{"id":"a36a5e74-b3c0-4415-99f9-27413f7ef28d","arxiv_id":"1908.03833","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves the first space-time error estimates for deep ReLU network approximations of Euler approximations of perturbed differential equations.","lead":"This mathematics paper proves explicit error bounds for deep neural networks that approximate the Euler time-stepping scheme of certain differential equations, including bounds on the network size needed. It is a rigorous step toward understanding when deep learning PDE solvers can be trusted.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's verdict of ACCEPT with high confidence is well supported. The proof is a lengthy but conventional sequence of explicit estimates, and I could not find a load-bearing gap. The reader's weakest_assumption points to the linear growth bound on the drift network; this is indeed used essentially in Lemma 3.11 to obtain the a priori trajectory bound g_n, but for ReLU networks this bound is automatically satisfied (every ReLU network is piecewise linear and globally Lipschitz), so the assumption is not a restrictive or fragile hypothesis. The more substantive limitations are the restriction to Euler approximations (rather than the actual differential equation solution) and the very large network size exponent d^{16+8d}, both of which are explicitly stated in the paper and do not contradict the central claim. Thus my stress-test pass does not alter the reader's verdict.","tokens_in":74400,"tokens_out":28077,"duration_ms":278750,"concrete_test":"Specialize the full construction to the simplest nontrivial case N=1, d=1, y=0, with a single-layer affine drift network R(Φ)(x)=x, and symbolically verify that the network Ψ_0 produced by Propositions 2.32, 3.7, and 3.10 realizes Γ(Π_0(t),x)+Γ(Π_1(t),x+T·x) with the stated ε error and parameter count; this isolates the hat-function interpolation, the product approximation, and the sum-composition calculus in one concrete instance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a conditional existence result: given any ReLU drift network with linear growth and bounded parameter count, the paper constructs a ReLU network approximating the corresponding Euler scheme in space-time with explicit error and size bounds. I checked the main chain of reasoning: the ANN calculus lemmas (compositions, parallelizations, sums) are consistent; the product approximation built from the square-function approximant satisfies the stated error bounds; the hat-function interpolation combined with the product network yields the a posteriori estimate in Proposition 3.10; and the discrete Gronwall argument in Lemma 3.11 correctly produces the a priori bound g_n used to convert a posteriori to a priori estimates. The reader's identified weakest assumption, the linear growth bound ||R(Φ)(x)|| ≤ C(1+||x||), is not actually a fragile premise: every fixed-depth ReLU network is piecewise linear and hence globally Lipschitz, so this bound holds automatically with C equal to the network's Lipschitz constant. The value of C affects the error constants through exp(CT), but no delicate cancellation is required. The main limitation, that the estimates apply to Euler approximations rather than to the PDE/ODE solution itself, is explicitly stated in the introduction and abstract, and it does not undermine the stated central claim. I found no internal inconsistency, circular step, or missing proof in the presented argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to provide the first space-time error estimates for deep ReLU network approximations of Euler approximations of perturbed differential equations. The main result, Theorem 3.12, asserts that for an Euler scheme driven by a ReLU drift network with linear growth, there exist ReLU networks whose space-time map approximates the Euler flow with explicit error bounds of the form eps times a polynomial in d, N, and the a priori growth bound, together with explicit parameter-count bounds polynomial in N and d and quadratic in |log eps|. The proof develops an ANN calculus for compositions, parallelizations, and sums of ReLU networks; constructs product-approximation networks from a square-function approximant; and combines a posteriori error estimates via time hat-function interpolation with a discrete Gronwall a priori bound. Corollary 3.13 and Theorem 1.1 give simplified closed-form bounds, including the y=0 case stated in the introduction. The paper explicitly states that the estimates concern Euler approximations rather than the differential equation or PDE solution itself.","tokens_in":74593,"tokens_out":11181,"duration_ms":115001,"significance":"If correct, the result is a useful foundational contribution: it appears to be the first space-time error estimate, rather than a purely spatial one, for DNN approximation of an Euler scheme, and it supplies reusable ingredients with fully written proofs and explicitly tracked constants. The central statement is a conditional existence result with no free parameters: given any ReLU drift network, the constructed network and the error and size bounds are explicit. The weakest-assumption concern raised in review, namely the linear growth bound on the drift network, is not fragile: every fixed-depth ReLU network is globally Lipschitz, so the bound holds automatically with C equal to the network's Lipschitz constant, and the constant only enters through exp(CT). The main limitation, that the result concerns Euler approximations and not the exact ODE/PDE solution, is stated honestly in the abstract and introduction and does not undermine the central claim.","major_comments":[],"minor_comments":[{"comment":"The displayed intermediate bound in Proposition 3.5(v) appears as \"+ 1/(q-2) - 252\", while the proof around Eq. (306) establishes the same expression with \"+ 720/(q-2) - 252\". The final bound is unaffected, but the statement should be corrected for consistency.","section":"Proposition 3.5(v), Section 3.2.1"},{"comment":"The paper says that every result in Section 2 is proved in detail, yet Proposition 3.10 invokes the existence of identity networks I_d with D(I_d)=(d,2d,d) and R(I_d)=id by citing [25, Lemma 5.4], and Lemma 2.14 treats such networks as an assumption. Since the construction is elementary, either include it or explicitly acknowledge this external result in the self-containedness statement.","section":"Section 2 and Proposition 3.10"},{"comment":"The final constant is stated as \"C in R\", but the displayed inequalities are only meaningful when the constant is nonnegative; the statements should read \"C in [0, infinity)\".","section":"Theorem 1.1 and Corollary 3.13"},{"comment":"The title refers to \"differential equations\", but the theorem concerns Euler approximations rather than solutions of the differential equation; although the abstract and introduction make this clear, the title could be more precise, e.g., by mentioning Euler schemes.","section":"Title and abstract"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is heavily self-referential in its citations, but the cited statements are used as external proved lemmas and the dependence is transparent. The scope is narrow, covering Euler approximations rather than PDE or ODE solutions themselves, but this is clearly stated and is a legitimate first step toward space-time PDE error estimates. The paper is very long, yet the proof is essentially complete and the identified issues are local presentation points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the main result is real and it is new. Prior work on ReLU approximation of PDE-related quantities only had spatial estimates; this paper gives the first space-time estimates, applied to Euler schemes for ODEs, with explicit rates and network-size bounds. I checked the main chain and the proof is sound. The ANN calculus (compositions, parallelizations, sums) is developed carefully, the product approximation is built correctly from the square approximant, the hat-function interpolation gives a clean a posteriori bound, and the discrete Gronwall argument converts it to an a priori bound without any hidden assumption. The paper also genuinely improves on its predecessor [25] in the parameter estimates of Lemma 2.27(iv) and Proposition 2.28(iv).\n\nThe most likely worry is the linear-growth assumption on the drift network. The reader flagged it as the weakest premise, and the stress-test note says it is automatic for fixed-depth ReLU networks. The stress-test note is right. A fixed ReLU network is piecewise linear and therefore globally Lipschitz, so the bound ||R(Phi)(x)|| <= C(1+||x||) costs nothing. The constant C only enters through exp(CT), which is exactly what a Gronwall argument should produce. No fragile cancellation is needed.\n\nSoft spots, in proportion: the estimates are for the Euler-approximated trajectory, not for the solution of the underlying ODE or PDE. The authors state this plainly in the introduction and even say the PDE transfer is future work. That is a real limitation for somebody hoping for a ready-made PDE result, but it is not a flaw in the theorem as stated. The parameter bound in Theorem 1.1 has d^{16+8d}, which is unusably large in high dimension; this is a qualitative contribution, not a practical recipe. The paper is also extremely notationally heavy, with every elementary composition lemma proved in full. That is a stylistic choice, costly for the reader but not wrong.\n\nThe citation pattern is heavily self-referential, but the relied-on results from [25] are independently proved and are not restatements of the target theorem. No circularity. No code or data, which is standard for this kind of proof-driven paper.\n\nWho gets value: researchers in DNN approximation theory and numerical analysis who want a template for space-time estimates, plus anyone working on the Jentzen group's program. Practitioners will not find this useful.\n\nMy verdict: this deserves a serious referee and likely acceptance after minor revisions. I would send it out without hesitation. The result is a correct, reusable step forward, and the limitations are explicit rather than hidden.","headline":"Solid, fully written-up first space-time error estimates for ReLU-network Euler schemes; the central proof holds up, with the usual caveat that it is about Eulerization, not PDEs.","tokens_in":75127,"tokens_out":1801,"would_cite":true,"duration_ms":24490,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D15","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"ReLU networks approximate Euler trajectories in time and space.","keywords":["deep neural networks","ReLU","Euler approximation","space-time error estimates","ANN calculus","differential equations","rectified linear units","approximation theory"],"falsifier":"Check the claim for d=1, N=1, T=1, with drift R(Phi)(x)=max{x,0}. The Euler map is Y_{t,x}=x+t*max{x,0}, so the network must realize the piecewise-quadratic map (t,x) -> x+t*max{x,0} on [0,1] x R. Compute the minimal ReLU network parameters needed to approximate this map to error epsilon and compare with the asserted bound P(Psi) <= C(1+|ln(epsilon)|^2); if the required parameter count grows faster than log(1/epsilon) as epsilon tends to 0, the epsilon-dependence in Theorem 1.1 fails.","tokens_in":74201,"feed_emoji":"🧠","tokens_out":5731,"duration_ms":65096,"temperature":0.7,"pith_summary":"This paper proves that rectified deep neural networks can approximate the space-time evolution of a large class of differential equations: for any prescribed tolerance, dimension, and number of Euler steps, there is an explicit ReLU network whose output is uniformly close to the Euler approximation of the solution across the whole time interval and all of R^d, not just at a fixed time. The error is controlled simultaneously in space and time by a single parameter epsilon, with the network parameter count bounded by an explicit polynomial in the number of Euler steps and a logarithmic factor in 1/epsilon. This closes a gap left by earlier rigorous results, which gave spatial error estimates for neural-network PDE approximations but no temporal component. If the estimates hold, they supply a first rigorous consistency check for using deep networks as space-time solvers rather than as spatial function approximators.","feed_headline":"ReLU networks approximate Euler trajectories in time and space","feed_subtitle":"For any tolerance, one explicit network reproduces the Euler map on [0,T] x R^d with bounded size.","key_machinery":"The argument is carried by an ANN calculus, a compositional algebra for building networks realizing sums, parallelizations, compositions, and Euler updates, together with explicit ReLU approximations of the square function x -> $x^{2}$ and of the product (t,x) -> tx. The product network is the space-time bridge: it converts each Euler step into a network operation, hat functions interpolate between grid times, and a discrete Gronwall estimate keeps the trajectory growth polynomial in ||x||. The rectifier activation a(x)=max{x,0} is used throughout.","core_discovery":"The main result, Theorem 3.12 with the special case Theorem 1.1, asserts that for every epsilon in (0,1], every dimension d, and every number N of Euler steps, there exist ReLU networks $Psi^{{epsilon,d,N}}$ such that for all t in [0,T] and x in R^d, ||$Y^{{d,N}}$_{t,x} - R($Psi^{{epsilon,d,N}}$)(t,x)|| <= C $d^{{1/2}}$ $N^{{3/2}}$ epsilon (1+||x||^3), while the network size satisfies P($Psi^{{epsilon,d,N}}$) <= C $d^{{16+8d}}$ $N^{6}$ (1+|ln(epsilon)|^2). Here $Y^{{d,N}}$ is the Euler approximation of an ordinary differential equation whose drift is itself given by a network with linear growth. The approximation is space-time in the strong sense: the same network realizes the solution map (t,x) -> Y_{t,x} continuously on [0,T] x R^d, and the error is uniform in t and x.","pith_inferences":["The paper stops at Euler approximations of ODEs; the announced next step is to feed these space-time networks into PDE approximation schemes. A testable extension is to insert the constructed network into a splitting or backward-solver pipeline and check whether the observed error obeys the same epsilon-dependence.","The size bound d^{16+8d} N^6 (1+|ln(epsilon)|^2) grows exponentially in d, so the result is a consistency and expressivity statement for fixed dimension, not a curse-of-dimensionality breakthrough; a follow-up would need to reduce this to a polynomial-in-d bound.","Because the construction is explicit and deterministic, it provides a training-free baseline for neural emulation of differential-equation solvers: the weights can be hard-coded and evaluation cost compared against classical Euler integration.","The requirement that the drift network grow at most linearly is inherited from the Gronwall step; weakening it would likely need a different a priori control and would change the polynomial-in-N factors in the error bound."],"forward_implications":["For every tolerance epsilon, dimension d, and number of Euler steps N, a ReLU network of explicitly bounded size realizes the Euler map uniformly on [0,T] x R^d.","The time direction is controlled by the same accuracy parameter epsilon as the spatial direction, closing the gap left by purely spatial estimates.","The construction is explicit enough to be read as an algorithm: compose networks for the square, the product, hat functions, and the Euler update.","The network size grows polynomially in N and logarithmically in 1/epsilon, but exponentially in the dimension d.","The a priori bounds used in the proof are uniform in the perturbation sequence y, so the same network family covers perturbed Euler schemes in one stroke."],"supporting_citations":[{"why":"Supplies the ANN calculus and the Euler-step representation that the constructive proof is mainly inspired by.","marker":"[25]"},{"why":"Provides spatial DNN expression-rate analysis and the square-function approximation ideas that are refined here.","marker":"[11]"},{"why":"Provides the ReLU approximation scheme for the square function that feeds into the product construction.","marker":"[37]"},{"why":"Provides the reformulated ANN definitions, realizations, and parameter counts that Section 2 builds on.","marker":"[32]"},{"why":"Supplies the discrete Gronwall inequality used in Lemma 3.11 to bound the Euler trajectory.","marker":"[23]"}],"fun_headline_variants":["ReLU nets uniformly approximate Euler trajectories in space-time","Explicit space-time error bounds for ReLU Euler approximations","One ReLU network reproduces the Euler map on [0,T]xR^d","Any tolerance: one ReLU net reproduces Euler map in space-time","ReLU networks: strong space-time convergence to Euler schemes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The drift network must grow at most linearly, with ||R(Phi)(x)|| <= C(1+||x||), because that bound is the input to the Gronwall step that keeps the Euler trajectories finite and makes the a priori bounds enter the error estimate; if it is violated, the claimed uniform bounds can blow up.","fun_headline_variants_meta":{"raw":{"variants":["ReLU nets uniformly approximate Euler trajectories in space-time","Explicit space-time error bounds for ReLU Euler approximations","One ReLU network reproduces the Euler map on [0,T]xR^d","Any tolerance: one ReLU net reproduces Euler map in space-time","ReLU networks: strong space-time convergence to Euler schemes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000789,"raw_usage":{"total_tokens":3504,"prompt_tokens":994,"completion_tokens":2510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2420}},"tokens_in":610,"tokens_out":2510,"duration_ms":23627,"temperature":1.0,"reasoning_tokens":2420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:00:45.756602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the claim for d=1, N=1, T=1, with drift R(Phi)(x)=max{x,0}. The Euler map is Y_{t,x}=x+t*max{x,0}, so the network must realize the piecewise-quadratic map (t,x) -> x+t*max{x,0} on [0,1] x R. Compute the minimal ReLU network parameters needed to approximate this map to error epsilon and compare with the asserted bound P(Psi) <= C(1+|ln(epsilon)|^2); if the required parameter count grows faster than log(1/epsilon) as epsilon tends to 0, the epsilon-dependence in Theorem 1.1 fails.","supporting_citations":[{"cited_title":"DNN Expression Rate Analysis of High-dimensional PDEs: Application to Option Pricing","cited_arxiv_id":"1809.07669","evidence_quote":"Provides spatial DNN expression-rate analysis and the square-function approximation ideas that are refined here."},{"cited_title":"Optimal approximation of piecewise smooth functions using deep ReLU neural networks","cited_arxiv_id":"1709.05289","evidence_quote":"Provides the reformulated ANN definitions, realizations, and parameter counts that Section 2 builds on."},{"cited_title":"Overcoming the curse of dimensionality in the approximative pricing of financial derivatives with default risks","cited_arxiv_id":"1903.05985","evidence_quote":"Supplies the discrete Gronwall inequality used in Lemma 3.11 to bound the Euler trajectory."}],"review_version":1}