{"id":"417f76cb-3797-4690-a936-80175ebf4970","arxiv_id":"2607.07680","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Random sampling maps (with-replacement, binning, species) induce metrics that give uniform any-dimensional generalization and sketching rates for continuous functions on sequences, graphs and tensors.","lead":"This paper gives a unified sampling framework to compare and approximate machine-learning inputs of different sizes (point clouds, graphs, token sequences). It yields explicit rates for how well models trained on small inputs generalize to large ones and how to sketch large inputs for cheap evaluation.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly isolates the two standing hypotheses (compact closed sets and Lipschitz continuity w.r.t. the sampling metric) as the only non-automatic ingredients. Those hypotheses are not hidden; they are the natural price of obtaining uniform rates over infinite-dimensional spaces, and the paper supplies both the abstract compactness criterion (Theorem 2.1) and concrete verifications for the architectures of interest. The improved rates for polynomials and transformers rest on the sampling representation theorems of the companion work [15], which are cited and used correctly. Because the logical chain from sampling concentration to sketching/generalization bounds is complete and the constants are explicit, no load-bearing gap remains that would warrant changing the ACCEPT verdict.","tokens_in":49781,"tokens_out":526,"duration_ms":8163,"concrete_test":"Independently re-derive the sub-Gaussian constant and bias term for W1(Rk(x),Rk∘Rn(x)) in Proposition 3.11 from the coupling of Stam (1978) and the bounded-difference inequality alone, without invoking the later covering-number argument of Lemma 3.12; if the constants match those used in Theorem 3.1, the concentration foundation is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (explicit high-probability sketching and uniform generalization rates for L-Lipschitz functions on compact sets closed under compatible sampling of finite degree) is supported by complete, self-contained arguments. Theorems 2.1 and 2.3 reduce the problem to precompactness of the sampling metric space; Theorems 3.1/4.2 establish the required concentration of dsamp(x,Sn(x)) via bounded-difference inequalities and covering-number estimates (or the elementary empirical-measure bound (27)); Theorems 3.7/4.5 then transfer those rates to Lipschitz functions, with the improved O(n^{-1}) and O(n^{-1/2}) rates following once the function factors through a fixed-dimensional law or moment (Corollaries 3.8–3.10, 4.6–4.9). The standing hypotheses (compactness + closedness under sampling, Lipschitz continuity) are stated explicitly before the main theorems and are verified for the concrete families that motivate the paper. No internal inconsistency, missing step, or unstated assumption that would invalidate the rates was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a unified sampling-based framework for any-dimensional generalization and sketching of functions defined on objects of varying size (sequences, graphs, tensors, etc.). It introduces three families of sampling maps—sampling with replacement, random binning, and species sampling—via FinSet-compatible actions on index sets of finite degree, and defines a sampling metric via Wasserstein distances of the resulting samples. Compactness of the resulting metric spaces (Prop. 2.3, Thm. 2.1) is shown to be equivalent to uniform approximation, sketching, and generalization properties. Explicit high-probability sketching rates and uniform generalization rates e_∞(f,ĝ)≤e_n(f,ĝ)+R_n are proved for L-Lipschitz functions on compact sets closed under the sampling maps (Thms. 3.1, 3.7, 4.2, 4.5); the rates improve to O(n^{-1}) or O(n^{-1/2}) when the function factors through a fixed-dimensional law or moment (Cors. 3.8–3.10, 4.6–4.9). Concrete applications include moment polynomials, graph homomorphism densities/numbers, permutation-invariant transformers, DeepSets, PointNet, and graph neural networks. Any-dimensional data distributions are characterized via projection- or equipartition-consistent sequences (Thms. 3.4, 4.3).","tokens_in":50037,"tokens_out":823,"duration_ms":20712,"significance":"The work supplies a clean, first-principles reduction of any-dimensional learning to classical finite-dimensional generalization once a suitable sampling metric is fixed, together with explicit, non-asymptotic rates that improve substantially on earlier bounds for the same function classes (e.g., O(n^{-1}) versus O(n^{-1/d}) or O((log n)^{-1/4}) for polynomials). The FinSet-action formalism unifies several previously separate literatures (graphons, exchangeable arrays, random quotients, species sampling) and yields immediately usable sketching maps whose size depends only on the target accuracy, not on ambient dimension. The proofs are complete and self-contained (bounded-difference concentration, covering-number estimates, finite de Finetti lemmas from the authors’ prior work), and the standing hypotheses are stated clearly and verified for the motivating families. This is a solid contribution to the mathematical foundations of size-generalization and sketching.","major_comments":[],"minor_comments":[{"comment":"The informal statements of Theorems 1.5 and 1.8 in the introduction omit the precise dependence of the constants on the degree D and the ambient radius r; a short parenthetical reference to the formal statements would help readers who stop at the introduction.","section":null},{"comment":"In Corollary 3.10 the infinite-precision self-attention and mean-pooling assumptions are stated only in a footnote; elevating them to the main text would make the applicability conditions more transparent.","section":null},{"comment":"The covering-number argument in the proof of Theorem 3.1 (after Lemma 3.12) invokes a generic bound from Kolmogorov–Tikhomirov; a one-line remark that the same rates hold under any entropy integral of order o(n^{1/2}) would clarify robustness.","section":null},{"comment":"A few typographical inconsistencies appear (e.g., “Wehavestated” on p. 2, missing spaces after periods in several places). A light copy-edit pass would remove them.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is long and technical, but the core contribution is clean and the proofs are complete. It is a natural fit for a strong probability/statistics or theoretical ML venue. No novelty or citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a solid theory paper that does something useful: it gives a single sampling-based language for comparing objects of different sizes, then turns that into explicit high-probability sketching bounds and uniform any-dimensional generalization rates. The main technical move is treating sampling-with-replacement, random binning, and species sampling as actions of FinSet on index sets of finite degree. Once you have that, precompactness of the sampling metric space (Thm 2.1 + Prop 2.3) reduces both problems to concentration of dsamp(x, Sn(x)), which they get via bounded differences and covering numbers (or the elementary empirical L2 bound for species sampling). The rates then transfer immediately to L-Lipschitz functions, and improve to O(n^{-1}) or O(n^{-1/2}) once the function only sees a fixed-dimensional law or moment.\n\nWhat is new is the unification plus the improved rates. Earlier transferability work (including their own) gave slower rates for the same classes; here moment polynomials and homomorphism densities get O(n^{-1}), transformers get the expected Wasserstein rate O(n^{-1/d}), and DeepSets/GNNs get clean n^{-1/2} under the natural zero-padding symmetries. The de Finetti-type characterizations of the induced any-dimensional data distributions are clean and match the classical pictures for exchangeable arrays and graphons. Proofs are complete and self-contained; the finite de Finetti lemmas are cited from their prior work but stated independently.\n\nSoft spots are real but proportional. The standing hypotheses (compact sets closed under the sampling maps, Lipschitz continuity in the sampling metric) have to be checked case-by-case; they do it for hypercubes, simplices, unweighted graphs, and the listed architectures, but it is not automatic. The general Lipschitz rates are slow (log-log type for high-degree actions), which is expected from regularity lemmas, and the paper is honest about that. No circularity, no missing steps, no load-bearing gaps.\n\nThis is for people who already care about graphons, exchangeable arrays, or size generalization of transformers/GNNs. It is not a rewrite of a field, but it is the cleanest quantitative package I have seen for these questions. I would send it to referees without hesitation; the math is careful and the rates are immediately usable.","headline":"Clean, usable theory of size generalization and sketching via FinSet sampling maps, with improved rates for the architectures people actually use.","tokens_in":50597,"tokens_out":581,"would_cite":true,"duration_ms":8504,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","68T05","60B10","05C80"],"pacs":[],"model":"grok-4.5","headline":"Random sampling maps turn size differences into a metric, giving explicit rates for training small and evaluating large any-dimensional models.","keywords":["any-dimensional learning","sampling metric","generalization","sketching","exchangeability","random binning","species sampling","graph limits"],"falsifier":"Construct a sequence of compact sets closed under sampling and an L-Lipschitz function for which the observed |f(x)−f(S_n(x))| fails to decay at the claimed rate (or fails to concentrate) as n grows, or exhibit a natural any-dimensional model whose continuity modulus with respect to the sampling metric is infinite.","tokens_in":50711,"feed_emoji":"🎲","tokens_out":1023,"duration_ms":19220,"temperature":0.7,"pith_summary":"Machine-learning models for point clouds, token sequences, and graphs must work on inputs of every size, yet they are trained only on bounded sizes and are expensive on large ones. The paper shows that both problems reduce to comparing objects of different sizes by drawing random samples of them and measuring Wasserstein distance between the sample laws. Three families of sampling maps—generalized sampling with replacement, random binning, and species sampling—are matched to the natural symmetries of each domain (duplication versus zero-padding). On compact sets closed under the chosen sampling, every Lipschitz function of the resulting sampling metric admits high-probability sketching bounds and uniform generalization bounds that vanish with sketch size; the rates improve to 1/n or 1/√n once the function depends only on a fixed-dimensional sample or moment. Concrete rates are recovered for moment polynomials, graph homomorphism densities and numbers, permutation-invariant transformers, DeepSets, PointNet, and graph neural networks.","feed_headline":"Sampling turns size gaps into rates for growing models","feed_subtitle":"Train on small inputs, sketch large ones; explicit bounds follow for transformers, GNNs and graph densities","key_machinery":"The sampling metric dsamp(x,y)=∑_k 2^{-k} W_1(S_k(x),S_k(y)), built from random sampling maps S_k that arise from actions of maps between finite sets on index sets of finite degree; compactness of the completion plus concentration of dsamp(x,S_n(x)) produce the rates.","core_discovery":"If a sequence of compact sets is closed under a compatible sampling map of finite degree, the induced sampling metric makes the space of all sizes precompact; every function that is Lipschitz continuous with respect to that metric then satisfies explicit high-probability sketching bounds |f(x)−f(S_n(x))| and uniform generalization bounds e_∞(f,ĝ)≤e_n(f,ĝ)+R_n with R_n→0. The rates become O(n^{-1}) or O(n^{-1/2}) as soon as f depends only on the law or moments of a fixed-size sample.","pith_inferences":["The same sampling metric may furnish a practical early-stopping or curriculum criterion: keep increasing training size until the empirical remainder R_n falls below a target tolerance.","Domains whose natural symmetries mix duplication and zero-padding (e.g., sparse graphs with occasional dense cores) may require hybrid sampling maps not yet covered by the three families.","Because the rates are uniform over all Lipschitz functions, they immediately supply dimension-free sketching guarantees for any future architecture proved continuous in one of the three sampling metrics."],"forward_implications":["A model trained only up to size n automatically controls error on every larger size once the training error and the explicit remainder R_n are known.","Evaluating a permutation-invariant transformer or GNN on an N-token or N-node input can be replaced, with high probability, by evaluation on a fixed-size random sketch whose size depends only on the desired accuracy.","Any-dimensional data distributions are precisely the sequences obtained by sampling a single random limit object; average-error generalization rates follow for free.","Moment polynomials, graph densities/numbers, DeepSets, PointNet and polynomial GNNs inherit the improved 1/n or 1/√n rates without further analysis."],"fun_headline_variants":["Sampling maps turn size gaps into explicit rates for any-dim models","Random sampling yields sketching and generalization bounds across sizes","Compatible sampling makes size-varying domains precompact for Lipschitz f","Finite-degree sampling gives O(1/n) rates for graphs, sequences and tensors","Sketch large inputs by sampling to match small-train continuous functions"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The sets of objects of each size must be compact and closed under the chosen sampling maps, and the functions of interest must be Lipschitz continuous with respect to the resulting sampling metric.","fun_headline_variants_meta":{"raw":{"variants":["Sampling maps turn size gaps into explicit rates for any-dim models","Random sampling yields sketching and generalization bounds across sizes","Compatible sampling makes size-varying domains precompact for Lipschitz f","Finite-degree sampling gives O(1/n) rates for graphs, sequences and tensors","Sketch large inputs by sampling to match small-train continuous functions"]},"model":"grok-4.5","effort":"low","cost_usd":0.005916,"raw_usage":{"total_tokens":1604,"prompt_tokens":831,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":59160000,"prompt_tokens_details":{"text_tokens":831,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":680,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":831,"tokens_out":93,"duration_ms":6602,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T18:19:12.602159+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct a sequence of compact sets closed under sampling and an L-Lipschitz function for which the observed |f(x)−f(S_n(x))| fails to decay at the claimed rate (or fails to concentrate) as n grows, or exhibit a natural any-dimensional model whose continuity modulus with respect to the sampling metric is infinite.","supporting_citations":[],"review_version":2}