{"id":"644bf1ea-3840-4379-823c-6d29eb7bce32","arxiv_id":"2507.09156","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Energy-distance-minimizing support points and projected support points are shown to be robust Gaussian process emulation designs, with a theory linking them to maximum-entropy, minimax, and maximin designs.","lead":"This paper argues that support points and projected support points, known representative point sets, also make robust designs for Gaussian process emulation of computer experiments. It connects energy distance to emulation error and shows in simulations that these designs often beat minimax, maximin, and MaxPro designs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.2's 'near-maximal information' claim for support points rests on the unquantified surrogate (11)-(13): minimizing the sum of correlation entries is not shown to approximate maximum-entropy designs, and Gershgorin's theorem alone does not justify it.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing link: equations (11)-(13) replace the maximum-entropy objective with an unquantified lower-bound surrogate. My stress-test confirms this is the most serious gap in the paper's theoretical argument. The paper's other contributions—Proposition 1, Theorem 3, Theorem 4, and the minimax/maximin trade-off discussion—are either correct identities or useful heuristic connections; Proposition 2 is informal but not central to the robustness claim. The practical recommendation that SPs/PSPs perform well is supported by simulations, so this is not grounds for rejection. However, the abstract's 'near-maximal information' claim is stronger than what the theory establishes: Theorem 2 only equates energy distance with a particular sum of projected-kernel entries, and the bridge from that sum to maximum entropy is the unvalidated surrogate. Because the reader already assigned CONDITIONAL and my analysis reinforces that judgment without identifying a distinct fatal flaw, the verdict should remain CONDITIONAL. The proposed numerical test would settle whether the surrogate is merely unproven or actually false in concrete cases; if it fails, the theory section should be revised to weaken the information-theoretic claim or supply a tightness argument.","tokens_in":23343,"tokens_out":8946,"duration_ms":103597,"concrete_test":"Check the surrogate (13) numerically in the simplest nontrivial case: p=1, n=3, Gaussian kernel r(x,y)=exp{−10(x−y)^2}. Enumerate designs x1≤x2≤x3 on a fine grid, compute S(D)=Σ_{i,j}r(x_i,x_j) and det R(D). Identify the design minimizing S and the design maximizing det R. If the S-minimizer's determinant falls more than a few percent below the maximum, then (13) is not a faithful surrogate and the inference 'near-maximal information' in Section 3.2 is unsupported. Repeat for several θ values (e.g., θ=1,10,100) to confirm the failure is not isolated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim of Section 3.2—that SPs provide near-maximal information on the variation term P⊥_G f—depends on the chain (11)→(13): maximum-entropy designs are approximated by designs minimizing Σ_{i,j} rθ(x_i,x_j). This approximation is not justified. Equation (11) is not a consequence of Gershgorin's theorem in the direction claimed: Gershgorin bounds eigenvalues, hence gives a lower bound on det(R_n), but replacing the determinant by a linear function of row sums requires that the bound be tight or order-preserving over the design space, and no such argument is supplied. Moreover, the step from (12) to (13) replaces the objective by a lower bound 2nr − Σ r_ij; maximizing a lower bound is not equivalent to maximizing the original quantity when the bound is loose. The subsequent substitution of the projected kernel rθ,G in place of rθ is also heuristic. Consequently Theorem 2, while a correct identity for the kernel rθ(x,y) ∝ ||x||+||y||−||x−y||, only shows that energy distance equals a particular sum of projected-kernel entries; it does not establish that minimizing energy distance maximizes Shannon information unless the surrogate (13) is faithful. A concrete failure mode exists even for n=3: for a fixed sum of off-diagonal correlations, det(R) varies (equal pairwise correlations give a much larger determinant than one large and two small correlations with the same sum), so minimizing the sum cannot distinguish designs with different entropies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies two classes of experimental designs, support points (SPs) and projected support points (PSPs), for Gaussian-process emulation of computer experiments. It argues that SPs are robust designs because they provide (i) a tight Koksma–Hlawka bound on the integration error for the mean term of the response, and (ii) near-maximal Shannon information on the variation term via a maximum-entropy surrogate argument involving the energy distance. The paper also derives connections between SPs and minimax/maximin designs, connects PSPs to MaxPro designs through a closed-form objective under an exponential prior, proposes algorithmic improvements (informed initialization and randomized Sobol' batch sampling), and reports numerical experiments on distance metrics, an IRMSE-based efficiency criterion, and four test functions. The theoretical sections contain several correct and useful identities, but the central robustness claim hinges on an unquantified surrogate-objective approximation (eqs. (11)–(13)) and the small-q approximation in Proposition 2 has uncontrolled errors.","tokens_in":23676,"tokens_out":4744,"duration_ms":44411,"significance":"If the central robustness claim holds, the paper offers a practically attractive answer to a well-known problem: GP emulation designs are typically sensitive to the assumed correlation parameters, whereas SPs/PSPs are cheap to generate in high dimensions and empirically competitive across smooth and rough surfaces. The manuscript contains several correct and useful calculations: Lemma 1 is a standard projection-kernel identity and is proved cleanly; Theorem 3 (integration error of fBm sample paths equals q-energy distance) is a correct and elegant identity; Theorem 4 gives a genuinely informative closed-form connection between PSPs and MaxPro designs with a boundary-correction term; and the numerical study, including the efficiency metric Eff(D;Theta) and the four test functions, is a strength because the comparisons are measured rather than fitted. The algorithmic claims are supported by the cited prior works and the new initialization/Sobol-batch suggestions are motivated, though their added benefit is reported only informally.","major_comments":[{"comment":"The central robustness claim that SPs provide 'near-maximal information gain' on P_perp_G f relies on replacing the maximum-entropy objective det(R_n) by the sum of correlation entries sum_{i,j} r_theta(x_i,x_j). The text states that this follows from Gershgorin's circle theorem, but Gershgorin only gives a lower bound on the determinant in terms of row sums; maximizing that lower bound is not equivalent to maximizing det(R_n) unless tightness or order-preservation over the design space is established, and no such argument is given. The step from (12) to (13) additionally replaces the objective by the lower bound 2nr - sum r_ij, and maximizing a lower bound is not equivalent to maximizing the original quantity when the bound is loose. A concrete failure mode exists even at n = 3: for a fixed sum of off-diagonal correlations, det(R) depends on the configuration (equal pairwise correlations give a larger determinant than one large and two small correlations with the same sum), so minimizing the sum cannot distinguish designs with different entropies. Because Theorem 2 only proves an exact identity for a specific kernel after the surrogate has been adopted, the chain (11)->(13) is the load-bearing step for the information-theoretic interpretation. I recommend either proving a tightness or approximation bound for the surrogate, or reframing the theoretical claim as a heuristic motivation and relying on the numerical evidence for robustness.","section":"Section 3.2, eqs. (11)–(13)"},{"comment":"The claim that SPs are 'two-parts minimax and one-part maximin' rests on approximations (19) and (20), whose errors are uncontrolled and may be of the same order as the quantities being estimated as q -> 0+. For finite q, the right-hand side of (21) approximates E_q(F,F_n) with an error that depends on the variation of h^q over Voronoi regions and on the gap between the q-th power mean and the minimum over pairs; no bound is supplied, and the approximation is used to draw a design-principle conclusion (greater weight on minimax than maximin). This is not a fatal flaw for the paper's practical claim, but the statement 'SPs can be viewed as designs which are two-parts minimax and one-part maximin' in Section 4.2 should be softened or accompanied by a quantitative error analysis, since it is used in the discussion of nominal versus numeric error control.","section":"Section 4.2, Proposition 2 and eqs. (19)–(21)"},{"comment":"The assumption r_theta(x,y) proportional to ||x||+||y||-||x-y|| is a very specific correlation structure (an fBm(1/2)-type kernel), and Theorem 2 is stated only for this kernel. The robustness claim in Table 1, however, is phrased broadly for the variation term under 'expected 1/2-Holder continuity'. The passage from the surrogate approximation (13) with a general kernel r_theta to the energy-distance identity (14) for this one kernel is not justified for other kernels, so the breadth of the claim as summarized in Table 1 is larger than what the proof establishes. Please clarify the role of the specific kernel in the theorem and state more precisely which kernels the near-maximal-information claim covers.","section":"Section 3.2, Theorem 2 and Table 1"}],"minor_comments":[{"comment":"Equation (11) is introduced as an 'approximate formulation' citing Section 5 of [18], but the connection to Gershgorin's theorem is not derived in the text; a short derivation or a more precise citation (equation number in [18]) would help the reader assess the approximation.","section":"Section 3.2, eq. (11)"},{"comment":"The notation in the proof of Theorem 2 sums over 'j=i' in the first expression of eq. (14), whereas the intended sum appears to be over all i,j = 1,...,n; this appears to be a typo that should be corrected.","section":"Section 2.1, Definition 2"},{"comment":"The approximation of the average of ||x_i - x_j||^q by the minimum [M_m(D)]^q requires justification beyond eq. (20), since the minimum is not a representative value of the distribution of pairwise distances; at minimum, the statement 'sufficiently small q' should be accompanied by an explicit error term.","section":"Section 4.2, eq. (23)"},{"comment":"In the proof, the use of eq. (20) to replace the average of 2/n sum_i ||z_i - x_i||^q by 2 max_i ||z_i - x_i||^q is not obvious, since the maximum is not the average of the values; this step should be clarified or corrected.","section":"Section 4, Proposition 2 proof"},{"comment":"The statement 'PSPs correct the boundary attraction behavior of MaxPro designs' is supported for the exponential-prior anisotropic kernel, but the PSPs used in the simulations use the POD prior (29); the transfer of the theoretical boundary-correction insight to the simulation setting should be stated as an interpretation, not as a proven property of the POD-based PSPs.","section":"Section 5.2, Theorem 4"},{"comment":"The timing comparison in Figure 3 is reported without a description of the software/hardware setup for the other designs (only the processor is mentioned for SPs/PSPs); adding the implementation details for minimax, MmLHD, MaxPro, and UD would make the computational comparison reproducible.","section":"Section 6 and Figure 3"},{"comment":"The efficiency metric Eff(D;Theta) uses D*(theta) estimated as the lowest IRMSE among the six designs, which makes the reported efficiencies relative to a finite set of designs rather than the true optimum; this is fine for ranking the six designs, but the phrase 'D*(theta)' in eq. (31) should be defined consistently as the best design among those considered, not as the global argmin.","section":"Section 7, Table 2"},{"comment":"The fEff values are reported only as averages over n for each function; reporting standard errors or per-n full values (beyond the underline convention) would allow the reader to assess the variability of the comparisons.","section":"Section 8, Table 4"},{"comment":"There are a few typographical issues, e.g., 'absense' in Section 7.3.1, 'Alterative' in Table 3, and the incomplete sentence in the proof of Theorem 2 ('this completes the proof' following an incorrect sum index); a careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is solid and the theoretical identities in Lemma 1, Theorem 3, and Theorem 4 are correct and useful. The main risk is that the paper's central theoretical claim—that SPs provide near-maximal information on the variation term—is presented in the abstract and introduction as proven, whereas the proof chain (11)-(13) is a heuristic approximation that is not quantified. If the authors can either prove a bound on the approximation error, or explicitly reframe the information-theoretic argument as motivation and move the empirical evidence to the foreground, I would be comfortable with the paper's claims. The small-q approximation in Proposition 2 is a secondary but related issue. I do not see grounds for rejection: the numerical results and the exact identities are valuable, and the heuristic nature of the surrogate objective can be fixed within the manuscript's scope by careful rewriting. The citation pattern is appropriate; the paper builds on the authors' own prior work, which is correctly cited and independently established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful paper, but it does not deliver the full-strength robustness proof it advertises. What is genuinely new is the framing of support points and projected support points as GP emulation designs, Theorem 3 (expected squared integration error for fBm equals the q-energy distance), Theorem 4 connecting PSPs to MaxPro under exponential priors, and the minimax/maximin compromise story. The calculations in Theorems 2-4 check out, and the empirical section is competently run. I believe the practical conclusion—that SP/PSP designs are competitive and often best across smooth/rough and full/partial active-factor settings—is likely correct.\n\nThe soft spot is exactly where the stress-test lands. The \"near-maximal information\" claim for the variation term depends on replacing the maximum-entropy design criterion (max det R_n) with minimization of the sum of entries of R_n via equations (11)-(13). Gershgorin's theorem gives a lower bound on the determinant, not an approximate equivalence, and no argument shows the bound is tight or order-preserving over the design space. Maximizing a lower bound is not the same as maximizing the original quantity, and the n=3 correlation-matrix example is enough to show the surrogate cannot distinguish designs with different entropies: equal pairwise correlations and one-large/two-small correlations can have the same sum of entries but very different determinants. So Theorem 2 is a correct identity for a particular kernel, but the information-theoretic conclusion built on it is not established. The same problem appears in Proposition 2: the small-q approximations are uncontrolled, with errors that can be the same order as the energy distance itself. The minimax/maximin trade-off is a plausible heuristic, not a proved theorem.\n\nSmaller issues: the efficiency oracle in Section 7 takes the best of the six designs as D*(theta), so the reported efficiencies are comparative rather than absolute; that is fine but should not be over-read. No code is shipped with the manuscript; the R package is promised but not referenced in a verifiable way. The reliance on the authors' own prior work for definitions and algorithms is legitimate, since those results are independently established.\n\nBottom line: for anyone working on computer experiment design, the paper is worth engaging. The empirical comparisons and the two clean identities (Theorems 3 and 4) give it real value. I would send it to a serious referee, but with the expectation that Section 3.2 needs major revision or an honest weakening of the central claim.","headline":"A competent extension of the support-points framework to GP emulation, with useful identities and strong empirical evidence, but the central robustness claim rests on an unquantified surrogate approximation.","tokens_in":24189,"tokens_out":3878,"would_cite":true,"duration_ms":48794,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","62K20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that support points—designs minimizing the energy distance between their empirical distribution and the uniform hypercube—are robust Gaussian-process emulation designs that perform near-optimally for both smooth and…","keywords":["computer experiments","Gaussian process emulation","support points","projected support points","energy distance","robust design","difference-of-convex programming","Koksma-Hlawka inequality"],"falsifier":"For a fixed small design in two dimensions, compute the true maximum-entropy design (maximizing $\\det R_n$) for a Gaussian correlation with moderate correlation length and compare its integrated prediction error with that of the support-point design across a grid of test functions satisfying the Hölder condition. If support points are consistently and substantially worse than the true maximum-entropy design, the surrogate-objective step is not doing the claimed work. A second check is whether the lower bound in equation (12) is tight: if the design minimizing the sum of correlation entries does not track the design maximizing $\\det R_n$ as $n$ grows, the approximation fails.","tokens_in":23108,"feed_emoji":"🎯","tokens_out":5618,"duration_ms":63348,"temperature":0.7,"pith_summary":"This paper argues that a design chosen to minimize the energy distance between its empirical distribution and the uniform distribution on the cube—the support points, and their projected variant—is a good default design for Gaussian process emulation. The claim is not that these designs are optimal for any one correlation model, but that they perform nearly optimally across a wide class of smooth and rugged response surfaces. The argument splits emulation into estimating the mean of the surface and estimating the variation around that mean. For the mean, a classical Koksma–Hlawka bound ties integration error to the energy distance; for the variation, an information-theoretic chain ties the sum of entries of a projected correlation matrix to the same energy distance under a mild Hölder-continuity assumption. The payoff is a design that avoids committing to correlation parameters in advance while matching or beating model-specific and distance-based competitors in the paper's numerical experiments.","feed_headline":"Support points beat model-tuned designs for emulation","feed_subtitle":"Minimizing energy distance to the uniform cube keeps GP predictions accurate for smooth and rugged surfaces alike.","key_machinery":"The central object is the energy distance $E(F,F_n) = 2\\mathbb{E}\\|X-Y\\| - \\mathbb{E}\\|Y-Y'\\| - \\mathbb{E}\\|X-X'\\|$ between the target uniform distribution $F$ and the empirical distribution $F_n$ of a design. It simultaneously plays three roles: its minimization defines support points; it appears in the Koksma–Hlawka upper bound for the mean-estimation error; and, through Theorem 2, it serves as a surrogate for the maximum-entropy objective for variation estimation under the assumed correlation form. The projected support-point objective replaces the Euclidean norm with a generalized Gaussian kernel weighted by a prior on scale parameters, encoding effect sparsity, hierarchy, and heredity.","core_discovery":"Under the correlation model $r_\\theta(x,y) \\propto \\|x\\|^2 + \\|y\\|^2 - \\|x-y\\|^2$, the entry-wise sum of the projected correlation matrix equals, up to a constant, the energy distance $E(F,F_n)$ between the uniform distribution $F$ and the empirical distribution of the design. Since minimizing the energy distance is exactly the support-point criterion, support points maximize information about the variation term of a Gaussian process while simultaneously optimizing a tight Koksma–Hlawka bound for the mean term, assuming integrable derivatives. The paper further shows that support points approximate a two-parts-minimax, one-part-maximin compromise, and that projected support points under an exponential sparsity prior recover the MaxPro criterion plus a correction that pulls points away from the boundary.","pith_inferences":["Because the theoretical argument depends only on a Hölder-type continuity condition rather than specific correlation values, the same energy-distance design should transfer to Gaussian process models with estimated or unknown correlation lengths; this is a natural extension the paper does not prove.","The two-parts-minimax, one-part-maximin weighting is derived in a small-$q$ limit, so whether a different weighting depending on design size or dimension would improve emulation remains open and is testable.","Minimizing the energy distance is equivalent to minimizing a negative-definite kernel discrepancy, so support points inherit quasi-Monte Carlo integration guarantees for the mean term and may serve double duty as an integration rule in the same experiment.","For non-uniform target distributions $F$, the same construction should yield designs weighted toward regions where prediction accuracy matters most, a direction the paper notes but does not test."],"forward_implications":["Support points and projected support points can be generated efficiently by difference-of-convex programming, so large designs in high dimensions become computationally practical.","In the paper's numerical suite, support points and projected support points achieved the best efficiency in nearly every smooth-versus-rough and all-versus-some-active-factors combination, with projected support points especially strong when only some factors are active.","Support points provide an explicit trade-off between minimax and maximin behavior, with roughly double weight on the minimax (fill-distance) component, jointly addressing nominal and numeric interpolation errors.","Projected support points recover MaxPro designs with an added boundary-repulsion correction, improving projected-space uniformity and emulation for functions with sparse active effects.","The connection to fractional Brownian motion with Hurst index $1/2$ suggests the design supports non-stationary emulation where predictions do not revert to the mean."],"supporting_citations":[{"why":"Defines support points and supplies the support-point theory, including the convergence and integration-error results that Proposition 1 adapts.","marker":"[27]"},{"why":"Defines projected support points and the prior that encodes effect hierarchy, heredity, and sparsity in the generalized Gaussian kernel.","marker":"[25]"},{"why":"Introduces the energy distance that underlies the support-point objective and its goodness-of-fit interpretation.","marker":"[45]"},{"why":"Provides the Koksma–Hlawka inequality framework used to bound the mean-estimation error in terms of the energy distance.","marker":"[7]"},{"why":"Supplies the maximum-entropy sampling objective that Theorem 2 connects to the energy distance via the projected kernel.","marker":"[44]"},{"why":"Defines the MaxPro criterion and the Gershgorin-based surrogate used to approximate maximum-entropy designs, which Theorem 4 recovers from projected support points.","marker":"[18]"},{"why":"Introduces fractional Brownian field emulation, providing the correlation form and the robustness motivation used in Section 3.3.","marker":"[55]"},{"why":"Defines minimax and maximin distance designs and their optimality properties, which Proposition 2 reconciles with support points.","marker":"[17]"}],"fun_headline_variants":["Support points outperform model-tuned designs","Energy distance designs ensure robust GP emulation","Support points: robust emulation without model tuning","Support points optimize GP emulation via energy distance","Robust designs via support points for computer experiments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper replaces the maximum-entropy design objective with a lower bound, assuming that minimizing the sum of all correlation-matrix entries closely approximates maximizing the determinant, but no tightness bound for this lower bound is proved.","fun_headline_variants_meta":{"raw":{"variants":["Support points outperform model-tuned designs","Energy distance designs ensure robust GP emulation","Support points: robust emulation without model tuning","Support points optimize GP emulation via energy distance","Robust designs via support points for computer experiments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00034,"raw_usage":{"total_tokens":1804,"prompt_tokens":802,"completion_tokens":1002,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":934}},"tokens_in":418,"tokens_out":1002,"duration_ms":10770,"temperature":1.0,"reasoning_tokens":934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:03:21.687503+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed small design in two dimensions, compute the true maximum-entropy design (maximizing $\\det R_n$) for a Gaussian correlation with moderate correlation length and compare its integrated prediction error with that of the support-point design across a grid of test functions satisfying the Hölder condition. If support points are consistently and substantially worse than the true maximum-entropy design, the surrogate-objective step is not doing the claimed work. A second check is whether the lower bound in equation (12) is tight: if the design minimizing the sum of correlation entries does not track the design maximizing $\\det R_n$ as $n$ grows, the approximation fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines support points and supplies the support-point theory, including the convergence and integration-error results that Proposition 1 adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the energy distance that underlies the support-point objective and its goodness-of-fit interpretation."},{"cited_title":"Y., and Sloan, I","cited_arxiv_id":null,"evidence_quote":"Provides the Koksma–Hlawka inequality framework used to bound the mean-estimation error in terms of the energy distance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-entropy sampling objective that Theorem 2 connects to the energy distance via the projected kernel."},{"cited_title":"R., Gul, E., and Ba, S","cited_arxiv_id":null,"evidence_quote":"Defines the MaxPro criterion and the Gershgorin-based surrogate used to approximate maximum-entropy designs, which Theorem 4 recovers from projected support points."},{"cited_title":"and Apley, D","cited_arxiv_id":null,"evidence_quote":"Introduces fractional Brownian field emulation, providing the correlation form and the robustness motivation used in Section 3.3."},{"cited_title":"E., Moore, L","cited_arxiv_id":null,"evidence_quote":"Defines minimax and maximin distance designs and their optimality properties, which Proposition 2 reconciles with support points."}],"review_version":1}