Pith. sign in

REVIEW 3 major objections 4 minor 300 references

Every square-integrable predictable process can be approximated by a causal neural network that sees only finitely many Brownian samples, with optimal rates for sparse processes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:18 UTC pith:YJA43STW

load-bearing objection Useful causal architecture and a clean density/rates theorem, but the abstract's claim that compressibility is typical overreaches — natural non-degenerate Gaussian priors fail S-compressibility for all useful S. the 3 major comments →

arxiv 2607.14361 v1 pith:YJA43STW submitted 2026-07-15 math.PR cs.LGq-fin.CPstat.ML

NeuralChaos: Optimal Adapted Approximation of Square Integrable Predictable Processes

Anastasis Kratsios , Giulia Livieri , Philipp Schmocker This is my paper
classification math.PR cs.LGq-fin.CPstat.ML MSC 60G0560H0768T0741A3541A5091G80
keywords predictable processesWiener chaoschaosletsbest N-term approximationMalliavin–Sobolev regularitycompressibilityneural operatorscausal architecture
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces NeuralChaos, a family of neural-network models that generate square-integrable predictable processes from finitely many samples of the driving Brownian motion. The main result proves that this family is dense in the space H^2_T(R^d): every such process can be approximated arbitrarily well in the mean-square sense, without computing iterated stochastic integrals. For processes that are sparse (compressible) in the chaoslet basis and possess Malliavin–Sobolev regularity, the approximation error decays like N^{-(S-1/2)} + (1+P)^{-s/2}, using only O(N(P+1)) grid points and O(N[P+d+1+S_H]) parameters. The paper also shows that compressibility is typical for random processes under non-degenerate sub-Gaussian sampling, whereas finite-dimensional Markovian Euler–Maruyama-type (neural SDE) parametrizations are meagre and Gaussian-null in H^2_T. This gives a practical, causality-preserving, trainable representation for controls and hedging strategies in continuous-time problems.

Core claim

On the paper's own terms, the central discovery is that the infinite-dimensional Wiener chaos of predictable processes can be captured by a finite-sampling architecture. NeuralChaos—lower-triangular linear lifts on Brownian samples, row-wise ReLU heads, and causal time masks—is dense in H^2_T(R^d). For S-compressible X in D^{s,2:d}_T, the error is at most C_X(N^{-(S-1/2)}+(1+P)^{-s/2}+ε) using O(N(P+1)) grid points and O(N[P+d+1+S_H]) parameters. Chaoslets (time-Haar atoms times Hermite polynomials in Gaussian Haar coordinates) make coefficients computable from finitely many samples. Separately, random chaoslet-series models make compressibility almost sure, while finite-dimensional Markovia

What carries the argument

The load-bearing object is the chaoslet basis of H^2_T(R^d): each chaoslet is the product of a deterministic Haar wavelet localized in time and a tensorized normalized Hermite polynomial evaluated at finitely many Gaussian Haar coordinates Z^r_{i,k} of the Brownian path. These Gaussian coordinates are linear combinations of three Brownian samples at dyadic times, so any finite chaoslet expansion can be evaluated without iterated integrals. The NeuralChaos architecture implements such expansions by using lower-triangular matrices to compute the Haar coordinates in an adapted way, row-wise ReLU networks to approximate the Hermite factors, and causal time masks to assemble the rows into a predi

Load-bearing premise

The quantitative rates and the claim that compressibility is generic rest on the target process having coefficients in the paper's special basis that decay like a power law; if a real process lacks this sparsity, only the abstract density statement remains and the advertised speeds disappear.

What would settle it

For a centred non-degenerate Gaussian measure on H^2_T whose law satisfies Assumption 4.2, draw many sample paths and sort their chaoslet coefficients; if, for some such measure, a positive fraction of samples fail to have power-law decaying sorted coefficients (i.e., are not S-compressible for any S>1/2), then the paper's genericity claim (Proposition 4.3) is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Universal approximation: for any target in H^2_T(R^d) and any ε>0, a NeuralChaos model achieves error < ε, so the architecture can serve as a universal parameterization for controls and hedging strategies.
  • Rates: for compressible and Malliavin–Sobolev regular targets, the error is O(N^{-(S-1/2)} + (1+P)^{-s/2} + ε), matching the best N-term chaoslet rate; the parameter count grows linearly in N and P.
  • Genericity: random processes sampled from non-degenerate sub-Gaussian chaoslet-series laws are almost surely S-compressible, so sparse structure is typical rather than special.
  • Markovian limitation: finite-dimensional Euler–Maruyama-type neural SDE models form a meagre and Gaussian-null subset of H^2_T, so they cannot represent typical predictable processes, regardless of discretization.
  • Practical computation: the architecture avoids iterated integral computations, using only Brownian samples and standard ReLU MLPs, and is compatible with gradient-based training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If compressibility is indeed typical of predictable processes, then sparse chaoslet-type representations could become a natural inductive bias in stochastic control and finance; one could test this by computing chaoslet coefficients of known optimal strategies and checking for power-law decay.
  • The theorem's two-parameter trade-off (N controls within-level sparsity, P controls degree tail) suggests that adaptive procedures should jointly select the number of terms and the maximum chaos degree from data; the paper leaves this selection unspecified.
  • The density result is abstract; extending it to statistical learning guarantees (finite-sample risk) would bridge approximation and learning, which the paper does not attempt.
  • The architecture is one concrete way to realize chaoslet coefficients; other causal, finite-memory networks (e.g., recurrent models with strict causal masks) might achieve similar rates, offering design alternatives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces NeuralChaos, a neural-operator architecture for approximating R^d-valued predictable square-integrable processes. The architecture samples the driving Brownian motion at finitely many times, applies lower-triangular linear lifts and row-wise ReLU heads, and assembles the rows through causal time masks, preserving predictability by construction. Theorem 4.1 claims (i) density of NeuralChaos in H^2_T(R^d) and (ii) for S-compressible and Malliavin–Sobolev regular targets, an error bound of order N^{-(S-1/2)}+(1+P)^{-s/2}+ε with explicit grid and parameter counts. Proposition 4.3 claims that compressibility is generic under a random chaoslet model satisfying Assumption 4.2, while Proposition 4.5 shows that generalized predictable Euler–Maruyama-type processes are meagre and Gaussian-null in H^2_T. Numerical experiments on stochastic control and hedging illustrate the architecture.

Significance. If the main theorem is correct, the paper provides a useful and constructive bridge between Wiener-chaos theory and trainable causal neural parametrizations, avoiding explicit iterated integrals. The density result is plausible, and the quantitative rate for compressible targets is a genuine contribution to non-linear approximation for predictable processes. The negative results for finite-dimensional Markovian surrogate classes are also informative. However, the advertised 'genericity' of compressibility is not established for natural non-degenerate laws, and the proof of the causality-preserving lift has a technical gap. The paper's utility depends on the S-compressibility assumption, which the authors themselves show is orthogonal to Malliavin smoothness; the scope of the quantitative claims should be made more precise.

major comments (3)
  1. [Proposition 4.3 / Assumption 4.2 and Abstract] The claim that compressibility is 'typical under non-degenerate sub-Gaussian sampling' is not supported. Assumption 4.2 assumes the coefficient sequence ϑ lies in ℓ^{p_cmp} with p_cmp<2 and that the unitary matrix A_U is ℓ^{p_cmp}-bounded. By Proposition A.2, this assumption directly yields sorted-coefficient decay of order k^{-1/p_cmp}, i.e. the advertised S-compressibility. The argument is essentially a restatement of the assumption. The centered Gaussian law with covariance eigenvalues λ_n=1/(n log^2 n) in the chaoslet basis is non-degenerate and sub-Gaussian, yet its sampled coefficients fail to be in ℓ^p for any p<2 and fail S-compressibility for all S>1/2. Thus the abstract's genericity claim is false in that natural setting, and the proposition should be reframed as a statement about a parametric family satisfying a sparsity-in-coefficients condition, not as generic non-degenerate
  2. [Lemma B.3] The proof that A_{I_H,J_H,i_0} is lower-triangular in the rectangular sense is incomplete. The argument shows that every nonzero entry of row ℓ occurs at a column whose time is no later than the terminal information time θ_ℓ. But rectangular lower triangularity requires A_{ℓ,m}=0 for every m>ℓ, i.e. the column index must be at most the row index. If the sampling grid contains unselected points lying between the selected Haar sampling times and θ_ℓ, the column index of a needed sample can exceed ℓ. This breaks the hypothesis of Lemma B.2, and hence the causality guarantee for the constructed NeuralChaos realization. The problem is fixable by choosing the grid and the row ordering more carefully (e.g. assigning each chaoslet block to the grid position of its information time and ensuring all needed samples appear among earlier columns), but the current proof does not establish the claimed
  3. [Theorem 4.1 / Definition 2.1] The 'optimal' approximation rate is a direct consequence of the definition of S-compressibility: polynomial decay of sorted coefficients by definition gives best-N-term error N^{-(S-1/2)}. The paper does not provide lower bounds or identify a broader class for which this rate is minimax. The phrasing 'optimal' should be understood as 'matching the best N-term rate under the stated assumption' rather than an intrinsic minimax optimality statement. This is a framing issue, but it matters for how the contribution is read.
minor comments (4)
  1. [Section 2.1 and throughout] The notation [N] is used both for {0,...,N} and [N]+ for {1,...,N}; this is standard but the distinction is easy to miss. Also '≲_X' is used before being formally defined; please define the dependence on the constant explicitly.
  2. [Lemma B.1 / Lemma B.3] The grid count in Lemma B.1 is correct, but the relationship between the dyadic indices (i,k) and the terminal information time θ_ℓ in Lemma B.3 is delicate; please clarify how ties in θ_ℓ are ordered and how the row index relates to the column index of the grid points, especially when multiple chaoslets share information times.
  3. [Section 5] The numerical experiments are illustrative but do not provide quantitative evidence for the convergence rates in Theorem 4.1. Reporting empirical rates versus N and P would strengthen the practical claims. Also, the caption of Figure 5 references 'train/test set' but the split is only described in the text.
  4. [Appendix B.4] In the proof of Theorem 4.1, the sentence 'we choose a deterministic grid = (t_i)_{i=1}^M, 0=t_0<...<t_M=T ap,q and all Brownian sampling times needed...' is grammatically incomplete and should be rewritten.

Circularity Check

2 steps flagged

Quantitative rates and genericity claims reduce to the compressibility/ℓ^{p_cmp} assumptions; the density result is independent but the advertised optimal-rate and typicality narrative is partly circular.

specific steps
  1. self definitional [Definition 2.1, Proposition 2.2, Theorem 4.1]
    "Definition 2.1: 'if one retains only the N largest chaoslet coefficients, then an S-compressible process admits a best N-term approximation error of order N^{-(S-1/2)}'. Proposition 2.2: 'inf ... ≲_X N^{-(S-1/2)} + (1+P)^{-s/2}'. Theorem 4.1: '∥X−X̂_{N,P,ε}∥_{H^2_T(R^d)} ≲_X N^{-(S-1/2)} + (1+P)^{-s/2} + ε'."

    S-compressibility is defined as sorted chaoslet coefficients decaying as C_X r^{-S}. The best-N-term error bound N^{-(S-1/2)} is the standard, essentially immediate consequence of that decay, and the paper itself states this consequence inside Definition 2.1. Proposition 2.2 and Theorem 4.1 then restate that same bound as the advertised quantitative guarantee. The rate is therefore not derived from an independent structural property; it is built into the definition of the assumption.

  2. fitted input called prediction [Assumption 4.2 and Proposition 4.3]
    "Assumption 4.2: 'Suppose that there exist 0<p_cmp<2, ϑ=(ϑ_k)∈ℓ^{p_cmp}, and a unitary operator U... such that A_U:ℓ^{p_cmp}→ℓ^{p_cmp} is bounded... We define X·=Σ_{k=1}^∞ ϑ_k ξ_k U φ_k'. Proposition 4.3: 'If Law(X·)=µ_{ϑ,U}, then X· is (S=1/p_cmp)-compressible and belongs to D^{s,2:d}_T µ-almost surely.'"

    The 'generic' process is constructed from a coefficient sequence ϑ that is already required to lie in ℓ^{p_cmp}. Membership in ℓ^{p_cmp} is precisely what forces the sorted coefficient decay r^{-1/p_cmp} that defines 1/p_cmp-compressibility; Proposition A.2 only shows that ℓ^{p_cmp}-stability of U preserves this decay. Thus the compressibility conclusion is essentially assumed in the coefficient model, not shown to follow from non-degenerate sub-Gaussian sampling alone. Indeed, a centered Gaussian law with diagonal eigenvalues λ_n=1/(n log^2(n+1)) is non-degenerate but its coefficients are not in any ℓ^{p_cmp} with p_cmp<2, so it is not S-compressible for S>1/2.

full rationale

The paper's constructive density argument for NeuralChaos is not circular: it builds chaoslet approximations from Brownian samples via causal lifts, ReLU heads, and masks, and the finite-parameter complexity bounds follow from explicit approximation lemmas. The numerical control and hedging experiments are demonstrations, not derivations. The central circularity is in the quantitative/genericity narrative. The advertised rate N^{-(S-1/2)} is a direct restatement of Definition 2.1's coefficient-decay condition, which the paper itself acknowledges in the same definition; Theorem 4.1 inherits this condition through Proposition 2.2. More significantly, Proposition 4.3's 'compressible processes are generic' claim depends on Assumption 4.2, which already assumes ϑ∈ℓ^{p_cmp}; this is essentially the same power-law sparsity that defines compressibility. A non-degenerate Gaussian with slowly decaying spectrum shows the broad typicality claim is false without that extra ℓ^{p_cmp} assumption. The paper's conclusion honestly says the rates 'are inherited from best N-term chaoslet approximation and therefore rely on compressibility and Malliavin–Sobolev regularity assumptions,' which mitigates the issue but does not remove the fact that the headline optimal-rate and typicality results reduce by construction to the compressibility/ℓ^{p_cmp} inputs. The reliance on [4] for the chaoslet basis is a self-citation, but it is a stated external basis theorem rather than a fitted parameter or a conclusion-equivalent input, so I do not count it as load-bearing circularity. Overall, the density theorem is independent, but the quantitative and genericity claims are substantially pre-encoded in their assumptions: score 6.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central quantitative claims depend on assumed regularity exponents S, s, and p_cmp, none of which are measured from data or derived from dynamics. The most consequential imported fact is the chaoslet basis from [4], which the paper should have audited more explicitly because of author overlap. The only clearly ad hoc assumption is Assumption 4.2, where the generic-compressibility conclusion is largely baked into the ℓ^p coefficient hypothesis.

free parameters (3)
  • S (compressibility exponent)
    Assumed in Definition 2.1 and Theorem 4.1. The advertised rate N^{-(S-1/2)} is fully determined by S, which is not measured or derived from process dynamics.
  • s (Malliavin-Sobolev smoothness)
    Assumed in Theorem 4.1. Controls the chaos-degree truncation error (1+P)^{-s/2}.
  • p_cmp (ℓ^p coefficient exponent)
    Assumed in Assumption 4.2. Determines the generic compressibility exponent S = 1/p_cmp in Proposition 4.3.
axioms (5)
  • domain assumption The chaoslet family C_all defined in (7) is an orthonormal basis of H^2_T, cited from [4, Lemma 19].
    Used as the foundation for density and for all N-term rates. The cited paper shares an author with this paper, so the dependency is self-referential even if the result is likely correct.
  • standard math ReLU-MLPs can approximate univariate and tensorized Hermite polynomials on compact cubes with the stated width/depth/non-zero-parameter bounds, via [97, Lemma 7.5], Lemma B.5, Lemma B.8, and Lemma B.10.
    Imported approximation-theoretic tools used in Proposition B.4 and Theorem 4.1. Not proven in the paper but standard in the ReLU approximation literature.
  • standard math The Brownian Haar coordinates Z^r_{i,k} in (5) are standard Gaussian, orthogonal across coordinates, and computable from finitely many Brownian samples.
    Used in Steps 1 and 2 of the proof to connect chaoslets to the finite-sampling architecture. Follows from the orthogonal Haar wavelet system.
  • ad hoc to paper Assumption 4.2: the random coefficient sequence ϑ lies in ℓ^{p_cmp} and the unitary change-of-basis matrix A_U is bounded on ℓ^{p_cmp}.
    This is the engine of Proposition 4.3. The conclusion that X is 1/p_cmp-compressible is essentially an unpacking of the assumption that its coefficients already decay like an ℓ^{p_cmp} sequence.
  • domain assumption Malliavin-Sobolev process spaces D^{s,2:d}_T are characterized by weighted chaoslet projections, with weights (1+n)^s.
    Used to control the chaos-degree tail in Proposition 2.2 and Theorem 4.1. This is a process-level adaptation of the standard random-variable Malliavin-Sobolev norm.

pith-pipeline@v1.3.0-alltime-deepseek · 53485 in / 17896 out tokens · 179930 ms · 2026-08-02T02:18:27.600047+00:00 · methodology

0 comments
read the original abstract

We address fundamental challenges in representing and computing $\mathbb{R}^{d}$-valued predictable square-integrable processes over $[0,T]$, collected in the space $\mathcal{H}^2_T(\mathbb{R}^{d})$. These processes are central to continuous-time stochastic control, reinforcement learning, and mathematical finance. Although Wiener-chaos expansions offer strong theoretical tools, traditional computational methods are hindered by the need for large chaos dictionaries and high-order iterated integrals. To overcome these obstacles, we introduce NeuralChaos -- a neural operator architecture that produces elements of $\mathcal{H}^2_T(\mathbb{R}^{d})$ using only finitely many evaluations of the driving Brownian motion, while preserving predictability and square-integrability. We prove that NeuralChaos is dense in $\mathcal{H}^2_T(\mathbb{R}^{d})$ and achieves the best $N$-term chaoslet approximation rates for compressible and Malliavin--Sobolev regular processes. Moreover, compressibility is shown to be typical for processes from $\mathcal{H}^2_T(\mathbb{R}^{d})$ under non-degenerate sub-Gaussian sampling. In contrast, we show that finite-dimensional Markovian neural SDE models constitute a meagre and Gaussian-null subset in $\mathcal{H}^2_T(\mathbb{R}^{d})$, regardless of discretization, whereas compressible processes are generic. Numerical experiments on a stochastic optimal control problem and dynamic hedging highlight the practical effectiveness of our approach. Our results enable more efficient and expressive modelling in stochastic analysis and mathematical finance.

Figures

Figures reproduced from arXiv: 2607.14361 by Anastasis Kratsios, Giulia Livieri, Philipp Schmocker.

Figure 1
Figure 1. Figure 1: Our NeuralChaos architecture (for details cf. Section 3): The input Brownian path is first sampled at the time-points 0 = t0 < t1 < · · · < tM = T. (cf. Figure 4a), producing the path-wise evaluations Wt0 , . . . , WtM−1 ; notice that no iterated integrals are explicitly computed, only implicitly. Lifting Channels: These evaluations are mapped, via a sequence of K lower-triangular matrices, into a random f… view at source ↗
Figure 2
Figure 2. Figure 2: A process which is easy for adaptive approximation methods to approximate but hard for non-adaptive methods. Evidently, at most two parameters are needed to exactly represent any of the illustrated processes: one parameter for the Brownian factor W· and one pa￾rameter for the higher-order chaos factor X˜ ∗ ·; equivalently, each process is a two-term degree-10 polynomial in the Wiener-chaos sense. A non-ada… view at source ↗
Figure 3
Figure 3. Figure 3: Compressibility vs. Incompressibility. A compressible process is one whose chaoslet expansion, cf. (9), can be rearranged to exhibit at least power-law coefficient decay, cf. (10). The left subfigure illustrates the expansion in the default chaoslet ordering of C all, cf. (7), while the middle subfigure illustrates the same expansion after sorting the chaoslets by coefficient magnitude; i.e. in the spirit … view at source ↗
Figure 4
Figure 4. Figure 4: Inputs and causal assembly in the NeuralChaos architecture. Panel 4a shows the sto￾chastic input: the Brownian path is sampled at a finite ordered grid 0 = t0 < t1 < · · · < tM = T producing the causally ordered samples These samples are processed by lower-triangular lifting matrices and row-wise neural heads, so that the mth random output is measurable with respect to the information available at the corr… view at source ↗
Figure 5
Figure 5. Figure 5: Stochastic optimal control: Learning the optimal control of the cost functional J (α·) in (15). Left-hand side: Value of J (α·) and J (α ⋆ · ) as well as mean squared error (MSE) 1 SM PS s=1 PM m=1 |α ⋆ tk (ωs) − αtk (ωs)| 2 on train/test set. Right-hand side: Comparison of learned control (dotted) and true optimal control (continuous) for three samples of the train/test set. 5.2. Dynamic hedging in the Bl… view at source ↗
Figure 6
Figure 6. Figure 6: Hedging: Learning the hedging strategy of a European call option in a Black-Scholes model. Left-hand side: Empirical distribution of the profit-and-loss (PnL) process C(t, St) − [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 5 canonical work pages

  1. [1]

    On consistency of signature using

    Guo, Xin and Wang, Binnan and Zhang, Ruixun and Zhao, Chaoyi , journal=. On consistency of signature using. 2025 , publisher=

    2025

  2. [2]

    Optimal approximation of continuous functions by very deep

    Yarotsky, Dmitry , booktitle=. Optimal approximation of continuous functions by very deep. 2018 , organization=

    2018

  3. [3]

    Preprint arXiv:2407.18384 , year=

    Mathematical theory of deep learning , author=. Preprint arXiv:2407.18384 , year=

    arXiv

  4. [4]

    Foundations of Computational Mathematics , volume=

    Do log factors matter? On optimal wavelet approximation and the foundations of compressed sensing , author=. Foundations of Computational Mathematics , volume=. 2022 , publisher=

    2022

  5. [5]

    Foundations of Computational Mathematics , pages=

    Optimal sampling for least-squares approximation , author=. Foundations of Computational Mathematics , pages=. 2025 , publisher=

    2025

  6. [6]

    Foundations of Computational Mathematics , volume=

    Sparse harmonic transforms: a new class of sublinear-time algorithms for learning functions of many variables , author=. Foundations of Computational Mathematics , volume=. 2021 , publisher=

    2021

  7. [7]

    Foundations of Computational Mathematics , volume=

    Function values are enough for L 2-approximation , author=. Foundations of Computational Mathematics , volume=. 2021 , publisher=

    2021

  8. [8]

    Quantitative Finance , volume=

    Randomized signature methods in optimal portfolio selection , author=. Quantitative Finance , volume=. 2025 , publisher=

    2025

  9. [9]

    Signature Methods in Finance: An Introduction with Computational Applications , pages=

    Signature-Based Models in Finance , author=. Signature Methods in Finance: An Introduction with Computational Applications , pages=. 2025 , publisher=

    2025

  10. [10]

    Advances in neural information processing systems , volume=

    Deep signature transforms , author=. Advances in neural information processing systems , volume=

  11. [11]

    Annals of Mathematics , pages=

    Uniqueness for the signature of a path of bounded variation and the reduced path group , author=. Annals of Mathematics , pages=. 2010 , publisher=

    2010

  12. [12]

    Journal of Machine Learning Research , volume=

    Signature moments to characterize laws of stochastic processes , author=. Journal of Machine Learning Research , volume=

  13. [13]

    The Annals of Applied Probability , volume=

    Adapted topologies and higher rank signatures , author=. The Annals of Applied Probability , volume=. 2023 , publisher=

    2023

  14. [14]

    Universal approximation property of

    Neufeld, Ariel and Schmocker, Philipp , journal=. Universal approximation property of

  15. [15]

    Bridging the gap between approximation and learning via optimal approximation by

    Hong, Ruiyang and Kratsios, Anastasis , journal=. Bridging the gap between approximation and learning via optimal approximation by

  16. [16]

    Closed-loop equilibria for

    Hern. Closed-loop equilibria for. Preprint arXiv:2406.19607 , year=

    Pith/arXiv arXiv

  17. [17]

    Journal of the European Mathematical Society , volume=

    Dynamic programming for stochastic target problems and geometric flows , author=. Journal of the European Mathematical Society , volume=

  18. [18]

    , author=

    Adaptive subgradient methods for online learning and stochastic optimization. , author=. Journal of machine learning research , volume=

  19. [19]

    Operations Research Letters , volume=

    Mirror descent and nonlinear projected subgradient methods for convex optimization , author=. Operations Research Letters , volume=. 2003 , publisher=

    2003

  20. [20]

    Proceedings of Thirty Fourth Conference on Learning Theory , pages =

    Optimizing Optimizers: Regret-optimal gradient descent algorithms , author =. Proceedings of Thirty Fourth Conference on Learning Theory , pages =. 2021 , editor =

    2021

  21. [21]

    Dokl akad nauk Sssr , volume=

    A method for solving the convex programming problem with convergence rate O (1/k2) , author=. Dokl akad nauk Sssr , volume=

  22. [22]

    2018 , url=

    Stochastic Hyperparameter Optimization through Hypernetworks , author=. 2018 , url=

    2018

  23. [23]

    Journal of Machine Learning Research , volume=

    Optimization with momentum: Dynamical, control-theoretic, and symplectic perspectives , author=. Journal of Machine Learning Research , volume=

  24. [24]

    Optimization Letters , volume=

    The forward--backward splitting method and its convergence rate for the minimization of the sum of two functions in Banach spaces , author=. Optimization Letters , volume=. 2021 , publisher=

    2021

  25. [25]

    1957 , publisher=

    Dynamic Programming , author=. 1957 , publisher=

    1957

  26. [26]

    2006 , publisher=

    Controlled Markov processes and viscosity solutions , author=. 2006 , publisher=

    2006

  27. [27]

    1999 , publisher =

    Stochastic Controls: Hamiltonian Systems and HJB Equations , author =. 1999 , publisher =

    1999

  28. [28]

    2009 , isbn =

    Continuous‐Time Stochastic Control and Optimization with Financial Applications , author =. 2009 , isbn =. doi:10.1007/978‐3‐540‐89500‐8 , url =

    doi:10.1007/978 2009

  29. [29]

    2013 , isbn =

    Optimal Stochastic Control, Stochastic Target Problems, and Backward SDE , author =. 2013 , isbn =. doi:10.1007/978-1-4614-4286-8 , url =

    work page doi:10.1007/978-1-4614-4286-8 2013

  30. [30]

    Preprint arXiv:2509.00904 , year=

    Convergence Rates of Time Discretization in Extended Mean Field Control , author=. Preprint arXiv:2509.00904 , year=

    Pith/arXiv arXiv

  31. [31]

    Applied Mathematics & Optimization , volume=

    On the rate of convergence of difference approximations for uniformly nondegenerate elliptic Bellman’s equations , author=. Applied Mathematics & Optimization , volume=. 2014 , publisher=

    2014

  32. [32]

    Proceedings of Thirty Third Conference on Learning Theory , pages =

    Finite-Time Analysis of Asynchronous Stochastic Approximation and Q -Learning , author =. Proceedings of Thirty Third Conference on Learning Theory , pages =. 2020 , editor =

    2020

  33. [33]

    2008 , publisher=

    Stochastic approximation: a dynamical systems viewpoint , author=. 2008 , publisher=

    2008

  34. [34]

    and Peng, S

    El Karoui, N. and Peng, S. and Quenez, M. C. , TITLE =. Mathematical Finance , FJOURNAL =. 1997 , NUMBER =

    1997

  35. [35]

    and Peng, S

    Pardoux, \'E. and Peng, S. G. , TITLE =. Systems Control Lett. , FJOURNAL =. 1990 , NUMBER =. doi:10.1016/0167-6911(90)90082-6 , URL =

    doi:10.1016/0167-6911(90)90082-6 1990

  36. [36]

    1999 , PAGES =

    Revuz, Daniel and Yor, Marc , TITLE =. 1999 , PAGES =. doi:10.1007/978-3-662-06400-9 , URL =

    doi:10.1007/978-3-662-06400-9 1999

  37. [37]

    Convex analysis and monotone operator theory in Hilbert spaces , pages=

    Correction to: convex analysis and monotone operator theory in Hilbert spaces , author=. Convex analysis and monotone operator theory in Hilbert spaces , pages=. 2020 , publisher=

    2020

  38. [38]

    1990 , PAGES =

    Meyer, Yves , TITLE =. 1990 , PAGES =

    1990

  39. [39]

    Neural Operators Can Play Dynamic

    Alvarez, Guillermo and Ekren, Ibrahim and Kratsios, Anastasis and Yang, Xuwei , journal=. Neural Operators Can Play Dynamic

  40. [40]

    2002 , edition =

    Foundations of Modern Probability , author =. 2002 , edition =

    2002

  41. [41]

    Advances in neural information processing systems , volume=

    Monotone operator equilibrium networks , author=. Advances in neural information processing systems , volume=

  42. [42]

    and Combettes, Patrick L

    Bauschke, Heinz H. and Combettes, Patrick L. , TITLE =. 2017 , PAGES =. doi:10.1007/978-3-319-48311-5 , URL =

    doi:10.1007/978-3-319-48311-5 2017

  43. [43]

    arXiv , author =:2101.09740 , primaryclass =

    On the oracle complexity of smooth strongly convex minimization , year =. arXiv , author =:2101.09740 , primaryclass =

    Pith/arXiv arXiv

  44. [44]

    arXiv , author =:2101.09741 , primaryclass =

    An optimal gradient method for smooth strongly convex minimization , year =. arXiv , author =:2101.09741 , primaryclass =

    Pith/arXiv arXiv

  45. [45]

    Stochastic first-order methods: non-asymptotic and computer-aided analyses via potential functions , volume =

    Taylor, Adrien and Bach, Francis , booktitle =. Stochastic first-order methods: non-asymptotic and computer-aided analyses via potential functions , volume =

  46. [46]

    Neural Networks , volume=

    Fading memory echo state networks are universal , author=. Neural Networks , volume=. 2021 , publisher=

    2021

  47. [47]

    2001 , institution=

    The ``Echo State'' Approach to Analysing and Training Recurrent Neural Networks , author=. 2001 , institution=

    2001

  48. [48]

    Neural Computation , volume=

    Real-Time Computing Without Stable States: A New Framework for Neural Computation Based on Perturbations , author=. Neural Computation , volume=. 2002 , publisher=

    2002

  49. [49]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Random Features for Large-Scale Kernel Machines , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2007 , url=

    2007

  50. [50]

    Maximum Principle Based Algorithms for Deep Learning , url =

    Qianxiao Li and Long Chen and Cheng Tai and Weinan E , journal =. Maximum Principle Based Algorithms for Deep Learning , url =. 2018 , Bdsk-Url-1 =

    2018

  51. [51]

    Smooth strongly convex interpolation and exact worst-case performance of first-order methods , volume =

    Adrien Taylor and Julien Hendrickx and Fran. Smooth strongly convex interpolation and exact worst-case performance of first-order methods , volume =. Mathematical Programming , month =

  52. [52]

    An automatic method for finding the greatest or least value of a function , volume =

    Rosenbrock, HoHo , date-added =. An automatic method for finding the greatest or least value of a function , volume =. The Computer Journal , number =

  53. [53]

    The Maximum Principle: Discrete Time , year =

    Sethi, Suresh P , booktitle =. The Maximum Principle: Discrete Time , year =

  54. [54]

    Infinite-horizon optimal control in the discrete-time framework , year =

    Blot, Jo. Infinite-horizon optimal control in the discrete-time framework , year =

  55. [55]

    Introductory lectures on convex optimization: A basic course , volume =

    Nesterov, Yurii , date-added =. Introductory lectures on convex optimization: A basic course , volume =

  56. [56]

    , publisher =

    Tonelli, L. , publisher =. Opere scelte , year =

  57. [57]

    , mrclass =

    Kuratowski, K. , mrclass =. Topology

  58. [58]

    Set-valued analysis , url =

    Aubin, Jean-Pierre and Frankowska, H\'. Set-valued analysis , url =. 2009 , Bdsk-Url-1 =. doi:10.1007/978-0-8176-4848-0 , isbn =

    doi:10.1007/978-0-8176-4848-0 2009

  59. [59]

    , isbn =

    Munkres, James R. , isbn =. Topology , year =

  60. [60]

    Sequences and series in

    Diestel, Joseph , doi =. Sequences and series in. 1984 , Bdsk-Url-1 =

    1984

  61. [61]

    and Lorentz, George G

    DeVore, Ronald A. and Lorentz, George G. , doi =. Constructive approximation , url =. 1993 , Bdsk-Url-1 =

    1993

  62. [62]

    Preprint arXiv:1905.01208 , title =

    Gribonval, R. Preprint arXiv:1905.01208 , title =

    Pith/arXiv arXiv 1905

  63. [63]

    Regularity problems for a new functional in the calculus of variations , volume =

    De Giorgi, Ennio and Congedo, Giuseppe and Tamanini, Italo , fjournal =. Regularity problems for a new functional in the calculus of variations , volume =. Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Nat. (8) , mrclass =

  64. [64]

    Wimmer , date-added =

    Harald K. Wimmer , date-added =. The algebraic Riccati equation: conditions for the existence and uniqueness of solutions , url =. Linear Algebra and its Applications , pages =. 1984 , Bdsk-Url-1 =. doi:https://doi.org/10.1016/0024-3795(84)90224-6 , issn =

    work page doi:10.1016/0024-3795(84)90224-6 1984

  65. [65]

    Optimal control theory: an introduction , year =

    Kirk, Donald E , date-added =. Optimal control theory: an introduction , year =

  66. [66]

    Tyrell Rockafellar , date-added =

    R. Tyrell Rockafellar , date-added =. Convex Analysis , url =. 1970 , Bdsk-Url-1 =

    1970

  67. [67]

    An introduction to nonsmooth analysis , year =

    Ferrera, Juan , date-added =. An introduction to nonsmooth analysis , year =

  68. [68]

    Preprint arXiv:1802.09419 , title =

    Lorraine, Jonathan and Duvenaud, David , date-added =. Preprint arXiv:1802.09419 , title =

    Pith/arXiv arXiv

  69. [69]

    Learned Optimizers that Scale and Generalize , year =

    Wichrowska, Olga and Maheswaranathan, Niru and Hoffman, Matthew W and Colmenarejo, Sergio G. Learned Optimizers that Scale and Generalize , year =. International Conference on Machine Learning , date-added =

  70. [70]

    Generic methods for optimization-based modeling , year =

    Domke, Justin , booktitle =. Generic methods for optimization-based modeling , year =

  71. [71]

    Preprint arXiv:2002.12493 , title =

    Muehlebach, Michael and Jordan, Michael I , date-added =. Preprint arXiv:2002.12493 , title =

    Pith/arXiv arXiv 2002

  72. [72]

    A general analysis of the convergence of ADMM , year =

    Nishihara, Robert and Lessard, Laurent and Recht, Ben and Packard, Andrew and Jordan, Michael , booktitle =. A general analysis of the convergence of ADMM , year =

  73. [73]

    A Unified Analysis of Stochastic Optimization Methods Using Jump System Theory and Quadratic Constraints , volume =

    Hu, Bin and Seiler, Peter and Rantzer, Anders , date-added =. A Unified Analysis of Stochastic Optimization Methods Using Jump System Theory and Quadratic Constraints , volume =. Proceedings of Machine Learning Research vol , pages =

  74. [74]

    Analysis and design of optimization algorithms via integral quadratic constraints , volume =

    Lessard, Laurent and Recht, Benjamin and Packard, Andrew , date-added =. Analysis and design of optimization algorithms via integral quadratic constraints , volume =. SIAM Journal on Optimization , number =

  75. [75]

    Preprint arXiv:2004.06840 , title =

    Fran. Preprint arXiv:2004.06840 , title =

    Pith/arXiv arXiv 2004

  76. [76]

    Acceleration via symplectic discretization of high-resolution differential equations , year =

    Shi, Bin and Du, Simon S and Su, Weijie and Jordan, Michael I , booktitle =. Acceleration via symplectic discretization of high-resolution differential equations , year =

  77. [77]

    Accelerating rescaled gradient descent: Fast optimization of smooth functions , year =

    Wilson, Ashia C and Mackey, Lester and Wibisono, Andre , booktitle =. Accelerating rescaled gradient descent: Fast optimization of smooth functions , year =

  78. [78]

    Dual averaging for distributed optimization: Convergence analysis and network scaling , volume =

    Duchi, John C and Agarwal, Alekh and Wainwright, Martin J , date-added =. Dual averaging for distributed optimization: Convergence analysis and network scaling , volume =. IEEE Transactions on Automatic control , number =

  79. [79]

    Dual Averaging Methods for Regularized Stochastic Learning and Online Optimization , volume =

    Xiao, Lin , date-added =. Dual Averaging Methods for Regularized Stochastic Learning and Online Optimization , volume =. Journal of Machine Learning Research , pages =

  80. [80]

    Convex optimization , year =

    Boyd, Stephen and Boyd, Stephen P and Vandenberghe, Lieven , date-added =. Convex optimization , year =

Showing first 80 references.