{"paper_id":"2311.15333","title":"Asymptotic Error Analysis of Multilevel Stochastic Approximations for the Value-at-Risk and Expected Shortfall","authors":["Stéphane Crépey","Noufel Frikha","Azar Louzi","Gilles Pagès"],"paper_date":"2023-11-26","paper_type":"theoretical","paper_url":"https://arxiv.org/pdf/2311.15333v4","page_url":"https://thequant.space/flowcharts/asymptotic-error-analysis-of-multilevel-stochastic-approxima/","scores":{"empirical_rigor":4,"math_complexity":9,"quadrant":"Lab Rats"},"proposes":{"status":"available","statement":"The paper establishes central limit theorems for the renormalized estimation errors of nested stochastic approximation (NSA) and multilevel stochastic approximation (MLSA) algorithms, as well as for their Polyak‑Ruppert averaged versions.","quote":"We hereby establish central limit theorems for the renormalized estimation errors associated with both algorithms as well as their averaged versions.","summary":"Crépey, Frikha, and Louzi (2023) introduced a nested stochastic approximation algorithm and its multilevel acceleration to compute the value-at-risk and expected shortfall of a random financial loss. We hereby establish central limit theorems for the renormalized estimation errors associated with…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"For any bias parameter h, the cdf of X_h admits a first‑order Taylor expansion with remainder terms that vanish as h→0.","type":"statistical","stated":"explicit","quote":"(i) For any h ∈H, FXh admits the first order Taylor expansion FXh(ξ) −FX0(ξ) = v(ξ)h + w(ξ, h)h, ξ ∈R, for some functions v, w( · , h): R →R satisfying, for any ξ0 ⋆∈Argmin V0, Z ∞ ξ0⋆ v(ξ)dξ < ∞, lim H∋h↓0 w(ξ0 ⋆, h) = lim H∋h↓0 Z ∞ ξ0⋆ w(ξ, h)dξ = 0.","support":"full"},{"assumption":"For any h, the law of X_h has a continuous probability density function f_{X_h} that converges locally uniformly to f_{X_0}.","type":"statistical","stated":"explicit","quote":"(ii) For any h ∈H∪{0}, the law of Xh admits a continuous pdf fXh with respect to the Lebesgue measure. Moreover, the pdf {fXh, h ∈H} converge locally uniformly to fX0.","support":"full"},{"assumption":"The densities f_{X_h} are uniformly bounded and uniformly Lipschitz over h∈H∪{0}.","type":"statistical","stated":"explicit","quote":"(iv) The pdf {fXh, h ∈H ∪{0}} are uniformly bounded and uniformly Lipschitz, namely, sup h∈H∪{0} (∥fXh∥∞+ [fXh]Lip) < ∞, where [fXh]Lip denotes the Lipschitz constant of fXh, h ∈H ∪{0}. Remark 1.2.","support":"full"},{"assumption":"For any radius R, the densities f_{X_h} are uniformly positive on the ball B(ξ*_0,R).","type":"statistical","stated":"explicit","quote":"(iii) For any R > 0, inf h∈H∪{0} ξ∈B(ξ0 ⋆,R) fXh(ξ) > 0.","support":"full"},{"assumption":"The bias parameter h satisfies sup_{ξ∈K}|f_{X_h}(ξ)-f_{X_0}(ξ)| ≤ C h^{1/4+δ_0} for any compact K.","type":"statistical","stated":"explicit","quote":"There exist positive constants C, δ0 < ∞such that, for any h ∈H and any compact set K ⊂R, sup ξ∈K |fXh(ξ) −fX0(ξ)| ≤Ch 1 4 +δ0.","support":"full"},{"assumption":"Learning rates are of the form γ_n = γ_1 n^{-β} with γ_1>0 and β∈(1/2,1].","type":"execution","stated":"explicit","quote":"and advancing at the positive learning rate {γn, n ≥1} such that P∞ n=1 γn = ∞and P∞ n=1 γ2 n < ∞. Given a learning rate {γn = γ1n−β, n ≥1}, γ1 > 0, β ∈ \u0000 1 2, 1 \u0003 , this complexity is achieved by MLSA for β = 1 under the constraint λγ1 > 1, where the constant λ > 0 in (1.6) is explicit but tedious to compute.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"Numerical case study","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Section 5 presents a numerical case study supporting the paper’s theoretical findings. The present article complements the aforementioned one by analyzing the corresponding asymptotic error distributions, as required for delimiting VaR and ES trust regions and confidence 19 intervals.","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"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":"Central limit theorems are proved for the renormalized errors of NSA, MLSA and their averaged counterparts.","kind":"theoretical","quote":"We hereby establish central limit theorems for the renormalized estimation errors associated with both algorithms as well as their averaged versions. According to context, we use the notation Nℓ to designate both Nℓand \u0006 Nℓ \u0007 interchangeably. Theorem 3.5.","support":"full"},{"finding":"A numerical example confirms the theoretical CLTs and shows the Gaussian fit of the renormalized (VaR, ES) errors.","kind":"other","quote":"Our findings are substantiated through a numerical example. Figure 1 plots, for each algorithm, a 95%-confidence ellipse and Gaussian marginals of the joint distributions of the renormalized (VaR, ES) estimation errors.","support":"full"},{"finding":"The multilevel scheme reduces computational complexity to O(ε^{-5/2}) for a prescribed accuracy ε.","kind":"theoretical","quote":"For some prescribed accuracy ε > 0, the optimal complexity attained by the presented multilevel algorithms is in O(ε−5 2 ). Given a learning rate {γn = γ1n−β, n ≥1}, γ1 > 0, β ∈ \u0000 1 2, 1 \u0003 , this complexity is achieved by MLSA for β = 1 under the constraint λγ1 > 1, where the constant λ > 0 in (1.6) is explicit but tedious to compute.","support":"full"},{"finding":"Averaged versions of the algorithms achieve better convergence rates and increased numerical stability compared to non‑averaged schemes.","kind":"theoretical","quote":"To bypass this tuning issue, we explore the effect of the Polyak-Ruppert averaging principle [30, 18, 33] on the VaR NSA scheme. Further averaged extensions of these algorithms are also presented and shown to achieve better convergence rates than their original counterparts.","support":"full"},{"finding":"Optimal complexity O(ε^{-5/2}) is also attained by the averaged MLSA (AMLSA) for β∈(8/9,1] without constraints on γ_1.","kind":"theoretical","quote":"This complexity is also achieved by AMLSA for β ∈ \u0000 8 9, 1 \u0001 , without any constraint on γ1.","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":19,"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":true,"code_url":"https://github.com/azarlouzi/avg_mlsa","note":null},"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":"2026-11-14T20:00:14Z"},"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/"}