{"id":"47556a05-cebe-4995-b939-f24e28dec4c8","arxiv_id":"2411.18218","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A general radial-update recipe that provably makes MCMC samplers converge exponentially on non-compact spaces, with a near-optimal step size scaling as the inverse square root of the dimension.","lead":"This paper shows how to add a radial update to almost any Monte Carlo sampler so that sampling from distributions on unbounded spaces is guaranteed to converge exponentially fast. It gives a simple recipe, including an optimal step size, and demonstrates enormous speedups on heavy-tailed distributions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's proof does not cover the paper's polynomial-potential recipe: with r=e^z the effective potential grows exponentially only as z→+∞, not as z→−∞, so Lemma 3.12's symmetric condition (21) is not met.","rationale":"The paper's central contribution is a universal recipe (Theorem 3) for exponential convergence via radial updates, with the polynomial case as its flagship application (Algorithm 1, Figures 1, 4, 5). The proof of Theorem 3 hinges entirely on Lemma 3.12, which assumes symmetric exponential growth of the effective potential. For r=e^z, the effective potential is V_eff=c e^{a z}-d z; it is exponential as z→∞ but only linear as z→−∞. Thus the lemma is not applicable on the negative tail, and the proof's assertion that the update is 'asymptotically equivalent' to a condition-(21) substitution ignores this tail. This is a genuine proof gap: the theorem as stated does not cover the paper's headline recipe. However, the gap is likely repairable: the negative tail in z corresponds to r→0, a compact subset of the radial manifold, and Theorem 2 only needs the SGDC off a compact set. The conclusion is therefore probably true, but the paper needs a compact-tail lemma or an explicit one-sided condition. This matches the reader's weakest assumption exactly. Since the fix does not change the practical claims or the need for a revised proof, the conditional verdict stands.","tokens_in":24148,"tokens_out":23652,"duration_ms":213150,"concrete_test":"Analytic check: replace Eq. (21) in Theorem 3 by the one-sided condition V_eff(z,θ)=c(θ)e^{a(θ)z+o(z)} as z→∞ together with f(z)→0 as z→−∞; adapt Lemma B.11 instead of Lemma 3.12 to obtain SGDC for V_eff on z>R, then verify that the SGDC transfers to V on X\\X_0 with X_0={x:||x||≤R_0} by bounding E[V(z')]≤α V(z)+K using V=V_eff+d z and d z=o(V_eff) for large z. If this proof goes through, Theorem 3 holds for the polynomial recipe as intended.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3 is proven by reducing to Lemma 3.12, which establishes the SGDC for potentials of the form V(z)=c e^{a|z|+o(|z|)}. The paper's central polynomial-potential recipe (Algorithm 1, Corollary 3.14) uses r=e^z, giving V_eff(z)=c e^{a z}-d z. This is c e^{a z+o(z)} as z→+∞ but only d|z|+O(1) as z→−∞, so it does not satisfy Eq. (21) on the negative tail. The proof of Corollary 3.14 asserts the update is 'asymptotically equivalent' to a substitution satisfying Eq. (21), but that equivalence holds only for z→+∞; the z→−∞ tail is a compact region r→0 in the original manifold and can plausibly be absorbed into the compact set X0 of Theorem 2, but that argument is not given. As stated, the theorem does not cover the paper's main example, and the proof of the general recipe needs a one-sided condition or an explicit compact-tail lemma.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops radial updates for MCMC sampling on non-compact radial Riemannian manifolds. It proposes to substitute the radius r=f(z) so that the effective potential V_eff(z,θ)=V(f(z),θ)-ln|det J(z,θ)| grows exponentially in |z|, then apply a symmetric additive update to z. Theorem 3 claims exponential convergence of the combined algorithm under this condition; Corollary 3.14 specializes this to log-normal updates for exponential, polynomial, and logarithmic potentials; Theorem 4 classifies convergence rates for slower or badly chosen radial updates; Theorem 5 derives a dimension scaling σ≈√(2/(ad)) for polynomial potentials. Numerical examples show sampling of very heavy-tailed distributions and autocorrelation measurements supporting the scaling claim.","tokens_in":24333,"tokens_out":21016,"duration_ms":201697,"significance":"If the central theorem is correct, the paper gives an actionable and essentially parameter-free recipe for exponential convergence of MCMC on non-compact spaces, generalizing the Kennedy-Yu construction [2] and removing most tuning from the radial-update step. The numerical demonstrations are convincing: the heavy-tailed examples reach radii as large as 10^200, and the autocorrelation data confirm the 1/√d scaling to within the expected O(1) factor. The paper also ships public code and data, which is a strength. However, the proof of the main theorem has a gap with respect to the paper's own Algorithm 1: condition (21) is two-sided, while the polynomial recipe only satisfies it on the outward tail.","major_comments":[{"comment":"Theorem 3's condition (21) requires V_eff(z,θ)=c(θ)e^{a(θ)|z|+o(|z|)} on both tails. For the polynomial-potential recipe in Corollary 3.14 and Algorithm 1, the substitution r=e^z gives V_eff(z)=c e^{a z}-d z, which grows exponentially for z→+∞ but only linearly for z→−∞. The Appendix A proof of Corollary 3.14 dismisses the linear term as 'negligible in the e^{o(z)}-sense,' which is valid only on the positive tail. Since z→−∞ corresponds to r→0, a compact region, the conclusion can be repaired by absorbing this tail into the compact set X_0 of Theorem 2, but the manuscript does not supply that argument. As stated, Theorem 3 does not cover the paper's main algorithm; please add a one-sided condition with a compact-tail lemma, or prove that the log-normal update is equivalent to a substitution such as r=exp(z-e^{-z}) for which (21) holds on both tails.","section":"§3.1, Eq. (21) and Corollary 3.14"},{"comment":"The proof of Lemma 3.12 assumes that for large z the acceptance probability satisfies A(z,z+γ)=1 for all γ≤0. This is false when z+γ<−z (more precisely γ<−2z), because then |z+γ|>z and V(z+γ)>V(z), so the proposal is not automatically accepted. For heavy-tailed symmetric additive kernels this region has nonzero probability, and the displayed calculation replaces the true (small) accepted contribution by an expression that is not the correct V(z+γ)/V(z) ratio. The lemma is probably true and the missing contribution is bounded by the Metropolis acceptance factor, but a complete proof needs a separate treatment of the γ<−2z region; as written, this is a gap in a central statement.","section":"§3.1, Lemma 3.12 proof, Eq. (8)"},{"comment":"Lemma 3.12 gives the SGDC for the effective potential V_eff in the auxiliary variable z, whereas Theorem 2 requires the SGDC for the original potential V on X\\X_0. The proof of Theorem 3 moves directly from one to the other. This transfer is not automatic because V and V_eff differ by the log-Jacobian term; one must show that V_eff and V are comparable outside a compact set (for the substitutions used here, V_eff/V→1 on the outward tail). Without this comparability, the reduction to Theorem 2 is incomplete; adding this argument would also clarify which tails may be placed into X_0.","section":"§3.1, proof of Theorem 3"},{"comment":"The claim that the HMC without radial updates converges with the rates listed in Corollary 3.15 is not established by the given proof. The short-trajectory ('Langevin regime') argument covers only a limiting regime, and the step to arbitrary trajectory lengths rests on the assertion that 'the drift that guarantees convergence can only increase and never decrease,' which is not proved and does not follow from symplecticity alone. Since this corollary is advertised as a generalization, either supply a rigorous argument or explicitly label it as a heuristic/conjecture.","section":"Appendix B, proof of Corollary 3.15"}],"minor_comments":[{"comment":"The Landau symbols are defined for the one-sided limit x_0=(∞,θ_0), while condition (21) uses |z|→∞; the two-sided convention should be stated explicitly where o(|z|) is used.","section":"Table 1 and Eq. (21)"},{"comment":"The line z_i ← f^{-1}(||x_i||) is undefined at ||x_i||=0; since this is a measure-zero set, a short sentence on handling the origin would remove ambiguity.","section":"Algorithm 2"},{"comment":"The theorem is phrased as giving the 'optimal' standard deviation, but the proof relies on the heuristic Estimate 3.17 and the numerical fit in Fig. 5 gives σ_min=1.528(7)/√d rather than 1/√d; consider stating the result as near-optimal or asymptotically optimal up to an O(1) factor.","section":"Theorem 5"}],"recommendation":"major_revision","confidential_remarks":"The one-sided-tail gap in Theorem 3 relative to Algorithm 1 is the main load-bearing issue, and it appears repairable with a compact-tail argument or by switching the default substitution to r=exp(z-e^{-z}). I do not see grounds for rejection: the numerical work is credible and the conceptual recipe is likely correct. The proof of Corollary 3.15 is weaker than the surrounding theorems and should be explicitly downgraded if not proved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ostmeyer's radial-update paper is worth reading: it takes the Kennedy-Yu result, which was HMC-specific, and turns it into a near-universal recipe for any MCMC method on non-compact spaces. The core idea is simple: substitute r=f(z) so that the effective potential (potential minus log Jacobian) grows exponentially in |z|, then any symmetric additive update in z gives exponential convergence. That is a genuinely useful generalization, and the paper proves it carefully. The classification of convergence for badly chosen updates (theorem 4) and the dimension-scaling law for sigma are also new. The numerics are convincing: sampling from a distribution with mass out at r ~ 10^200 is a nice demonstration, and the code and data are public.\n\nThe main soft spot is a real mismatch between theorem 3 as stated and the paper's own flagship example. Theorem 3 requires Veff(z,theta)=c(theta)e^{a(theta)|z|+o(|z|)}, symmetric in |z|. But the polynomial-potential recipe uses r=e^z, giving Veff(z)=c e^{a z}-d z, which grows exponentially only as z→+infinity. As z→-infinity it grows linearly. The proof of corollary 3.14 says the linear term is negligible in the e^{o(z)} sense, but that's only true on the positive tail. The z→-infinity tail corresponds to r→0, a compact region that can be absorbed into the compact set X0 of theorem 2, so the conclusion is probably right. But as written, the theorem does not cover the main example. This needs a one-sided condition or an explicit compact-tail lemma. It is a moderate, fixable gap, not a fatal one.\n\nThe sigma scaling result is honest: the sqrt(2) prefactor is from a heuristic estimate, and their own fits find sigma_min ≈ 1.53/sqrt(d), not 1/sqrt(d). They say the constant can vary by O(1), which is fair. The scaling with 1/sqrt(d) is confirmed and that's the useful part.\n\nThis paper deserves a serious referee. The contribution is real, the proofs are mostly careful, and the practical value is high. A referee should push for a corrected statement of theorem 3/corollary 3.14 and a slightly more careful claim about the sigma prefactor. I'd accept it after minor revision.","headline":"Generalizes Kennedy-Yu radial updates into a practical universal recipe, but theorem 3's stated condition doesn't cover its own polynomial example; a fixable gap, still worth peer review.","tokens_in":24859,"tokens_out":2791,"would_cite":true,"duration_ms":24849,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J22","65C05","65C40"],"pacs":["02.70.Uu","05.10.Ln"],"model":"deepseek-v4-flash","headline":"Radial updates guarantee exponential convergence for MCMC on non-compact spaces.","keywords":["radial updates","Monte Carlo convergence","geometric ergodicity","effective potential","heavy-tailed distributions","hybrid Monte Carlo","autocorrelation time"],"falsifier":"Take the polynomial target $V(r) = r$ in $d = 1$ and implement algorithm 1 with $r = e^z$ and a normal additive update in $z$, then measure the distance to stationarity over a long simulation; if the decay is not exponential, the theorem's guarantee fails for this substitution. A more direct check is to evaluate the drift inequality $(PV)(z) \\le \\alpha V(z) + K$ for the effective potential $V_{\\mathrm{eff}}(z) = e^z - z$ on the negative tail, where the potential is linear in $|z|$ rather than exponential; the lemma's proof does not cover this region, so that numerical evaluation decides whether the missing compact-tail argument is essential.","tokens_in":23906,"feed_emoji":"🎲","tokens_out":12822,"duration_ms":98743,"temperature":0.7,"pith_summary":"Monte Carlo sampling on unbounded spaces often fails because local updates diffuse through flat, low-probability regions and take astronomically long to reach the bulk of the distribution. This paper proves that adding a radial update — an update of the distance from the origin, performed in a reparameterised coordinate z with r = f(z) — turns this diffusive behaviour into guaranteed exponential convergence. The single requirement is that the effective potential, $V_{\\mathrm{eff}}(z, \\theta) = V(f(z), \\theta) - \\ln|\\det J|$, grows exponentially in $|z|$; then any non-trivial symmetric additive step in $z$ (a normal increment, for example) produces the drift needed for exponential convergence. The paper also gives the near-optimal step size $\\sigma \\simeq \\sqrt{2/(ad)}$ for polynomial potentials and demonstrates speedups of many orders of magnitude on heavy-tailed targets.","feed_headline":"Radial updates make Monte Carlo sampling exponentially convergent","feed_subtitle":"Rewriting the radius so the effective potential grows exponentially makes unbounded-space sampling provably fast.","key_machinery":"The load-bearing object is the effective potential $V_{\\mathrm{eff}}(z, \\theta) := V(f(z), \\theta) - \\ln|\\det J(z, \\theta)|$, where $f: \\mathbb{R} \\to [0, \\infty)$ is a diffeomorphism reparameterising the radius and $J$ is its Jacobian. This substitution converts multiplicative steps in the radius (which become vanishingly small in flat regions) into additive steps in $z$, where a local step always changes the effective potential by an exponential factor. The proof then reduces to lemma 3.12, which shows that a non-trivial symmetric additive kernel with Metropolis–Hastings acceptance satisfies the strong geometric drift condition for any potential of the form $c e^{a|z|+o(|z|)}$. That lemma is what turns the coordinate substitution into a provable exponential convergence rate.","core_discovery":"The central claim is theorem 3: on a radial Riemannian manifold, any algorithm with stationary distribution $p(x) \\propto e^{-V(x)}$ that satisfies the compact Doeblin condition and the weak geometric drift condition converges exponentially when combined with a radial update of the form $z \\mapsto z + \\gamma$, provided the substitution $r = f(z)$ makes the effective potential $V_{\\mathrm{eff}}(z, \\theta) = c(\\theta)\\, e^{a(\\theta)|z| + o(|z|)}$. The mechanism is lemma 3.12: for a one-dimensional potential of exactly exponential form $V(z) = c e^{a|z|+o(|z|)}$, every non-trivial symmetric additive kernel with Metropolis–Hastings acceptance satisfies $(PV)(z) \\le \\alpha V(z) + K$ with $\\alpha < 1$, which is the strong geometric drift condition. The radial update therefore repairs the missing drift ingredient that prevents convergence on non-compact spaces, and for polynomial potentials the paper's recipe becomes $r \\mapsto r e^{\\gamma}$ with $\\gamma \\sim \\mathcal{N}(0, \\sigma^2)$ and $\\sigma = \\sqrt{2/(ad)}$.","pith_inferences":["Because the radial update is independent of the angular sampler, the same substitution recipe should transfer to any future MCMC method that satisfies the compact Doeblin and weak drift conditions; the paper proves the theorem generally but demonstrates tuning only on one-dimensional radial problems, so high-dimensional non-radially-symmetric targets are a natural stress test.","The proof's treatment of the polynomial case leaves one tail of the effective potential growing only linearly, so in practice the substitution needs to tame the potential only outside a compact region; this suggests simpler, easier-to-invert substitutions could be used without losing exponential convergence, although the stated theorem requires the stronger exponential-tail condition.","The radial update's documented ability to jump over potential barriers (the paper cites its use to cure an ergodicity problem in a Hubbard-model simulation) implies the method may also help with metastability in compact problems where the radial direction is bounded but barriers are large; that use is not proven in this paper."],"forward_implications":["Any MCMC algorithm that samples well on compact spaces can be extended to non-compact spaces with guaranteed exponential convergence by combining it with a radial update built from a substitution that makes the effective potential grow exponentially.","For asymptotically polynomial potentials $V(x) \\simeq c|x|^a$ in $d$ dimensions, the radial update $r \\mapsto r e^{\\gamma}$ with $\\gamma \\sim \\mathcal{N}(0, \\sigma^2)$ and $\\sigma = \\sqrt{2/(ad)}$ gives near-optimal autocorrelation; the $\\sigma \\propto 1/\\sqrt{d}$ scaling is universal.","Heavy-tailed targets such as $p(r) \\propto 1/(1+r^{1.01})$ become tractable: the sampler reliably visits radii as large as $10^{200}$, whereas a conventional additive updater would need times $\\gg 10^{400}$ to reach such regions.","Even a badly chosen radial update still leads to convergence; theorem 4 classifies the approach as exponential, polynomial, or diffusive depending on whether $\\log V$ grows at least linearly in $r$, $V$ grows at least linearly in $r$, or merely $e^{-V}$ is normalisable.","The hybrid Monte Carlo algorithm alone is guaranteed to converge on radial Riemannian manifolds, with the same classification of approach rates, but the radial update accelerates it by many orders of magnitude."],"supporting_citations":[{"why":"Proves HMC converges on compact manifolds and that a radial update satisfying the SGDC yields exponential convergence on non-compact manifolds; the paper's theorem 2 builds directly on this.","marker":"[2]"},{"why":"Supplies the Harris ergodic theorem used as theorem 1, linking the compact Doeblin and strong drift conditions to exponential convergence.","marker":"[14]"},{"why":"Introduces the Metropolis accept/reject step that every radial update uses to maintain detailed balance.","marker":"[15]"},{"why":"Generalises the acceptance step to asymmetric proposals; the paper's symmetric additive kernel simplifies to the Metropolis form.","marker":"[16]"}],"fun_headline_variants":["Radial updates fix slow Monte Carlo on unbounded spaces","Exponential convergence for MCMC via radial substitution","Radial trick makes MCMC converge exponentially fast","Radial updates: orders-of-magnitude speedup for heavy-tailed sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction stands on the claim that a non-trivial symmetric additive update in a variable whose effective potential grows exponentially is guaranteed to pull the chain back toward the bulk at a constant rate; in the main polynomial case the effective potential grows only linearly on one side, and the proof covers that side by asserting it is a harmless compact region rather than by direct verification.","fun_headline_variants_meta":{"raw":{"variants":["Radial updates fix slow Monte Carlo on unbounded spaces","Exponential convergence for MCMC via radial substitution","Radial trick makes MCMC converge exponentially fast","Radial updates: orders-of-magnitude speedup for heavy-tailed sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1469,"prompt_tokens":994,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":610,"tokens_out":475,"duration_ms":4416,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:24:34.644933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the polynomial target $V(r) = r$ in $d = 1$ and implement algorithm 1 with $r = e^z$ and a normal additive update in $z$, then measure the distance to stationarity over a long simulation; if the decay is not exponential, the theorem's guarantee fails for this substitution. A more direct check is to evaluate the drift inequality $(PV)(z) \\le \\alpha V(z) + K$ for the effective potential $V_{\\mathrm{eff}}(z) = e^z - z$ on the negative tail, where the potential is linear in $|z|$ rather than exponential; the lemma's proof does not cover this region, so that numerical evaluation decides whether the missing compact-tail argument is essential.","supporting_citations":[],"review_version":1}