{"id":"e2633b92-5525-4f94-8cb5-970e78b12cca","arxiv_id":"2607.22449","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The Zig-Zag and Bouncy Particle samplers have known dimension, anisotropy, and big-data scaling limits that make their efficiency costs predictable, and this survey assembles them with tuning heuristics.","lead":"This survey collects scaling-limit results for two continuous-time MCMC samplers — the Zig-Zag process and the Bouncy Particle Sampler — showing how their cost scales with dimension, with narrow ridges, and with dataset size. A reader gets a compact comparison of when these modern samplers beat classical algorithms like Random Walk Metropolis.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Anisotropy complexity claim appears off by √ε: Eq. (1) switching rates scale as ε^{-1/2}, not ε^{-1}, so §6's O(ε^{-1}) vs O(ε^{-2}) event counts need correction.","rationale":"The reader's weakest assumption—the unproved adiabatic reduction in Section 6—correctly identifies the anisotropy section as load-bearing. My independent reading locates a more concrete problem: the survey's own rate equation (1) appears to contradict the event-rate scaling (ε^{-1}) used to derive the central complexity comparison. This is an internal consistency issue, not a challenge to the external literature. The high-dimensional Gaussian section (Theorem 5.1) is supported by the stated scaling limits and the per-event cost analysis; the big-data section (Section 7) is explicit about the control-variate setup cost. Section 6, however, is where the quantitative claims are least secure. The discrepancy is easy to test and does not require access to [14]: compute the expected switching rate from the displayed formulas. If the rate is indeed O(ε^{-1/2}), the complexity numbers for anisotropic targets must be revised, and the survey should be accepted only after that correction or after a clear reconciliation with [14]. This does not reject the survey's overall contribution—the scaling-limits framework and the other sections remain useful—but it does change the basis for the current CONDITIONAL verdict.","tokens_in":12262,"tokens_out":35131,"duration_ms":371160,"concrete_test":"For a generic angle, e.g. θ=π/8, compute the stationary expected switching rate E[λ_1+λ_2] for the ZZP and E[λ_b] for BPS under U_{θ,ε} using Eq. (1), either by exact Gaussian integration or by a short simulation at ε=10^{-4}. If the rate is O(ε^{-1/2}) rather than O(ε^{-1}), the prose after Theorem 6.1 is incorrect. Additionally, simulate the ZZP at several ε and measure the autocorrelation time of the slow coordinate Y_1; if it scales as ε^{-1/2} rather than ε^{-1}, the stated time-rescaling (or the survey's statement of it) must be corrected. This directly settles whether the §6 event-count comparison should be O(ε^{-1}) vs O(ε^{-2}) or O(ε^{-1/2}) vs O(ε^{-3/2}).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The survey's headline anisotropy conclusion—BPS requires O(ε^{-1}) switching events and ZZP O(ε^{-2}) per independent sample—rests on the assertion after Theorem 6.1 that 'the number of switching events per unit time of both processes scales as ε^{-1}'. This is inconsistent with the survey's own switching-rate formula (1) applied to U_{θ,ε}. For generic θ, ∇_x U = R(-θ) Λ_ε^{-1/2} y with y_2 ~ N(0,1), so ∂_{x_i}U contains a term sinθ ε^{-1/2} y_2 (or cosθ ε^{-1/2} y_2). Hence E[λ_i] = ½ E[|∂_{x_i}U|] = O(ε^{-1/2}), not O(ε^{-1}). The same holds for the BPS bounce rate, since E[(⟨v,∇U⟩)_+] = O(||∇U||) = O(ε^{-1/2}). Moreover, the slow-coordinate velocity V_1 = v_1 cosθ + v_2 sinθ changes at the total flip rate O(ε^{-1/2}), implying a velocity autocorrelation time ~ ε^{1/2} and a natural diffusive mixing time for y_1 of O(ε^{-1/2}), not O(ε^{-1}). If the published theorem in [14] genuinely yields an ε^{-1} rescaling, the survey needs to explain how the ε^{-1/2} event rate produces an O(ε^{-1}) timescale; as written, the complexity numbers in §6 are unsupported. The reader's adiabatic-reduction concern is related, but the sharper, more concrete issue is this factor-√ε discrepancy in the rate scaling.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey reviews scaling results for two piecewise deterministic Monte Carlo samplers: the Bouncy Particle Sampler (BPS) and the Zig-Zag Process (ZZP). It collects results on ergodicity, functional CLTs, high-dimensional Gaussian scaling limits, scaling under anisotropic/ridged targets, and scaling in the big-data regime with subsampling. The paper reports quantitative complexity comparisons: in high dimension, BPS needs O(d) time speed-up and O(d) operations per event, while ZZP needs O(d) events per unit time but O(1) cost per event; for ε-ridged targets it claims BPS needs O(ε^{-1}) switching events per independent sample and ZZP O(ε^{-2}); and in the big-data setting, control-variate subsampling gives the same convergence rate as the canonical sampler at O(1) per-event cost instead of O(n). The theorems are stated without proofs, with references to the author's prior work.","tokens_in":12722,"tokens_out":15512,"duration_ms":177908,"significance":"If the reported scaling results are correct, the survey provides a valuable synthesis of quantitative efficiency results for two important PDMC algorithms. Its strengths include a clear organizational structure, honest disclosure of an unresolved exceptional case (θ=kπ/4 with odd k in Section 6), and explicit complexity tables that make the comparisons concrete. The big-data control-variate result, if accurate, is significant for Bayesian computation. However, the Section 6 complexity conclusions are not supported by the survey's own equations as written, and the central anisotropic scaling claim must be reconciled with the displayed switching-rate formula. Because the survey's main contribution is precisely these scaling conclusions, this issue is load-bearing.","major_comments":[{"comment":"The statement that \"the number of switching events per unit time of both processes scales as ε^{-1}\" is inconsistent with Eq. (1). For U_{θ,ε}, write Y=Λ_ε^{-1/2}R_θ X; under π, Y∼N(0,I). Then ∇_xU = R(-θ)Λ_ε^{-1/2}Y. Hence E[λ_i] = 1/2 E|(R(-θ)Λ_ε^{-1/2}Y)_i| = O(ε^{-1/2}) for generic θ, and the BPS bounce rate E[(⟨v,∇U⟩)_+] is also O(ε^{-1/2}). Consequently the conclusions that \"BPS requires O(ε^{-1}) switching events\" and \"ZZP requires O(ε^{-2}) switching events\" are unsupported as stated. The ε^{-1} time rescaling in Theorem 6.1(ii) also does not follow from a velocity-flip process with flip rate O(ε^{-1/2}); the natural diffusive time scale would be O(ε^{-1/2}). Please reconcile with [14] or correct the statements.","section":"Section 6, Eq. (1) and paragraph after Theorem 6.1"},{"comment":"The limit theorem rests on the assertion that the fast coordinate y² \"quickly converges to its stationary distribution\" before the slow coordinate moves. No hypotheses or proof sketch for this adiabatic/averaging step are given; the survey defers to [14]. Since this two-timescale separation is load-bearing for the OU limit (at speed Ω_θ) and for the BPS ODE-with-jumps limit, the survey should state the precise condition from [14] or explicitly mark the reduction as heuristic. The disclosed exceptional angles θ=kπ/4 (odd k) show that the separation is not uniform, so more care is needed.","section":"Section 6, paragraph before Theorem 6.1"}],"minor_comments":[{"comment":"\"ZZO\" appears to be a typo for \"ZZ\" or \"ZZP\".","section":"Table 3"},{"comment":"The values λ_r=1.424 and the refreshment-jump ratio 0.7812 are quoted without the promised \"semi-analytical expression.\" Please include the expression or a precise pointer to [13]. The further claim that the ratio is a \"suitable criterion\" for general tuning needs justification or qualification.","section":"Section 5, Figure 1"},{"comment":"The figure lacks axis labels and units, and the \"explicit (but complicated)\" expression for Ω_θ is not given. Either include the formula or provide a precise reference to [14].","section":"Section 6, Figure 2"},{"comment":"There is a typographical backslash in \"\\nabla U(x)\"; it should read \"\\nabla U(x)\" without the stray backslash.","section":"Section 7, first paragraph"},{"comment":"The caveat for θ=kπ/4 with odd k appears only after the main complexity conclusions. Please restate the O(ε^{-2}) claim as valid for generic θ, and give details of the numerical experiment that reports \"up to an order of magnitude faster\" convergence for the exceptional angles.","section":"Section 6, Theorem 6.1 and following paragraph"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the internal inconsistency in Section 6 between the switching-rate formula (1) and the claimed ε^{-1} scaling/time rescaling. I did not consult [14] directly; my check is based only on the manuscript's own definitions. If [14] indeed proves an ε^{-1} rescaling, the survey must explain how this arises from an O(ε^{-1/2}) event rate; otherwise the complexity tables in Section 6 need correction. The rest of the survey is clearly written and the organization is good."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the survey is a clean, honest restatement of the author's scaling results for BPS and ZZP, and it will be handy for people entering the area. But Section 6 has a problem: the claim that the per-time switching rate is O(ε^{-1}) is inconsistent with the paper's own rate definition, and the complexity table built on it needs scrutiny.\n\nWhat's good: the paper assembles the FCLT, high-dimensional Gaussian scaling, anisotropy, and subsampling results in one place, with crisp theorem statements and honest disclosure of open cases (the θ=kπ/4 gaps, transients deferred to [1]). The tables comparing BPS, ZZP, RWM, and MALA are useful. For a proceedings survey, that's solid work.\n\nThe soft spots. First, the 0.7812 refreshment-jump ratio is presented as a general BPS tuning rule. It is an observation from the Gaussian analysis, and generalizing it to arbitrary targets is a leap. The author should label it as a heuristic, not a criterion. Minor.\n\nSecond, and more seriously, §6's complexity conclusions rest on 'the number of switching events per unit time of both processes scales as ε^{-1}'. But applying the survey's own switching rate (1) to U_{θ,ε} for generic θ, each gradient component contains a term of size ε^{-1/2} (e.g., sinθ ε^{-1/2} y_2). With y_2 at stationarity, the average rate is O(ε^{-1/2}), not O(ε^{-1}). So the O(ε^{-1}) vs O(ε^{-2}) event counts look off by a factor √ε. If the published theorem in [14] genuinely yields an ε^{-1} timescale, the survey needs to explain the mechanism; as written, the derivation is missing. The adiabatic averaging assumption is also just stated, not justified.\n\nIf the failure is real, the table's numbers are wrong but the conceptual message — BPS is cheaper than ZZP in this regime — may survive. The survey needs a check against [14] before publication.\n\nWho it's for: PDMC practitioners and people who want a single entry point to the scaling literature. It deserves a serious referee; the referee should focus on §6 and on verifying the tuning heuristic against the original papers.","headline":"Useful survey of PDMC scaling results, but the anisotropy complexity claims don't match the survey's own rate formula and should be checked against the source theorem.","tokens_in":13210,"tokens_out":15847,"would_cite":true,"duration_ms":164349,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J25","65C05","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey uses scaling-limit theorems to turn the question of how fast BPS and Zig-Zag samplers mix into concrete complexity numbers: O(d²) versus O(d) per sample on Gaussian targets, and O(ε⁻¹) versus O(ε⁻²) events on ridges.","keywords":["piecewise deterministic Monte Carlo","Bouncy Particle Sampler","Zig-Zag process","scaling limits","high-dimensional scaling","anisotropic targets","subsampling","control variates"],"falsifier":"Simulate the Zig-Zag process on rotated Gaussian ridges U(x) = ½ xᵀ R(−θ) Λ⁻¹_ε R(θ) x for ε from 10⁻² down to 10⁻⁶ and several θ, recording the autocorrelation time of the slow coordinate; if ε² times the number of switching events to decorrelate does not converge to a finite, angle-dependent constant, the OU limit of the ridge theorem, and the averaging reduction behind it, is falsified.","tokens_in":12117,"feed_emoji":"🎲","tokens_out":5479,"duration_ms":58586,"temperature":0.7,"pith_summary":"This survey collects quantitative scaling-limit results for two continuous-time samplers, the Bouncy Particle Sampler and the Zig-Zag process. Its central message is that the two algorithms have complementary strengths: on isotropic high-dimensional Gaussians the Zig-Zag process costs O(d) per independent sample while the Bouncy Particle Sampler costs O(d²); on narrow ridge targets the ranking flips, with BPS needing only O(ε⁻¹) switching events compared with Zig-Zag's O(ε⁻²). The survey also covers a functional central limit theorem, exponential ergodicity conditions, and a big-data analysis showing that a control-variate subsampled Zig-Zag process reproduces the canonical sampler's convergence at O(1) per-event cost rather than O(n). A sympathetic reader should take away that PDMC samplers now come with precise dimension, anisotropy, and dataset-size complexity statements, not just heuristic speed comparisons.","feed_headline":"On Gaussians, Zig-Zag beats BPS: O(d) vs O(d²) per sample","feed_subtitle":"But ridge targets flip it: BPS needs only O(ε⁻¹) events to Zig-Zag's O(ε⁻²), and control variates make big-data ZZP O(1).","key_machinery":"The unifying mechanism is scaling-limit analysis: instead of tracking an exact finite-dimensional sampler, the theorems pass to a limit (d→∞, ε↓0, n→∞) and identify the rescaled process with a tractable limit—Ornstein-Uhlenbeck diffusions for coordinates and log-density, an ODE with jumps for ridge BPS, or fluid-limit ODEs for transient big-data ZZP. The load-bearing device in the ridge analysis is a two-timescale averaging reduction: the fast ridge coordinate is asserted to settle to its stationary distribution before the slow coordinate moves, reducing the fast dynamics to an averaged drift and allowing an OU/ODE limit for the slow coordinate. In the big-data analysis, Poisson thinning wit","core_discovery":"The paper's own claims, stated as results: for d-dimensional standard Gaussians, the BPS first-coordinate process converges after time acceleration by d to an Ornstein-Uhlenbeck process and its log-density process likewise accelerates by d, giving O(d²) operations per nearly independent sample; the ZZP's coordinates already move at unit speed, giving O(d) (or O(d²) without a coordinate-exploiting implementation). For anisotropic Gaussian targets with ridge width ε, the BPS reaches its limit without time rescaling and costs O(ε⁻¹) events, while the ZZP needs acceleration ε⁻¹ and costs O(ε⁻²) events, with an explicit angle-dependent speed except at special angles. In the big-data regime, the Z","pith_inferences":["If the two-timescale separation is not universal, the ridge ranking may be a Gaussian artifact; a natural extension is to test the O(ε⁻¹) versus O(ε⁻²) event counts on non-Gaussian ridge targets (e.g., curved or heavy-tailed ridges), where the fast coordinate may not equilibrate before the slow coordinate moves.","The high-dimensional Zig-Zag O(d) result is implementation-coupled; for correlated targets with naive coordinate-wise implementation, the per-sample cost may scale with the number of active coordinates, suggesting that preconditioned or adaptive versions deserve their own scaling analysis.","The control-variate result suggests a broader recipe: any unbiased gradient estimator whose variance vanishes near the mode inherits the canonical sampler's convergence; this could be tested with batched or sketch-based gradient estimators beyond single-datum subsampling.","The 0.7812 refreshment-fraction rule, derived for Gaussians, is plausibly a robust BPS tuning heuristic for other targets, since it is expressed in scale-invariant event fractions rather than in absolute rates; verifying this on multimodal targets would be a cheap experiment."],"forward_implications":["On high-dimensional Gaussian targets, an optimized Zig-Zag implementation needs O(d) computations per approximately independent sample, versus O(d²) for BPS; plain implementations erase Zig-Zag's advantage.","On ε-narrow ridge targets, BPS requires O(ε⁻¹) switching events per independent sample versus O(ε⁻²) for Zig-Zag, matching the random-walk Metropolis complexity; so BPS, not ZZP, is the PDMC choice for strongly anisotropic posteriors.","For Zig-Zag with control-variate subsampling in the big-data regime, the stationary scaling is the same as the canonical full-data Zig-Zag, with per-event cost O(1) instead of O(n): a factor-O(n) speedup at no loss of mixing.","The refreshment rate of BPS has a scale-invariant tuning criterion: the expected fraction of refreshment jumps that maximizes speed is 0.7812 (at the numerically optimal refreshment rate for Gaussian targets).","A functional central limit theorem holds under exponential ergodicity conditions, so ergodic averages of these samplers are asymptotically normal, and the asymptotic variance is accessible through the Poisson equation (explicitly for one-dimensional Zig-Zag)."],"fun_headline_variants":["Zig-Zag beats BPS on Gaussians, but ridge targets flip the race","Ridge targets reverse PDMC scaling: BPS O(ε⁻¹), Zig-Zag O(ε⁻²)","Gaussians: Zig-Zag O(d) vs BPS O(d²); ridges flip it","Zig-Zag scales O(d) on Gaussians; BPS O(ε⁻¹) on ridges","Piecewise Monte Carlo scaling: Zig-Zag wins on Gaussians, BPS on ridges"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The ridge-scaling results rely on the assumption that the fast ridge coordinate reaches stationarity before the slow coordinate changes, so the slow dynamics can be averaged over that fast equilibrium; if this separation fails, the stated Ornstein-Uhlenbeck and ODE limits need correction.","fun_headline_variants_meta":{"raw":{"variants":["Zig-Zag beats BPS on Gaussians, but ridge targets flip the race","Ridge targets reverse PDMC scaling: BPS O(ε⁻¹), Zig-Zag O(ε⁻²)","Gaussians: Zig-Zag O(d) vs BPS O(d²); ridges flip it","Zig-Zag scales O(d) on Gaussians; BPS O(ε⁻¹) on ridges","Piecewise Monte Carlo scaling: Zig-Zag wins on Gaussians, BPS on ridges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001319,"raw_usage":{"total_tokens":5176,"prompt_tokens":681,"completion_tokens":4495,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":4368}},"tokens_in":425,"tokens_out":4495,"duration_ms":34126,"temperature":1.0,"reasoning_tokens":4368,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:40:48.793341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the Zig-Zag process on rotated Gaussian ridges U(x) = ½ xᵀ R(−θ) Λ⁻¹_ε R(θ) x for ε from 10⁻² down to 10⁻⁶ and several θ, recording the autocorrelation time of the slow coordinate; if ε² times the number of switching events to decorrelate does not converge to a finite, angle-dependent constant, the OU limit of the ridge theorem, and the averaging reduction behind it, is falsified.","supporting_citations":[],"review_version":1}