{"id":"c0e621cb-62ae-4a6a-b908-d077586cf368","arxiv_id":"2608.04284","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DCS pulls heavy-tailed targets back to a Euclidean ball and proves uniform ergodicity for any polynomial tail, with O(d^3) Hit-and-Run mixing for Student-t targets.","lead":"This paper introduces the Diffeomorphic Contraction Sampler (DCS), an MCMC method that squeezes heavy-tailed distributions on R^d into a ball, samples there, then stretches the results back. It proves uniform ergodicity for any polynomial tail and gives O(d^3) mixing bounds in special cases, while reporting large numerical gains over NUTS and stereographic samplers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fast-mixing guarantees require an oracle transport; the implemented VI-tuned Möbius automorphism cannot make a multimodal pull-back log-concave, so Theorem 3.6 does not cover the practical algorithm on multimodal or funnel targets.","rationale":"I read the paper in good faith. Theorem 3.1's uniform ergodicity proof is internally consistent: Lemma A.4 gives two-sided bounds on pi_B, and Lemma A.5 constructs explicit minorizations for the listed kernels. Example B.1 correctly verifies (Conv), (Iso), and (Cold) for Student-t under the oracle T_omega, so the O(d^3) bounds are valid for that idealized chain. The load-bearing gap is exactly the bridge from the oracle automorphism to the implemented algorithm: the VI step in Section 2.3.1 and Appendix D optimizes only the Moebius parameter delta, but the non-asymptotic theory needs a transport that makes pi_B log-concave and isotropic. For multimodal targets, no Moebius map can achieve log-concavity because it is a homeomorphism and preserves the number of modes. Thus the practical algorithm's fast-mixing guarantee is unsupported in the multimodal regime, which is the regime emphasized in the numerical sections. The reader's weakest_assumption identifies the same gap, and the requested code/artifacts would help but would not close the analytical gap unless the VI family is replaced or the theory is extended to the actual tuned class. The verdict of CONDITIONAL remains appropriate; no change is needed.","tokens_in":32751,"tokens_out":14265,"duration_ms":128544,"concrete_test":"Run Algorithm (DCS) with the Appendix D VI/Moebius tuning on the two-component Student-t mixture of Example B.2 in d=10, and evaluate the minimum eigenvalue of nabla^2 U_B at the midpoint of the segment joining the two modes; a negative eigenvalue confirms (Conv) fails, so the O(d^3) bound of Theorem 3.6 cannot be invoked for the VI-tuned algorithm. An equivalent analytical check: because each Moebius map is a homeomorphism of B(1), it preserves the number of superlevel-set components, so the pull-back of a bimodal target is bimodal and therefore not log-concave.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central fast-mixing claim (Theorem 3.6, Example B.1) is proved for a specific oracle automorphism T_omega that makes pi_B log-concave and isotropic; for Student-t this T_omega is given in closed form, while for multimodal mixtures Example B.2 invokes a non-constructive Knothe-Rosenblatt transport (Remark B.2). The actual tuning in Algorithm (VI) of Appendix D optimizes only the Moebius parameter delta in (8), a family of ball diffeomorphisms with a single hyperbolic kick. Since every Moebius map is a homeomorphism of B(1), it preserves the number and topology of the modes: the pull-back of a two-mode target under any Moebius map remains two-mode. Assumption (Conv) of Theorem 3.6 requires U_B = -log pi_B to be convex, which forces pi_B to be log-concave and in particular unimodal. Hence for any genuinely multimodal target the implemented Moebius-tuned DCS cannot satisfy (Conv), and the O(d^3) warm-start and cold-start bounds of Theorem 3.6 are not applicable. Uniform ergodicity (Theorem 3.1) is unaffected because it only needs the mild tail condition (A-beta). The numerical sections on PosteriorDB funnel targets (Section 4.2) are explicitly outside the full-space theory, yet the abstract and contribution (II) present the method as fast-mixing for heavy-tailed posteriors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Diffeomorphic Contraction Sampler (DCS), which maps a target on R^d to a density on a Euclidean ball via a radially contracting diffeomorphism composed with an automorphism of the ball, runs a convex-body sampler (Ball Walk, Hit-and-Run, RWM, Dikin walk, or projected Langevin) on the pull-back density, and pushes the samples forward. Theorem 3.1 establishes uniform ergodicity for every continuous target satisfying limsup_{|x|→∞} π(x)|x|^{d+β} < ∞, for the HnR, BW, RWM, and covariance-floored Dikin kernels. Theorem 3.6 imports non-asymptotic mixing bounds under convexity/log-concavity, isotropy, and warm/cold start conditions on the pull-back density. Example B.1 verifies these conditions for Student-t targets and yields an O(d^3) cold-start mixing bound for DCS-HnR. Numerical experiments compare DCS favourably with NUTS and with stereographic samplers on skewed Student-t, PosteriorDB funnel-type, and infinite-moment targets.","tokens_in":32975,"tokens_out":20100,"duration_ms":179777,"significance":"The uniform ergodicity theorem is a clean and useful result: Lemma A.4 bounds the pull-back density under the moment condition, Lemma A.5 gives m-step minorization from an arbitrary starting point, and Theorem 3.1 follows by a standard minorization argument. The total-variation invariance observation (Proposition A.1, Corollary A.3) correctly reduces the analysis of the pushed-forward chain to sampling on the ball and is a strong structural simplification. The explicit Student-t verification of Assumptions (Conv), (Iso), and (Cold) in Example B.1, with the resulting O(d^3) cold-start Hit-and-Run bound, is a concrete and useful advance over the exponential-in-d RWM bound. The main weakness is that the fast-mixing theory for multimodal targets is existential rather than constructive, and the practical Möbius-tuned algorithm is not covered by Theorem 3.6 for multimodal or funnel targets; the abstract and contribution list should be reworded so that this scope limitation is explicit.","major_comments":[{"comment":"The fast-mixing claim for multimodal targets is not supported by a constructive transport. Example B.2 invokes a Knothe–Rosenblatt diffeomorphism whose existence is guaranteed by [6] but for which neither an explicit form nor closed-ball boundary regularity is supplied; Remark B.2 and Open Problem (IIb) concede that only existence is known. The practical tuning procedure in Algorithm (VI) optimizes only (μ, R, δ) within the Möbius family (8), and a Möbius automorphism is a homeomorphism of B(1) that cannot make a genuinely multimodal pull-back density log-concave, because Assumption (Conv) forces log-concavity and hence unimodality. Therefore Theorem 3.6 does not cover the implemented VI-tuned algorithm on multimodal or funnel targets, and the numerical results in Section 4.2 are, as the text admits, outside the full-space theory. The abstract and contribution (II) should attribute the fast non-asymptotic guarantee specifically to the Student-t construction in Example B.1 and, conditionally, to target classes for which a suitable automorphism is actually available, rather than to general heavy-tailed or multimodal targets.","section":"§3.2, Example B.2, Remark B.2, §5(IIb)"},{"comment":"The statement of the DCS-LMC bound is ill-typed: the chain X_N lives on R^d and targets ν, but the displayed bound compares L_x(X_N) with ν_B, which is a measure on B(R). The proof also refers to a chain 'started at 0', while the statement takes X_0 = RC_total(0). This should be corrected by writing either the bound for the transformed chain on B(R) or the push-forward bound against ν. The error is local and easy to fix, but it currently makes the formal statement of part (d) unreadable.","section":"§3.2, Theorem 3.6(d)"}],"minor_comments":[{"comment":"Remark 3.11 states that choosing β too large results in 'failure of uniform ergodicity'; Theorem 3.1 gives only a sufficient condition for uniform ergodicity, so the remark should say that condition (A-β) fails rather than asserting non-uniform ergodicity.","section":"§3.2.1, Remark 3.11"},{"comment":"The numerical experiments in Section 4.2 implement Hit-and-Run using a univariate slice-sampling transition, while Lemma A.5 and Theorem 3.1 concern the exact-conditional HnR kernel. Appendix C notes this, but the main text should explicitly state that the implemented slice-based kernel is not covered by the uniform-ergodicity or mixing theorems.","section":"Appendix C.2 and §4.2"},{"comment":"The variational objective minimizes reverse KL with respect to the uniform measure on B(1); it does not directly promote Assumptions (Conv), (Iso), or (Cold). A sentence stating that the tuned Möbius family can satisfy these assumptions for unimodal approximately isotropic targets but not in general for multimodal targets would help align the practical tuning step with the theoretical assumptions.","section":"§2.3.1, Algorithm (VI)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely publishable after revision. The uniform ergodicity part is solid, and the Student-t fast-mixing example is a concrete contribution. The main risk is not an error in the proofs but an over-broad claim about fast mixing for general heavy-tailed or multimodal targets. I would ask the authors to make the conditional nature of Theorem 3.6 for multimodal targets explicit in the abstract and contribution list, and to state clearly that the Möbius-tuned implementation is not covered by the fast-mixing theorem for multimodal or funnel targets. The numerical experiments can remain as empirical evidence, but they should not be presented as consequences of the non-asymptotic theory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core new idea is a radial diffeomorphism that pulls a heavy-tailed target on R^d back onto a Euclidean ball, with a Jacobian singularity of order d+β. That gives a clean sufficient condition (A-β) for uniform ergodicity, covering arbitrarily heavy polynomial tails. The proof in Appendix A is internally consistent: the density bounds in Lemma A.4, the m-step minorization in Lemma A.5, and the standard small-set argument all hold up. The explicit Student-t analysis in Example B.1, with a closed-form automorphism and O(d^3) cold-start mixing for DCS-HnR, is a real result. I also appreciate that the paper clearly embeds the spherical samplers of Yang et al. and Grazzi et al. as special cases of its framework.\n\nThe soft spot is the gap between the fast-mixing theory and the implemented algorithm. Theorem 3.6 needs the pull-back density to be log-concave. The Möbius automorphism tuned by VI (Appendix D) is a single-parameter hyperbolic kick; it is a homeomorphism of the ball, so it cannot turn a two-mode target into a unimodal one. Example B.2 proves existence of a suitable transport via Knothe–Rosenblatt, but the paper itself admits (Remark B.2, Open Problem IIb) that this transport is non-constructive and a richer automorphism class is needed. So the abstract's phrasing that DCS gives fast non-asymptotic mixing for heavy-tailed posteriors, including multimodal ones, is not covered by the theory for the actually implemented algorithm. Uniform ergodicity is unaffected, and the Student-t bounds stand. The PosteriorDB funnel experiments are also explicitly outside the full-space theory, and the comparisons use a slice-sampling variant of HnR rather than the ideal kernel analyzed in the theorem. That is a real mismatch between what is proved and what is benchmarked.\n\nMinor issues: no code or data is shipped, so the numerical claims cannot be independently checked, and there is a small typo in Theorem 3.6(d) (ν_B where ν is meant). Neither is damaging by itself, but together with the coverage gap they make the practical claims weaker than the paper's framing suggests.\n\nWho gets value: researchers working on heavy-tailed MCMC and transformation-based samplers. The uniform ergodicity theorem and the Student-t bounds deserve a serious referee, and the paper is honest about many of its own limitations in Sections 3.2.1 and 5. My recommendation: accept for peer review, but request code and a sharper scope statement — restrict the fast-mixing guarantees to what the automorphism class can actually deliver, or show a constructive richer family.","headline":"Solid uniform-ergodicity theory for heavy tails via a radial pull-back to a ball, but the fast-mixing claims for multimodal targets outrun what the implemented Möbius tuning can actually deliver; still deserves a serious referee.","tokens_in":33623,"tokens_out":1961,"would_cite":true,"duration_ms":20540,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J10","60J20","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A radial contraction onto a ball turns heavy-tailed MCMC into uniformly ergodic sampling on a bounded domain, with $O(d^3)$ mixing for Student-$t$ targets.","keywords":["heavy-tailed distributions","uniform ergodicity","non-asymptotic mixing time","Hit-and-Run","Ball Walk","diffeomorphic contraction sampler","variational inference","log-concave sampling"],"falsifier":"For a Student-$t$ target in $d$ dimensions with $v\\ge2$, run DCS-HnR from a cold start with the paper's recommended $R^2=d+v$ and measure the total-variation mixing time $\\tau_X(\\varepsilon,\\delta_0)$; if $\\tau_X$ grows faster than a polynomial of degree $d^3$ in $d$, the central non-asymptotic claim fails. A second check is to evaluate the Hessian $\\nabla^2 U_B$ of the pulled-back potential at points approaching $\\partial B(R)$: a negative eigenvalue anywhere would violate Assumption (Conv) and place the example outside Theorem 3.6.","tokens_in":32456,"feed_emoji":"🎯","tokens_out":6912,"duration_ms":55940,"temperature":0.7,"pith_summary":"The paper introduces the Diffeomorphic Contraction Sampler (DCS), which pulls a target density on $\\mathbb{R}^d$ back onto a Euclidean ball $B(R)$ through a radial diffeomorphism and then runs a standard sampler, such as Hit-and-Run or Ball Walk, on the bounded ball. Its central claim is that whenever the target has a finite polynomial moment in the sense $\\limsup_{|x|\\to\\infty}\\pi(x)|x|^{d+\\beta}<\\infty$ for some $\\beta>0$, DCS is uniformly ergodic on $\\mathbb{R}^d$ for several standard ball kernels. For Student-$t$ targets, a stronger set of assumptions on the pull-back density yields a cold-start mixing time of order $d^3$ for the Hit-and-Run variant, in contrast to the exponential-in-$d$ mixing of Gaussian Random Walk Metropolis. The paper also reports numerical evidence that DCS outperforms the No-U-Turn Sampler and spherical projection samplers on heavy-tailed and funnel-geometry benchmarks. The reason this matters is that heavy tails combined with high dimension make standard samplers mix slowly and can break central limit theorems for ergodic averages.","feed_headline":"Pull heavy tails onto a ball, then sample fast","feed_subtitle":"DCS maps any polynomial-tailed target in R^d onto a ball; Student-t targets mix in O(d^3).","key_machinery":"The load-bearing object is the radial diffeomorphism $RC_\\beta(y)=y/(1-|y|^\\beta)^{1/\\beta}$ from the unit ball to $\\mathbb{R}^d$, whose Jacobian has a singularity of order $d+\\beta$ at the boundary. Composed with a ball automorphism $T_\\omega$ (in examples, a Möbius transformation) and parameters $(\\mu,R)$, it defines $RC_{\\mathrm{total}}(z)=\\mu+R\\,RC_\\beta(T_\\omega(z/R))$, and the pull-back density is $\\pi_B(z)=\\pi(RC_{\\mathrm{total}}(z))\\,J_{RC_{\\mathrm{total}}}(z)$. The Jacobian singularity cancels polynomial tails, making $\\pi_B$ bounded on $B(R)$; the automorphism preconditions skew and multimodal geometry. Total-variation distance is invariant under this diffeomorphism, so mixing on the ball transfers exactly to mixing on $\\mathbb{R}^d$.","core_discovery":"On the paper's own terms, the discovery is a class of samplers whose mixing behaviour is governed by the boundary singularity of a chosen radial contraction rather than by the tail decay of the target. Theorem 3.1 states that if $\\pi$ is continuous and satisfies $\\limsup_{|x|\\to\\infty}\\pi(x)|x|^{d+\\beta}<\\infty$ for some $\\beta>0$, then Algorithm (DCS) using the DCS-HnR, DCS-BW, DCS-RWM, or covariance-floored DCS-DW kernel is uniformly ergodic on $\\mathbb{R}^d$. Under additional assumptions on the pulled-back density $\\pi_B$ — convexity of $U_B=-\\log\\pi_B$, isotropy, and warm or cold start conditions — Theorem 3.6 gives non-asymptotic mixing-time bounds, and Example B.1 verifies these assumptions for Student-$t$ targets with $R^2=d+v$, yielding a mixing time of order $d^3$ for DCS-HnR from a cold start.","pith_inferences":["Because the TV distance is unchanged by the diffeomorphism, any future sampler with tighter bounds on the ball would automatically transfer those bounds to the full heavy-tailed problem; the paper's use of existing ball samplers is not the only possible choice.","The cold-start guarantee depends on the automorphism $T_\\omega$ being found; for multimodal targets the paper guarantees existence via triangular transport but not a constructive recipe, so an empirical check of whether VI finds such a map on a given target is a direct test of the $O(d^3)$ claim.","The tail-decay trichotomy in Remark 3.11 suggests an adaptive rule: estimate the tail index of $\\pi$ and choose $\\beta$ slightly smaller, so that (A-$\\beta$) and asymptotic convexity hold; the paper notes this preference but does not propose an estimator.","One could extend DCS to gradient-based samplers by handling the boundary singularity of $\\nabla U_B$; the paper lists this as an open problem, but its resolution would likely improve practical performance in regimes where random-walk proposals are diffusive."],"forward_implications":["Uniform ergodicity holds for every target satisfying (A-$\\beta$), regardless of how slow the polynomial tail is, so the usual moment restrictions of stereographic samplers are removed.","For Student-$t$ targets, DCS-HnR mixes in $O(d^3)$ steps from a cold start, whereas Gaussian Random Walk Metropolis has an exponential-in-$d$ bound; this is the first polynomial cold-start bound of this type.","Any improvement in sampling algorithms on the Euclidean ball immediately becomes an improvement for DCS, because the pull-back reduces the problem to sampling on a ball.","The VI-tuned Möbius automorphism is meant to make the pull-back density near-uniform; when it succeeds, the non-asymptotic bounds of Theorem 3.6 apply even for skewed or multimodal targets.","Numerical experiments indicate the method reaches comparable accuracy to tuned NUTS or spherical samplers in a fraction of the wall-clock time on heavy-tailed posteriors and constrained GARCH-type boundaries."],"supporting_citations":[{"why":"Establishes why heavy tails break central limit theorems and quantifies slow mixing of vanilla samplers, motivating DCS.","marker":"[8]"},{"why":"Sub-Cauchy sampler used as the main numerical baseline and as the comparator in the uniform-ergodicity discussion.","marker":"[19]"},{"why":"Stereographic MCMC baseline whose moment restrictions DCS removes.","marker":"[47]"},{"why":"Earlier variable-transformation approach that achieves geometric but not uniform ergodicity, the gap DCS closes.","marker":"[25]"},{"why":"Supplies cold-start Hit-and-Run mixing bounds used in Theorem 3.6(c).","marker":"[32]"},{"why":"Supplies warm-start Ball Walk and Hit-and-Run mixing bounds used in Theorem 3.6(a) and (b).","marker":"[33]"},{"why":"Supplies Dikin Walk mixing bounds used in Theorem 3.6(e).","marker":"[28]"},{"why":"Supplies projected Langevin Monte Carlo finite-time bounds used in Theorem 3.6(d).","marker":"[11]"},{"why":"Provides the triangular transport existence used for multimodal targets in Example B.2.","marker":"[6]"},{"why":"Gives the exponential-in-$d$ RWM bound for Student-$t$ targets that DCS-HnR improves on.","marker":"[1]"}],"fun_headline_variants":["Diffeomorphic MCMC tames heavy tails with ball mapping","Heavy-tailed targets: map to a ball, mix in O(d^3)","Radial contraction trick speeds MCMC for heavy tails","Ball pullback yields fast mixing for polynomial tails","Student-t mixing in O(d^3) via diffeomorphic contraction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the advertised $O(d^3)$ cold-start guarantee, the pull-back density $\\pi_B$ must be log-concave, near-isotropic, and satisfy cold-start level-set conditions; these are verified for Student-$t$ targets in the paper, but for general or multimodal targets a suitable preconditioning automorphism is only shown to exist, not constructed.","fun_headline_variants_meta":{"raw":{"variants":["Diffeomorphic MCMC tames heavy tails with ball mapping","Heavy-tailed targets: map to a ball, mix in O(d^3)","Radial contraction trick speeds MCMC for heavy tails","Ball pullback yields fast mixing for polynomial tails","Student-t mixing in O(d^3) via diffeomorphic contraction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1684,"prompt_tokens":1026,"completion_tokens":658,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":571}},"tokens_in":642,"tokens_out":658,"duration_ms":5935,"temperature":1.0,"reasoning_tokens":571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:05:53.334939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a Student-$t$ target in $d$ dimensions with $v\\ge2$, run DCS-HnR from a cold start with the paper's recommended $R^2=d+v$ and measure the total-variation mixing time $\\tau_X(\\varepsilon,\\delta_0)$; if $\\tau_X$ grows faster than a polynomial of degree $d^3$ in $d$, the central non-asymptotic claim fails. A second check is to evaluate the Hessian $\\nabla^2 U_B$ of the pulled-back potential at points approaching $\\partial B(R)$: a negative eigenvalue anywhere would violate Assumption (Conv) and place the example outside Theorem 3.6.","supporting_citations":[{"cited_title":"Roberts,Stereographic Markov chain Monte Carlo, Ann","cited_arxiv_id":null,"evidence_quote":"Stereographic MCMC baseline whose moment restrictions DCS removes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes why heavy tails break central limit theorems and quantifies slow mixing of vanilla samplers, motivating DCS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sub-Cauchy sampler used as the main numerical baseline and as the comparator in the uniform-ergodicity discussion."},{"cited_title":"Johnson and Charles J","cited_arxiv_id":null,"evidence_quote":"Earlier variable-transformation approach that achieves geometric but not uniform ergodicity, the gap DCS closes."},{"cited_title":"Comput.35(2006), no","cited_arxiv_id":null,"evidence_quote":"Supplies cold-start Hit-and-Run mixing bounds used in Theorem 3.6(c)."},{"cited_title":"3, 307–358","cited_arxiv_id":null,"evidence_quote":"Supplies warm-start Ball Walk and Hit-and-Run mixing bounds used in Theorem 3.6(a) and (b)."},{"cited_title":"3137–3240","cited_arxiv_id":null,"evidence_quote":"Supplies Dikin Walk mixing bounds used in Theorem 3.6(e)."},{"cited_title":"Geom.59(2018), no","cited_arxiv_id":null,"evidence_quote":"Supplies projected Langevin Monte Carlo finite-time bounds used in Theorem 3.6(d)."},{"cited_title":"3, 309–335","cited_arxiv_id":null,"evidence_quote":"Provides the triangular transport existence used for multimodal targets in Example B.2."},{"cited_title":"Wang,Weak Poincar´ e inequalities for Markov chains: theory and applications, Ann","cited_arxiv_id":null,"evidence_quote":"Gives the exponential-in-$d$ RWM bound for Student-$t$ targets that DCS-HnR improves on."}],"review_version":1}