":"The negative binomial urn scheme (Definition 3.1) is the central object: given an array (Y_{n,m}) of Bernoulli processes, it forms Y_n as the union of the supports of the first ⌈r⌉ rows and assigns to each atom the count of Bernoulli successes before r failures (integer case), or the output of the negative binomial factory (Algorithm 4.1) for general r>0. The negative binomial factory is a rejection sampler whose proposal is NB(⌈r⌉,p), accepted with probability (r)_W/(⌈r⌉)_W; for the diffuse part of the base measure the argument uses the fact that the ordinary components of the input processes are independent Poisson processes whose supports are almost surely disjoint, so the rejection step reduces to a Poisson thinning that multiplies the intensity by r/⌈r⌉. The proof works through the Laplace functional (2.6), using Lemma A.1 to identify each fixed-atom mass as NB(r,B{s}) and the thinning calculation to identify the diffuse part as a Poisson process of intensity r times the diffuse component of B.","core_discovery":"The central claim is Theorem 3.1 (integer r) and Theorem 4.1 (any r>0): if B is a random element of M0, the space of measures whose atoms are all <1, and (Y_{n,m}) is an exchangeable array of Bernoulli processes directed by B, then the negative binomial urn scheme outputs a sequence (X_n) that is conditionally i.i.d. NBP(r,B). For integer r, each X_n{γ} at an atom γ is the number of successes before r failures in the Bernoulli trials Y_{n,1}{γ}, Y_{n,2}{γ},...; for non-integer r, a rejection sampler called the negative binomial factory proposes NB(⌈r⌉,p) variates from the same Bernoulli stream and accepts with probability (r)_W/(⌈r⌉)_W. The proof shows that the Laplace functional of X_n equals the NBP(r,B) functional of Proposition 2.1: fixed atoms independently receive NB(r,B{s}) masses, and the union of the diffuse ordinary parts, after the rejection step, is thinned to a Poisson process with intensity r times the diffuse part of B. Therefore the scheme extracts the negative binomial process directly from Bernoulli-process inputs, for any base measure in M0 that admits an exchangeable Bernoulli-process representation.","pith_inferences":["The rejection sampler in Algorithm 4.1 is effectively a Bernoulli factory for the negative binomial distribution; if used as a subroutine in a larger exact sampling scheme, its expected iteration count (1-p)^{⌈r⌉-r} suggests runtimes become sensitive when atoms have success probability p close to 1.","The a.s. disjointness of the ordinary components is what makes the diffuse part work; this suggests a robustness condition for implementation: if the input Bernoulli processes are not exactly Poisson in their ordinary parts (e.g., numerical approximations that share atoms), the resulting process may deviate from NBP(r,B) in a way concentrated on the diffuse component.","Because the construction is black-box in B, it may be combined with future Bernoulli-process constructions for new base measures beyond the beta/generalized-beta families; the only requirement is an exchangeable Bernoulli sequence, not a completely random or purely atomic B."],"forward_implications":["For any random base measure with atoms <1 that admits an exchangeable Bernoulli process construction, one obtains an exact finitary sampler for the corresponding negative binomial process, without truncating the support of B.","Applying the scheme to known Bernoulli-process constructions yields NBP samplers for beta processes, stable beta processes, Gibbs-type beta processes, and arbitrarily deep hierarchies of these.","The construction never represents B explicitly, so it removes the need for MCMC subroutines that correct for truncation in stick-breaking or inverse Lévy representations.","The finite-termination guarantee holds whenever B has finite total mass, meaning each output X_n is almost surely determined by a finite prefix of the input array."],"supporting_citations":[{"why":"Supplies the theory of completely random measures and the disintegration theorem used to reduce the law of X_n, given B, to a Laplace functional calculation.","marker":"Kallenberg, 2002"},{"why":"Defines the negative binomial process ordinary component used here and gives the finitary beta-process construction that this paper generalizes.","marker":"Heaukulani and Roy, 2016"},{"why":"Defines the fixed component of the negative binomial process and the beta-negative binomial conjugacy.","marker":"Broderick et al., 2014"},{"why":"Gives the exchangeable Bernoulli process sequence directed by a beta process (and hierarchies) that serves as an input example.","marker":"Thibaux and Jordan, 2007"},{"why":"Provides the one-parameter process construction producing exchangeable Bernoulli sequences directed by generalized beta processes and hierarchies, which the black-box scheme consumes.","marker":"Roy, 2014"},{"why":"Supplies the rejection sampling framework underlying the negative binomial factory in Algorithm 4.1.","marker":"Robert and Casella, 1999"},{"why":"Provides the foundational representation of completely random measures as sums of diffuse, fixed, and Poisson ordinary components.","marker":"Kingman, 1967"}],"fun_headline_variants":["Black-box urn turns Bernoulli into negative binomial","Any base measure, exact NBP via Bernoulli streams","Urn scheme: Bernoulli inputs to exact NBP outputs","Black-box urn: Bernoulli in, negative binomial out","Exact NBP from any Bernoulli-exchangeable base"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that B lies in M0, so every atom of B has mass strictly less than 1, and that the input Bernoulli processes have the standard complete-random decomposition with a diffuse Poisson ordinary component; if an atom had mass 1, the 'count before r failures' would never terminate, and if the ordinary components could share atoms, the Poisson thinning step would break.","fun_headline_variants_meta":{"raw":{"variants":["Black-box urn turns Bernoulli into negative binomial","Any base measure, exact NBP via Bernoulli streams","Urn scheme: Bernoulli inputs to exact NBP outputs","Black-box urn: Bernoulli in, negative binomial out","Exact NBP from any Bernoulli-exchangeable base"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3097,"prompt_tokens":974,"completion_tokens":2123,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2048}},"tokens_in":590,"tokens_out":2123,"duration_ms":14314,"temperature":1.0,"reasoning_tokens":2048,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:50:24.952582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take B with a single atom of mass p in (0,1) and no diffuse part, simulate the integer-r urn scheme from an exchangeable Bernoulli array directed by B, and compare the empirical distribution of X_1 at that atom to NB(r,p): a mismatch at any p would disprove Theorem 3.1. A sharper check is to add a diffuse component and test whether the ordinary part of X_1 is Poisson with intensity r times the diffuse part; if the Bernoulli inputs are modified so their ordinary components ever coincide, that Poisson character should fail, exposing the reliance on disjoint supports.","supporting_citations":[],"review_version":1}