{"id":"a8e9af72-bd17-4ac4-afaf-fc09b2d604e6","arxiv_id":"2509.18712","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Sparse-grid Gauss-Hermite quadrature is provably suboptimal at N^{-alpha/2}, while some QMC methods with change of variables achieve the optimal N^{-alpha} (ln N)^{(d-1)/2} rate in Gaussian Sobolev spaces.","lead":"The paper proves that the standard sparse-grid Gauss-Hermite rule converges at half the optimal rate in Gaussian Sobolev spaces, and that several quasi-Monte Carlo methods reach the full optimal rate. The strict lower bound explains earlier numerical observations, though the abstract's claim about weight modifications is stronger than what the theorems prove.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lower bound for sparse-grid Gauss–Hermite depends entirely on a cited, unproved univariate inequality; the multivariate reduction is sound but the N^{-alpha/2} claim inherits all its force from [41, Thm 3.2].","rationale":"I agree with the reader's primary concern. The novel and central claim is the suboptimal N^{-α/2} lower bound for the Smolyak–Gauss–Hermite rule. The multivariate telescoping argument is clean, and the QMC upper bounds use standard embeddings and established CBC/digital-net results. But the decisive step is a one-dimensional estimate imported from a self-cited previous paper, and the current manuscript supplies no proof or independent check. If that univariate inequality fails, the main theorem collapses regardless of the correctness of the surrounding multivariate analysis. The proposed numerical re-computation and proof re-derivation would settle whether the concern lands. The reader's conditional verdict is therefore appropriate; I would not change it, but the univariate estimate should be verified before accepting the suboptimality result as fully established.","tokens_in":25584,"tokens_out":19812,"duration_ms":171632,"concrete_test":"Independently recompute R_n for n = 2^k, k = 6,...,16, and α = 1,2,3, using the roots of H_n (e.g., via Golub–Welsch) and high-precision quadrature on each interval [ξ_j, ξ_{j+1}], including direct quadrature of ∫ p_n ρ and of ||D^r p_n||^2_{L2_ρ}. Fit log R_n against log n. The central claim is confirmed if the fitted slope is ≤ -α/2 with a positive intercept, i.e., R_n ≥ c n^{-α/2} for all tested n. If the slope is materially steeper than -α/2, Theorem 3.1 remains true but is not sharp; if it is shallower for large n, the suboptimality claim is unsupported. Additionally, re-derive the proof of [41, Thm 3.2] from its Lemma 3.1 to check that c_α is indeed n-independent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 3.1 constructs a univariate fooling function p_n supported on the gaps between roots of H_n and reduces the multivariate worst-case error to the ratio R_n = |∫ p_n ρ| / (Σ_{r=0}^{α} ||D^r p_n||^2_{L2_ρ})^{1/2}. The paper then imports the assertion R_n ≥ c_α n^{-α/2} from the authors' earlier paper [41, Theorem 3.2] without reproducing the argument. This is the load-bearing step: the multivariate analysis in Equations (4)–(5) only establishes the reduction; the polynomial rate N^{-α/2} in Theorem 3.1 and the logarithmic refinement in Theorem 3.2 are exactly the univariate rate. If the constant in [41, Thm 3.2] were not independent of n, or if the H^α_ρ norm of p_n grew faster than n^{α/2} relative to its integral, the central suboptimality claim would collapse. The paper's closing remarks emphasize the fooling-function construction, but the key estimate is not shown. This is not an allegation that the cited theorem is wrong; it is a request for independent verification of the one step on which the central claim entirely depends. The weight-independence remark in the abstract is a separate issue and does not affect this lower-bound argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies worst-case numerical integration over R^d with respect to the standard Gaussian measure in Gaussian Sobolev spaces H^α_ρ of dominating mixed smoothness. Its main results are: (i) a lower bound e(S_Λ) ≳ N^{-α/2} for sparse-grid Gauss–Hermite quadrature for downward-closed index sets, and N^{-α/2}(ln N)^{α(d-1)/2} for the classical isotropic Smolyak index set, matched up to a logarithmic factor by an upper bound; and (ii) upper bounds N^{-α}(ln N)^γ for four QMC constructions (affine-mapped rank-1 lattices and higher-order digital nets, and cotangent/Möbius-mapped rank-1 lattices and higher-order digital nets), with γ=(d-1)/2 for the last construction, matching the information-theoretic lower bound of Dick et al. [11]. The proofs use a univariate fooling function from the authors' earlier paper [41], a new mixed-Sobolev embedding (Prop. 2.3), Smolyak theory, and a chain-rule analysis for the cotangent transform.","tokens_in":25914,"tokens_out":16983,"duration_ms":144111,"significance":"If the results hold, the paper rigorously explains the observed gap between sparse-grid Gauss–Hermite and QMC methods in a clean model setting: sparse-grid Gauss–Hermite is suboptimal by a square root in N, while QMC, especially higher-order digital nets with the cotangent transform, is optimal including the logarithmic factor. The paper is carefully written and contains several self-contained ingredients, notably the Sobolev embedding in Prop. 2.3 and the chain-rule analysis of §4.2. However, the central lower bound is imported from the authors' earlier [41, Thm 3.2] without proof, and the proof of Theorem 3.3 contains a mismatch between quadrature and interpolation nodes. These issues need to be addressed before the claims are fully supported.","major_comments":[{"comment":"The proof asserts the equality |∫ p_n ρ − Q^uni_{ℓ̄_1}(p_n)| = |∫ (p_n − I^uni_{ℓ̄_1}(p_n)) ρ| because I^uni is interpolatory. This equality requires Q^uni_{ℓ̄_1}(p_n) = ∫ I^uni_{ℓ̄_1}(p_n) ρ. But Q^uni_ℓ uses the n_ℓ roots of H_{n_ℓ}, while I^uni_ℓ is defined to use the zeros of H_{n_ℓ+1}. These node sets differ, and in general Q^uni(p_n) − ∫ I^uni(p_n)ρ = Σ_j w_j [p_n(ξ_j) − I^uni(p_n)(ξ_j)] ≠ 0. Thus the reduction in the proof does not follow. Please either change the definition of I^uni_ℓ to use the same nodes as Q^uni_ℓ, or supply a different argument. This issue does not affect Theorems 3.1–3.2, but it leaves Theorem 3.3 unsupported as written.","section":"§3.2, Theorem 3.3"},{"comment":"The multivariate lower-bound proof correctly reduces the worst-case error to the univariate ratio in Eq. (5), but the N^{-α/2} rate is then taken verbatim from [41, Theorem 3.2] without statement or proof. Since that univariate estimate is the load-bearing step for the paper's central suboptimality claim and is a self-citation, the manuscript should state the theorem and include a proof or at least a complete proof sketch, in particular the n-dependence of ||p_n||_{H^α_ρ} and the constant in the lower bound. This is not an accusation of circularity, but a request for self-contained verification of the only step that gives the polynomial rate.","section":"§3.1, Theorems 3.1 and 3.2"}],"minor_comments":[{"comment":"The denominator on the right-hand side of Eq. (5) is missing the square root: it should be (Σ_{r=0}^α ||D^r p_n||^2_{L^2_ρ})^{1/2}, consistent with the definition of the H^α_ρ norm.","section":"Eq. (5)"},{"comment":"The logarithmic exponent in the statements is written as (ln N)^{a(d-1)/2} with 'a' instead of α. The proofs also use 'a' in the same way; this should be corrected.","section":"Theorems 3.2 and 3.3"},{"comment":"After the telescoping identity S_Λ(h_Λ) = Q^uni_{ℓ̄_1}(p_n), it would help to state explicitly that lower-level Gauss–Hermite nodes cancel and never appear. This is what justifies the abstract's claim that the lower bound is independent of quadrature weights.","section":"§3.1"},{"comment":"The sentence explaining that a continuous linear functional can be viewed as an operator into H^α_ρ(R) is confusing: a functional maps into R, not into H^α_ρ(R). The intended embedding via constants should be spelled out more clearly.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The main claim of the paper is likely correct, but the gap in the proof of Theorem 3.3 and the heavy reliance on the unproved [41, Thm 3.2] make the current version unsuitable for acceptance. The Theorem 3.3 issue is local and fixable; if the authors also provide a proof or detailed statement of the univariate estimate, the paper would be in much better shape. I would not reject on these grounds, since neither issue appears to invalidate the central comparison between sparse-grid Gauss–Hermite and QMC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know about this paper if you care about high-dimensional Gaussian integration. It closes a known gap: the sparse-grid Gauss–Hermite rule is shown to be stuck at N^{-α/2} in the Gaussian Sobolev space H^α_ρ(R^d), while a Möbius-transformed higher-order digital net achieves the optimal N^{-α}(log N)^{(d-1)/2}. The results are mostly what the abstract promises.\n\nThe genuinely new part is the multivariate lower bound for sparse-grid Gauss–Hermite, Theorems 3.1 and 3.2. The reduction from d dimensions to a univariate fooling function is clean, and it works because the Smolyak differences telescope: only the highest-level univariate rule sees the fooling function. This also means the abstract's 'weight-independence' statement is correct, contrary to a worry I had going in—the weights of the lower-level quadratures cancel, and p_n vanishes at the top-level Hermite nodes.\n\nThe QMC side is mostly assembly of published machinery. The new Sobolev embedding (Prop 2.3) is useful and the treatment of the cotangent transform's boundary behavior (Prop 4.7) is careful; the endpoint result—optimal rate including the logarithmic factor—is real and well-sourced.\n\nThe main caveat is that the central lower bound inherits its entire rate from the authors' earlier univariate theorem [41, Thm 3.2], which is cited without proof. That is normal for a follow-up, and [41] is peer-reviewed, but a referee will want to confirm that the constant is n-independent; the multivariate analysis carries no weight beyond that reduction. If [41] is sound—and I have no reason to doubt it—the paper's central claims hold.\n\nMinor issues: the sparse-grid upper bound's log exponent is weaker than the lower bound's, so 'up to a logarithmic factor' is doing some work. The proof of Prop 4.1 leans heavily on [55], and a reader may want the CBC constants stated more explicitly.\n\nBottom line: this is a solid, useful paper for numerical analysts working on QMC versus sparse grids. It deserves a serious referee and, assuming the univariate dependency checks out, publication in a good numerical analysis venue. I would bring it to a reading group and would cite it if I wrote about Gaussian integration.","headline":"Sparse-grid Gauss–Hermite is suboptimal at N^{-α/2}; a Möbius-transformed digital net hits the optimal rate—the paper is right, with the main caveat being dependence on the authors' earlier univariate theorem.","tokens_in":26416,"tokens_out":7130,"would_cite":true,"duration_ms":59918,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D30","65D32","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves sparse-grid Gauss–Hermite quadrature converges only at half the optimal rate for Gaussian Sobolev spaces, while several quasi-Monte Carlo rules, one with a cotangent change of variables, achieve the optimal rate.","keywords":["quasi-Monte Carlo","Gauss–Hermite quadrature","sparse grids","Gaussian Sobolev space","worst-case error","cotangent transform","digital nets","rank-1 lattices"],"falsifier":"Take p_n supported on gaps between Hermite roots for α=1,2,3 and compute the ratio R_n = |∫ p_n ρ|/(Σ_{r=0}^α ||D^r p_n||^2_{L2ρ})^{1/2} numerically for n=10,...,10^4; if R_n decays like n^{-α} or faster rather than n^{-α/2}, the lower-bound engine fails.","tokens_in":25471,"feed_emoji":"🧮","tokens_out":7633,"duration_ms":63076,"temperature":0.7,"pith_summary":"This paper asks how fast any numerical integration rule can converge when integrating functions of finite smoothness α against the standard Gaussian measure on R^d. It proves a sharp gap between two popular methods: sparse-grid quadrature built from Gauss–Hermite rules achieves, up to logarithms, only the rate N^{-α/2} with N function evaluations, whereas several quasi-Monte Carlo rules achieve the optimal rate N^{-α} with a logarithmic factor. One construction—a higher-order digital net combined with a cotangent change of variables—attains the optimal rate including the exact logarithmic exponent (d−1)/2. The lower bound for the sparse-grid rule holds for any choice of quadrature weights, so reweighting the Gauss–Hermite nodes cannot rescue it. A sympathetic reader would take this as a definitive comparison: for Gaussian integration of finite smoothness, QMC with a suitable change of variables is the rate-optimal choice, and sparse-grid Gauss–Hermite is not.","feed_headline":"Sparse-grid Gauss-Hermite converges half as fast as optimal QMC","feed_subtitle":"For Gaussian integrals with smoothness α, sparse grids top out at N^{-α/2}; a cotangent-mapped digital net reaches the optimal N^{-α}.","key_machinery":"The lower bound rests on a univariate fooling function p_n: a piecewise polynomial supported on the gaps between consecutive Hermite roots, defined as ((t−ξ_j)/(ξ_{j+1}−ξ_j))^α (1−(t−ξ_j)/(ξ_{j+1}−ξ_j))^α on the j-th gap and zero elsewhere. Because it vanishes at all n Gauss–Hermite nodes, any quadrature using those points integrates it to zero, and a cited univariate estimate gives |∫ p_n ρ| / ||p_n||_{H^α_ρ} ≥ c n^{-α/2}, which is the engine of the lower bound. The upper bounds for QMC use a componentwise cotangent change of variables Ψ(t)=−cot(πt) that maps the unit cube to R^d; under this map the transformed integrand extends continuously by zero to the boundary and lies in mixed-derivat","core_discovery":"On the Gaussian Sobolev space H^α_ρ(R^d), the worst-case error of sparse-grid Gauss–Hermite quadrature is bounded below by c N^{-α/2} (up to logarithms) for any downward-closed index set, and bounded above by C N^{-α/2} (ln N)^{(d−1)(1+α/2)} for the isotropic sparse grid, so the true rate is N^{-α/2} up to a log factor. In contrast, quasi-Monte Carlo rules with affine or cotangent changes of variables achieve worst-case error of order N^{-α} times a power of ln N; the cotangent-transformed higher-order digital net attains N^{-α}(ln N)^{(d−1)/2}, matching the known lower bound for all algorithms. The key to the suboptimality is a fooling function that vanishes at every Gauss–Hermite node, for","pith_inferences":["The fooling-function construction depends only on the distribution of the quadrature nodes, not on the weights, so the same approach could be used to test other node families (e.g., Leja points) for suboptimality, as the paper itself suggests.","The rate gap N^{-α/2} versus N^{-α} means the relative penalty grows with smoothness α; for α=1 the sparse grid has error ~N^{-1/2}, far worse than the ~N^{-1} of optimal QMC.","One might conjecture that the cotangent transform is the 'right' universal change of variables for Gaussian measures in this Sobolev setting; a testable extension would be to prove that only transformations whose derivatives grow like polynomials can yield the optimal log exponent.","The dimension-dependent constants in the error bounds may limit practical use in very high dimensions; the paper's own remark points to weighted Sobolev spaces as the standard route to dimension-independent constants."],"forward_implications":["Users of sparse-grid Gauss–Hermite quadrature for Gaussian integrals of functions with finite Sobolev smoothness cannot expect better than N^{-α/2} worst-case error, no matter how the weights are chosen.","The cotangent-transformed higher-order digital net achieves the sharp rate N^{-α}(ln N)^{(d−1)/2}, so equal-weight rules with positive weights suffice for optimal Gaussian integration.","The boundary extension result means boundary points of the unit cube—where the cotangent transform is undefined—can be assigned value zero with no loss of rate.","The same fooling-function argument yields analogous lower bounds for sparse-grid Gauss–Hermite interpolation in the L1ρ norm.","An affine-mapped lattice and affine-mapped digital net also reach the optimal polynomial rate, but with larger logarithmic exponents than the cotangent construction."],"fun_headline_variants":["Sparse-grid Gauss-Hermite only achieves half the optimal rate","Optimal QMC doubles the rate of sparse-grid Gauss-Hermite","Sparse-grid Gauss-Hermite is suboptimal; QMC with change of variables wins","Half the rate: sparse-grid Gauss-Hermite can't keep up with QMC","Gauss-Hermite sparse grid: half the optimal QMC rate"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The sparse-grid lower bound rests on a cited univariate inequality for the fooling function built between consecutive Hermite roots: that |∫ p_n ρ|/||p_n||_{H^α_ρ} ≥ c n^{-α/2}; the paper does not reproduce the proof, and if that inequality were wrong the suboptimality claim would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Sparse-grid Gauss-Hermite only achieves half the optimal rate","Optimal QMC doubles the rate of sparse-grid Gauss-Hermite","Sparse-grid Gauss-Hermite is suboptimal; QMC with change of variables wins","Half the rate: sparse-grid Gauss-Hermite can't keep up with QMC","Gauss-Hermite sparse grid: half the optimal QMC rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000941,"raw_usage":{"total_tokens":3884,"prompt_tokens":794,"completion_tokens":3090,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2988}},"tokens_in":538,"tokens_out":3090,"duration_ms":20994,"temperature":1.0,"reasoning_tokens":2988,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:41:52.386631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take p_n supported on gaps between Hermite roots for α=1,2,3 and compute the ratio R_n = |∫ p_n ρ|/(Σ_{r=0}^α ||D^r p_n||^2_{L2ρ})^{1/2} numerically for n=10,...,10^4; if R_n decays like n^{-α} or faster rather than n^{-α/2}, the lower-bound engine fails.","supporting_citations":[],"review_version":1}