{"id":"68889653-1ecd-4d6b-bffa-3652eff5f64c","arxiv_id":"2505.06098","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A new sampler for Fourier Basis Densities discretizes the density into an ancestor distribution, adds triangular B-spline noise, and proves O(1/K^2) total variation and Wasserstein error bounds.","lead":"This paper gives a fast way to draw random samples from the Fourier Basis Density Model, a flexible probability model that previously had no efficient sampler. The method puts the density on a grid, samples one grid point, and adds simple random noise, with proven error bounds that shrink as the grid grows.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the boundary-wrap step in Proposition 2 checks out, so the O(1/K^2) bounds stand.","rationale":"The paper's central claim is Theorem 1, and its proof is self-contained apart from the boundary case in Proposition 2. I checked that case: the modulo operation in Algorithm 1 is exactly the periodic wrap needed, and the edge interval yields the same linear interpolant as the interior intervals. Hence the O(1/K^2) TV and W1 bounds cover all of [-1, 1]. No load-bearing flaw found. The garbled proof of Property 3 (minimum zero spacing) is a transcription artifact in the arXiv text, but that property is not used in the proof of Theorem 1, so it does not affect the central claim. The W1 expression in Appendix E is the line metric, which dominates the circular W1, so the bound is conservative rather than incorrect. The empirical W1 estimates and the MCMC refinement claims are not formal parts of the theorem and are adequately flagged as empirical. I agree with the reader's ACCEPT verdict; the reader's identified boundary assumption is the right place to look but, on inspection, it holds.","tokens_in":13242,"tokens_out":22177,"duration_ms":207229,"concrete_test":"Compute q(x) on the edge interval [1 - 2/K, 1] directly from Algorithm 1's definition: q(x) = p[K-1](K/2)w1(K/2(x - x_{K-1})) + p[0](K/2)w1(K/2(x - 1)), with p[K-1] = (2/K)p(1 - 2/K) and p[0] = (2/K)p(-1). Substituting w1(t) = max(0, 1 - |t|) yields (1-s)p(x_{K-1}) + s p(-1) with s = (K/2)(x - (1 - 2/K)), which is the linear interpolant at x. Running this check for a random FBM (e.g., N = 10, K = 30) at x = 1 - 1/K and x = 1 verifies equality to machine precision and closes the omitted boundary case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 1) rests on identifying q from Algorithm 1 with the piecewise linear interpolant of p at x_k = -1 + 2k/K. The only place this could fail is the wrapped edge interval, because Appendix C proves the interior intervals [x_k, x_{k+1}] and only asserts \"appropriately wrapping around\" at the boundaries. A direct check closes the gap. For x in [x_{K-1}, 1], Algorithm 1 receives mass only from the right tail of the kernel at n = K-1 and the wrapped left tail of the kernel at n = 0. Writing s = (x - x_{K-1})/(2/K), the conditional densities give q(x) = p[K-1](1-s) + p[0]s. Since p[K-1] = (2/K)p(x_{K-1}), p[0] = (2/K)p(-1), and p(1) = p(-1) by periodicity, this is exactly the linear interpolant between p(x_{K-1}) and p(-1) at x = 1, matching the boundary interval. The same reasoning applies at x = -1 by identification. Thus the pointwise bound of Property 7 covers the whole circle, and the TV and W1 bounds in Appendix E are valid. The only substantive deficiency is presentational: Appendix C should state x_K = x_0 + 2, p[K] = p[0], and verify the last interval explicitly. This does not affect the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Discretized Approximate Ancestral Sampling (DAAS), a two-step sampler for Fourier Basis Density Models (FBMs). DAAS evaluates the FBM density p at K equally spaced grid points, samples an index from the discrete weights p[k] = (2/K)p(x_k), and adds i.i.d. triangular interpolation noise, with an optional ULA/MALA refinement step. The central theoretical claim is Theorem 1: for the triangle-kernel version, the total variation and Wasserstein-1 distances between the target FBM density p and the DAAS sample density q are O(1/K^2), with explicit constants given in Appendix E. The proof route is Proposition 1 (exact quadrature identity), Proposition 2 (the compound distribution equals the periodic piecewise linear interpolant of p on the grid), Property 7 (interpolation error bound from a second-derivative bound), and Appendix E (integral estimates). The experiments show the expected decrease of divergence as K grows, compare the triangle kernel with other B-spline kernels, and compare DAAS with rejection sampling and Langevin-based MCMC baselines.","tokens_in":13501,"tokens_out":16709,"duration_ms":170775,"significance":"If the main theorem is correct, DAAS is a practically attractive sampler for FBMs: it has explicit, non-asymptotic accuracy bounds, no free parameters, and a cost of O(S + K log K) for producing S samples once the K grid values are computed. The explicit constants in Appendix E, pi^2 N(N+1)(2N+1)/(12K^2) for total variation and twice that for Wasserstein-1, are directly usable for choosing K. The derivation is clean and non-circular: Proposition 1 is an exact quadrature identity, Proposition 2 identifies the triangle-kernel compound with piecewise linear interpolation, and the error bounds follow from standard interpolation estimates. The stress-test concern about the boundary wrap does not invalidate the theorem; a direct calculation closes the gap, so I treat it as a presentation issue rather than a soundness issue. The paper gives credit to the prior FBM model and to the thesis that first formulated discretization-based sampling for band-limited circular densities, and the empirical study, while limited, is consistent with the theory.","major_comments":[],"minor_comments":[{"comment":"The proof verifies the interpolation identity on interior intervals only and asserts that the triangle kernels 'appropriately wrap around' the boundaries without displaying the edge calculation. Please add x_K = x_0 + 2, p[K] = p[0], and verify the last interval explicitly: for s = K(x - x_{K-1})/2, the wrapped kernels give q(x) = p(x_{K-1})(1-s) + p(x_0)s, matching the periodic linear interpolant. The direct calculation closes the gap, but it should be written in the paper.","section":"Appendix C (Proposition 2)"},{"comment":"The displayed proof of Property 3 is corrupted: Equation (43) contains unreadable tokens, and the step 'In particular, with max[x1,x2]p(x)=||p(x)||_infinity' is not justified for arbitrary pairs of zeros, so the lower bound x2 - x1 >= 2/(pi N) does not follow as stated. Since Property 3 is not used in the proof of Theorem 1, this does not affect the main claim, but the property and its proof should be corrected or its scope explicitly restricted.","section":"Appendix D (Property 3)"},{"comment":"The caption says 'considering 10 randomly initialized FBMs' whereas Section V-A reports 20 different trials for each configuration; these numbers should be aligned.","section":"Figure 3 caption"},{"comment":"References [4] and [5] are the same article and should be consolidated into a single bibliographic entry.","section":"References"},{"comment":"The theorem statement asserts the existence of constants C1 and C2 without indicating their dependence on N; since Appendix E provides the explicit values, stating them in the theorem itself would improve readability and would make the hyperparameter guidance in the abstract fully self-contained.","section":"Theorem 1"}],"recommendation":"minor_revision","confidential_remarks":"The main claims are sound and the required changes are local and presentational; I do not see grounds for rejection or for a major reworking. The most important revision is to make the boundary interval in Proposition 2 explicit and to repair the corrupted proof of Property 3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: DAAS is a clean, useful sampler for the Fourier Basis Density Model, and the main theorem holds. The idea is to discretize the band-limited density on K grid points, sample the resulting categorical, then add a triangular B-spline kernel. Proposition 2 connects that compound distribution to piecewise linear interpolation; the interpolation error bound with explicit derivative bounds from Appendix D gives the O(1/K^2) TV and W1 bounds. I checked the one place where the proof gets hand-wavy—the wrapped boundary intervals. It works out: the triangle kernels wrap exactly to give linear interpolation across the seam, so the theorem is safe.\n\nWhat is genuinely new: the previous FBM paper had no sampler, and Olofsson's thesis used Féjer kernels with rejection. Using simple B-spline kernels, especially the triangle, and bounding the approximation error explicitly is not in prior work. The complexity story is nice: O(K log K) setup, O(S) sample generation asymptotically.\n\nSoft spots, in order of severity. One, Appendix C says the triangle kernels 'appropriately wrap around' but doesn't show it. The stress-test verification closes the gap, so this is a presentation fix, not a flaw. Two, Property 3's proof is garbled in the arXiv version—there is corrupted text in equation (43). That property is peripheral to Theorem 1, but it should be corrected before publication. Three, the MCMC refinement (ULA/MALA) is offered with empirical evidence only; there are no convergence bounds from the DAAS init. The paper is upfront about that, so it is a limitation, not an error. Four, no code is released; the experiments are simple enough to reproduce, and the derivations are self-contained.\n\nThe citation pattern is honest. The authors credit Olofsson's thesis and the FBM predecessor, and the related work on inverse transform, rejection, Langevin is standard. There is no overclaiming: the paper says this is for a specialized density family, and the bounds are asymptotic in K.\n\nWho is this for: anyone working with FBM or band-limited circular densities, and people interested in practical approximate samplers with explicit error control. It deserves a serious referee. I would accept it for review, with a request to fix the boundary exposition and the corrupted proof.","headline":"A correct, clean sampler for Fourier basis densities with explicit O(1/K^2) error bounds; the one hand-wave in the boundary proof checks out, so this deserves a serious referee.","tokens_in":14049,"tokens_out":2184,"would_cite":true,"duration_ms":22387,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B10","65C05","42A10","41A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces DAAS, a sampling method for Fourier Basis Density Models that reaches any target accuracy at $O(1/K^2)$ error and costs $O(S+K \\log K)$ for $S$ samples.","keywords":["Fourier Basis Density Model","band-limited density","ancestral sampling","discretization interpolation","triangular kernel","total variation distance","Wasserstein-1 distance","piecewise linear interpolation"],"falsifier":"Run Algorithm 1 for a simple FBM with $N=1$ and $K=3$, compute the exact density of the output distribution on the edge interval $[-1, -1+2/3)$ by convolving the discrete ancestor masses with the triangle kernel under the modulo map, and compare it pointwise with the piecewise linear interpolant of $p$ on that interval; any pointwise difference, or a violation of $|p(x)-q(x)| \\le \\pi^2/(12K^2)$ at the edges, would contradict Proposition 2 and Theorem 1.","tokens_in":13041,"feed_emoji":"🎲","tokens_out":9747,"duration_ms":87849,"temperature":0.7,"pith_summary":"The Fourier Basis Density Model (FBM) is a flexible band-limited density model but has lacked an efficient way to draw samples. This paper introduces Discretized Approximate Ancestral Sampling (DAAS): evaluate the density at $K$ equally spaced points to form a discrete ancestor distribution, draw an index, then add triangular noise. The resulting continuous distribution is exactly the piecewise linear interpolation of the target density, and the paper proves that its total variation and Wasserstein-1 distances to the target are bounded by explicit constants divided by $K^2$. Because the algorithm costs $O(S+K \\log K)$ to produce $S$ samples, it offers a practical route to high-quality sampling from FBMs, with guaranteed accuracy by choosing $K$ large enough.","feed_headline":"A K-point trick samples smooth densities with 1/K^2 error","feed_subtitle":"Discretize the Fourier density, draw an ancestor, add triangle noise: explicit accuracy bounds, near-linear cost.","key_machinery":"The triangular kernel $w_1(x) = \\max(0, 1 - |x|)$ acts as the interpolation filter. The compound distribution $q(x) = \\sum_{k=0}^{K-1} p[k] \\frac{K}{2} w_1(\\frac{K}{2}(x-x_k))$ coincides with the piecewise linear interpolation of $p$ at $x_k = -1 + 2k/K$, and this coincidence is what turns a sampling procedure into an interpolation error problem. The argument also relies on the band-limited structure of the FBM: the second-derivative bound $\\max |p''(x)| \\le \\pi^2 N(N+1)(2N+1)/6$ follows from the coefficient properties of the truncated Fourier series, and this bound feeds directly into the standard $\\frac{M}{2K^2}$ interpolation error estimate.","core_discovery":"The central claim is Theorem 1: for an FBM density $p$ with $N$ frequency terms and the triangle-kernel DAAS distribution $q$ on $K$ grid points, the divergences satisfy $D_{TV}(p,q) \\le \\pi^2 N(N+1)(2N+1)/(12K^2)$ and $D_{W_1}(p,q) \\le \\pi^2 N(N+1)(2N+1)/(6K^2)$, with the constants given in Appendix E. The key step is Proposition 2, which shows that $q$ is the piecewise linear interpolation of $p$ at the grid points: within each interval $[x_k,x_{k+1}]$ only the neighboring triangular kernels contribute, and their weighted sum reproduces the linear interpolant. Proposition 1 guarantees that the discrete masses $2p(x_k)/K$ form a valid probability distribution for $K > 2N$ equally spaced points, so Algorithm 1 is well-defined. The $1/K^2$ rate then follows from the classical interpolation error bound once the second derivative of $p$ is bounded by $\\pi^2 N(N+1)(2N+1)/6$.","pith_inferences":["The proof never uses the specific FBM parameterization beyond boundedness of the second derivative, so the same $O(1/K^2)$ guarantee and the same algorithm should apply to any circular density with bounded second derivative, with the constant replaced by the corresponding bound.","The boundary condition is the delicate point: the paper asserts that the triangle kernels 'appropriately wrap around' at $\\pm1$, but does not prove that the modulo operation in Algorithm 1 produces exactly that wrapped kernel; a mismatch would require a corrected constant on the edge intervals while likely preserving the overall rate.","Using B-spline kernels of degree $D$ should yield $O(1/K^{D+1})$ error by standard spline interpolation theory; the paper's experiments with the quadratic kernel hint at this, though no such bound is stated.","Since sampling cost is independent of $N$ once $K$ is fixed, DAAS may make FBM practical as a latent prior in deep generative pipelines where fast approximate sampling matters more than exactness."],"forward_implications":["For any FBM and any tolerance $\\epsilon$, choosing $K > \\sqrt{\\pi^2 N(N+1)(2N+1)/(12\\epsilon)}$ guarantees the sampled distribution is within $\\epsilon$ total variation of the target.","The cost is $O(K\\log K)$ for the initial FFT evaluation of the grid plus $O(S)$ for the $S$ samples themselves, giving $O(S + K\\log K)$ overall and linear time when $S \\gg K$.","The explicit constants remove guesswork from hyperparameter selection: users can compute $K$ directly from the model's frequency count and their accuracy requirement.","DAAS can be used to warm-start ULA or MALA chains, combining the speed of the approximate sampler with the asymptotic exactness of MCMC at the price of additional score-function evaluations.","The analysis gives a quantitative argument for triangular noise over uniform or higher-order B-spline kernels, matching the empirical finding that $w_1$ minimizes KL divergence for a given $K$."],"supporting_citations":[{"why":"Defines the FBM, its autocorrelation parameterization that guarantees non-negative densities, and its normalization $Z=2c_0$.","marker":"[1]"},{"why":"Herglotz's theorem links positive semi-definite Fourier coefficients to non-negativity of the density, validating the FBM construction.","marker":"[12]"},{"why":"Supplies the discrete sum identity (Proposition 1) that makes the ancestor probabilities sum to 1.","marker":"[13]"},{"why":"Alias sampling gives $O(1)$ per-sample complexity after $O(K)$ setup, backing the claimed $O(S)$ sampling cost.","marker":"[14]"},{"why":"Standard piecewise linear interpolation error bound that, combined with the FBM second-derivative bound, yields the $O(1/K^2)$ constants.","marker":"[16]"}],"fun_headline_variants":["Sample Fourier densities with 1/K^2 error bounds","Triangle-kernel sampling: 1/K^2 accuracy for FBM","Proven error bounds for sampling band-limited densities","Ancestral sampling on a grid: 1/K^2 accuracy","Discretize-interpolate sampling with quadratic convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that the compound distribution equals piecewise linear interpolation assumes the triangular kernels 'appropriately wrap around' at the boundaries of $[-1,1)$, but does not verify that the modulo operation in Algorithm 1 produces exactly the same wrapped triangles at the edges; if it does not, the $O(1/K^2)$ interpolation bound would not cover the edge intervals.","fun_headline_variants_meta":{"raw":{"variants":["Sample Fourier densities with 1/K^2 error bounds","Triangle-kernel sampling: 1/K^2 accuracy for FBM","Proven error bounds for sampling band-limited densities","Ancestral sampling on a grid: 1/K^2 accuracy","Discretize-interpolate sampling with quadratic convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000487,"raw_usage":{"total_tokens":2406,"prompt_tokens":958,"completion_tokens":1448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1363}},"tokens_in":574,"tokens_out":1448,"duration_ms":10725,"temperature":1.0,"reasoning_tokens":1363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:48:51.667029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 for a simple FBM with $N=1$ and $K=3$, compute the exact density of the output distribution on the edge interval $[-1, -1+2/3)$ by convolving the discrete ancestor masses with the triangle kernel under the modulo map, and compare it pointwise with the piecewise linear interpolant of $p$ on that interval; any pointwise difference, or a violation of $|p(x)-q(x)| \\le \\pi^2/(12K^2)$ at the edges, would contradict Proposition 2 and Theorem 1.","supporting_citations":[{"cited_title":"Fourier basis density model,","cited_arxiv_id":null,"evidence_quote":"Defines the FBM, its autocorrelation parameterization that guarantees non-negative densities, and its normalization $Z=2c_0$."},{"cited_title":"Brockwell and R","cited_arxiv_id":null,"evidence_quote":"Herglotz's theorem links positive semi-definite Fourier coefficients to non-negativity of the density, validating the FBM construction."},{"cited_title":"An efficient method for generating dis- crete random variables with general distributions,","cited_arxiv_id":null,"evidence_quote":"Alias sampling gives $O(1)$ per-sample complexity after $O(K)$ setup, backing the claimed $O(S)$ sampling cost."},{"cited_title":"Ralston and P","cited_arxiv_id":null,"evidence_quote":"Standard piecewise linear interpolation error bound that, combined with the FBM second-derivative bound, yields the $O(1/K^2)$ constants."}],"review_version":1}