Pith. sign in

REVIEW 2 major objections 3 minor 1 cited by

This paper proves a randomized resampling estimator can reach true ε-stationarity for stochastic convex problems, where the actual subdifferential contains a small element, without Lipschitz or smoothness assumptions.

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 08:09 UTC pith:4YTDSWD3

load-bearing objection Novel and credible route to genuine ε-stationarity without Lipschitz gradients; the proof has a real but repairable gap in Lemma 4.6 that the authors must fix. the 2 major comments →

arxiv 2607.06883 v3 pith:4YTDSWD3 submitted 2026-07-08 stat.ML cs.LGmath.OC

Finding a stationary point of a stochastic convex problem

Felipe Areces , John Duchi , Malo Sommers This is my paper
classification stat.ML cs.LGmath.OC MSC 65K1090C1547N10
keywords stationary pointsstochastic convex optimizationsubdifferentialmonotone operatorsproximal point methoddimension theoryHausdorff measureresampling
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.

The paper asks for a stronger solution concept in stochastic convex optimization: not merely a point with small excess risk or a point near a minimizer, but a point where the population objective's subdifferential actually contains a small element—the genuine first-order stationarity condition. It proves that a simple randomized resampling scheme achieves this: run an empirical proximal-point step from many random starting points, then report the candidate with the smallest empirical subdifferential residual. The argument requires only a local moment bound and maximal monotonicity of the population operator; it does not need Lipschitz or single-valuedness assumptions. The technical engine is a dimension-theoretic decomposition of the graph of a maximal monotone operator, showing that a positive-measure set of starting points maps into the ε-stationary set. The guarantee is non-quantitative: the authors state that rates and distribution-free uniformity remain open, and they prove a naive neighborhood-mapping version is impossible.

Core claim

The core claim is that for a stochastic convex problem—or more generally for a stochastic maximal monotone operator A_P (a set-valued map whose graph is maximal among monotone maps), with the normal cone of the constraint X added—the output of the proposed resampling procedure satisfies dist(0, A_P(bx_n)+N_X(bx_n)) → 0 in probability. This means the actual subdifferential of the population objective contains a vector of vanishing norm, the genuine first-order stationarity condition, not a surrogate like a small Moreau-envelope gradient. The proof shows that a positive-measure set of random starting points map to ε-stationary points under the empirical proximal-point update, so sampling enoug

What carries the argument

The central object is the Minty parameterization of the graph of a maximal monotone operator, M_{λ,A}(x) = (prox_{λA}(x), Y_λA(x)), a bi-Lipschitz homeomorphism between R^d and the graph. Its work is to turn the stationarity condition 0 ∈ A(x)+N_X(x) into a measure-preservation statement: the graph piece over the ε-stationary set is d-dimensional, and a dimension-theory lemma (Lemma 4.6) partitions it into pieces of complementary dimension, of which at least one has a positive-measure base. This lets the authors show that a positive fraction of random centers u map via the empirical proximal point to ε-stationary points, making randomized resampling effective.

Load-bearing premise

The load-bearing premise is the graph-decomposition Lemma 4.6(iii): for some k, the set of points whose subdifferential fiber contains a k-simplex of positive measure must itself have positive (d−k)-dimensional Hausdorff measure; if this positive-measure base fails to exist, the proof that a positive fraction of random centers land in the stationary pre-image collapses, with the secondary premise being that A_P+N_X is maximal monotone.

What would settle it

Construct a maximal monotone operator A (e.g., the subdifferential of a convex function) satisfying Assumption A.1 for which, for every k and ε>0, the set X_{k,ε} of Lemma 4.6 has zero (d−k)-dimensional Hausdorff measure—then the claimed positive-probability sampling event has no measure-theoretic support and the proof fails. Alternatively, simulate Algorithm 2 on a finite-dimensional analogue of the Section 1.3 example and check whether the selected point's population subdifferential residual fails to converge in probability.

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

If this is right

  • For convex losses, Algorithm 2 returns a point whose population subdifferential contains a vector of norm at most ε with probability tending to one (Corollary 3.1).
  • The same conclusion holds for stochastic variational inequalities and general maximal monotone operators satisfying Assumption A.1.
  • In conformal prediction, ε-stationarity of the quantile loss implies coverage at level 1−α−ε, so the result yields distribution-free prediction sets certified by stationarity rather than by optimality.
  • The authors prove that no neighborhood of the optimum maps entirely into the stationary set under the empirical proximal mapping, so some randomization is essential (Proposition 3).
  • Rates of convergence and distribution-free guarantees remain open (Questions 1 and 2); the guarantee is asymptotic and depends on constants not explicitly computable from the problem data.

Where Pith is reading between the lines

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

  • Because the proof needs only a positive fraction of starting points, a practical heuristic is to run the resampling with a modest number of draws and select by empirical subdifferential; if the dimension-theoretic constants were ever made quantitative, the same algorithm would immediately inherit a finite-sample rate.
  • The graph decomposition suggests a link between stationarity and the geometry of singular sets of convex functions: functions whose subdifferential fibers are k-dimensional only on zero-measure bases would defeat the resampling argument, so the class of 'bad' functions may be topologically small.
  • For conformal prediction, the result implies that any method that finds an ε-stationary point of the quantile loss—not necessarily a minimizer—yields valid coverage, opening a new design axis for conditional coverage.
  • The impossibility result and the open Question 2 together suggest that any finite-sample rate must degrade with dimension, and that a rate-free statement that the empirical prox maps almost all points near the optimum into the stationary set is false.

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

2 major / 3 minor

Summary. The paper studies the problem of finding points satisfying true ε-stationarity, dist(0, ∂F_P(x)+N_X(x)) ≤ ε, for stochastic convex optimization, and the analogous formulation for stochastic maximal monotone operators. Because uniform Hausdorff convergence of subdifferentials does not hold in general, the authors develop a dimension-theoretic decomposition of the graph of the population subdifferential/operator. Using the Minty parameterization, they argue that positive-measure pieces of the graph survive under sampling, so that the preimage of the ε-stationary set under the empirical proximal map has positive Lebesgue measure. This leads to randomized resampling algorithms whose output satisfies the stationary residual convergence in probability (Theorems 1–2 and Corollary 3.1). The results are asymptotic and non-quantitative, and the paper is explicit about the lack of rates and about several open questions.

Significance. If the proof is completed as indicated, the result is significant: it shows that genuine ε-stationarity — not a Moreau-envelope or proximity surrogate — is statistically attainable for nonsmooth stochastic convex problems under only local moment and maximal monotonicity assumptions, without individual differentiability or Lipschitz continuity. The paper is self-contained and does not fit parameters to data; the constants arise from geometric measure theory and are non-quantitative, a limitation the authors state openly. The examples involving medians and conformal prediction make the notion of stationarity concrete and motivate the target. The graph-decomposition technique goes beyond existing singular-set results for convex functions and appears to be a novel and useful proof strategy, though one part of it currently needs repair.

major comments (2)
  1. [§4.4, Lemma 4.6] The lemma opens with 'Assume G has inductive dimension indG=d (and hence Hausdorff measure 0<H^d(G)<∞).' The implication indG=d ⇒ H^d(G)<∞ is false for general compact subsets of R^{2d}; a compact set of Hausdorff dimension d can have infinite d-dimensional Hausdorff measure. The finiteness is load-bearing: Eq. (10) uses the Fubini-type bound H^{d−k}(π(K)) ≤ C_{k,d−k}H^d(K)/ε, and Section 4.5 applies Lemma 4.6 to G_P without having established H^d(G_P)<∞. The gap is repairable via the Minty parameterization (7): since G_P is contained in the bi-Lipschitz image of a bounded subset of R^d, H^d(G_P) ≤ Lip(M)^d Leb(M^{-1}(G_P)) < ∞. This repair needs to be written into the proof; as stated, Lemma 4.6 asserts a false implication.
  2. [§3.1, proof of Theorem 2] The proof states that applying Theorem 1 shows P(Leb(X^-_{0,n})≥c)→1 whenever B_{1/2}⊂\hat B. But Theorem 1's conclusion is stated for a set of good x restricted to ||x−x_0||≤4λε and for pre-images landing in S(ε)∩B, with B=x_0+bB. X^-_{0,n} is not a subset of that restricted set, and \hat B⊃B_{1/2} does not by itself imply that \hat B contains the radius-4λε ball around x_0. The step can be repaired by taking ε≤b/8 and applying Theorem 1 with the neighborhood size b/2 (which is valid because Assumption A.1(iii) holds at smaller radii), but this argument is currently missing.
minor comments (3)
  1. [§4.3, Lemma 4.5(ii)] The proof that the graph piece G has dimension d is terse. Since G contains M_{λ,A}(K), the lower bound dim G ≥ d is clear, but the upper bound requires noting that M_{λ,A}^{-1}(G) is a bounded subset of R^d (because prox_{λA}(x)∈prox_{λA}(K) and Y_{λA}(x)∈rB), so ind M^{-1}(G)≤d. Please spell this out.
  2. [§3.1, Algorithm 1] The loop 'for i=1,...,√n' should state that √n is rounded to an integer, or use m_n=⌊√n⌋; otherwise the pseudo-code is not well-defined for general n.
  3. [§4.7, Example 6] In the decomposition for f(x)=|x|, the sentence about H^0({1})=H^0({−1})=1 seems to justify the 0-dimensional base set but is not connected to the subsequent use of H^1. Clarify the choice of base sets X_0 and X_1.

Circularity Check

0 steps flagged

No significant circularity: the stationarity target is not an input, Assumption A.1 is a stated domain condition, and no fitted parameter is relabeled as a prediction.

full rationale

The paper's derivation chain is self-contained: Theorem 2 and Corollary 3.1 reduce to Theorem 1, which is proved through the Minty parameterization (Section 4.3), the dimension-theoretic decomposition Lemma 4.6, the sampling survival Lemma 4.9, and the measure-preservation argument in Section 4.6. Assumption A.1 is explicitly a domain assumption on the population operator, not the conclusion; in the convex subdifferential case it is automatic by standard external facts about convex analysis. Lemma 4.1, the pointwise empirical convergence of Hausdorff distances, is proved in Appendix B.1 via symmetrization and concentration, rather than assumed or fitted. The algorithm's selection step uses an independent sample P_n' and Lemma 3.3 only as a statistical estimator of the population subdifferential; no fitted constant is later called a prediction. The self-citations [7] and [24] appear in motivational examples and in a discussion of distribution-free optimization, and neither is load-bearing for Theorem 1, Theorem 2, or Corollary 3.1. The only notable issue is a non-circular correctness gap: Lemma 4.6 states "Assume G has inductive dimension indG = d (and hence Hausdorff measure 0 < H^d(G) < ∞)", but the implication indG = d ⇒ H^d(G) < ∞ is false for general compact sets, and H^d(G) < ∞ is used in inequality (10). This is a mathematical proof gap, repairable via the Minty bi-Lipschitz parameterization and compactness, not a reduction of the result to its own inputs. Hence no circularity step is present.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No fitted constants or invented entities. The paper's constants (c, c', ε, λ, b, r, α_n) are existential or user-chosen, not fit to data. The load-bearing mathematical assumptions are standard theorems in monotone operator theory and geometric measure theory plus the stated Assumption A.1.

axioms (5)
  • standard math Maximal monotone operators have a bi-Lipschitz Minty parameterization (Rockafellar–Wets Thm 12.15).
    Used in Section 2.2 and Lemma 4.5 to identify d-dimensional graph pieces; the key geometric engine.
  • standard math Va˘inˇste˘in's first theorem, the sum theorem, and the volume lower bound indX≥k ⇒ H^k(X)>0.
    Used in Lemma 4.6 to force the existence of a positive-measure base X_{k,ε}; this is the non-quantitative core.
  • standard math Federer's Fubini-type inequality for Hausdorff measures (Thm 2.10.25).
    Used in Eq. (10) to bound H^{d−k}(π(K)) from above and in Section 4.6 to lower-bound H^d(G_n).
  • domain assumption For convex f_z, ∂F_P(x)=E[∂f_z(x)] and ∂f is maximal monotone.
    Section 1.2.1; maps the convex problem into the monotone-operator setting; standard under their measurability/integrability conventions.
  • domain assumption Assumption A.1: A_P+N_X maximal monotone, nonempty solution set, p≥2 moments near x0.
    Standing assumptions stated before Theorem 1; used to make prox well-defined and to get pointwise Hausdorff convergence.

pith-pipeline@v1.3.0-alltime-deepseek · 29286 in / 18293 out tokens · 170932 ms · 2026-08-02T08:09:44.839089+00:00 · methodology

0 comments
read the original abstract

We consider the problem of finding stationary points for stochastic convex optimization problems. Rather than surrogates to stationarity, such as a proximity-to-stationarity guarantee or small gradient of the Moreau envelope, we ask for a stronger notion: that the subdifferential of the objective actually contains a small element. This criterion is non-trivial, because subdifferentials of convex functions fail to converge uniformly, even in arbitrarily small neighborhoods of the optimum. Our convergence guarantees rely on dimension theory to decompose the graph of the subdifferential of a convex function, showing how stochastic sampling preserves "pieces" of these graphs, and allowing effective application of proximal-point-like methods.

Figures

Figures reproduced from arXiv: 2607.06883 by Felipe Areces, John Duchi, Malo Sommers.

Figure 1
Figure 1. Figure 1: Illustrations of partitions of gr ∂f. (a) The graph gr ∂f of f(x) = |x| as in Example 6. The set K1 = {0} × [−1, 1] = {0} × ∂f(0) corresponds to the vertical segment, while K0 = R>0 × {1} ∪ R<0 × {−1} consists of the two partially open horizontal segments. (b) The function f(x) = max{f0(x), f1(x)} for f0(x) = ∥x − e1∥ 2 2 and f1(x) = 1 2 ∥x + e1∥ 2 2 in Example 8. On the illustrated 1-dimensional sphere S … view at source ↗
Figure 1
Figure 1. Figure 1: Illustrations of partitions of gr ∂f. (a) The graph gr ∂f of f(x) = |x| as in Example 6. The set K1 = {0} × [−1, 1] = {0} × ∂f(0) corresponds to the vertical segment, while K0 = R>0 × {1} ∪ R<0 × {−1} consists of the two partially open horizontal segments. (b) The function f(x) = max{f0(x), f1(x)} for f0(x) = ∥x − e1∥ 2 2 and f1(x) = 1 2 ∥x + e1∥ 2 2 in Example 8. On the illustrated 1-dimensional sphere S … view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Lipschitzian SLLNs for random functions

    math.OC 2026-07 accept novelty 8.0

    Empirical averages of random locally Lipschitz functions converge to their expectation in the Lipschitz pseudometric under separability or NIP-definability conditions, giving uniform convergence of subdifferentials.

Reference graph

Works this paper leans on

53 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Agarwal, P

    A. Agarwal, P. L. Bartlett, P. Ravikumar, and M. J. Wainwright. Information-theoretic lower bounds on the oracle complexity of convex optimization.IEEE Transactions on Information Theory, 58(5):3235–3249, 2012

    2012

  2. [2]

    G. Alberti. On the structure of singular sets of convex functions.Calculus of Variations and Partial Differential Equations, 2(1):17–27, 1994. doi: 10.1007/BF01234313

    doi:10.1007/bf01234313 1994

  3. [3]

    Alberti and L

    G. Alberti and L. Ambrosio. A geometrical approach to monotone functions inR n. Mathematische Zeitschrift, 230(2):259–316, 1999

    1999

  4. [4]

    Alberti, L

    G. Alberti, L. Ambrosio, and P. Cannarsa. On the singularities of convex functions. Manuscripta Mathematica, 76(3–4):421–435, 1992. doi: 10.1007/BF02567770. 31

    work page doi:10.1007/bf02567770 1992

  5. [5]

    Ambrosio, P

    L. Ambrosio, P. Cannarsa, and H. M. Soner. On the propagation of singularities of semi- convex functions.Annali della Scuola Normale Superiore di Pisa - Classe di Scienze, Serie 4, 20(4):597–616, 1993

    1993

  6. [6]

    A. N. Angelopoulos and S. Bates. Conformal prediction: A gentle introduction.Foun- dations and Trends in Machine Learning, 16(4), 2023

    2023

  7. [7]

    Areces and J

    F. Areces and J. C. Duchi. Distribution free M-estimation.arXiv:2505.22807 [math.ST], 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Asi and J

    H. Asi and J. C. Duchi. The importance of better models in stochastic optimization. Proceedings of the National Academy of Sciences, 116(46):22924–22930, 2019

    2019

  9. [9]

    Aubin and H

    J.-P. Aubin and H. Frankowska.Set-Valued Analysis. Birkh¨ auser, 1990

    1990

  10. [10]

    R. F. Barber, E. J. Cand` es, A. Ramdas, and R. J. Tibshirani. The limits of distribution- free conditional predictive inference.Information and Inference, 10(2):455–482, 2021

    2021

  11. [11]

    H. H. Bauschke and P. L. Combettes.Convex Analysis and Monotone Operator Theory in Hilbert Spaces. Springer, Second edition, 2017

    2017

  12. [12]

    D. P. Bertsekas. Stochastic optimization problems with nondifferentiable cost functionals. Journal of Optimization Theory and Applications, 12(2):218–231, 1973

    1973

  13. [13]

    D. P. Bertsekas.Convex Optimization Algorithms. Athena Scientific, 2015

    2015

  14. [14]

    P. Bianchi. Ergodic convergence of a stochastic proximal point algorithm.SIAM Journal on Optimization, 26(4):2235–2260, 2016

    2016

  15. [15]

    Boyd and L

    S. Boyd and L. Vandenberghe.Convex Optimization. Cambridge University Press, 2004

    2004

  16. [16]

    Brøndsted and R

    A. Brøndsted and R. T. Rockafellar. On the subdifferentiability of convex functions. Proceedings of the American Mathematical Society, 16:605–611, 1965

    1965

  17. [17]

    Cauchois, S

    M. Cauchois, S. Gupta, and J. Duchi. Knowing what you know: valid and validated confidence sets in multiclass and multilabel prediction.Journal of Machine Learning Research, 22(81):1–42, 2021

    2021

  18. [18]

    P. L. Combettes and J. I. Madariaga. Asymptotic analysis of stochastic splitting methods for multivariate monotone inclusions.arXiv:2512.03023 [math.OC], 2025

    arXiv 2025

  19. [19]

    Davis and D

    D. Davis and D. Drusvyatskiy. Stochastic model-based minimization of weakly convex functions.SIAM Journal on Optimization, 29(1):207–239, 2019

    2019

  20. [20]

    Davis and D

    D. Davis and D. Drusvyatskiy. Graphical convergence of subgradients in nonconvex optimization and learning.Mathematics of Operations Research, 47(1):209–231, 2022

    2022

  21. [21]

    Davis and B

    D. Davis and B. Grimmer. Proximally guided stochastic subgradient method for nons- mooth, nonconvex problems.SIAM Journal on Optimization, 29(3):1908–1930, 2019

    1908

  22. [22]

    V. H. de la Pe˜ na and E. Gin´ e.Decoupling: From Dependence to Independence. Springer, 1999

    1999

  23. [23]

    Drusvyatskiy and C

    D. Drusvyatskiy and C. Paquette. Efficiency of minimizing compositions of convex func- tions and smooth maps.Mathematical Programming, Series A, 178(1–2):503–558, 2019. 32

    2019

  24. [24]

    J. C. Duchi. Sample-conditional coverage in conformal prediction. InAdvances in Neural Information Processing Systems 38, 2025

    2025

  25. [25]

    I. Ekeland. On the variational principle.Journal of Mathematical Analysis and Applica- tions, 47(2):324–353, 1974

    1974

  26. [26]

    Engelking.Dimension Theory

    R. Engelking.Dimension Theory. North-Holland Publishing Company, 1978

    1978

  27. [27]

    Federer.Geometric Measure Theory

    H. Federer.Geometric Measure Theory. Springer Berlin, Heidelberg, 1969

    1969

  28. [28]

    D. J. Foster, A. Sekhari, O. Shamir, N. Srebro, K. Sridharan, and B. Woodworth. The complexity of making the gradient small in stochastic convex optimization. InProceedings of the Thirty Second Annual Conference on Computational Learning Theory, 2019

    2019

  29. [29]

    Gibbs, J

    I. Gibbs, J. Cherian, and E. Cand` es. Conformal prediction with conditional guarantees. Journal of the Royal Statistical Society, Series B, 87(4):1100–1126, 2025

    2025

  30. [30]

    Hiriart-Urruty and C

    J. Hiriart-Urruty and C. Lemar´ echal.Convex Analysis and Minimization Algorithms I. Springer, New York, 1993

    1993

  31. [31]

    Hurewicz and H

    W. Hurewicz and H. Wallman.Dimension Theory. Princeton University Press, 1941

    1941

  32. [32]

    Juditsky, A

    A. Juditsky, A. Nemirovski, and C. Tauvel. Solving variational inequalities with the stochastic mirror-prox algorithm.Stochastic Systems, 1(1):17–58, 2011

    2011

  33. [33]

    Kenderov

    P. Kenderov. Multivalued monotone mappings are almost everywhere single-valued.Stu- dia Mathematica, 56(3):199–203, 1976

    1976

  34. [34]

    Kotsalis, G

    G. Kotsalis, G. Lan, and T. Li. Simple and optimal methods for stochastic variational inequalities, I: operator extrapolation.SIAM Journal on Optimization, 32(3):2041–2073, 2022

    2041

  35. [35]

    J. Lei. Classification with confidence.Biometrika, 101(4):755–769, 2014

    2014

  36. [36]

    Lei and L

    J. Lei and L. Wasserman. Distribution-free prediction bands for non-parametric regres- sion.Journal of the Royal Statistical Society, Series B, 76(1):71–96, 2014

    2014

  37. [37]

    Maggi.Sets of Finite Perimeter and Geometric Variational Problems: An Introduction to Geometric Measure Theory

    F. Maggi.Sets of Finite Perimeter and Geometric Variational Problems: An Introduction to Geometric Measure Theory. Cambridge University Press, 2012

    2012

  38. [38]

    J.-J. Moreau. Proximit´ e et dualit´ e dans un espace Hilbertian.Bulletin de la Soci´ et´ e Math´ ematique de France, 93:273–299, 1965

    1965

  39. [39]

    Nemirovski, A

    A. Nemirovski, A. Juditsky, G. Lan, and A. Shapiro. Robust stochastic approximation approach to stochastic programming.SIAM Journal on Optimization, 19(4):1574–1609, 2009

    2009

  40. [40]

    Nesterov.Introductory Lectures on Convex Optimization

    Y. Nesterov.Introductory Lectures on Convex Optimization. Kluwer Academic Publish- ers, 2004

    2004

  41. [41]

    R. T. Rockafellar. Measurable dependence of convex sets and functions on parameters. Journal of Mathematical Analysis and Applications, 28:4–25, 1969

    1969

  42. [42]

    R. T. Rockafellar.Convex Analysis. Princeton University Press, 1970. 33

    1970

  43. [43]

    R. T. Rockafellar. On the maximality of sums of nonlinear monotone operators.Trans- actions of the American Mathematical Society, 149(1):75–88, 1970

    1970

  44. [44]

    R. T. Rockafellar. Monotone operators and the proximal point algorithm.SIAM Journal on Control and Optimization, 14:877–898, 1976

    1976

  45. [45]

    R. T. Rockafellar and R. J. B. Wets.Variational Analysis. Springer, New York, 1998

    1998

  46. [46]

    Romano, E

    Y. Romano, E. Patterson, and E. J. Cand` es. Conformalized quantile regression. In Advances in Neural Information Processing Systems 32, 2019

    2019

  47. [47]

    F. Ruan. On the uniform convergence of subdifferentials in stochastic optimization and learning.Mathematics of Operations Research, To Appear, 2025. doi: 10.1287/moor. 2024.0533

    arXiv 2025

  48. [48]

    Shalev-Shwartz, O

    S. Shalev-Shwartz, O. Shamir, N. Srebro, and K. Sridharan. Stochastic convex opti- mization. InProceedings of the Twenty Second Annual Conference on Computational Learning Theory, 2009

    2009

  49. [49]

    Shapiro and H

    A. Shapiro and H. Xu. Uniform laws of large numbers for set-valued mappings and subdifferentials of random functions.Journal of Mathematical Analysis and Applications, 325(2):1390–1399, 2007

    2007

  50. [50]

    Tian and J

    L. Tian and J. Royset. Failure of uniform laws of large numbers for subdifferentials and beyond.arXiv:2511.16568 [math.OC], 2025

    arXiv 2025

  51. [51]

    A. W. van der Vaart and J. A. Wellner.Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York, 1996

    1996

  52. [52]

    V. Vovk, A. Grammerman, and C. Saunders. Machine-learning applications of algorith- mic randomness. InProceedings of the Sixteenth International Conference on Machine Learning, 1999

    1999

  53. [53]

    M. J. Wainwright.High-Dimensional Statistics: A Non-Asymptotic Viewpoint. Cam- bridge University Press, 2019. 34

    2019