{"id":"c94c8ec6-21ae-49f8-86c2-5aeb374cbc64","arxiv_id":"2411.11983","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The occlusion process is a parallel, variational-assisted wrapper that decorrelates MCMC samples and provably inherits LLN, normed convergence, and geometric ergodicity, with empirical variance reduction when the variational proposal is good.","lead":"This paper introduces the occlusion process, a wrapper around an MCMC sampler that occasionally replaces correlated chain samples with independent samples drawn from the target distribution restricted to the region the chain is in, using a variational approximation and extra CPU threads. It aims to cut the autocorrelation that inflates MCMC estimator variance, and it proves the wrapper inherits key convergence properties of the original sampler.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's post-hoc uniform allocation does not realize the kernel (5), so Proposition 7's variance formula does not cover the simulated estimator.","rationale":"The reader's weakest assumption focused on the variational Q's acceptance probability, which is a practical limitation. The more fundamental issue is that the algorithm's post-processing allocation is not the process analyzed in Propositions 7 and 22. This does not invalidate the kernel (5) as a theoretical construction, nor the unbiasedness or LLN-style inheritances if a corrected simulator were used, but it means the variance reduction claims and the 'inherited properties' do not currently apply to the experiments. Since this is fixable (e.g., generate S_t online as independent Bernoulli draws with the realized empirical α, or derive the variance under the without-replacement scheme), the appropriate verdict remains CONDITIONAL, not REJECT. I therefore keep the reader's verdict but for a different reason, hence 'partial' agreement on the weakest assumption.","tokens_in":33685,"tokens_out":12404,"duration_ms":126400,"concrete_test":"Take R=2 with T_1=2, T_2=0, N_1=1, and f(X_1)=a, f(X_2)=b (within region 1), with the chain positions fixed and only the allocation random. Compute Var(μocc) exactly for (i) Algorithm 1's uniform selection of one of the two times and (ii) the kernel (5) with α(1)=1/2 (independent Bernoulli). If these variances differ, Proposition 7's formula (8) does not describe Algorithm 1. A further check is to numerically confirm the variance of the actual Algorithm 1 output for a simple Metropolis chain on a two-point or Gaussian target and compare it to (8); a mismatch invalidates the claimed theoretical backing for the simulation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is between the theoretical object (kernel (5)) and the simulated estimator (Algorithm 1). In (5), S_t are independent Bernoulli draws with fixed probabilities α(ρ(X_t)), and the proof of Proposition 7 explicitly uses the conditional independence {S_t}⊥{S_t'} | {ρ(X_t)} (Section 7.1.4). Algorithm 1 instead allocates, for each region i, exactly min{N_i,T_i} occlusions by sampling a uniform subset of the T_i times in that region. The joint law of {S_t} given the region sequence is hypergeometric, not independent Bernoulli; the paper's claim that this is 'equivalent in distribution to occluding at each step with probability min{1,N_i/T_i}' holds only for each marginal S_t, not for the joint law. Consequently, the variance expression (8), and all variance-reduction conclusions derived from it (Proposition 5, Section 5.2.5), are not proven for the estimator actually simulated. Moreover, α(i) := E[min{1,N_i/T_i}] is data-dependent, whereas (5) requires fixed α; the actual process is not a time-homogeneous Markov chain of the form Kocc. The empirical variance reductions may survive, but the theoretical support for the implementation is missing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes the occlusion process, a wrapper around a P-invariant Markov chain that replaces some chain states with independent samples from P restricted to the region currently visited, obtained by rejection sampling with a variational proposal Q. The paper proves unbiasedness and a variance decomposition for the idealized estimator (Proposition 7), derives inheritance of a Law of Large Numbers, convergence in normed spaces, geometric ergodicity, and a CLT (Theorems 10, 12, 17, 20, Corollary 21), and presents a threaded implementation (Algorithm 1) claimed to run in the same wall-clock time as the underlying sampler. Two experiments, on a bimodal Gaussian mixture and on the Ising model over stochastic-block-model graphs, illustrate variance reduction, and the paper honestly reports cases where the method fails, such as the low-temperature N=100 Ising setting and the Wolff-algorithm comparison.","tokens_in":33958,"tokens_out":6559,"duration_ms":71926,"significance":"The theoretical core around the idealized kernel (5) is a genuine contribution: it connects MCMC variance reduction to stratified sampling and variational inference, and the inheritance theorems are cleanly proved through the operator identity K_occ^t g = K^t f_alpha. The paper is also honest about counterexamples and open questions, which strengthens credibility. However, the implementation in Algorithm 1 does not actually realize the kernel (5), so the variance and inheritance results do not currently cover the estimator that is simulated. This gap is load-bearing for the central claim that the method reduces variance while inheriting the underlying chain's properties.","major_comments":[{"comment":"The theory defines K_occ with independent Bernoulli indicators S_t conditional on the region sequence, and the proof of Proposition 7 explicitly uses the conditional independence {S_t} independent of {S_t'} given {rho(X_t)}. Algorithm 1 instead allocates, for each region i, exactly min{N_i, T_i} occlusions by sampling a uniform subset of the T_i visits. Conditional on T_i and N_i, the joint law of the indicators is hypergeometric, not independent Bernoulli: for example with T_i=2 and N_i=1, the event that both times are occluded has probability 0, whereas independent Bernoulli draws give probability 1/4. Moreover, alpha(i) := E[min{1,N_i/T_i}] is data-dependent, whereas (5) requires fixed alpha; the simulated process is therefore not a time-homogeneous Markov chain of the form K_occ. Consequently the variance formula (8), and all variance-reduction conclusions derived from it (Proposition 5 and Section 5.2.5), are not proven for the estimator actually simulated. The empirical variance reductions may survive, but the theoretical support for the implementation is missing.","section":"Section 4.2.1, Algorithm 1; Section 2.2.2 Eq. (5); Proposition 7; Section 7.1.4"},{"comment":"The claim that the high-temperature Ising experiment satisfies condition 2 of Proposition 5 is not adequately supported. First, the paper infers that the Markov kernel K is positive from positive lag-1 autocorrelations of the magnetisation functional; positivity of one autocovariance at lag 1 does not imply positivity of the spectrum of K on L^2(P), which is what Proposition 5 requires. Second, the argument that alpha is identically 1 is inferred from occlusion proportions in Figure 7, but proportions close to 1 do not imply alpha(i)=1 exactly for every region in every replication. The theoretical explanation of the variance reduction is therefore not established, even setting aside the mismatch between the theory and Algorithm 1.","section":"Section 5.2.5"},{"comment":"The empirical proof-of-concept is weakened by the use of oracle information in constructing the variational regions. The clusters V_i are set to the true SBM communities, while the text acknowledges that 'in practice we will not have access to the actual latent cluster structure' (Section 5.2.3). No experiment uses an estimated clustering, and the threshold constants C_1, C_2 and the parameters epsilon and beta_tilde are chosen with per-temperature values, so the claim in Section 5.2.4 that the parameters were 'not tuned' is hard to reconcile with the setup. The paper's own Section 5.2.4 also reports occlusion proportions near zero in the low-temperature N=100 case; this is an explicit limitation that should be discussed as restricting the method's domain rather than presented only as a tuning issue.","section":"Section 5.2.3 and Section 5.2.4"}],"minor_comments":[{"comment":"The notation P(A) is used both for the probability of a set A under the target measure and, a few lines later, for the power set of A. Please use a different symbol, such as 2^A, for the power set.","section":"Section 1.1"},{"comment":"The text refers to 'the top right graph' for both the lag-1 autocorrelation coefficients and the occlusion proportions, but the figure caption and surrounding discussion suggest these are different panels (top left and top right). Please align the textual references with the figure.","section":"Section 5.2.4 and Figure 7"},{"comment":"The proof of Theorem 17 cites '[section 6.3.1]', which does not exist in the manuscript; this cross-reference should be replaced with the actual derivation or a specific equation number.","section":"Section 7.2.6"},{"comment":"There is a typo: 'We would like to that Francois Perron' should read 'We would like to thank Francois Perron'.","section":"Acknowledgements"},{"comment":"The claim of 'no additional time-complexity' is only true in an asymptotic wall-clock sense with a sufficient number of threads and comparable per-step costs. As written, it could be misread as no additional computational work; please qualify the statement.","section":"Abstract and Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the gap between the idealized kernel (5) and Algorithm 1. The theoretical content is sound for the idealized process, but the simulated estimator is not that process, so the variance and inheritance theorems do not currently support the empirical claims. A revision should either modify Algorithm 1 to actually realize independent Bernoulli occlusions with a fixed or data-dependent alpha for which the theory is proved, or develop a variance analysis for the hypergeometric allocation. The use of true SBM community labels in the Ising experiments is an additional weakness that should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The occlusion process is a real new idea: a wrapper that replaces some MCMC states with independent draws from the target restricted to the visited region, without needing known region masses or regeneration. It is clearly distinct from stratified sampling and from Kick-Kac samplers, and the inheritance results — LLN, normed convergence, geometric ergodicity — are genuinely useful and the proofs are careful. Proposition 7's variance decomposition is transparent, and the paper is honest about cases where occlusion hurts, like low-temperature Wolff.\n\nThe soft spot is load-bearing. In kernel (5), the S_t are independent Bernoulli draws with fixed α(ρ(X_t)), and the proof of Proposition 7 explicitly uses conditional independence of the S_t given the region sequence. Algorithm 1 instead fixes N_i and then chooses exactly min{N_i,T_i} times uniformly within each region. The paper's claim that this is 'equivalent in distribution' to per-step Bernoulli occlusion is true only for each marginal S_t, not for the joint law: the S_t are hypergeometric within a region, hence negatively correlated. So the variance formula (8), and the variance-reduction discussion in Section 5.2.5, are not proven for the estimator actually simulated. Unbiasedness may survive, but it is not established by Proposition 7, and the Kocc inheritance theorems do not cover a time-inhomogeneous post-hoc allocation rule. This is not a cosmetic mismatch; it is the bridge between the theory and the experiments.\n\nThe other weaknesses are smaller but real: the Ising regions are tuned from a pilot Wolff run, the variational clusters use the true SBM community labels, there is no code or data, and the variance comparisons lack error bars. All fixable.\n\nRecommendation: send it to peer review, but ask for major revision. The easiest fix is to change Algorithm 1 so that occlusion decisions are made online with fixed probabilities α(i), matching kernel (5); alternatively, redo the variance analysis for the hypergeometric allocation. As written, the paper overclaims coverage of the method it evaluates.","headline":"A genuinely new variance-reduction construction with careful inheritance theorems, but the paper's main applied claim overreaches: Algorithm 1 does not realize the kernel that the variance formulas analyze.","tokens_in":34475,"tokens_out":4436,"would_cite":true,"duration_ms":47721,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Wrapping an MCMC sampler with a parallel rejection layer that swaps autocorrelated chain states for independent region-restricted draws yields an unbiased estimator with lower variance and the chain's convergence guarantees.","keywords":["Markov chain Monte Carlo","variance reduction","variational inference","parallel computation","occlusion process","stratified sampling","Ising model","geometric ergodicity"],"falsifier":"Run the occlusion process on a target where $Q$ is deliberately chosen so that $(Z_P/Z_Q)(1/C_i)P(X_i)$ predicts near-certain acceptance on the regions the chain visits, with a functional $f$ that is piecewise constant on those regions, and measure the variance ratio $\\mathrm{Var}(\\hat{\\mu}_{\\mathrm{occ}})/\\mathrm{Var}(\\hat{\\mu}_{\\mathrm{chain}})$. Proposition 5 and equation (8) predict a ratio at or below 1; a measured ratio above 1 in this regime would falsify the variance-reduction claim.","tokens_in":33470,"feed_emoji":"🔀","tokens_out":16369,"duration_ms":132947,"temperature":0.7,"pith_summary":"The paper introduces the occlusion process, a wrapper that sits on top of any $P$-invariant Markov chain Monte Carlo (MCMC) sampler and replaces some of the chain's autocorrelated states with independent draws from the target restricted to the region the chain currently occupies. The claim is that this yields an unbiased estimator whose variance can be smaller than the plain chain's whenever the partition makes the target functional nearly piecewise constant on regions, or when the chain's kernel has positive spectrum and the functional is orthogonal to the region means (Proposition 5). The wrapped process inherits the underlying chain's law of large numbers, normed-space convergence, and geometric ergodicity, and it can be simulated in the same wall-clock time as the chain by running rejection samplers with a variational proposal $Q$ on spare threads. Two numerical studies, a bimodal Gaussian mixture and the Ising model, show decorrelation and variance reduction, with the caveat that a poor $Q$ or an anticorrelated baseline chain can neutralize or reverse the gain. The paper matters because it converts otherwise idle parallel threads into variance reduction while preserving MCMC's unbiasedness and consistency guarantees.","feed_headline":"Occlusion process cuts MCMC variance at zero extra time","feed_subtitle":"Autocorrelated chain states get swapped for independent region-restricted draws, with no loss of consistency.","key_machinery":"The load-bearing object is the occlusion kernel $K_{\\mathrm{occ}}$ (equation 5), which extends the chain's kernel $K$ with a per-region success indicator $S_t$ of probability $\\alpha(\\rho(X_t))$ and an independent replacement sample $Y_t$ drawn from $P$ restricted to the region $\\rho(X_t)$; the estimator (6) then averages $f$ over whichever variable, chain state or occluding draw, is active at each time. The variance analysis is carried by the resolution operator $\\overrightarrow{P}f(x) := \\sum_{i=1}^{R} \\mu_i \\mathbf{1}\\{x \\in X_i\\}$, the piecewise-constant function of region means: it splits $\\mathrm{Var}_P(f)$ into between-region variance $\\mathrm{Var}_P(\\overrightarrow{P}f)$ and within-region variance $\\mathrm{Var}_P(\\overleftarrow{P}f)$, and occlusion replaces the autocorrelated within-region variation with independent region-restricted draws. The region-restricted draws come from a rejection sampler with variational proposal $Q$, with the regions defined as level sets of the Radon-Nikodym derivative $\\mathrm{d}\\widetilde{P}/\\mathrm{d}\\widetilde{Q}$ and per-region acceptance probability $(Z_P/Z_Q)(1/C_i)P(X_i)$; spare threads run these samplers concurrently with the base chain, which is why the wall-clock cost does not rise.","core_discovery":"On the paper's own terms, the central discovery is that an MCMC estimator can be improved by a layer of occlusion: for a partition of the state space into regions, whenever a parallel rejection sampler built from a variational approximation $Q$ succeeds in drawing a sample from the target $P$ restricted to the region the chain has just entered, that independent sample replaces the chain's state in the Monte Carlo average. Proposition 7 states that the resulting estimator is unbiased and gives its exact variance as the stratified-proportional-allocation variance plus a covariance term for the residual autocorrelation of partially occluded states (equation 8). The occluded process does not merely keep its consistency for free: Theorems 10, 12, and 20 show it inherits a law of large numbers, convergence in normed function spaces, and geometric ergodicity from the base chain, and Corollary 21 adds a central limit theorem under reversibility. The paper also sets limits on the idea: Fact 2 shows that full occlusion can raise variance relative to the plain chain when the kernel has negative spectral part, and it leaves open whether that can happen with a partition into more than one region.","pith_inferences":["My reading: the paper's variance formula (8) suggests a principled region-selection rule analogous to Neyman allocation, choosing the level-set constants $C_i$ to balance the acceptance probability $(Z_P/Z_Q)(1/C_i)P(X_i)$ against the within-region variances $\\sigma_i^2$; the authors leave such tuning to future work.","My reading: the open multi-region counterexample question could plausibly be settled spectrally; if $\\mathrm{Var}(\\hat{\\mu}_{\\mathrm{ideal}}) \\leq \\mathrm{Var}(\\hat{\\mu}_{\\mathrm{chain}})$ holds whenever $K$ has nonnegative spectrum, then the positivity condition of Proposition 5 is not just sufficient but necessary for guaranteed improvement.","My reading: the method's weakest practical link is the fixed, pre-trained $Q$; an online variant that re-fits $Q$ from accepted occluding samples as the chain runs would directly attack the near-zero occlusion proportions reported for the low-temperature Ising $N=100$ case.","My reading: the chain states discarded by occlusion could be recycled rather than erased, for instance as control variates or Rao-Blackwellized combinations of $X_t$ and $Y_t$, converting the method from sample replacement into information reuse."],"forward_implications":["Any existing $P$-invariant Markov chain can be wrapped in the occlusion layer, and the standard convergence toolkit (LLN, normed-space convergence, geometric ergodicity, CLT under reversibility) transfers unchanged, so users keep the guarantees they already paid for.","On a threaded machine the occluded estimator runs in the same wall-clock time as the plain chain, so the variance reduction is essentially free whenever a reasonable variational approximation $Q$ is available.","Variance reduction is guaranteed in two regimes: functionals that are piecewise constant on the regions, and functionals such as the Ising magnetisation (odd function, symmetric target, symmetric regions) averaged over a kernel with positive spectrum; the high-temperature Ising experiments instantiate the second.","When the base chain already produces anticorrelated states, as Wolff's algorithm does at low temperature, occlusion's decorrelation can raise variance instead of lowering it, so the method should be aimed at positively autocorrelated samplers.","The occluded estimator is not universally better than the plain chain: Fact 2 exhibits a one-region counterexample, and it remains open whether any multi-region partition can make the fully occluded estimator lose to the chain."],"supporting_citations":[{"why":"It supplies the stratified sampling with proportional allocation framework whose variance the fully occluded estimator mirrors in Proposition 1.","marker":"[Cochran, 1977]"},{"why":"It defines the closely related Kick-Kac samplers that the occlusion process is compared against and contributes the Proposition 3.5 used to prove the LLN inheritance in Theorem 10.","marker":"[Douc et al., 2023]"},{"why":"It provides the rejection-sampling mechanism (Section 2.3.4) that yields exact samples from the target restricted to a region.","marker":"[Tierney, 1994]"},{"why":"It grounds the variational inference paradigm that supplies the proposal distribution $Q$ the whole implementation depends on.","marker":"[Blei et al., 2017]"},{"why":"It is the Metropolis algorithm baseline whose high autocorrelation the Ising experiments show occlusion removing.","marker":"[Metropolis et al., 1953]"},{"why":"It is the cluster-update baseline whose anticorrelated low-temperature states make occlusion increase variance instead of reducing it.","marker":"[Wolff, 1989]"},{"why":"It underwrites the CLT for reversible geometrically ergodic chains that Corollary 21 invokes for the occluded estimator.","marker":"[Roberts and Rosenthal, 1997]"},{"why":"It supplies the antithetic-chain example whose negative spectral part witnesses Fact 2 that full occlusion can lose to the plain chain.","marker":"[Neal and Rosenthal, 2024]"},{"why":"It gives the drift-plus-small-set characterization of geometric ergodicity used in the proof of Theorem 20.","marker":"[Gallegos-Herrada et al., 2024]"}],"fun_headline_variants":["Occlusion swap: replace MCMC state with independent draw, cut variance","Occlusion process: decorrelated draws cut MCMC variance, no time cost","Occlusion process replaces chain samples to cut variance at zero time","Occlusion: parallel swap cuts MCMC variance without slowing sampler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a variational distribution $Q$ can be found whose rejection-sampler acceptance probability $(Z_P/Z_Q)(1/C_i)P(X_i)$ is non-negligible on the regions the chain actually visits; when $Q$ misses that mass the occlusion rate collapses and the estimator degenerates to the plain chain, as the authors' own low-temperature Ising run at $N=100$ shows.","fun_headline_variants_meta":{"raw":{"variants":["Occlusion swap: replace MCMC state with independent draw, cut variance","Occlusion process: decorrelated draws cut MCMC variance, no time cost","Occlusion process replaces chain samples to cut variance at zero time","Occlusion: parallel swap cuts MCMC variance without slowing sampler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001027,"raw_usage":{"total_tokens":4327,"prompt_tokens":943,"completion_tokens":3384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3304}},"tokens_in":559,"tokens_out":3384,"duration_ms":24570,"temperature":1.0,"reasoning_tokens":3304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:03:28.229158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the occlusion process on a target where $Q$ is deliberately chosen so that $(Z_P/Z_Q)(1/C_i)P(X_i)$ predicts near-certain acceptance on the regions the chain visits, with a functional $f$ that is piecewise constant on those regions, and measure the variance ratio $\\mathrm{Var}(\\hat{\\mu}_{\\mathrm{occ}})/\\mathrm{Var}(\\hat{\\mu}_{\\mathrm{chain}})$. Proposition 5 and equation (8) predict a ratio at or below 1; a measured ratio above 1 in this regime would falsify the variance-reduction claim.","supporting_citations":[{"cited_title":"Boost your favorite Markov Chain Monte Carlo sampler using Kac's theorem: the Kick-Kac teleportation algorithm","cited_arxiv_id":"2201.05002","evidence_quote":"It defines the closely related Kick-Kac samplers that the occlusion process is compared against and contributes the Proposition 3.5 used to prove the LLN inheritance in Theorem 10."},{"cited_title":"Markov chains for exploring posterior distributions","cited_arxiv_id":null,"evidence_quote":"It provides the rejection-sampling mechanism (Section 2.3.4) that yields exact samples from the target restricted to a region."},{"cited_title":"Collective monte carlo updating for spin systems","cited_arxiv_id":null,"evidence_quote":"It is the cluster-update baseline whose anticorrelated low-temperature states make occlusion increase variance instead of reducing it."},{"cited_title":"Geometric Ergodicity and Hybrid Markov Chains","cited_arxiv_id":null,"evidence_quote":"It underwrites the CLT for reversible geometrically ergodic chains that Corollary 21 invokes for the occluded estimator."}],"review_version":1}