{"id":"3f5fee08-9916-4f5a-ae0c-672eac4f2517","arxiv_id":"2501.03395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A grid-based hidden Markov model approximation is used to build importance proposals inside particle Gibbs with ancestor sampling, cutting compute on regime-switching state-space models.","lead":"This paper introduces a new way to run particle Gibbs samplers for state-space models, using a coarse grid to build a hidden Markov model approximation and guide particles toward high-probability regions. The method cuts computation time on hard regime-switching models, including a COVID-era tourism demand example, compared with standard particle Gibbs with ancestor sampling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central efficiency claim rests on a fixed equal-spaced grid approximating the HMM proposal well enough; the paper's own tourism result shows this fails for wide-range latent states, so the claimed gains are not established as general.","rationale":"The paper proposes a valid PGAS variant: using an arbitrary importance distribution inside conditional SMC preserves the exact-Gibbs property, so the grid proposal is not a correctness threat. The simulated stochastic volatility results provide genuine evidence that the grid proposal can reduce sample impoverishment for fixed computation time, and the paper deserves credit for that positive demonstration. However, the abstract's claim of substantial computational gains is an efficiency claim, and efficiency rests on the uncharacterized approximation error of the deterministic grid. That error is controlled by cell width relative to the posterior scale, and the paper's own tourism case shows the failure mode: for u_t with range [-300, 1000], pure GPGAS is dominated by conditional PGAS, and only the hybrid GPGAS+PGAS variant improves on it. The discussion section concedes this limitation, but the abstract and title do not qualify it. A fixed grid that is coarse for wide-range states is not a robust general recipe, and no theoretical or practical criterion is given for when the grid is fine enough. The reader's weakest-assumption identification is accurate, and the conditional verdict is appropriate. The proposed concrete test would distinguish whether the problem is the equal-spaced grid geometry or the grid-proposal concept itself, and would inform whether a grid-placement or grid-adaptation fix can rescue the claimed gains.","tokens_in":22580,"tokens_out":19302,"duration_ms":197585,"concrete_test":"Re-run the Section 4.2 tourism analysis with pure GPGAS, keeping N=200, M=100 and the same model, but with the finite grid for u_t set to a pilot-estimated 5-95% posterior quantile interval instead of the full [-300, 1000] range, retaining infinite outer cells as in Section 3.4c. If pure GPGAS then matches or beats the GPGAS+PGAS row of Table 1 (MRAE <= 0.029, ESS >= 5800), the fixed wide equal-spaced grid is the identified cause of the underperformance; if it still underperforms, the grid proposal's cell-based approximation, not just the chosen range, is the limitation for wide-range states.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise of the efficiency claim is that a fixed, equal-spaced grid with midpoint-integrated HMM probabilities (Section 3.1, Equation (8)) produces an importance distribution close enough to p(xt|yt, xt-1, theta) that the conditional SMC proposal is efficient. This premise is not proven and is scale-fragile: the grid proposal conditions only on the sampled cell B_{t-1}, not on the exact previous particle, and uses uniform within-cell draws, so its quality degrades as cell width grows relative to the transition scale. The paper's own real-data result in Section 4.2.2 (Table 1) is direct evidence: pure GPGAS (MRAE 0.044, ESS 1100) is less efficient than conditional PGAS (MRAE 0.039, ESS 3300), attributed by the authors to the wide range [-300, 1000] of u_t requiring too many grid cells. Thus the abstract's 'substantial computational gains' hold only when the grid is fine relative to the posterior range; the paper gives no criterion for when that holds, and the fixed-grid choice is not robust to exactly the regime where grid methods are supposed to help. This is an approximation-quality assumption, not an internal inconsistency, but it is load-bearing because the headline claim is an efficiency claim. The empirical reporting also lacks code, data, and uncertainty bars, and excludes non-convergent runs without a stated criterion, which further limits the strength of the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Grid Particle Gibbs with Ancestor Sampling (GPGAS), an extension of particle Gibbs with ancestor sampling for general state-space models. The method partitions the continuous state space into a fixed, equal-sized grid and constructs an approximate hidden Markov model using midpoint-integrated transition and observation probabilities (Section 3.1). The approximate HMM is used to form SMC importance proposal distributions that first sample a grid cell and then draw a continuous particle within that cell (Section 3.2). The resulting algorithm, Algorithm 6, is applied to a simulated regime-switching stochastic volatility model and to a real post-COVID tourism demand model for Edinburgh. The paper claims substantial computational gains over standard PGAS and reports favorable error-versus-time results in both settings, with the real-data study carried out on a hybrid GPGAS+PGAS update scheme.","tokens_in":22944,"tokens_out":6599,"duration_ms":64936,"significance":"If the efficiency gains are reproducible, GPGAS is a useful contribution to the toolkit for fitting nonlinear and non-Gaussian state-space models, especially regime-switching models where sample impoverishment is severe. Strengths of the paper include the clear algorithmic presentation, the practical computational strategies listed in Section 3.4, and the use of high-particle PGAS ground-truth runs in both the simulation and the real-data study. However, the empirical evidence is more qualified than the abstract suggests: the headline real-data gains come from a hybrid that applies GPGAS to only part of the state vector, and the simulated study is reported without uncertainty bars and without a stated convergence-exclusion criterion. The central efficiency claim is plausible but is not yet established at the level of generality claimed.","major_comments":[{"comment":"The captions of Figures 3 and 4 state that non-convergent implementations are excluded, but no convergence criterion is defined anywhere in the manuscript. If exclusion is based on the error relative to the ground truth, the comparison can be biased in favor of whichever algorithm survives the filter. Please report the exact diagnostic used, the number of excluded runs per configuration, and ideally show results with all runs included as a sensitivity check. This is load-bearing because the simulated-study efficiency claim rests on these plots.","section":"Section 4.1.2 / Figures 3-4"},{"comment":"The real-data results show that the pure GPGAS algorithm is not more efficient than conditional PGAS: GPGAS has MRAE 0.044 and ESS 1100, compared with conditional PGAS MRAE 0.039 and ESS 3300. The best result in Table 1, GPGAS+PGAS, replaces the GPGAS update for u_{1:T} with a standard PGAS update. The abstract and the discussion nevertheless claim substantial computational gains of the proposed method without this qualification. The claim should be restricted to the setting where the fixed equal-spaced grid is a reasonable approximation, or to the hybrid algorithm, and the paper should provide a practical criterion (for example, the ratio of posterior range to transition scale) for when pure GPGAS is expected to be efficient. Without such a criterion, the generality of the headline claim is not established.","section":"Section 4.2.2 / Table 1"},{"comment":"The efficiency comparisons are reported as single points per configuration from 10 runs, with no Monte Carlo error bars, and each comparison uses a single ground-truth run (around 89 hours for the simulation and 97 hours for the tourism data). Since the central claim is an efficiency claim, the absence of uncertainty quantification makes it difficult to assess whether the observed differences are systematic or within Monte Carlo noise. Please add error bars or interval estimates for the reported MRAE/ESS values and, if feasible, quantify the Monte Carlo error of the ground-truth estimates.","section":"Section 4.1.2 and Section 4.2.2"}],"minor_comments":[{"comment":"In the displayed weight formula, the denominator for the time-1 weight is written as q(x_m^1, B_1 = b_m^1 | y_t, θ); it should condition on y_1 rather than y_t.","section":"Section 3.3"},{"comment":"The denominator of the normalized weight is typeset as 'Pm_{k=1} w_{1:t}' which should be a summation symbol; please correct the typesetting and clarify the notation for the sets w_{1:t}^{1:M} and W_{1:t}^{1:M}.","section":"Section 2.1 / Equation (3)"},{"comment":"The text contains duplicated words: 'Figure 6 shows the the results' and 'Figure 7 shows the the results' should be 'shows the results'.","section":"Appendix B"},{"comment":"The caption says the metric is the mean relative absolute error (MRAE) of the posterior predictive mean and variance, then reports 'average RRMSE' for credible intervals; please define MRAE and RRMSE and make the notation consistent, since RRMSE is not defined elsewhere.","section":"Table 1"},{"comment":"The HMM approximation is fixed after a pilot-chosen iteration tilde_s (2000 in the simulation, 1500 in the tourism example) using posterior means, but no sensitivity analysis is provided for this choice or for the grid ranges. A short robustness discussion or table would help readers judge how sensitive the method is to these tuning decisions.","section":"Section 3.4(a) / Sections 4.1.1 and 4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is squarely within the scope of a computational statistics journal, and the hybrid GPGAS+PGAS idea is publishable after revision. The main issue is that the abstract and conclusion overstate the real-data support for the proposed method as a standalone algorithm; the paper's own Table 1 shows that pure GPGAS is less efficient than conditional PGAS in the tourism example. The simulated-study evidence would also be strengthened by reporting the convergence-exclusion criterion and Monte Carlo uncertainty. I did not find evidence of circularity: the comparison targets standard PGAS and a high-particle ground truth, not the authors' own earlier method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a genuine, if incremental, methodological contribution—putting a deterministic-grid HMM proposal inside PGAS—and the simulated stochastic volatility results are encouraging. But the headline efficiency claim is broader than the evidence, and the reporting leaves some gaps a referee should close.\n\nWhat's new: the GPGAS algorithm uses a coarse HMM on a fixed grid to build importance proposals; the computational tricks in Section 3.4 (one transition matrix per iteration, computing only for occupied cells, batch sampling within cells) are sensible and could help others. The simulated regime-switching SV comparison is the strongest part: GPGAS reaches lower MRAE than PGAS for fixed wall-clock time across a range of N and M, and the paper honestly breaks down errors by switching/non-switching states. I also give credit for reporting the tourism example where pure GPGAS loses to conditional PGAS—that's the right kind of self-criticism, though it undercuts the abstract's sweeping claim.\n\nSoft spots, in order of importance:\n\n1. The grid approximation is load-bearing. The proposal is built from a fixed equal-spaced grid with midpoint integration; its quality depends on cell width relative to the posterior scale. The tourism result is direct evidence: for u_t with range [-300,1000], pure GPGAS has higher MRAE (0.044 vs 0.039) and lower ESS (1100 vs 3300) than conditional PGAS. The paper gives no criterion for when the grid is fine enough, and the abstract claims 'substantial computational gains' without this caveat.\n\n2. Empirical reporting is thin. No code or data, no uncertainty bars on the efficiency curves, and 'non-convergent implementations are excluded' with no stated convergence criterion. That matters because the method has several tuning parameters (N, grid range, fixing iteration, resampling threshold), and the reader can't assess robustness without sensitivity analysis.\n\n3. The theoretical grounding is limited to a heuristic argument about reducing sample impoverishment; there is no result on approximation error or convergence rate. That's not a fatal flaw for an applied methods paper, but it should be acknowledged more explicitly.\n\nOverall, the algorithm is coherent, the authors know the literature, and the limitation they found is real and reported. This paper deserves peer review—not desk rejection—but the revision should include code/data (or a clear availability statement), a convergence criterion, uncertainty quantification, and a more measured abstract.\n\nRecommendation: send to review, with a statistical referee who uses PGAS in practice.\n\nBest","headline":"Solid incremental method with a plausible efficiency gain on simulated SV, but the abstract overclaims and the reporting needs more rigor before it can be recommended unconditionally.","tokens_in":23442,"tokens_out":3091,"would_cite":true,"duration_ms":29180,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","65C05","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A coarse hidden Markov model on a deterministic grid steers SMC particles into high-posterior regions, giving particle Gibbs with ancestor sampling lower estimation error per unit computation time.","keywords":["particle Gibbs with ancestor sampling","state-space models","hidden Markov model approximation","deterministic grid proposal","sample impoverishment","regime-switching state-space models","sequential Monte Carlo","Bayesian inference"],"falsifier":"Run GPGAS and conditional PGAS on the regime-switching stochastic volatility model of Section 4.1, but let a new regime appear after the pilot iterations whose latent-state values lie outside the pilot-tuned grid range $[-12, 12]$. If the frozen equal-sized grid cannot direct particles into the new high-posterior region and GPGAS's error per unit computation time is no better than conditional PGAS, the approximation-quality assumption, and with it the claimed computational gain, fails. The paper already reports this failure mode for the wide-support trend state $u_t$ in the tourism model (Table 1).","tokens_in":22393,"feed_emoji":"📊","tokens_out":20587,"duration_ms":153848,"temperature":0.7,"pith_summary":"This paper claims that the most expensive part of particle Gibbs with ancestor sampling—the sequential Monte Carlo sweeps that propose new latent-state paths—can be made much cheaper by proposing from a coarse, deterministic grid instead of from the model's raw transition noise. The authors partition the state space into equally sized intervals, evaluate the model's transition and observation densities at the interval midpoints, and use the resulting approximate hidden Markov model to sample each particle from a grid cell that the approximate filter favors. Because the proposals land in regions of high posterior mass, fewer particles are wasted in resampling, so the same accuracy is reached with far fewer particles. On a simulated regime-switching stochastic volatility model and on a post-COVID tourism demand model, they report lower estimation error per unit computation time than standard particle Gibbs with ancestor sampling, with the best results coming from a hybrid that applies the grid only to states with sufficiently narrow posterior support.","feed_headline":"Grid-guided particles cut error per second on switching models","feed_subtitle":"GPGAS proposes latent states from a coarse HMM grid, reporting lower error than standard PGAS at equal wall-clock time.","key_machinery":"The engine of the method is the deterministic-grid HMM approximation. The state space is divided into $N$ equally sized intervals $I^{(1)}, \\dots, I^{(N)}$ with two infinite outer cells; each interval has midpoint $\\xi^{(n)}$ and length $L^{(n)}$. Midpoint integration turns the state-space model into a discrete HMM with $\\hat P(B_1 = n \\mid \\theta) \\propto L^{(n)} p(\\xi^{(n)} \\mid \\theta)$, $\\hat P(B_t = n \\mid B_{t-1} = k, \\theta) \\propto L^{(n)}L^{(k)} p(\\xi^{(n)} \\mid \\xi^{(k)}, \\theta)$, and $\\hat p(y_t \\mid B_t = n, \\theta) \\propto L^{(n)} p(y_t \\mid \\xi^{(n)}, \\theta)$, each normalized over cells. From these, the approximate optimal importance distribution $\\hat P(B_t = n \\mid y_t, B_{t-1} = k, \\theta) \\propto \\hat P(B_t = n \\mid B_{t-1} = k, \\theta)\\, \\hat p(y_t \\mid B_t = n, \\theta)$ is formed only for grid cells occupied at the previous time point, so the per-iteration cost drops from $O(N^2 T)$ to $O(N^2 + N \\sum_t \\tilde N_{t-1})$. This object carries the argument: it directs particles toward high-posterior regions while keeping the computation near-linear in the number of occupied cells.","core_discovery":"The central claim is that a cheap hidden Markov model (HMM) approximation of a general state-space model can serve as the proposal inside particle Gibbs with ancestor sampling (PGAS, the MCMC scheme that updates the whole latent-state path with a conditional sequential Monte Carlo sweep), and that this removes much of the sample impoverishment—the loss of particle diversity from resampling—that forces PGAS to use many particles. The approximation partitions the continuous state space into $N$ equal cells with midpoints $\\xi^{(n)}$ and lengths $L^{(n)}$, then replaces the intractable integrals in the HMM's initial, transition, and observation probabilities by one-point midpoint quadrature: $\\hat P(B_t = n \\mid B_{t-1} = k, \\theta) \\propto L^{(n)}L^{(k)} p(\\xi^{(n)} \\mid \\xi^{(k)}, \\theta)$ and $\\hat p(y_t \\mid B_t = n, \\theta) \\propto L^{(n)} p(y_t \\mid \\xi^{(n)}, \\theta)$. At each sequential Monte Carlo step a grid cell index is drawn from the discrete filtered distribution $\\hat P(B_t = n \\mid y_t, B_{t-1} = k, \\theta) \\propto \\hat P(B_t = n \\mid B_{t-1} = k, \\theta)\\, \\hat p(y_t \\mid B_t = n, \\theta)$, and a continuous particle is then drawn uniformly inside that cell, or from a truncated Gaussian in the outer cells. This tractable proposal mimics the optimal importance distribution and reduces sample impoverishment; the paper reports 11-50% fewer states left unupdated in the SMC sweep and clear improvements in mean relative absolute error per unit time across the regime-switching examples.","pith_inferences":["I expect the grid-proposal idea to transfer beyond MCMC to plain particle filtering with fixed parameters, where the HMM matrices are computed once rather than per iteration, making the overhead even smaller; the paper itself notes filtering as a natural extension.","The equal-sized grid is a convenience rather than a necessity: a quantile-informed grid that allocates more cells where the posterior has mass could remove the wide-support failure mode the paper reports for $u_t$, at the cost of some simplicity.","Because the HMM is frozen after pilot iterations, GPGAS should be sensitive to posterior movement later in the run; a testable extension is to compare the fixed version against one that recomputes the grid on a slow schedule, weighing robustness against wall-clock gains.","The reported 11-50% reductions in unupdated states suggest the gain scales with switching frequency; a natural experiment is to sweep the regime-persistence parameter $\\pi_{11}$ and measure when the GPGAS advantage over PGAS disappears."],"forward_implications":["GPGAS can reach a given level of latent-state estimation error with substantially fewer particles than PGAS: in the stochastic volatility experiments, the average number of states not updated in the SMC sweep falls by 11-50%, so the accuracy gain is not bought with extra particles.","The per-iteration cost scales with the number of occupied grid cells rather than with $N^2$ per time point, so adding grid resolution to cover the posterior is much cheaper than adding particles.","On the tourism demand model, the efficient configuration is hybrid: apply GPGAS to the regime labels and the narrow-support level state, and keep standard PGAS for the wide-support trend state; this raises the effective sample size to roughly 5800 versus 3100-3300 for pure PGAS variants, with more iterations completed per hour.","Because the grid only changes the SMC proposal and not the target distribution, the resulting MCMC samples still target the joint posterior $p(x_{1:T}, \\theta \\mid y_{1:T})$, so the gains carry over without changing the interpretation of the posterior estimates.","The method's advantage is largest when the posterior mass is concentrated enough for an equal-sized grid to cover it with moderate $N$; the paper documents that pure GPGAS loses to conditional PGAS on the wide-support trend state $u_t$."],"supporting_citations":[{"why":"Establishes the particle Gibbs algorithm and its convergence to the joint posterior, the framework that GPGAS inherits.","marker":"Andrieu et al. (2010)"},{"why":"Proposes particle Gibbs with ancestor sampling, whose conditional SMC steps the grid proposal replaces.","marker":"Lindsten et al. (2014)"},{"why":"Provides the point-mass HMM grid construction and midpoint-integration scheme that Step 1 of GPGAS follows.","marker":"Llewellyn et al. (2023a)"},{"why":"Introduces the point-mass HMM approximation of continuous state-space models that the grid builds on.","marker":"Kitagawa (1987)"},{"why":"Defines the auxiliary particle filter, the alternative optimal-proposal approximation used in the PGAS-APF comparison baseline.","marker":"Pitt and Shephard (1999)"},{"why":"Defines the optimal importance distribution that the grid-based proposal approximates.","marker":"Doucet and Johansen (2009)"}],"fun_headline_variants":["Grid HMM proposal trims particle waste","Particle Gibbs gets a grid shortcut","Grid-based proposal reduces sample impoverishment","Grid particles improve accuracy per second","GPGAS: grid proposal for efficient latent state sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one coarse, fixed, equal-sized grid, with probabilities computed by one-point midpoint integration and frozen after a short pilot run, continues to approximate the posterior propagation well enough to steer particles into high-probability regions for every time step of the whole MCMC run; the paper's own tourism result shows this can fail when a latent state's posterior is spread over a wide range, where pure GPGAS is less efficient than conditional PGAS.","fun_headline_variants_meta":{"raw":{"variants":["Grid HMM proposal trims particle waste","Particle Gibbs gets a grid shortcut","Grid-based proposal reduces sample impoverishment","Grid particles improve accuracy per second","GPGAS: grid proposal for efficient latent state sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1453,"prompt_tokens":1040,"completion_tokens":413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":348}},"tokens_in":656,"tokens_out":413,"duration_ms":51085,"temperature":1.0,"reasoning_tokens":348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:56.536642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GPGAS and conditional PGAS on the regime-switching stochastic volatility model of Section 4.1, but let a new regime appear after the pilot iterations whose latent-state values lie outside the pilot-tuned grid range $[-12, 12]$. If the frozen equal-sized grid cannot direct particles into the new high-posterior region and GPGAS's error per unit computation time is no better than conditional PGAS, the approximation-quality assumption, and with it the claimed computational gain, fails. The paper already reports this failure mode for the wide-support trend state $u_t$ in the tourism model (Table 1).","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the particle Gibbs algorithm and its convergence to the joint posterior, the framework that GPGAS inherits."},{"cited_title":"I., and Schön, T","cited_arxiv_id":null,"evidence_quote":"Proposes particle Gibbs with ancestor sampling, whose conditional SMC steps the grid proposal replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the point-mass HMM approximation of continuous state-space models that the grid builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the auxiliary particle filter, the alternative optimal-proposal approximation used in the PGAS-APF comparison baseline."},{"cited_title":"and Johansen, A","cited_arxiv_id":null,"evidence_quote":"Defines the optimal importance distribution that the grid-based proposal approximates."}],"review_version":1}