{"id":"c031dde5-d5a9-4ca2-9069-2370e7c8ce59","arxiv_id":"2502.05623","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The Proximal Sampler has exponential convergence in relative Fisher information for strongly log-concave targets, matching the rate of continuous-time Langevin dynamics.","lead":"This paper proves that the Proximal Sampler, a standard sampling algorithm, converges exponentially fast when the error is measured by relative Fisher information, provided the target distribution is strongly log-concave. It also proves a strong data processing inequality for relative Fisher information along Gaussian channels, a result that may be useful beyond sampling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6 relies on Lemma 2's smoothness/decay hypothesis, which is neither stated in the theorems nor supplied via an approximation argument; the central contraction is not proven under FI<∞ alone.","rationale":"The reader's weakest_assumption is exactly the smoothness/decay hypothesis of Lemma 2, and I agree. I elevate this from a minor note to a conditional-verdict concern because every contraction in the paper flows through Lemma 2: Theorem 3(ii) uses it to get the forward-step SDPI (24a), the backward-step DPI (24b) uses the same formula with b_t=∇logν_t, and Theorem 6 composes these steps. The theorem statements replace Lemma 2's smoothness with 'FI<∞', but no approximation argument is supplied; the blanket regularity sentence in §2.2 is not a proof. Since ν_X is only assumed α-SLC, it may have a non-smooth log-density, and the reverse channel reaches ν_X at the endpoint t=η. The Gaussian examples in the paper confirm the rates in the smooth case and make the result plausible, but they do not close the gap. The concrete test above uses a non-smooth but SLC target to probe whether the omitted boundary terms vanish; if they do not vanish, the central claim fails as stated, and even if they do, the paper should add an explicit approximation argument before the theorem is fully rigorous. Hence I recommend CONDITIONAL: accept only after the regularity/approximation step is supplied or the theorems are restricted to smooth targets.","tokens_in":36418,"tokens_out":18588,"duration_ms":167861,"concrete_test":"Compute the right time-derivative of FI(ρ_t||ν_t) at t=0 for d=1, ρ0=N(0,1), ν0(x)∝exp(−x²/2−|x|) (a 1-SLC target with a cusp at 0). The smooth-formula (43) predicts d/dt FI ≤ −2 FI(ρ0||ν0) = −2; evaluate the exact derivative by differentiating the explicit convolution ν_t=ν0*N(0,t), and check the boundary terms that Lemma 2 discards. If the derivative is >−2 or positive, Theorem 3(ii) is false in the stated generality and Theorem 6 collapses; if it equals the predicted value, a smoothing or approximation argument is still needed before the central rate (25) is rigorous.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper states Theorems 3, 4, and 6 under only FI(ρ0||ν0)<∞ (plus α-SLC), but Lemma 2 — the sole source of the time-derivative formula (17) — is proved under the explicit hypothesis in Appendix F.1 that ρ_t and ν_t are smooth and decay fast enough that all integrations by parts have vanishing boundary terms. The proof of Theorem 3(ii) applies formula (43) at t=0 and integrates the differential inequality (44) from 0; however ν0 is only α-SLC and need not be C^2 (e.g., ν0(x)∝exp(−x²/2−|x|) is 1-SLC but its log-density has a cusp at 0). Likewise, in the backward step of Theorem 6, the reverse-channel argument integrates d/dt FI ≤ 0 up to t=η, where ν_η=ν_X may be non-smooth, and endpoint continuity of FI is not justified. If the discarded boundary terms in Lemma 2 do not vanish for such targets, the formula (17) — and hence the SDPI (24a), the reverse-step DPI (24b), and the exponential rate (25) — could fail. No approximation or regularization argument is given, so the central claim is not established in the stated generality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the mixing time of the Proximal Sampler when the error is measured in relative Fisher information. Its main algorithmic result, Theorem 6, states that for an alpha-strongly log-concave target, each Proximal Sampler iteration contracts relative Fisher information by the factor (1 + alpha*eta)^-2, giving an exponential convergence rate that matches the continuous-time Langevin rate from Eq. (14). The proof is built on a time-derivative formula for relative Fisher information along simultaneous Fokker-Planck channels (Lemma 2), a strong data processing inequality along the Gaussian channel under strong log-concavity (Theorem 3), and a data processing inequality along the reverse Gaussian channel. The paper also gives an eventual SDPI along the Ornstein-Uhlenbeck channel, counterexamples showing that DPI can fail in general, and a high-accuracy iteration complexity corollary for log-smooth targets under rejection-sampling implementation of the restricted Gaussian oracle.","tokens_in":36641,"tokens_out":25800,"duration_ms":233420,"significance":"If the main result is fully established, it closes a natural gap: the Proximal Sampler is shown to inherit the continuous-time Langevin exponential convergence in relative Fisher information, a metric that is stronger than KL divergence under a log-Sobolev inequality. The proof is self-contained, the contraction constants are explicit and depend only on the strong log-concavity parameter, and the paper provides tight Gaussian examples and counterexamples. These are substantial strengths. The main caveat is a regularity gap between the proof of Lemma 2 and the stated assumptions of the theorems that use it.","major_comments":[{"comment":"","section":"Appendix F.1 and Theorem 6"},{"comment":"","section":"Appendix F.2, Theorem 3(iv)"}],"minor_comments":[{"comment":"","section":"Section 3.3.1, Example 4"},{"comment":"","section":"Section 4.3.1, Example 5"},{"comment":"","section":"Appendix F.1, proof of Lemma 2"}],"recommendation":"major_revision","confidential_remarks":"The regularity gap is the main obstacle to accepting the paper as is. It appears fixable by a standard approximation or limiting argument, and the rest of the proof structure is sound. The Theorem 3(iv) identity also needs verification. If both issues are resolved, I would support acceptance, as the main result is a valuable contribution to the sampling literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this one: it gives the first direct exponential convergence guarantee for the Proximal Sampler in relative Fisher information under strong log-concavity, matching the continuous-time Langevin rate. The route is a new strong data processing inequality for relative Fisher information along the Gaussian channel, plus DPI along the reverse Gaussian channel. The proof structure is clean: Lemma 2 gives a time-derivative formula, Theorem 3 extracts contraction, and Theorem 6 composes the channels. I checked the algebra in the appendices; it is detailed and the key steps are straightforward to verify. The examples showing that DPI fails in general are a nice touch.\n\nWhat is genuinely new: prior work on relative Fisher information was either low-accuracy and biased (ULA in Balasubramanian et al.) or indirect (Chewi et al. Appendix A.2 via heat flow after chi-square). This paper proves the rate directly and matches the proximal-gradient bound from optimization. No fitting, no circularity; the SDPI constants come from the strong log-concavity parameter.\n\nThe soft spot, as you might expect from the stress test, is regularity. Lemma 2 is proved under smoothness and sufficient decay so that boundary terms vanish. Theorems 3, 4, and 6 are stated under only FI(ρ0||ν0)<∞. The paper does not provide an approximation argument to bridge this. The stress-test example of a 1-SLC ν0 with a cusp in its log-density (exp(-x²/2-|x|)) is valid: such a target is not C², so the time-derivative formula is not directly applicable. This does not sink the paper—the result is certainly true for smooth targets, and it is very plausible that a standard smoothing argument extends it. But as written, the stated generality is not fully proven. That is a genuine referee-level issue, not a fatal flaw. There is also a harmless sign typo in Appendix F.4.\n\nBottom line: this is a strong, useful paper for the sampling theory community. It deserves a serious referee, and the referee should be asked to fill the regularity gap or restrict the statements to smooth densities. I would take it to a reading group and would cite it in future work on log-concave sampling.\n\nRecommendation: accept with the regularity issue as a required revision.","headline":"Direct exponential Fisher-information mixing for the Proximal Sampler, with a clean SDPI proof; watch the regularity gap between Lemma 2 and the theorems.","tokens_in":37173,"tokens_out":2421,"would_cite":true,"duration_ms":22627,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J60","94A17","60J25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The Proximal Sampler contracts relative Fisher information exponentially fast for strongly log-concave targets, at the same rate as continuous-time Langevin dynamics.","keywords":["relative Fisher information","Proximal Sampler","strong data processing inequality","strongly log-concave","Langevin dynamics","mixing time","Gaussian channel","restricted Gaussian oracle"],"falsifier":"Take a strongly log-concave target whose log-density is only once differentiable (or has slow-tail decay), initialize at a Gaussian, and compute by high-precision quadrature the relative Fisher information after one Gaussian-channel or Proximal Sampler step; if the ratio ever exceeds $(1+\\alpha\\eta)^{-2}$, the regularity assumption in Appendix F.1 is doing real work rather than serving as bookkeeping.","tokens_in":36188,"feed_emoji":"🎲","tokens_out":13385,"duration_ms":119682,"temperature":0.7,"pith_summary":"The paper proves that the Proximal Sampler—a Gibbs-sampling algorithm built from forward and reverse Gaussian steps—converges exponentially fast in relative Fisher information when the target distribution is strongly log-concave. Relative Fisher information measures the expected squared difference between the score functions (gradients of log-density) of the current and target distributions, so it is a stronger error notion than KL divergence. The per-iteration contraction is $FI(\\rho_{k+1}^X \\| \\nu_X) \\le FI(\\rho_k^X \\| \\nu_X)/(1+\\alpha\\eta)^2$ for any step size $\\eta$, exactly the rate of continuous-time Langevin dynamics evaluated at the same effective time. This fills a gap between continuous-time and discrete-time guarantees: previously the discrete Proximal Sampler had such a rate only for divergences like KL or $\\Phi$-divergence, not for relative Fisher information. With rejection sampling for the internal oracle, the result yields a high-accuracy iteration complexity whose iteration count grows logarithmically in $1/\\varepsilon$.","feed_headline":"Exponential Fisher-information mixing for the Proximal Sampler","feed_subtitle":"For strongly log-concave targets each step shrinks relative Fisher information by (1+αη)².","key_machinery":"The load-bearing object is a time-derivative identity for relative Fisher information along any Fokker-Planck channel (Lemma 2). It states that\n$$\\frac{d}{dt} FI(\\rho_t \\| \\nu_t) = -c\\, E_{\\rho_t}\\left[\\left\\|\\$nabla^{2}$ \\log\\frac{\\rho_t}{\\nu_t}\\right\\|_{\\mathrm{HS}}^2\\right] - 2\\, E_{\\rho_t}\\left[\\left\\|\\nabla \\log\\frac{\\rho_t}{\\nu_t}\\right\\|^2_{-c\\$nabla^{2}$ \\log \\nu_t + (\\nabla b_t)_{\\mathrm{sym}}}\\right],$$\nwhere $c$ is the noise coefficient and $b_t$ the drift of the channel. Along the Gaussian channel ($b_t=0$, $c=1$), strong log-concavity of $\\nu_0$ is preserved by convolution with modulus $\\alpha_t = \\alpha/(1+\\alpha t)$, making the second term at most $-2\\alpha_t FI$, which integrates to the $(1+\\alpha t)^{-2}$ contraction; along the reverse Gaussian channel the drift is chosen so the second term vanishes, leaving only the non-positive second-order Fisher information. This channel-wise contraction is exactly what composes into the Proximal Sampler rate.","core_discovery":"The central result, Theorem 6, states that if the target $\\nu_X$ is $\\alpha$-strongly log-concave, then every Proximal Sampler iteration obeys $FI(\\rho_{k+1}^X \\| \\nu_X) \\le FI(\\rho_k^X \\| \\nu_X)/(1+\\alpha\\eta)^2$, and hence $FI(\\rho_k^X \\| \\nu_X) \\le FI(\\rho_0^X \\| \\nu_X)/(1+\\alpha\\eta)^{2k}$. The proof splits the algorithm into its two channels: the forward Gaussian step $\\rho \\mapsto \\rho * N(0,\\eta I)$ satisfies a strong data processing inequality with contraction $(1+\\alpha\\eta)^{-2}$ when the target is $\\alpha$-strongly log-concave, while the backward reverse-Gaussian step is only non-expansive, giving no further contraction but no increase. The exponent $\\alpha\\eta$ makes the discrete rate match the continuous-time Langevin bound $e^{-2\\alpha t}$ in the small-step limit $\\eta k \\to t$. The paper also derives an eventual strong data processing inequality when the target is a log-Lipschitz perturbation of a strongly log-concave distribution, and exhibits Gaussian-channel examples where plain data-processing monotonicity of relative Fisher information fails, so some structural assumption on the second distribution is necessary.","pith_inferences":["The same channel-composition argument would extend to any sampler built from one strongly contractive forward channel and one non-expansive reverse channel, so other proximal or diffusion-based samplers may inherit a Fisher-information mixing bound under strong log-concavity; the paper develops this only for the Proximal Sampler.","Because the Gaussian-channel SDPI uses pointwise curvature of the second distribution, weakening the target assumption from strong log-concavity to a log-Sobolev inequality would require a genuinely new argument rather than a tighter constant.","A Fisher-information guarantee controls convergence of score functions, so it may sharpen downstream analyses that need good score estimates—for example mean-field Langevin systems—but this application is not explored in the paper."],"forward_implications":["For any $\\alpha$-strongly log-concave target, the relative Fisher information of the last iterate decays as $(1+\\alpha\\eta)^{-2k}$, making the iteration count to reach precision $\\varepsilon$ logarithmic in $1/\\varepsilon$.","With $\\eta=1/(dL)$ and rejection sampling for the Restricted Gaussian Oracle, each iteration uses $O(1)$ expected queries to the target log-density, so the total cost is $O((dL/\\alpha)\\log(dL/\\varepsilon))$ up to constants.","The rate matches continuous-time Langevin dynamics and, through the sampling-as-optimization analogy, the proximal gradient method for strongly convex objectives.","When the initial distribution is symmetric and satisfies a Poincaré inequality, the forward step contracts by $(1+\\beta\\eta)^{-1}(1+\\alpha\\eta)^{-2}$, yielding a faster overall mixing rate."],"supporting_citations":[{"why":"Defines the Proximal Sampler and its Restricted Gaussian Oracle, the algorithm whose per-iteration convergence is the paper's subject.","marker":"Lee et al., 2021"},{"why":"Supplies the forward/reverse Gaussian channel decomposition of the Proximal Sampler, the KL-divergence SDPI template, and the rejection-sampling implementation used in the complexity corollary.","marker":"Chen et al., 2022"},{"why":"Provides the preservation of strong log-concavity under convolution, which yields the modulus $\\alpha/(1+\\alpha t)$ along the Gaussian channel.","marker":"Saumard and Wellner, 2014"},{"why":"Gives the Otto-calculus identity for the relative Fisher information along Langevin dynamics, the continuous-time rate that Theorem 6 matches.","marker":"Villani, 2009b"},{"why":"Establishes the low-accuracy average-iterate ULA guarantee in relative Fisher information that the paper's high-accuracy last-iterate result is compared against.","marker":"Balasubramanian et al., 2022"},{"why":"Provides the heat-flow strong log-concavity estimate used for the eventual SDPI in Theorem 3(iv).","marker":"Brigati and Pedrotti, 2024"}],"fun_headline_variants":["Proximal Sampler achieves exponential Fisher mixing on log-concave targets","Strong data processing yields exponential Fisher mixing for Proximal Sampler","Proximal Sampler's Fisher mixing rate matches Langevin dynamics","Exponential Fisher mixing for Proximal Sampler via strong data processing","Strong log-concavity implies exponential Fisher mixing in Proximal Sampler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the regularity hypothesis stated in Appendix F.1—that the evolving densities are smooth and decay quickly enough at infinity for every integration by parts in Lemma 2 to have zero boundary terms—while the main theorems are stated under merely $FI(\\rho_0\\|\\nu_0)<\\infty$, so the missing approximation argument is what keeps the exponential rate from being fully general.","fun_headline_variants_meta":{"raw":{"variants":["Proximal Sampler achieves exponential Fisher mixing on log-concave targets","Strong data processing yields exponential Fisher mixing for Proximal Sampler","Proximal Sampler's Fisher mixing rate matches Langevin dynamics","Exponential Fisher mixing for Proximal Sampler via strong data processing","Strong log-concavity implies exponential Fisher mixing in Proximal Sampler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":3054,"prompt_tokens":1032,"completion_tokens":2022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1925}},"tokens_in":648,"tokens_out":2022,"duration_ms":14145,"temperature":1.0,"reasoning_tokens":1925,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:35:46.890802+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a strongly log-concave target whose log-density is only once differentiable (or has slow-tail decay), initialize at a Gaussian, and compute by high-precision quadrature the relative Fisher information after one Gaussian-channel or Proximal Sampler step; if the ratio ever exceeds $(1+\\alpha\\eta)^{-2}$, the regularity assumption in Appendix F.1 is doing real work rather than serving as bookkeeping.","supporting_citations":[{"cited_title":"A proximal algorithm for sampling from non-smooth potentials","cited_arxiv_id":null,"evidence_quote":"Supplies the forward/reverse Gaussian channel decomposition of the Proximal Sampler, the KL-divergence SDPI template, and the rejection-sampling implementation used in the complexity corollary."}],"review_version":1}