{"id":"33f2eb29-036c-4c6a-8e3d-59ea3baff249","arxiv_id":"2505.24066","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hierarchical prior on grid resolution makes grid-interpolated Gaussian process regression minimax-rate-adaptive over Holder classes, up to logarithmic factors.","lead":"This paper proves that grid-interpolated Gaussian processes, when given a heavy-tailed prior on the number of grid points, can learn smooth functions at the statistically fastest possible rate without knowing the smoothness in advance. It also shows that SPDE-based finite-element approximations have a fixed smoothness ceiling that prevents such adaptation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix B.2 proves Theorem 3.5 with d=1 small-ball exponents; for d>1 the displayed KL and entropy bounds are weaker than needed by a power of eps, so the multivariate adaptive rate is not established as written.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the most load-bearing weakness is not primarily the simulation/prior mismatch. That mismatch is real and means the numerics do not validate the theorem, but it does not threaten the theorem itself. The proof gap in Appendix B.2 is more serious: Theorem 3.5 is stated for general d, yet the displayed KL prior-concentration bound and the metric-entropy bound both use exponents appropriate to d=1. If those exponents are taken literally, the proof yields only the d=1 rate n^{-alpha/(2alpha+1)} for all d, slower than the claimed n^{-alpha/(2alpha+d)}. The construction may still be correct after restoring the missing d factors and redoing the balancing, which is why this warrants a conditional acceptance rather than rejection: the central idea is credible and the dimensional fix is plausible, but a corrected proof is required before the multivariate claim can be accepted. My test is deliberately minimal: track kappa^d through the integral and recompute the entropy term; this decides whether the issue is typographical or substantive. I agree partly with the reader because their identified weakest assumption (non-compliant simulations) is valid but secondary; the proof-dimensionality issue is the load-bearing concern.","tokens_in":37520,"tokens_out":5563,"duration_ms":58991,"concrete_test":"Re-derive the KL lower bound in Section B.2 with explicit dimension d: in the integral over kappa, keep the term kappa^d log^{1+d}(2kappa/eps) and substitute kappa ~ eps^{-1/alpha}; verify whether the surviving exponent is eps^{-d/alpha}. Then rebalance n eps^2 against -log of the resulting bound: if the rate is n^{-alpha/(2alpha+d)} up to log factors, the gap is typographical; if the exponent remains eps^{-1/alpha} (i.e., d is dropped), the proof does not support the theorem for d>=2. Independently recompute the metric-entropy bound for F_N^G using the d-dimensional entropy estimate of [39] and check whether it is eps^{-d/alpha} rather than eps^{-1/alpha}.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central theorem claims the GPI prior contracts at the minimax rate n^{-alpha/(2alpha+d)} for f* in C^alpha[0,1]^d. The proof in Appendix B.2, however, repeatedly uses one-dimensional small-ball exponents. The prior concentration estimate is first quoted as exp[-K1 kappa^d {log}^{1+d}], which is correct in d dimensions, but after integrating kappa over [(C0/eps)^{1/alpha}, 2(C0/eps)^{1/alpha}], the displayed lower bound collapses to exp(-C4 eps^{-1/alpha} log^C) instead of exp(-C4 eps^{-d/alpha} log^C). Likewise, the metric-entropy bound is quoted as log N(eps, F_N^G, ||.||_infty) ≲ eps^{-1/alpha}, whereas for a d-dimensional alpha-Holder class the correct d-dimensional statement from the cited framework is eps^{-d/alpha}. For d=1 the two exponents coincide, so the argument works; for d>1, balancing the written exponent eps^{-1/alpha} against n eps^2 gives only the rate n^{-alpha/(2alpha+1)}, which is slower than the claimed n^{-alpha/(2alpha+d)}. The theorem and proof therefore contain a real dimensional gap: either the displayed exponents are typos and the d factors must be restored throughout, or the multivariate statement is not proved. This is more central than the simulation/prior mismatch flagged by the reader, because it bears directly on the theorem's scope.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies finite-rank Gaussian process priors of the form f_N = sum_j w_j psi_j with dependent Gaussian coefficients and locally supported basis functions, focusing on two constructions: a finite-element/SPDE approximation of Matérn GPs and a regular-grid interpolation (GPI) prior induced by a parent GP. The main theoretical claims are that the SPDE construction has a suboptimal posterior contraction rate for sufficiently smooth true functions satisfying a uniform nonlinearity condition, while the GPI construction, equipped with independent hierarchical priors on the inverse bandwidth and the resolution parameter N, attains the adaptive minimax rate n^{-alpha/(2alpha+d)} up to logarithmic factors over alpha-Hölder classes. The paper also develops a Metropolis-within-Gibbs sampler for the hierarchical GPI model and reports simulations comparing the methods with full GP regression.","tokens_in":37896,"tokens_out":15635,"duration_ms":148941,"significance":"If the main results are correct, the paper makes a useful contribution: it shows that a finite-rank interpolation prior can inherit the adaptive minimax behavior of the full Gaussian process while reducing computational cost, and it provides quantitative evidence of the limitations of finite-element SPDE approximations. The paper contains substantial technical work, including explicit small-ball probability bounds for the SPDE prior, fluctuation bounds for interpolated Gaussian processes, and a concrete sampling scheme with Toeplitz/BTTB structure. The central rate result is built on external benchmarks from van der Vaart and van Zanten rather than on a circular use of the authors' own fitted quantities. However, the d-dimensional proof of Theorem 3.5 has a dimensional gap that currently limits the theorem to d=1, and the statement of Theorem 3.4 appears inconsistent with its proof; these issues need to be repaired before the central claims can be accepted.","major_comments":[{"comment":"The proof does not establish the claimed d-dimensional rate. The prior-concentration estimate is correctly quoted with a kappa^d factor, but after restricting kappa to the interval [(C0/epsilon)^{1/alpha}, 2(C0/epsilon)^{1/alpha}], the displayed lower bound becomes exp(-C4 epsilon^{-1/alpha} log^C epsilon) instead of exp(-C4 epsilon^{-d/alpha} log^C epsilon), because kappa^d is of order epsilon^{-d/alpha} on that interval. Similarly, the metric-entropy conclusion is stated as log N(epsilon, F_N^G, ||.||_infty) lesssim epsilon^{-1/alpha}; in the cited bound of [39] the factor is r^d, and with r as epsilon^{-1/alpha} this gives epsilon^{-d/alpha}. The complement-probability bound is also displayed with epsilon^{-1/alpha}. Balancing the written exponent epsilon^{-1/alpha} against n epsilon^2 yields only the rate n^{-alpha/(2alpha+1)}, which is slower than the claimed n^{-alpha/(2alpha+d)} for d>1. As written, the theorem is proved only for d=1; please restore the missing d factors throughout, or restrict Theorem 3.5 to d=1.","section":"B.2 (proof of Theorem 3.5)"},{"comment":"The statement of Theorem 3.4 and its proof appear to point in opposite directions. The proof derives a small prior-ratio bound P(||f_N - f*||_2 < xi_n)/P(||f_N - f*||_infty < epsilon_n) lesssim exp(-2 n epsilon_n^2) and concludes that xi_n is a lower bound on the contraction rate. In the Castillo-type lower-bound argument on which the proof is modeled, such a ratio bound implies that the posterior mass in the L2-ball of radius xi_n tends to zero, not to one. As printed, equation (9) states E_{f*} Pi_N(||f_N - f*||_2 < epsilon_n | D_n) -> 1, which is the opposite conclusion. Please correct (9) and the surrounding text so that the theorem's statement matches the lower-bound argument, and specify precisely whether the intended claim is that the posterior cannot contract faster than the displayed rates.","section":"Theorem 3.4 and B.1"},{"comment":"The numerical studies do not satisfy the assumptions of Theorem 3.5, so the claim that they support the theory is currently overstated. The GPI simulations use a Matérn parent kernel (stated in the first paragraph of Section 5), whereas Theorem 3.5 is proved for the squared-exponential kernel (10). In addition, the resolution priors are p(N) proportional to exp(-4N) on {4,8,12,20,30} in Section 5.1 and a discrete uniform prior on {6,8,10,14,18} in Section 5.3; neither satisfies the polynomial-tail condition (12). Please either run simulations within the theorem's scope or explicitly frame the numerics as an exploratory check rather than as a validation of the rate theorem.","section":"Section 5"}],"minor_comments":[{"comment":"There are several typographical and notational issues: equation (10) contains '||x|\\|_2^2', 'Topelitz' appears for 'Toeplitz' in Section 2.2, and Remark 3.1 contains 'appoach'. The Hölder space notation is also inconsistent in places (C^alpha versus C^{\\alpha}).","section":"Throughout"},{"comment":"The first sufficient condition is displayed as P(||f_N - f*||_infty < epsilon) greatersim exp{-epsilon^{d/(2alpha+d)} log^C(1/epsilon)}, which has a positive exponent and cannot be the intended lower bound; it should presumably read exp(-epsilon^{-d/(2alpha+d)} log^C(1/epsilon)) or an equivalent negative-exponent form.","section":"B.2, equation (49)"},{"comment":"Lemma 3.2 is stated only for d=1, while Theorem 3.5 is for general d. Please state whether a d-dimensional analogue holds, or clarify that the lemma is used only as motivation for the choice N ~ epsilon^{-1/(alpha wedge 2)}.","section":"Lemma 3.2"},{"comment":"The complexity statement O(n + N^{3d}) is correct only if N denotes the number of grid points per coordinate; please make this explicit, since the total number of basis functions is (N+1)^d and a reader could otherwise infer a different scaling.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The d-dimensional gap in the proof of Theorem 3.5 is the main obstacle. It appears fixable by restoring the missing d factors in the small-ball, entropy, and complement-probability bounds, and the d=1 case of the theorem seems sound. The direction of Theorem 3.4 also needs clarification, since the proof supports a lower-bound statement rather than the displayed (9). I would not recommend rejection, as the overall approach and most technical components are promising, but the manuscript requires a careful correction of these two load-bearing points and a revision of the numerical-support claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper has a genuinely useful d=1 result and an interesting SPDE suboptimality result, but the headline multivariate adaptive rate is not established. The proof in Appendix B.2 repeatedly uses one-dimensional small-ball exponents. The prior concentration lower bound after integrating out kappa comes out as exp(-C epsilon^{-1/alpha}) rather than exp(-C epsilon^{-d/alpha}); same for the metric entropy, which is quoted as epsilon^{-1/alpha} instead of epsilon^{-d/alpha}. Balancing those gives only the d=1 rate n^{-alpha/(2alpha+1)}. So Theorem 3.5 as stated for general d is unsupported. The d=1 special case likely works.\n\nWhat's good: the framework is sensible—dependent Gaussian coefficients on locally supported bases—and the paper correctly identifies that this is the gap in the literature. The SPDE lower bound (Theorem 3.4) is a real contribution; it shows that finite-element SPDE approximations inherit the full Matérn GP's suboptimality for smooth targets, plus a log factor. The sampler for the hierarchical grid interpolation model is also a useful methodological addition.\n\nThe soft spots beyond the proof gap: the simulations use a Matérn parent kernel and bounded or exponential priors on N, which violate the theorem's assumptions. That's a mismatch the authors should fix by either running the theorem-compliant priors (polynomial tail, squared-exponential) or explicitly labeling the numerics as heuristic. No code is provided. The UN condition is restrictive but the paper acknowledges it.\n\nBottom line: this deserves a serious referee, but the multivariate claim needs a careful check and possibly a correction. I'd send it back for revision pointing at B.2. If the authors restore the d factors, the paper becomes quite solid; if not, the d=1 version still has value.","headline":"The d=1 story is credible, but the multivariate version of Theorem 3.5 does not follow from the proof as written: the displayed exponents drop the d-dependence.","tokens_in":38377,"tokens_out":4392,"would_cite":false,"duration_ms":40734,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G20","62F15","60G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Putting a prior on grid resolution restores Gaussian-process optimality.","keywords":["Gaussian process","posterior contraction","adaptive inference","minimax rate","finite-rank approximation","SPDE","grid interpolation","hierarchical prior"],"falsifier":"Run the grid-interpolation posterior with a squared-exponential kernel on data from a very smooth function using a prior on $N$ with $P_2(N \\ge x) = K_1 x^{-K_2}$, and measure the empirical $L^2$ error at growing sample sizes; if the error does not decay at essentially $n^{-\\alpha/(2\\alpha+d)}$, Theorem 3.5 is false. Conversely, simulating with an exponentially tailed prior on $N$ and observing a slower rate would indicate condition (12) is not merely technical.","tokens_in":37341,"feed_emoji":"📈","tokens_out":6439,"duration_ms":57799,"temperature":0.7,"pith_summary":"This paper studies cheap approximations to Gaussian-process (GP) regression built from finitely many locally supported basis functions with dependent Gaussian coefficients, and asks whether such approximations can match the statistical efficiency of the full GP. For a grid-interpolation construction with a squared-exponential kernel, the authors prove that placing a hierarchical prior on the resolution $N$ and the inverse bandwidth $\\kappa$ yields the minimax-optimal posterior contraction rate $n^{-\\alpha/(2\\alpha+d)}$ up to logarithmic factors for every $\\alpha$-Hölder truth, without the user knowing $\\alpha$. For the SPDE-based finite-element construction with a fixed Matérn smoothness $\\beta$, they prove the opposite: the posterior concentrates at a strictly slower rate for sufficiently smooth functions regardless of bandwidth and truncation, so it cannot adapt to global smoothness beyond a threshold. The positive result matters because it indicates that a computationally cheap lattice interpolation can inherit the full GP's adaptive efficiency, provided the resolution itself is learned rather than fixed.","feed_headline":"A prior on grid resolution restores Gaussian-process optimality","feed_subtitle":"Finite-rank interpolation with a learned grid size matches the full GP's minimax rate without knowing the smoothness.","key_machinery":"The load-bearing object is the grid-interpolation GP $f_N(x)=\\sum_{0\\le j\\le N} w_j\\psi_j(x)$, where $\\psi_j$ are piecewise-linear tent functions on a regular lattice and the coefficient vector $(w_0,\\dots,w_N)$ is drawn from a parent squared-exponential GP evaluated at the lattice points, so the prior covariance is a Toeplitz or block-Toeplitz matrix. The argument runs through the GP concentration function: the proof lower-bounds the finite-rank prior's small-ball probability $P(\\|f^* - f_N\\|_\\infty < \\epsilon)$ by the full GP's small-ball probability $P(\\|f^* - f\\|_\\infty < \\epsilon/2)$ minus a fluctuation term $P(\\|f - f_N\\|_\\infty > \\epsilon/2)$, which is controlled by a sup-norm continuity estimate for the RBF GP over cells of width $1/N$. A polynomial-tailed prior on $N$ guarantees that with high probability there are enough grid points to make the interpolation error negligible at the relevant $\\epsilon$-scale, while the gamma-type prior on $\\kappa$ supplies the optimal concentration of the parent GP. The SPDE construction is analyzed by the same small-ball calculus, but there the fixed Matérn smoothness $\\beta$ caps the effective regularity and produces the slower lower-bound rate.","core_discovery":"The central claim is Theorem 3.5: for $f^* \\in C^\\alpha[0,1]^d$, if $f_N$ is the regular-grid linear interpolation of a squared-exponential GP with kernel $K(x)=\\exp(-\\kappa^2\\|x\\|^2)$, and independent priors are placed on $\\kappa$ satisfying a gamma-type tail condition and on $N$ satisfying a polynomial-tail condition $P_2(N \\ge x) \\ge K_1 x^{-K_2}$, then the posterior contracts at $\\epsilon_n = n^{-\\alpha/(2\\alpha+d)}(\\log n)^C$ in the empirical $L^2$ metric. This is the same minimax-optimal rate as the unapproximated GP with inverse-gamma bandwidth, obtained adaptively over all $\\alpha$. The mechanism is that for $N$ large enough relative to $\\kappa$, the interpolation error $f_N - f$ is of higher order than the GP's own concentration scale, so the finite-rank posterior inherits the full GP's prior concentration around $f^*$. The paper's contrasting result, Theorem 3.4, shows that the SPDE finite-element prior with Matérn smoothness $\\beta=2$ or even $\\beta \\ge 4$ concentrates at best at $n^{-(\\beta-1/2)/(2\\beta)}(\\log n)^{1/4}$, which is suboptimal for truths smoother than $\\beta - 1/2$; for $\\beta=2$ the lower bound is $(n/\\log n)^{-2/5}$. The paper also provides a Metropolis-within-Gibbs sampler that jointly updates $N$ and $\\kappa$ for the interpolation model, with per-iteration cost $O(n + N^{3d})$, and numerical studies in $d=1$ and $d=2$.","pith_inferences":["The theorem's tail condition on $N$ is at least polynomial, but the simulations use bounded or exponential priors on $N$; extending the proof to sub-polynomial tails would be needed before the numerics sit inside the theorem's scope.","The same interpolation-over-resolution mechanism might transfer to other covariance kernels with known small-ball probabilities, not just the squared-exponential kernel, since the fluctuation bound only needs a modulus-of-continuity control on the parent GP.","A practical reading is that practitioners should favor heavy-tailed or at-least-polynomial priors on the resolution when using grid GP emulators, because the theory ties adaptivity to that tail; discrete uniform or exponential priors may still work but are currently unsupported.","The suboptimality of the SPDE construction suggests that letting the Matérn smoothness itself be random, rather than only $\\kappa$ and $N$, would be a natural next step; the paper does not pursue that direction."],"forward_implications":["With a squared-exponential parent kernel, the grid-interpolation GP with hierarchical priors on $\\kappa$ and $N$ attains the minimax-optimal posterior contraction rate $n^{-\\alpha/(2\\alpha+d)}(\\log n)^C$ adaptively for every $\\alpha>0$, so no oracle knowledge of smoothness is needed.","The SPDE finite-element approximation with fixed Matérn smoothness $\\beta$ cannot adapt to functions smoother than $\\beta-1/2$: its contraction rate is suboptimal by a polynomial factor for such truths.","The hierarchical sampler updates resolution and bandwidth together at per-iteration cost $O(n + N^{3d})$, so the theoretical guarantees come with a tractable computational scheme whose cost is dominated by the resolution rather than the sample size.","Because posterior contraction at rate $\\epsilon_n$ implies existence of a point estimator with the same rate, the result supplies a practical estimation procedure that is both scalable and rate-optimal."],"supporting_citations":[{"why":"Supplies the full GP adaptive concentration result for the squared-exponential kernel that the grid-interpolation proof inherits.","marker":"[39]"},{"why":"Introduces the SPDE finite-element construction whose suboptimal contraction the paper establishes.","marker":"[22]"},{"why":"Introduces the regular-grid linear interpolation GP used as the grid-interpolation construction.","marker":"[23]"},{"why":"Provides the sufficient conditions and concentration-function framework used to prove contraction rates.","marker":"[36]"},{"why":"Gives the minimax benchmark and the Matérn GP rate against which the SPDE lower bound is compared.","marker":"[37]"},{"why":"Provides the independent-coefficient spline prior rate that the $\\beta=2$ SPDE result is compared with.","marker":"[11]"},{"why":"Supplies the lower-bound technique comparing $L^2$ and sup-norm small-ball probabilities used in Theorem 3.4.","marker":"[4]"},{"why":"Supplies the lower-bound lemma for GP posterior rates that underlies the SPDE suboptimality proof.","marker":"[6]"},{"why":"Defines the RKHS concentration function used throughout the theory.","marker":"[38]"}],"fun_headline_variants":["Prior on grid resolution restores minimax GP accuracy","Resolution prior makes finite-rank GPs minimax-optimal","Learned grid size gives full-GP minimax rates","Adaptive resolution fixes finite-rank GP posteriors","Grid-size prior yields optimal contraction for GPs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee of adaptivity rests on condition (12), that the prior on the resolution $N$ has at least a polynomial tail; the paper's own simulations use bounded or exponential priors on $N$, so those numerics do not fall within the theorem's assumptions and cannot directly confirm the rate.","fun_headline_variants_meta":{"raw":{"variants":["Prior on grid resolution restores minimax GP accuracy","Resolution prior makes finite-rank GPs minimax-optimal","Learned grid size gives full-GP minimax rates","Adaptive resolution fixes finite-rank GP posteriors","Grid-size prior yields optimal contraction for GPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1664,"prompt_tokens":1119,"completion_tokens":545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":483}},"tokens_in":735,"tokens_out":545,"duration_ms":5720,"temperature":1.0,"reasoning_tokens":483,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:38:44.829525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the grid-interpolation posterior with a squared-exponential kernel on data from a very smooth function using a prior on $N$ with $P_2(N \\ge x) = K_1 x^{-K_2}$, and measure the empirical $L^2$ error at growing sample sizes; if the error does not decay at essentially $n^{-\\alpha/(2\\alpha+d)}$, Theorem 3.5 is false. Conversely, simulating with an exponentially tailed prior on $N$ and observing a slower rate would indicate condition (12) is not merely technical.","supporting_citations":[{"cited_title":"Adaptive bayesian estimation using a gaussian random field with inverse gamma bandwidth","cited_arxiv_id":null,"evidence_quote":"Supplies the full GP adaptive concentration result for the squared-exponential kernel that the grid-interpolation proof inherits."},{"cited_title":"An explicit link between gaussian fields and gaussian markov random fields: the stochastic partial differential equation approach","cited_arxiv_id":null,"evidence_quote":"Introduces the SPDE finite-element construction whose suboptimal contraction the paper establishes."},{"cited_title":"Gaussian process emulators for computer experiments with inequality constraints","cited_arxiv_id":null,"evidence_quote":"Introduces the regular-grid linear interpolation GP used as the grid-interpolation construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the sufficient conditions and concentration-function framework used to prove contraction rates."},{"cited_title":"Information rates of nonparametric gaussian process methods","cited_arxiv_id":null,"evidence_quote":"Gives the minimax benchmark and the Matérn GP rate against which the SPDE lower bound is compared."},{"cited_title":"Adaptive estimation of multivariate functions using condi- tionally gaussian tensor-product spline priors","cited_arxiv_id":null,"evidence_quote":"Provides the independent-coefficient spline prior rate that the $\\beta=2$ SPDE result is compared with."},{"cited_title":"Anisotropic function estimation using multi-bandwidth gaussian processes","cited_arxiv_id":null,"evidence_quote":"Supplies the lower-bound technique comparing $L^2$ and sup-norm small-ball probabilities used in Theorem 3.4."},{"cited_title":"Lower bounds for posterior rates with gaussian process priors","cited_arxiv_id":null,"evidence_quote":"Supplies the lower-bound lemma for GP posterior rates that underlies the SPDE suboptimality proof."},{"cited_title":"Reproducing kernel hilbert spaces of gaussian priors","cited_arxiv_id":null,"evidence_quote":"Defines the RKHS concentration function used throughout the theory."}],"review_version":1}