{"id":"6f224dec-8f8c-42fd-8b30-b7f8058c8ec6","arxiv_id":"1908.10828","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A general theorem shows that neural networks inherit the absence of the curse of dimensionality from any discrete Monte Carlo scheme they can emulate, with applications to Kolmogorov PDEs.","lead":"This paper proves a general transfer theorem: if a function can be approximated by a Monte Carlo scheme without the curse of dimensionality, and if neural networks can mimic that scheme, then neural networks can approximate the function itself without the curse of dimensionality. The authors apply this to Kolmogorov PDEs, extending earlier results by giving explicit dimension and accuracy exponents.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Theorem 2.3 invalidly drops a positive ||x||^θ term when deriving the growth bound (2.23); the central moment estimates are not established.","rationale":"The reader's weakest_assumption was the composition-closure hypothesis, which is explicitly an assumption and is verified in the Kolmogorov PDE application. The concern I identify is more load-bearing because it is an invalid step in the proof of the central theorem itself. The paper's headline result, Theorem 2.3, is a conditional statement, but its proof uses (2.23) as a key growth bound for the true Monte Carlo step f_{N,d}. The derivation of (2.23) from (2.22) drops a positive term involving ||x||^θ, which is not dominated by the linear right-hand side when θ>1. This is not a matter of constants: a concrete choice of f and R(f^{ε,z}) satisfying (2.11)–(2.13) violates (2.23). The rest of the proof, including the Gronwall bounds on Y and X and the subsequent error estimates, relies on the linear growth of f. The application in Section 4 happens to have the needed structure because f_{N,d}(z,x)=z+x+(T/N)φ1,d(x), so the paper's PDE results may survive a repair, but the abstract theorem as stated lacks the necessary hypothesis. Therefore the central claim, as presented, is not proven. I recommend reject in current form, expecting a revised version with either a corrected proof or an additional growth assumption on f_{N,d}.","tokens_in":46916,"tokens_out":17045,"duration_ms":146423,"concrete_test":"Independently re-derive (2.23) from (2.11)–(2.13). The derivation must fail: exhibit the counterexample with f(z,x)=x+d^{d4}, R(f^{ε,z})(x)=x, d1=d2=0, C=1, N=1, ε=1, θ=2, d4>0, z=0, which satisfies (2.11)–(2.13) but violates (2.23) at x=0. Then check whether the later estimates of Theorem 2.3 can be recovered without (2.23) by controlling the ε ||Y_n||^θ term; if the resulting bound is exponential in N, the theorem needs an additional hypothesis on f_{N,d}.","verdict_should_be":"REJECT","load_bearing_attack":"In the proof of Theorem 2.3, equation (2.23) is asserted to follow from (2.22). However, (2.22) contains the additional nonnegative term ε C d^{d4}(d^{θ(d1+d2)} + ||x||^θ). Since θ ≥ 1 and ||x|| is unbounded, this term is not bounded by the right-hand side of (2.23), which is linear in ||x|| plus a constant. The subsequent moment bounds (2.26)–(2.28), and hence the error estimates (2.30), (2.38), and the final parameter bound (2.16), all depend on (2.23). The hypotheses (2.11)–(2.13) do not imply (2.23): take d1=d2=0, C=1, N=1, ε=1, θ=2, d4>0, f(z,x)=x+d^{d4}, R(f^{ε,z})(x)=x, and z=0; then (2.11)–(2.13) hold, but (2.23) fails at x=0. A correct proof would require an explicit growth assumption on f_{N,d}, such as ||f_{N,d}(z,x)|| ≤ (1+C/N)||x|| + C d^{d2}(d^{d1}+||z||). This is satisfied by the Euler scheme in Section 4 but is not among the hypotheses of Theorem 2.3. Without such an assumption, the abstract transfer theorem is not established as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an abstract transfer theorem (Theorem 2.3) asserting that if a function can be approximated by a suitable discrete Monte Carlo scheme without the curse of dimensionality, and if the steps of that scheme can be realized by DNNs without the curse of dimensionality, then the function itself can be approximated by DNNs without the curse of dimensionality. The theorem also provides explicit polynomial exponents in the dimension d and in the reciprocal accuracy ε. The main application is to Kolmogorov PDEs with constant diffusion and possibly nonlinear drift, for which the authors prove in Theorem 4.5 that viscosity solutions can be approximated by rectified DNNs without the curse of dimensionality. The proof combines discrete Gronwall-type moment bounds with a Monte Carlo Euler error analysis and an ANN calculus developed in Section 3.","tokens_in":47205,"tokens_out":7076,"duration_ms":69102,"significance":"If the central theorem were established, the paper would be a valuable contribution: it systematizes a widely used two-step argument (Monte Carlo scheme first, DNN realization second), gives explicit exponents in the parameter-count bounds, and extends prior Kolmogorov PDE results by making the exponents explicit. The ANN calculus in Lemmas 3.29–3.30 is coherent and useful, and the Kolmogorov PDE application is well structured. However, the main proof contains a gap in the derivation of the growth bound (2.23) from (2.22), and that gap is load-bearing for the moment estimates and the final conclusion. The abstract theorem is therefore not established as written, although the argument appears repairable by adding an explicit linear-growth hypothesis on the discrete scheme, which is satisfied by the Euler scheme used in Section 4.","major_comments":[{"comment":"The deduction of (2.23) from (2.22) is invalid: the term ε C d^{d4}(d^{θ(d1+d2)} + ‖x‖^θ) in (2.22) is nonnegative and is simply discarded. Since θ ≥ 1 and ‖x‖ is unbounded, this term is not bounded by the right-hand side of (2.23), which is affine in ‖x‖. The subsequent moment bounds (2.26)–(2.28), the perturbation estimate (2.30), the error bound (2.38), and the final parameter bound (2.16) all depend on (2.23). The hypotheses (2.11)–(2.13) do not imply (2.23); for instance, take d1 = d2 = 0, C = 1, N = 1, ε = 1, θ = 2, d4 > 0, f_{N,d}(z,x) = x + d^{d4}, R(f^{ε,z}_{N,d})(x) = x. Then (2.11)–(2.13) hold, but (2.23) fails at x = 0 for large d. The proof can be repaired by adding an explicit growth assumption such as ‖f_{N,d}(z,x)‖ ≤ (1 + C/N)‖x‖ + C d^{d2}(d^{d1} + ‖z‖), which is satisfied by the Euler scheme in Section 4; without such an assumption, Theorem 2.3 is not established as stated.","section":"§2.2, Eqs. (2.22)–(2.23)"}],"minor_comments":[{"comment":"The assumption on the composed networks is stated as \"D(φz) = D(φz)\" with two bound variables named z; this is tautological as written. It should presumably read D(φz) = D(φ_z̃) for all z, z̃ ∈ R^d, expressing that the depth does not depend on the point z.","section":"Theorem 2.3, composition-closure assumption"},{"comment":"The notation N_{d,ε} is used both for the set of networks in the theorem statement and for the integer step-size parameter in (2.18). This overloading makes the proof harder to follow; a different symbol, such as N̂_{d,ε}, would help.","section":"Proof of Theorem 2.3, Eq. (2.18)"},{"comment":"The width condition l_{2,L2−1} ≤ l_{1,L1−1} + i is essential for the composition to remain in the admissible class, but it is not explained. A short remark on why this condition is natural for the ANN calculus would improve readability.","section":"Lemma 3.30"},{"comment":"There are several typographical artifacts, such as \"www.univie.ac.at\" split as \"uni vie.ac.at\" and the notation \"(4eC+1C3)\" which should read \"(4e^{C+1} C^3)\". These should be corrected in a revised version.","section":"General typography"}],"recommendation":"major_revision","confidential_remarks":"The gap in Theorem 2.3 is genuine and central, but it is localized and likely repairable by adding a linear-growth hypothesis on f_{N,d} that the authors' own Euler scheme satisfies. If the authors fix this, the paper should be re-evaluated; I would not recommend rejection on the basis of this gap because the overall architecture and the Kolmogorov PDE application remain credible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this paper is the first to state a general meta-theorem that turns no-curse-of-dimensionality for Monte Carlo schemes into no-curse-of-dimensionality for DNN approximations, with explicit exponents. The ANN calculus in Section 3 is competent and reusable, and the Kolmogorov PDE application is a genuine generalization of earlier specialized results. Second, the proof of Theorem 2.3 has a load-bearing gap. At (2.22) they derive ||f|| ≤ ε C d^{d4}(d^{θ(d1+d2)} + ||x||^θ) + (1+C/N)||x|| + C d^{d2}(d^{d1}+||z||). Then (2.23) simply drops the first term. That term is nonnegative, so the inequality does not follow. The stress-test counterexample is correct: f(z,x)=x+d^{d4}, with R(f^{ε,z})(x)=x, satisfies (2.11)-(2.13) but violates (2.23). Since (2.26)-(2.28) are built on (2.23), the moment bounds and the final error and parameter estimates are not established. The fix is to add a hypothesis like ||f_{N,d}(z,x)|| ≤ (1+C/N)||x|| + C d^{d2}(d^{d1}+||z||); the Euler scheme in Section 4 satisfies this, so the application is safe.\n\nEverything else checks out. The transfer assumptions (2.9)-(2.15) are numerous but precisely stated; the composition-closure hypothesis is indeed the delicate one, and Section 4 does the work to verify it. The citations to prior specialized results are fair.\n\nWho is this for? People working on high-dimensional PDE approximation with networks will want the machinery, but they should not cite Theorem 2.3 as a black box until the missing growth condition is added. I would send it to a serious referee; the flaw is fixable and the meta-result is worth having. My own verdict would be revise, not accept.","headline":"The abstract transfer theorem is not proven as written—(2.23) drops a nonnegative term—but the ANN calculus and the Kolmogorov application are real work and worth a revision.","tokens_in":47782,"tokens_out":2968,"would_cite":false,"duration_ms":29658,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C05","65C30","68T07","35K15","41A25"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves a transfer theorem: if a discrete Monte Carlo scheme approximates a function without the curse of dimensionality and the scheme's components are DNN-approximable at polynomial cost, then the function is DNN-approximable…","keywords":["deep neural networks","Monte Carlo algorithms","curse of dimensionality","Kolmogorov PDEs","Euler scheme","rectifier networks","approximation rates","tractability"],"falsifier":"A concrete way to test the theorem: produce a discrete Monte Carlo scheme that satisfies the approximation, moment, and Lipschitz assumptions (2.9)–(2.13) and (2.14)–(2.15), but for which every DNN family approximating the update map forces the composition-closure parameter overhead to grow faster than any polynomial in $d$ or $\\varepsilon^{-1}$, while the scheme itself still has polynomial-in-$d$, polynomial-in-$1/\\varepsilon$ error. Such an example would show that the inheritance conclusion (2.16) fails in general, and it would isolate the composition-closure hypothesis as the obstruction.","tokens_in":46662,"feed_emoji":"🧮","tokens_out":12513,"duration_ms":108495,"temperature":0.7,"pith_summary":"Deep neural networks are known to approximate solutions of certain PDEs without the curse of dimensionality, and one recurring proof strategy is to show that a DNN can mimic a tractable Monte Carlo scheme. This paper extracts that strategy into a general transfer theorem. It proves that if a function can be approximated by a discrete Monte Carlo scheme with error bounded polynomially in the dimension and inverse accuracy, and if the scheme's update and averaging maps can themselves be realized by DNNs with polynomial parameter cost and a composition-closure property, then the function has DNN approximations with the same qualitative tractability. The theorem gives an explicit exponent for the number of parameters in terms of the scheme's exponents. Applied to Kolmogorov PDEs with constant diffusion and possibly nonlinear Lipschitz drift, it yields DNN approximations at $L^p$ error $\\varepsilon$ with parameter count bounded polynomially in $d$ and $\\varepsilon^{-1}$.","feed_headline":"If Monte Carlo beats the curse of dimensionality, DNNs inherit that","feed_subtitle":"A transfer theorem turns polynomial Monte Carlo error bounds into polynomial DNN parameter bounds.","key_machinery":"The machinery is an explicit ANN calculus: formal operations of composition, parallelization, sum, and scalar multiplication on feedforward networks with rectifier activations, each with controlled growth of the parameter count (Definitions 3.1–3.28 and Lemmas 3.8–3.29). The load-bearing object is the class $\\mathcal{N}_{d,\\varepsilon}$ of admissible networks, together with a composition-closure hypothesis: for every admissible $\\Phi$ and every sampled point $z$, there must be networks $\\varphi_z$ realizing the composition $(R(f_{N,d}^{\\varepsilon,z})) \\circ (R(\\Phi))$ with parameter overhead at most $C N^{n_1} d^{d_3} \\varepsilon^{-e}$ and depth independent of $z$. A second construction, Lemma 3.29, builds a network whose output is the average of $M_N$ independent network evaluations, and Lemma 3.30 verifies the closure condition for the Kolmogorov application by composing with an exact identity network and a scaled drift network while controlling the width of the composed network. Proposition 4.4 supplies the underlying Monte Carlo Euler error estimates.","core_discovery":"The central claim is Theorem 2.3: under the moment, approximation, Lipschitz, and composition-closure assumptions (2.9)–(2.15), for every $d \\in \\mathbb{N}$ and $\\varepsilon \\in (0,1]$ there are rectified DNNs $\\Psi_{d,\\varepsilon}$ with $\\left(\\int_{\\mathbb{R}^d} |u_d(x) - (R(\\Psi_{d,\\varepsilon}))(x)|^p \\, \\nu_d(dx)\\right)^{1/p} \\leq \\varepsilon$ and parameter count bounded by $c \\, d^{d_0(n_1+n_2+1)/n_0 + d_3 + e \\delta} \\varepsilon^{-(n_1+n_2+1)/n_0 - e}$, where $\\delta = \\max\\{d_5 + \\theta(d_1+d_2),\\, d_4 + d_6 + 2\\theta(d_1+d_2)\\}$. The proof replaces the Monte Carlo update $f_{N,d}$ by its DNN approximant, runs the recursion through composed networks, evaluates a DNN realization of the averaging function $g$, and chooses the number of steps $N$ and an internal accuracy $E_{d,\\varepsilon}$ so that the Monte Carlo error, the $g$-approximation error, and the perturbation error from replacing $f$ by a network all stay below $\\varepsilon$. The conclusion is a direct inheritance: the discrete scheme's polynomial-in-$d$, polynomial-in-$1/\\varepsilon$ error becomes a polynomial parameter bound for the DNN.","pith_inferences":["The transfer statement suggests a general recipe: any future Monte Carlo sampler that is tractable and whose components are network-friendly in the composition sense automatically yields DNN tractability; samplers such as multilevel or quasi-Monte Carlo variants could be compared through the induced parameter exponents.","The composition-closure hypothesis is likely the practical bottleneck: for nonlinear update maps it requires controlling the width of composed networks, not just their depth, so schemes with high-dimensional or ill-conditioned updates may fail the hypothesis even when they are otherwise tractable.","The explicit exponents imply that improving the scheme's accuracy in $\\varepsilon$, the dimension growth of the moments, or the Lipschitz constants directly improves the DNN parameter bound, giving quantitative targets for designing network-friendly samplers."],"forward_implications":["Any function tractably approximated by a discrete Monte Carlo scheme whose components admit polynomial-cost DNN representations inherits DNN tractability, with explicit parameter exponents (Theorem 2.3).","Solutions of Kolmogorov PDEs with constant diffusion matrix, possibly nonlinear Lipschitz drift, and terminal data that are Lipschitz and polynomially growing are DNN-approximable without the curse of dimensionality (Theorem 4.5).","On the unit cube with uniform measure, the parameter count for these PDE solutions is at most $c \\, \\varepsilon^{-(e+6)}$ times a polynomial in $d$ whose exponent is given in closed form by the data-approximation constants (Corollary 4.6).","The result makes explicit the dimension and accuracy exponents for classes previously known only through existence results, including Black-Scholes PDEs, semilinear heat equations, and nonsmooth value functions in zero-sum games."],"supporting_citations":[{"why":"Supplies the ANN calculus (compositions, parallelizations, sums, scalar multiplications) used to construct all networks in the transfer proof.","marker":"[22]"},{"why":"Establishes DNN tractability for Kolmogorov PDEs with constant diffusion and nonlinear drift; its Euler error estimates and Corollary 2.4 are used inside Theorem 2.3 and Theorem 4.5.","marker":"[32]"},{"why":"Provides the Black-Scholes DNN tractability result and the Monte Carlo error corollary that form the template for the transfer argument.","marker":"[21]"},{"why":"Proves DNN tractability for semilinear heat equations with an explicit accuracy exponent; Theorem 4.5 extends this to more general Kolmogorov PDEs with explicit dimension exponents.","marker":"[30]"},{"why":"Treats nonsmooth value functions in zero-sum games; Theorem 4.5 generalizes the class of Kolmogorov PDEs while making the exponents explicit.","marker":"[46]"}],"fun_headline_variants":["DNNs inherit Monte Carlo's curse-of-dimensionality escape","When Monte Carlo avoids the curse, DNNs can too","Transfer theorem: DNNs get Monte Carlo's dimension-free approximation","DNNs adopt Monte Carlo's polynomial error bounds","From Monte Carlo to DNNs: no curse of dimensionality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the composition-closure condition: every admissible network must remain admissible after composition with the DNN version of the Monte Carlo update, with parameter overhead at most $C N^{n_1} d^{d_3} \\varepsilon^{-e}$ and depth independent of the sampled point; if this width and depth control fails, the transfer theorem does not apply.","fun_headline_variants_meta":{"raw":{"variants":["DNNs inherit Monte Carlo's curse-of-dimensionality escape","When Monte Carlo avoids the curse, DNNs can too","Transfer theorem: DNNs get Monte Carlo's dimension-free approximation","DNNs adopt Monte Carlo's polynomial error bounds","From Monte Carlo to DNNs: no curse of dimensionality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3372,"prompt_tokens":1178,"completion_tokens":2194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":794,"completion_tokens_details":{"reasoning_tokens":2109}},"tokens_in":794,"tokens_out":2194,"duration_ms":15134,"temperature":1.0,"reasoning_tokens":2109,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:31:48.742844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete way to test the theorem: produce a discrete Monte Carlo scheme that satisfies the approximation, moment, and Lipschitz assumptions (2.9)–(2.13) and (2.14)–(2.15), but for which every DNN family approximating the update map forces the composition-closure parameter overhead to grow faster than any polynomial in $d$ or $\\varepsilon^{-1}$, while the scheme itself still has polynomial-in-$d$, polynomial-in-$1/\\varepsilon$ error. Such an example would show that the inheritance conclusion (2.16) fails in general, and it would isolate the composition-closure hypothesis as the obstruction.","supporting_citations":[{"cited_title":"A proof that rectified deep neural networks overcome the curse of dimensionality in the numerical approximation of semilinear heat equations","cited_arxiv_id":"1901.10854","evidence_quote":"Proves DNN tractability for semilinear heat equations with an explicit accuracy exponent; Theorem 4.5 extends this to more general Kolmogorov PDEs with explicit dimension exponents."}],"review_version":1}