{"id":"31749d43-dc1e-4d2a-8ad8-24d7871d1564","arxiv_id":"1908.07254","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A pseudo-marginal version of the PaRIS smoother computes online additive smoothing estimates with linear cost and constant memory, with exponential concentration, a central limit theorem, nearly linear variance growth, and an O(n epsilon) bias bound.","lead":"This paper builds an online algorithm that estimates smoothed averages of hidden states even when the model's transition rules cannot be computed exactly, only estimated. It proves the estimates converge, are asymptotically normal, stay stable over long time series, and have a bounded bias that grows at most linearly with time and estimation error.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(n ε) bias and O(n) variance results require (H3) Doeblin conditions that the paper's diffusion/OU examples violate; the Section 5 experiment uses ideal PaRIS, not Algorithm 6, leaving the advertised broad applicability under-supported.","rationale":"The reader's CONDITIONAL verdict is appropriate. The proofs appear careful and the CLT/concentration theorems are honestly qualified by (H1)-(H2); the pseudo-marginal construction is coherent. However, the strongest advertised applications, especially partially observed diffusions, and the only numerical section both lie outside the strong-mixing assumption on which the stability and bias theorems rest. The paper itself states that an exhaustive numerical analysis is beyond its scope and refers to [19] for illustrations of Algorithm 6; no code or data are shipped. Thus, the most load-bearing uncertainty is not whether the theorems are internally correct but whether the breadth claim is actually supported for the target models. A compact-state replication of Section 5 using the true Algorithm 6 would directly test this. This concern does not overturn the conditional verdict; it reinforces it, so no change is needed.","tokens_in":49426,"tokens_out":12355,"duration_ms":134403,"concrete_test":"Re-run the Section 5 experiment in a setting that satisfies (H3): truncate the OU state space to a compact interval [−A, A] with reflected dynamics, or use a finite-state Markov chain with transition probabilities bounded below by σ_- > 0, and implement Algorithm 6 (not Algorithm 3) with a biased density estimator. Use N = 200, M = 2, n up to 50, and ε ∈ {0, 0.05, ..., 0.5}, comparing to Kalman/analytic values. If the O(n ε) bias and roughly linear variance growth appear, the advertised applicability is supported; if they degrade, (H3) is confirmed as the decisive limitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 4.6 and 4.10, which justify the O(n) variance and O(n ε) bias claims, are explicitly conditional on Assumption (H3). This assumption requires uniform bounds σ_- ≤ ℓ_m ≤ σ_+ and σ_- ≤ ℓ^ε_m ≤ σ_+ for every m (Section 4.2), a Doeblin condition that, as the authors note, 'typically requires the state spaces (X_n) to be compact sets'. The paper's flagship applications, partially observed diffusions on R^{d_x} (Example 2), use Gaussian transition densities with unbounded support; the OU example in Section 5 has transition density q_δ that is Gaussian on R, so inf ℓ_m = 0 and (H3) fails. Yet Section 5 is presented as illustrating the O(n ε) bound of Theorem 4.10. Moreover, the experiment runs the ideal PaRIS (Algorithm 3), not the pseudo-marginal Algorithm 6, and uses Kalman smoothing as ground truth; Algorithm 6 itself is not numerically demonstrated in this paper, and no code or data are provided. Thus the central claim's advertised range of challenging path-space problems is supported neither by the stability/bias theorems in the diffusion regime nor by a simulation of the proposed algorithm under the stated assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pseudo-marginal extension of the PaRIS algorithm for online smoothing of additive functionals under path measures with intractable transition densities. Algorithm 6 replaces both forward and backward sampling operations by pseudo-marginal versions, admitting unbiased or biased density estimators. The main theoretical results are exponential concentration (Theorem 4.1), a central limit theorem with an explicit variance decomposition (Theorem 4.2), an O(n) bound on asymptotic variance under a strong-mixing condition (Theorem 4.6), and an O(n epsilon) bound on asymptotic bias under an additional Lipschitz-type condition on the estimated density (Theorem 4.10). The paper also derives by-product CLTs for the random-weight particle filter and for the ideal PaRIS in a general Feynman-Kac/auxiliary particle filter setting, and it presents a short numerical example based on an Ornstein-Uhlenbeck model.","tokens_in":49664,"tokens_out":8517,"duration_ms":95140,"significance":"If the advertised claims are taken at face value, the paper makes a substantial contribution: it extends the PaRIS algorithm to models whose transition densities are only estimable, preserves linear-in-N complexity and constant memory in n, and provides the first O(n epsilon) bias bound of this type. The proofs are detailed and organized, the assumptions H1-H4 are explicit, and there are no hidden fitted parameters in the derivation. The variance decomposition in Theorem 4.2 is interpretable, with distinct contributions from initialization, the ideal PaRIS dynamics, and the additional pseudo-marginal randomness. The by-product time-uniform variance bound for the random-weight particle filter (Proposition 4.9) is also valuable. The principal weakness is the mismatch between the advertised range of applications and the strong-mixing assumption needed for the long-horizon and bias results, together with a numerical section that does not actually run the proposed Algorithm 6.","major_comments":[{"comment":"The two long-horizon results that support the advertised practical scope both rest on Assumption H3, which requires uniform lower and upper bounds on the transition densities and their biased counterparts. As the authors themselves note, this is a Doeblin condition that typically requires compact state spaces. The diffusion model of Example 2 has Gaussian transition densities on R^{d_x}, and the Ornstein-Uhlenbeck example of Section 5 has a Gaussian transition density on R, so H3 fails in precisely those flagship applications. Consequently, for those models the paper provides no control on the long-run variance (Theorem 4.6) and no O(n epsilon) bound between the skew expectation and the true additive functional (Theorem 4.10); only the fixed-horizon CLT about the skew model (Theorem 4.2) applies. The abstract and introduction should qualify the claim that the algorithm applies to diffusion processes, or additional results for non-compact state spaces should be supplied.","section":"Section 4.2 (Assumption H3); Theorems 4.6 and 4.10"},{"comment":"The numerical section does not demonstrate the proposed algorithm. The PaRIS results in Figure 1 are generated by Algorithm 3 (ideal PaRIS) with the optimal importance function, not by the pseudo-marginal Algorithm 6, and the epsilon-skew is encoded in the observation density rather than in a biased estimator of an intractable transition density. Thus the experiment cannot validate the pseudo-marginal machinery of Section 3, nor does it quantify the additional variance term in Theorem 4.2. In addition, the OU model violates H3, and the constant c used to verify H4 in Section 5 depends on sup_n |y_n|, which is infinite almost surely for Gaussian observations; hence the example does not satisfy the assumptions of Theorem 4.10. A simulation of Algorithm 6 on a model satisfying H3-H4, with a comparison of the bias of the pseudo-marginal estimator, is needed to support the advertised claims.","section":"Section 5, Figure 1"},{"comment":"For epsilon > 0, Theorem 4.2 establishes asymptotic normality around the skew expectation phi^epsilon_{0:n} h_n, not around the target phi_{0:n} h_n. The bridge to the true target is provided only by Theorem 4.10, which requires H3-H4. Since H3 is not satisfied by the diffusion examples, the reader is left without end-to-end theoretical support for the headline applications. The manuscript should state this limitation explicitly in the abstract and introduction whenever the 'wide range' of diffusion and intractable-likelihood models is advertised; as written, the reader can reasonably infer that the convergence, stability, and bias guarantees apply to those models.","section":"Abstract and Introduction vs. Theorem 4.2"}],"minor_comments":[{"comment":"The initialization text says 'letting ... tau_i^n = 0', which appears to be a typo for tau_i^0 = 0; the same notation should be made consistent in Algorithms 3 and 6.","section":"Section 3.2.3, initialization of Algorithm 6"},{"comment":"The statement 'So is also w^{-1} : X0 -> dchi/dnu(x0)' is not well-formed notation; the boundedness condition should refer to the initial weight function w_0(x0) = dchi/dnu(x0) or to its reciprocal explicitly.","section":"Assumption H2"},{"comment":"Reference [24] is listed as Stochastic Analysis and Applications 32:397-420, 1998, but this volume and page range correspond to a much later year; the citation should be corrected to avoid bibliographic confusion.","section":"Reference [24]"},{"comment":"No code or data are provided for the numerical example; making a small reproducible script available would strengthen the computational claims and help readers assess the behavior of Algorithm 6.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core appears sound and the O(n epsilon) bias bound is a genuine contribution. The main weakness is the disconnect between the advertised diffusion scope and the strong-mixing assumptions under which the long-horizon and bias theorems hold, together with a numerical section that does not run the proposed algorithm. If the authors reframe the contributions accurately and add a small simulation of Algorithm 6 under H3-H4, I would support publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a substantial theory paper that does what it says in the theorem statements, but the advertised scope is broader than what the results actually cover. The genuinely new piece is Algorithm 6, a pseudo-marginal PaRIS with two separate randomization layers, and the accompanying CLT plus the O(n eps) bias bound. The proofs in the appendices are careful and the assumptions are explicit; this is real work, not a repackaging.\n\nWhat's good: the Feynman-Kac and auxiliary-filter formulation is a genuine generalization of the original PaRIS, and the CLT decomposition into ideal-PaRIS variance plus an extra pseudo-marginal variance term is a clean by-product. The O(n eps) bias theorem for biased density estimators is new, and the propagation argument in Appendix D is nontrivial. The paper also gives a time-uniform variance bound for the random-weight particle filter as a by-product, which is a nice bonus.\n\nWhere it gets soft: the main stability and bias theorems (4.6 and 4.10) sit on H3, a Doeblin condition that essentially requires compact state spaces. The diffusion examples that motivate the paper have Gaussian transitions on R^d, so H3 fails for them. The authors are honest about this in the text, but the numerical section still presents a Gaussian OU model as an illustration of the O(n eps) bound, a model where H3 also fails, and it runs ideal PaRIS (Algorithm 3), not the pseudo-marginal Algorithm 6 that the paper is about. So the empirical support for the advertised wide range of challenging path-space problems is thin. There is no code or data, and Algorithm 6 itself is never simulated.\n\nThat said, the theorems are stated with their assumptions; if your work falls inside H3, the results are there. The gap is between the advertised breadth and the verified scope, not a hole in the proofs. The paper deserves a serious referee, but the authors should be pushed to either soften the claims about diffusions or add a numerical demonstration of Algorithm 6 under the stated assumptions.","headline":"Substantial theory for a two-level pseudo-marginal PaRIS, but the advertised diffusion scope outruns the strong-mixing assumptions the main theorems actually require.","tokens_in":50234,"tokens_out":2194,"would_cite":true,"duration_ms":24120,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that online additive smoothing remains valid when transition densities are intractable but estimable, using a pseudo-marginal version of the PaRIS algorithm with linear complexity and bias/variance guarantees.","keywords":["central limit theorem","exponential concentration","partially observed diffusions","particle smoothing","pseudo-marginal methods","sequential Monte Carlo","online smoothing","additive functionals"],"falsifier":"Run Algorithm 6 with $M=2$ on a finite-state hidden Markov model that satisfies the uniform density bounds, using deliberately biased transition-density estimates of known size $\\varepsilon$. Theorems 4.6 and 4.10 predict that as $N$ grows the estimator's variance increases at most linearly in the path length $n$ and its bias stays below a constant times $n\\varepsilon$; observing superlinear variance growth, or bias that cannot be bounded by $C n\\varepsilon$ for any fixed $C$ at large $N$, would falsify the central claims.","tokens_in":49247,"feed_emoji":"🎲","tokens_out":10977,"duration_ms":104522,"temperature":0.7,"pith_summary":"This paper considers online estimates of expectations of additive path functionals—sums of terms along the hidden path—under general path distributions whose transition densities cannot be evaluated, but can be simulated or estimated, with or without bias. It proves that a pseudo-marginal version of the PaRIS particle smoother, which replaces every intractable density by a random estimate in both the forward and backward sampling steps, computes these estimates in $O(N)$ time per time step with constant memory. The paper establishes exponential concentration and a central limit theorem for this estimator, and shows that under strong mixing its asymptotic variance grows at most linearly in the path length $n$. If the density estimates are biased with size controlled by a parameter $\\varepsilon$, the asymptotic bias is at most $O(n\\varepsilon)$ for additive functionals, and $O(\\varepsilon)$ for filter marginals. A sympathetic reader would care because the result covers partially observed diffusions and approximate Bayesian computation, where exact transition densities are unavailable.","feed_headline":"Intractable transition densities no longer block online smoothing","feed_subtitle":"It swaps unknown transition densities for estimates with linear-time, constant-memory updates and bias bounds.","key_machinery":"The load-bearing object is two-level pseudo-marginalisation of the PaRIS update. PaRIS avoids particle-path degeneracy by alternating forward sampling, which propagates particles through the model, with backward sampling, which propagates statistics $\\tau_n^i$ approximating the backward-recursed additive functional $T_n h_n$. In the pseudo-marginal version, both levels use random estimates of the intractable density: the forward weights use $\\ell_n^{\\langle z\\rangle}$ directly, and the backward index kernel $\\bar\\Pi_n(i,\\cdot)$ samples ancestor indices with probabilities proportional to $\\omega_n^j\\ell_n^{\\langle z\\rangle}(\\xi_n^j,\\xi_{n+1}^i)$. The extended-target argument of pseudo-marginal methods keeps the algorithm consistent for the averaged densities $\\ell_n^\\varepsilon$, and the parameter $\\varepsilon$ measures the bias between those averages and the true densities. Rejection or Markov-chain sampling from the extended kernel keeps the per-step cost linear in the number of particles.","core_discovery":"The central claim is that Algorithm 6, the pseudo-marginal PaRIS update, is a valid online smoother for general path models of the form $\\varphi_{0:n}(dx_{0:n})\\propto\\chi(dx_0)\\prod_{m=0}^{n-1}L_m(x_m,dx_{m+1})$ when the kernels $L_m$ have densities that are intractable but estimable. Each transition density $\\ell_n$ is replaced by a random estimate $\\ell_n^{\\langle z\\rangle}$ with expectation $\\ell_n^\\varepsilon$, and the algorithm converges to the 'skew' expectation $\\varphi^\\varepsilon_{0:n}h_n$ under the averaged model. The paper proves a Hoeffding-type concentration inequality and a central limit theorem for the estimator, giving an explicit asymptotic variance with an additional nonnegative term caused by the pseudo-marginal randomness. Under a strong-mixing condition (H3) and with at least $M=2$ backward samples per particle, the asymptotic variance grows at most linearly in $n$, which is optimal for path-space estimators. Under a bias-control condition (H4), the skew target is within $O(n\\varepsilon)$ of the true target for additive functionals, with an $O(\\varepsilon)$ filter-marginal version as a corollary.","pith_inferences":["The $O(n\\varepsilon)$ bias bound suggests a practical discretisation rule the paper does not state: take the density-estimation bias $\\varepsilon$ of order $1/n$ or smaller so that the asymptotic bias stays bounded while the Monte Carlo variance accumulates linearly.","Because score functions and EM intermediate quantities are additive functionals, the same algorithm is a candidate for online parameter estimation in intractable models, though the paper does not run that experiment.","The central-limit-theorem variance could be used to allocate simulation effort between particles and density-estimation draws, in analogy with tuning pseudo-marginal MCMC; the paper leaves this allocation open.","If the uniform-density strong-mixing assumption could be relaxed to geometric ergodicity, the method might extend to noncompact state spaces such as Gaussian transition models; no such extension is proved here."],"forward_implications":["Online additive smoothing becomes feasible for partially observed diffusions and approximate-Bayesian-computation models at $O(N)$ work per time step with $O(1)$ memory, instead of quadratic-cost alternatives.","Biased density estimates are safe when their bias is controlled: choosing $\\varepsilon$ small relative to $1/n$ keeps the total asymptotic bias bounded while variance grows linearly.","The explicit central-limit-theorem variance, including the pseudo-marginal variance term, can guide choices of particle count $N$, backward sample size $M$, and estimation noise level.","The framework contains random-weight particle filters, the original PaRIS, and ABC smoothing as special cases, so their consistency and stability follow from the same theorems."],"supporting_citations":[{"why":"Supplies the PaRIS algorithm and its convergence theory, which the paper extends to general path models and pseudo-marginalisation.","marker":"[28]"},{"why":"Introduces random-weight particle filters for partially observed diffusions, the unbiased-estimation blueprint generalised here.","marker":"[17]"},{"why":"Extends PaRIS to diffusions with unbiased random weights; the present work generalises that extension to biased estimates and general path measures.","marker":"[19]"},{"why":"Provides the asymptotic-variance analysis for particle-based online smoothers that the central limit theorem here corrects and extends.","marker":"[29]"},{"why":"Supplies the Brownian-bridge importance estimator used to illustrate biased transition-density estimation satisfying the bias condition.","marker":"[16]"},{"why":"Provides the rejection-sampling complexity result and concentration lemmas used for the linear-cost backward sampling and exponential bounds.","marker":"[13]"},{"why":"Establishes the pseudo-marginal extended-target framework that justifies consistency when intractable densities are replaced by random estimates.","marker":"[1]"},{"why":"Defines the general interacting-particle path-model framework and mixing machinery that underlies the model class and stability arguments.","marker":"[9]"}],"fun_headline_variants":["Pseudo-marginal PaRIS extends online smoothing to intractable transitions","Online smoothing with pseudo-marginal trick for intractable densities","Pseudo-marginal upgrade lets PaRIS smooth online for intractable transitions","Pseudo-marginal trick unlocks online smoothing for intractable transitions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The long-term stability and bias bounds rely on every transition density, true and estimated, being bounded below and above by fixed positive constants, which in practice means compact state spaces; if that fails, the $O(n)$ variance and $O(n\\varepsilon)$ bias guarantees are not established.","fun_headline_variants_meta":{"raw":{"variants":["Pseudo-marginal PaRIS extends online smoothing to intractable transitions","Online smoothing with pseudo-marginal trick for intractable densities","Pseudo-marginal upgrade lets PaRIS smooth online for intractable transitions","Pseudo-marginal trick unlocks online smoothing for intractable transitions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3623,"prompt_tokens":1005,"completion_tokens":2618,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2541}},"tokens_in":621,"tokens_out":2618,"duration_ms":18168,"temperature":1.0,"reasoning_tokens":2541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:21:31.906270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 6 with $M=2$ on a finite-state hidden Markov model that satisfies the uniform density bounds, using deliberately biased transition-density estimates of known size $\\varepsilon$. Theorems 4.6 and 4.10 predict that as $N$ grows the estimator's variance increases at most linearly in the path length $n$ and its bias stays below a constant times $n\\varepsilon$; observing superlinear variance growth, or bias that cannot be bounded by $C n\\varepsilon$ for any fixed $C$ at large $N$, would falsify the central claims.","supporting_citations":[{"cited_title":"Olsson and J","cited_arxiv_id":null,"evidence_quote":"Supplies the PaRIS algorithm and its convergence theory, which the paper extends to general path models and pseudo-marginalisation."},{"cited_title":"Fearnhead, O","cited_arxiv_id":null,"evidence_quote":"Introduces random-weight particle filters for partially observed diffusions, the unbiased-estimation blueprint generalised here."},{"cited_title":"Gloaguen, M.-P","cited_arxiv_id":null,"evidence_quote":"Extends PaRIS to diffusions with unbiased random weights; the present work generalises that extension to biased estimates and general path measures."},{"cited_title":"Olsson and J","cited_arxiv_id":null,"evidence_quote":"Provides the asymptotic-variance analysis for particle-based online smoothers that the central limit theorem here corrects and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Brownian-bridge importance estimator used to illustrate biased transition-density estimation satisfying the bias condition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the rejection-sampling complexity result and concentration lemmas used for the linear-cost backward sampling and exponential bounds."},{"cited_title":"Andrieu and G","cited_arxiv_id":null,"evidence_quote":"Establishes the pseudo-marginal extended-target framework that justifies consistency when intractable densities are replaced by random estimates."},{"cited_title":"Del Moral","cited_arxiv_id":null,"evidence_quote":"Defines the general interacting-particle path-model framework and mixing machinery that underlies the model class and stability arguments."}],"review_version":1}