{"paper_id":"2512.25017","title":"Convergence of the generalization error for deep gradient flow methods for PDEs","authors":["Chenguang Liu","Antonis Papapantoleon","Jasper Rou"],"paper_date":"2025-12-31","paper_type":"theoretical","paper_url":"https://arxiv.org/pdf/2512.25017v1","page_url":"https://thequant.space/flowcharts/convergence-of-the-generalization-error-for-deep-gradient-fl/","scores":{"empirical_rigor":1.5,"math_complexity":9.5,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper proposes deep gradient flow methods (DGFMs) for solving high‑dimensional PDEs and provides a mathematical analysis showing that both the approximation and training errors vanish as the number of neurons and training time go to infinity.","quote":"The aim of this article is to provide a firm mathematical foundation for the application of deep gradient flow methods (DGFMs) for the solution of (high-dimensional) partial differential equations (PDEs). We decompose the generalization error of DGFMs into an approximation and a training error.","summary":"The aim of this article is to provide a firm mathematical foundation for the application of deep gradient flow methods (DGFMs) for the solution of (high-dimensional) partial differential equations (PDEs). We decompose the generalization error of DGFMs into an approximation and a training error. We…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The PDE is well‑posed and the non‑linear operator F satisfies the required conditions.","type":"model","stated":"explicit","quote":"Let us mention that the convergence results hold for certain non-linear operators F, assuming that the PDE is well-posed.","support":"full"},{"assumption":"The operators of the PDE satisfy Assumptions (CON) and (G∘A).","type":"model","stated":"explicit","quote":"Moreover, assume that the operators of the PDE (2.1)–(2.2) satisfy Assumptions (CON) and (G˚A). Then, the dynamic (4.3) converges to the gradient flow (4.4) as n →∞, i.e. for any T > 0, sup 0≤t≤T E h V n t −Vt H1 0 i −−−→ n→∞0. Lemma 4.2.","support":"full"},{"assumption":"The neural network initialization is random and independent of the number of neurons (Assumption NNI).","type":"model","stated":"explicit","quote":"We use a random initialization for this process, independent of n, denoted by: \u0010 βi,n 0 , αi,n 0 , ci,n 0 \u0011 = \u0000βi 0, αi 0, ci 0 \u0001 = θi 0. Assumption (NNI).","support":"full"},{"assumption":"Assumption (CON) holds for the PDE operators, ensuring continuity of the Fréchet derivative of the loss.","type":"model","stated":"explicit","quote":"Assume that the operators of the PDE (2.1)–(2.2) satisfy Assumption (CON).","support":"full"},{"assumption":"Assumption (G∘A) provides bounds λ1>0 and λ2≥0 for the operator L.","type":"model","stated":"explicit","quote":"Finally, note that the second derivative of ik equals \u0010 ik\u0011′′ (τ) = ∥v∥2 L2 + h ⟨Lv, v⟩H−1,H1 0 ≥(1 −hλ2) ∥v∥2 L2 + hλ1 ∥v∥2 H1 0 > 0, where we used Assumption (G ˚A). Moreover, we have that ⟨Lu, u⟩H−1,H1 0 ≥ \u0012σ2 2 + r \u0013 ∥u∥2 H1 0 , CONVERGENCE OF DEEP GRADIENT FLOW METHODS FOR PDE 27 therefore Assumption (G ˚A) is satisfied with λ1 = σ2 2 +r > 0 and λ2 = 0.","support":"full"}]},"datasets":{"status":"none_found","items":[],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"none"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"none","findings":[{"finding":"The generalization error of DGFMs tends to zero as the number of neurons and the training time tend to infinity.","kind":"theoretical","quote":"These results combined show that the generalization error of DGFMs tends to zero as the number of neurons and the training time tend to infinity.","support":"full"},{"finding":"For each fixed time step, the trained neural network converges to the true solution of the discretized PDE as the number of neurons and training time go to infinity.","kind":"theoretical","quote":"CONVERGENCE OF THE TRAINING ERROR In this section, we show that for each fixed time step k, the trained neural network converges to the true solution of the discretized PDE (3.1) as the number of neurons and the training time tend to infinity. Therefore, using the convergence of the time-stepping scheme, we can conclude the convergence of the training error. 4.1.","support":"full"},{"finding":"The training error converges to zero by virtue of the convergence of the time‑stepping scheme.","kind":"theoretical","quote":"Therefore, using the convergence of the time-stepping scheme, we can conclude the convergence of the training error. 4.1.","support":"full"},{"finding":"The gradient flow dynamics of the wide‑network limit converge to the true solution, yielding ‖V_t − w*‖_{H^1_0} → 0 as t → ∞.","kind":"theoretical","quote":"Hence, using Parseval’s identity again, we get ∥Vt −w∗∥2 e H1 0 = ∞ X i=1 \u0000hi t \u00012 = ∞ X i=1 e−2γit \u0000hi 0 \u00012 , which converges to 0 because γi > 0 and P∞ i=1 \u0000hi 0 \u00012 = ∥w∗∥2 e H1 0 < +∞. Finally, since the norm ∥·∥e H1 0 is equivalent to ∥·∥H1 0 , we conclude lim t→∞∥Vt −w∗∥H1 0 = 0. □ CONVERGENCE OF DEEP GRADIENT FLOW METHODS FOR PDE 21 APPENDIX A. AUXILIARY RESULTS A.1.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":17,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T19:44:35Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}