Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Drift Optimization of Regulated Stochastic Models Using Sample Average Approximation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper proves that solving a three-level discretized sample average approximation of a drift optimization problem by mirror descent yields an explicit optimality gap bounded by four additive error terms, one for optimization steps…

desk verdict Useful framework, but the main rate theorem's convexity assumption fails for the flagship Skorokhod example, and Theorem 2's proof overstates the Gaussian complexity rate. read the letter →

arxiv 2506.06723 v1 pith:IKKF24YS submitted 2025-06-07 math.OC stat.AP

classification math.OCstat.AP MSC 90C1593E2060H30
keywords driftoptimizationsampleaverageapproximationregulatedstochasticprocessesSkorokhodregulatormirrordescentoptimalitygapfunction-spacediscretizationprogramming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces a drift optimization formulation in which the decision variable is a function that additively shifts the paths of a stochastic process, possibly through a regulator that enforces path constraints, and the goal is to minimize an expected cost over that function space. Because the feasible set is infinite-dimensional, the authors build a Sample Average Approximation that simultaneously discretizes time, samples paths, and projects the feasible set onto a finite-dimensional subspace, then solves the resulting finite convex program with mirror descent. The central result (Theorem 3) is an explicit finite-time guarantee: for every number of optimization steps $k$, samples $N$, time-step $h$, and subspace dimension $n$, the expected gap between the computed objective value and the true optimum is at most $c_1/\sqrt{k} + c_2/\sqrt{N} + c_3 h^{\beta} + c_4 g(n)$. If correct, this is the first combined convergence-rate statement for drift optimization of regulated processes, and it turns the practical question of how to spend a fixed computational budget into a concrete allocation problem.

What carries the argument

The machinery has three pieces. First, the Skorokhod regulator map $\Gamma(y)(t)=y(t)+\sup_{s\le t}(-y(s))^+$ and its pathwise directional derivative, derived through Danskin's theorem: $D_u\Gamma(y)(t)=u(t)+\sup_{s\in\Phi_t(y)}\{-u(s)\}$ when the path has gone below zero or starts at zero with $u(0)<0$, and $D_u\Gamma(y)(t)=u(t)$ otherwise. Second, Gaussian complexity controls the supremum deviation of the SAA objective over the function space, giving consistency without discretizing paths. Third, mirror descent on a finite-dimensional subspace, with step size tuned to sampled Lipschitz constants, supplies the optimization error; its regret bound produces the $O(1/\sqrt{k})$ term, or $O(1/k)$ with pathwise gradients. The three discretization layers enter through the weak-convergence assumption on path sampling and the subspace-approximation rate $g(n)$.

What would settle it

Take a fixed reflected Brownian path (or an Euler-discretized version), choose two drift functions $F_1,F_2$ in a finite basis, and compare $J_{N,h}$ at the midpoint with the average of $J_{N,h}$ at the endpoints; a single path satisfying $J_{N,h}((F_1+F_2)/2) > (J_{N,h}(F_1)+J_{N,h}(F_2))/2$ disproves the convexity premise, and repeating over many paths can show whether Assumption 12 fails for the canonical queueing example.

Watch

Extended reading notes

Core claim

The paper claims that an implementable estimator for the infinite-dimensional drift problem can be built by solving a finite sample-average problem with mirror descent, and that the expected optimality gap of the resulting point $F^*_{N,n,k}$ decomposes additively into four sources: optimization error $O(1/\sqrt{k})$, Monte Carlo error $O(1/\sqrt{N})$, path-approximation error $O(h^{\beta})$, and subspace-projection error $O(g(n))$. This is Theorem 3. In the special case where the pathwise gradient is used in mirror descent, the optimization term improves to $O(1/k)$. Along the way the paper proves that the pathwise directional derivative of the cost is an unbiased estimator of the derivative of the expected cost, provides a Danskin-theorem derivation of the directional derivative of the Skorokhod regulator map, and establishes consistency of the functional SAA without early discretization.

Load-bearing premise

The entire rate bound rests on Assumption 12, which says that for each sample path the cost of the regulated process is convex in the drift function; if that fails, as it generally does when the regulator is the Skorokhod map, the stated finite-sample rates are unsupported.

Editorial extensions

If this is right

  • For a fixed computational budget $B$, solving the allocation problem with $h = nkN/B$ gives exponents for $k$, $N$, $n$, and $h$ that depend on the smoothness $\alpha$ of the function space and the weak order $\beta$ of the path sampler.
  • Rougher function spaces force the budget toward function approximation, while smoother spaces allow more effort in sampling and optimization.
  • The unbiased pathwise derivative estimator legitimizes gradient-based simulation optimization directly on regulated paths, without solving a Hamilton-Jacobi-Bellman equation.
  • Consistency of the functional SAA holds without discretization, extending finite-dimensional SAA results to decision variables in a Banach space of paths.
  • When the pathwise gradient is available, the optimization error drops to $O(1/k)$, making the mirror-descent step count less costly relative to the other three error sources.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The convexity assumption needed for Theorem 3 is not verified for the reflected-Brownian motivating examples; for the Skorokhod regulator, composition with a convex path cost is generally nonconvex in the drift function, so the stated finite-sample rates remain conditional.
  • The additive error decomposition suggests a testable diagnostic: estimate each gap term separately and compare its empirical order with the predicted exponents; a mismatch would indicate a violated Lipschitz, convexity, or weak-convergence condition.
  • A natural extension is to closed-loop or feedback drift policies, where the same SAA structure would require derivatives of the regulated process with respect to a parameterized policy rather than an additive path shift.
  • The Danskin derivation treats regulation as a minimax operator, which may yield pathwise derivative formulas for more general regulators beyond the Skorokhod map.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a sample average approximation (SAA) framework for drift optimization problems of the form min_{F in F} J(F) = ∫ J~∘Γ(z+F) π_x(dz), where Γ is a Lipschitz regulator such as the Skorokhod map. The authors derive pathwise directional derivatives for the Skorokhod regulator using Danskin's theorem, prove unbiasedness of the derivative estimator, establish consistency of a functional SAA via Gaussian complexity arguments, and give a finite-time optimality gap for a discretized SAA solved by mirror descent. The final gap bound in Theorem 3, Eq. (22), is a sum of terms for optimization steps, Monte Carlo error, path discretization error, and function-space approximation error, and Section 7 uses this bound to derive an asymptotically optimal budget allocation.

Significance. If the main results were fully supported, the paper would make a useful contribution: it is the first work to give explicit finite-time SAA guarantees for drift optimization of regulated stochastic processes, and the error decomposition plus budget-allocation analysis is novel. The connection between Danskin's theorem and directional derivatives of the Skorokhod regulator is also an interesting observation. However, the headline rate in Theorem 3 rests on Assumption 12, which is not verified for the paper's motivating Skorokhod-regulated example and, as shown below, is actually false for a simple instance of that setting. The consistency proof also contains an incorrect Gaussian-complexity rate. These issues materially narrow the scope of the claimed results as they stand.

major comments (3)
  1. [Section 6, Assumption 12 and Theorem 3, Eq. (22)] Assumption 12, which postulates pathwise convexity of F ↦ J~∘Γ(Z+F), is load-bearing for the mirror-descent bound and for the subgradient inequality used in Eq. (27), but it is neither verified for Example 1 nor implied by Assumptions 1–4. For a concrete failure, take d=1, T=1, Z≡0, F={ct : c∈[-1,1]}, and J~(y) = -y(1). The Skorokhod regulator gives Γ(F)(1) = max(c,0), so J~(Γ(F)) = -max(c,0), which is concave in c. Hence Assumption 12 is violated for a decreasing linear cost, and Theorem 3's rate (22) is unsupported for the flagship regulated-process setting. Moreover, the sentence after Assumption 4 claiming that J is convex 'as a straightforward implication' is not valid in general; convexity of J~ does not make J~∘Γ convex in F without a monotonicity condition on J~ relative to the order structure of Γ.
  2. [Section 5, Proposition 2 and proof of Theorem 2, Eq. (11)] The Gaussian complexity rate is miscomputed. Substituting Proposition 2's bound into the definition of R_N(F) in Eq. (10) gives R_N(F) ≤ C E[||K_Z||_p]/N, and since E[||K_Z||_p] scales like N^{1/p}, the correct intermediate rate is O(N^{1/p-1}), not O(1/N) as claimed in Eq. (11). The proof's statement that O(1/N) is 'a consequence of Assumption 3' ignores the N^{1/p} scaling of the p-norm of the Lipschitz constants. In addition, Lemma 5 uses a Hölder conjugate q ≥ 2 while Assumption 3 has p ≥ 2, which implies q ≤ 2; the sub-Gaussian constant L in that lemma is therefore not derived correctly. The final consistency rate in Theorem 2 may still hold for p ≥ 2, but the stated intermediate rate and its proof need correction.
  3. [Section 4, Lemma 4] The proof of L^p convergence in Lemma 4 is incomplete. Lemma 3 provides only a pointwise asymptotic statement of the form o(||u||) as ||u||→0, and the argument '1/ε o(||εu||) = o(||u||), and ||u|| is bounded, so by dominated convergence' does not supply the required dominated convergence hypothesis for the sequence of difference quotients as ε→0. A uniform integrable domination of the quotients, e.g., from the Lipschitz property of Γ, is needed, and the measure with respect to which the L^p convergence is claimed should be stated explicitly. This gap should be repaired because Lemma 4 is one of the stated technical foundations for the derivative-based SAA method.
minor comments (5)
  1. [Section 4, Lemma 1] Lemma 1 as stated omits Assumption 7 (Gâteaux differentiability of Γ), which is needed for the chain-rule identity; the surrounding text mentions Assumption 7, but the lemma statement should include it.
  2. [Section 5, Theorem 2] Theorem 2 states that Assumption 3 holds for some 1 ≤ p < ∞, whereas Assumption 3 is stated for 2 ≤ p < ∞; the ranges should be made consistent.
  3. [Section 6, Eq. (23)] Equation (23) displays the projection-error term with a minus sign, but the subsequent bound in Eq. (27) treats it as a positive upper bound; this sign inconsistency should be corrected.
  4. [Section 4, Proposition 1] There is a typo in the text preceding Proposition 1: 'direciotnal' should be 'directional'.
  5. [Section 6, Assumption 11] Assumption 11 postulates a uniform-in-F bias bound of order h^β; this is stronger than a standard weak-convergence-order statement for a fixed F, and the authors should either verify it for the listed examples or state explicitly how it follows from the cited approximation schemes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 3 is a conditional bound from stated assumptions; no fitted parameter is relabeled as a prediction, and self-citations are contextual only.

full rationale

The derivation chain is self-contained conditional on Assumptions 1-12. Theorem 2 uses Gaussian complexity and equiconvergence arguments; Theorem 3 decomposes the optimality gap into optimization, sampling, discretization, and projection errors, then applies standard mirror-descent bounds. The constants c1-c4 are read off from the assumptions (e.g., c3 = l1 from Assumption 11, c4 = E[K_Z] from Assumption 3) rather than fitted to data, so the boundedness of each error term is a direct consequence of an explicitly stated hypothesis, not a hidden reuse of the conclusion. The budget allocation in Section 7 is obtained by minimizing the derived bound subject to a computational budget constraint, so the allocation exponents are mathematical consequences of the theorem rather than inputs built into the theorem. Self-citations such as Honnappa et al. (2015), Selk et al. (2021), and Armony et al. (2019) appear as motivation, examples, or prior context and are not load-bearing in the proofs of the main results. Assumption 12 (pathwise convexity of F -> J~ o Gamma(Z + F)) is a strong condition, and whether it holds for the Skorokhod-regulated Example 1 is a legitimate correctness concern; however, an unverified or even false stated assumption is a soundness risk, not circularity. No equation in the paper is equivalent by construction to the result it purports to establish, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the conclusion.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The paper's main finite-time guarantee (Theorem 3) is conditional on a set of convexity and approximation assumptions. The most fragile is Assumption 12 (pathwise convexity), which is not verified for the Skorokhod-regulated examples and is likely violated because the reflection map introduces the positive part of a supremum. The other assumptions (Lipschitzness, covering numbers, weak convergence order, subspace rates) are standard for SAA and approximation theory, but their combination is strong.

assumptions (7)
  • domain assumption Assumption 12: The random functional F ↦ ˜J ∘ Γ(Z + F) is convex in F for each path Z.
    Used in the proof of Theorem 3 to apply the mirror descent complexity bound for convex objectives. Not verified for the Skorokhod-regulated examples; the regulator map introduces a positive-part sup, which generally breaks convexity in F.
  • domain assumption Assumption 11: There is a path approximation scheme with weak convergence order β > 0, i.e., sup_F |E[˜J∘Γ(F+Z_h)] - J(F)| ≤ ℓ1 h^β.
    This gives the c3 h^β term in the rate. The paper gives examples for Brownian motion, but the rate for the composed regulated objective is assumed, not proved.
  • domain assumption Assumption 9: The covering number of F satisfies log N(ε, F, ∥·∥∞) ≤ ε^{-1/α} for some α > 1.
    Used in Proposition 2 to bound the Gaussian complexity via Dudley's theorem. Reasonable for Sobolev-type spaces, but not verified for the specific feasible sets in the examples.
  • domain assumption Assumption 10: The finite-dimensional subspaces F_n satisfy sup_F ∥F - Π_{F_n}(F)∥ = O(g(n)) with g(n) → 0.
    Gives the c4 g(n) projection error. The specific rate g(n)=n^{-α} is assumed to derive budget allocation exponents.
  • domain assumption Assumptions 2 and 3: The cost is Lipschitz in the path (constant κ) and in the drift shift (constant K_z with E[K_z^p] < ∞).
    These Lipschitz conditions drive the sampling error and variance bounds in Lemma 7 and Proposition 3.
  • standard math Danskin's theorem (Lemma 2) for differentiating a supremum.
    Used in Proposition 1 to compute the directional derivative of the Skorokhod regulator as the derivative of a supremum over time points.
  • standard math Dudley's metric entropy bound and sub-Gaussian concentration inequalities (Kontorovich, Boucheron et al.).
    Used in the proof of Proposition 2 and Proposition 3 to bound the Gaussian complexity and deviation of the SAA objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Drift Optimization of Regulated Stochastic Models Using Sample Average Approximation." pith.science (2026). https://pith.science/paper/IKKF24YS

@misc{pith2026250606723,
  author       = {Pith},
  title        = {Pith review of: Drift Optimization of Regulated Stochastic Models Using Sample Average Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKKF24YS}},
  note         = {Machine review of arXiv:2506.06723}
}
read the original abstract

This paper introduces a drift optimization model of stochastic optimization problems driven by regulated stochastic processes. A broad range of problems across operations research, machine learning, and statistics can be viewed as optimizing the "drift" associated with a process by minimizing a cost functional, while respecting path constraints imposed by a Lipschitz continuous regulator. Towards an implementable solution to such infinite-dimensional problems, we develop the fundamentals of a Sample Average Approximation (SAA) method that incorporates (i) path discretization, (ii) function-space discretization, and (iii) Monte Carlo sampling, and that is solved using an optimization recursion such as mirror descent. We start by constructing pathwise directional derivatives for use within the SAA method, followed by consistency and complexity calculations. The characterized complexity is expressed as a function of the number of optimization steps, and the computational effort involved in (i)--(iii), leading to guidance on how to trade-off the computational effort allocated to optimization steps versus the "dimension reduction" steps in (i)--(iii).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 41 canonical work pages

  1. [1]

    Lectures on Stochastic Programming: Modeling and Theory, Third Edition

    Alexander Shapiro, Darinka Dentcheva, and Andrzej Ruszczynski. Lectures on Stochastic Programming: Modeling and Theory, Third Edition. Society for Industrial and Applied Mathematics, July 2021. ISBN 9781611976595. doi:10.1137/1.9781611976595. URL http://dx.doi.org/10.1137/1.9781611976595

  2. [2]

    Consistency of M onte C arlo Estimators for Risk-Neutral PDE -Constrained Optimization

    Johannes Milz. Consistency of M onte C arlo Estimators for Risk-Neutral PDE -Constrained Optimization . Applied Mathematics & Optimization , 87 0 (3): 0 57, Apr 2023 a . ISSN 1432-0606. doi:10.1007/s00245-023-09967-3. URL https://link.springer.com/content/pdf/10.1007/s00245-023-09967-3.pdf

  3. [3]

    Sample average approximations of strongly convex stochastic programs in H ilbert spaces

    Johannes Milz. Sample average approximations of strongly convex stochastic programs in H ilbert spaces. Optimization Letters, 17 0 (2): 0 471--492, Mar 2023 b . ISSN 1862-4480. doi:10.1007/s11590-022-01888-4. URL https://link.springer.com/content/pdf/10.1007/s11590-022-01888-4.pdf

  4. [4]

    Asymptotically optimal appointment schedules

    Mor Armony, Rami Atar, and Harsha Honnappa. Asymptotically optimal appointment schedules. Mathematics of Operations Research, 44 0 (4): 0 1345--1380, 2019. doi:10.1287/moor.2018.0973. URL https://doi.org/10.1287/moor.2018.0973

  5. [5]

    A survey of the S chr\"odinger problem and some of its connections with optimal transport

    Cédric L \'e onard. A survey of the S chr\"odinger problem and some of its connections with optimal transport. Discrete and Continuous Dynamical Systems - A, 34 0 (4): 0 1533--1574, 2014. doi:10.3934/dcds.2014.34.1533. URL https://doi.org/10.3934/dcds.2014.34.1533

  6. [6]

    Optimal Transport: Old and New, volume 338 of Grundlehren der mathematischen Wissenschaften

    Cédric Villani. Optimal Transport: Old and New, volume 338 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag Berlin Heidelberg, 2009. ISBN 978-3-540-71049-3. doi:10.1007/978-3-540-71050-9

  7. [7]

    Information projection on B anach spaces with applications to state independent KL -weighted optimal control

    Zachary Selk, William Haskell, and Harsha Honnappa. Information projection on B anach spaces with applications to state independent KL -weighted optimal control. Applied Mathematics and Optimization, 84: 0 S1081--S1115, 2021. doi:10.1007/s00245-021-09786-4. URL https://link.springer.com/article/10.1007/s00245-021-09786-4

  8. [8]

    Joris Bierkens and Hilbert J. Kappen. Explicit solution of relative entropy weighted control. Systems & Control Letters, 72: 0 36--43, 2014. doi:10.1016/j.sysconle.2014.08.001. URL https://doi.org/10.1016/j.sysconle.2014.08.001

Show all 55 references
  1. [9]

    Monge’s problem with a quadratic cost by the zero-noise limit of h-path processes

    Toshio Mikami. Monge’s problem with a quadratic cost by the zero-noise limit of h-path processes. Probability Theory and Related Fields, 129 0 (2): 0 245--260, 2004

  2. [10]

    L. C. G. Rogers and David Williams. Diffusions, Markov Processes, and Martingales: Volume 2, Itô Calculus. Cambridge University Press, 2nd edition, 2000. ISBN 9780521775941

  3. [11]

    Avi Mandelbaum and William A. Massey. Strong approximations for time-dependent queues. Mathematics of Operations Research, 20 0 (1): 0 33--64, 1995. doi:10.1287/moor.20.1.33. URL https://doi.org/10.1287/moor.20.1.33

  4. [12]

    Stochastic-Process Limits: An Introduction to Stochastic-Process Limits and Their Application to Queues

    Ward Whitt. Stochastic-Process Limits: An Introduction to Stochastic-Process Limits and Their Application to Queues. Springer New York, 2002 a . ISBN 9780387217482. doi:10.1007/b97479. URL http://dx.doi.org/10.1007/b97479

  5. [13]

    Harsha Honnappa, Rahul Jain, and Amy R. Ward. A queueing model with independent arrivals, and its fluid and diffusion limits. Queueing Systems, 80 0 (1): 0 71--103, Jun 2015. ISSN 1572-9443. doi:10.1007/s11134-014-9428-4. URL https://link.springer.com/content/pdf/10.1007/s1113...

  6. [14]

    Directional derivatives of oblique reflection maps

    Avi Mandelbaum and Kavita Ramanan. Directional derivatives of oblique reflection maps. Mathematics of Operations Research, 35 0 (3): 0 527--558, 2010. doi:10.1287/moor.1100.0453. URL https://doi.org/10.1287/moor.1100.0453

  7. [15]

    On directional derivatives of S korokhod maps in convex polyhedral domains

    David Lipshutz and Kavita Ramanan. On directional derivatives of S korokhod maps in convex polyhedral domains. The Annals of Applied Probability, 28 0 (2), April 2018. ISSN 1050-5164. doi:10.1214/17-aap1299. URL http://dx.doi.org/10.1214/17-AAP1299

  8. [16]

    Pathwise differentiability of reflected diffusions in convex polyhedral domains

    David Lipshutz and Kavita Ramanan. Pathwise differentiability of reflected diffusions in convex polyhedral domains. Annales de l'Institut Henri Poincar \'e , Probabilit \'e s et Statistiques , 55: 0 1439 -- 1476, 2019 a . doi:10.1214/18-AIHP924. URL https://doi.org/10.1214/18-AIHP924

  9. [17]

    A. Shapiro. M onte C arlo sampling methods. In A. Ruszczynski and Shapiro, editors, Stochastic Programming, Handbooks in Operations Research and Management Science, pages 353--426. Elsevier, 2004

  10. [18]

    A M onte C arlo method for estimating sensitivities of reflected diffusions in convex polyhedral domains

    David Lipshutz and Kavita Ramanan. A M onte C arlo method for estimating sensitivities of reflected diffusions in convex polyhedral domains. Stochastic Systems, 9 0 (2): 0 101--140, 2019 b . doi:10.1287/stsy.2019.0031. URL https://doi.org/10.1287/stsy.2019.0031

  11. [19]

    Gradient Estimation via Perturbation Analysis, volume 116

    Paul Glasserman. Gradient Estimation via Perturbation Analysis, volume 116. Springer Science & Business Media, 1990

  12. [20]

    R ademacher and G aussian complexities: Risk bounds and structural results

    Peter L Bartlett and Shahar Mendelson. R ademacher and G aussian complexities: Risk bounds and structural results. Journal of Machine Learning Research, 3 0 (Nov): 0 463--482, 2002

  13. [21]

    Asmussen and P.W

    S. Asmussen and P.W. Glynn. Stochastic simulation: Algorithms and analysis. Springer New York, 2007. ISBN 9780387690339. doi:10.1007/978-0-387-69033-9. URL http://dx.doi.org/10.1007/978-0-387-69033-9

  14. [22]

    A proposal on machine learning via dynamical systems

    Weinan E. A proposal on machine learning via dynamical systems. Communications in Mathematics and Statistics, 5 0 (1): 0 1--11, Mar 2017. ISSN 2194-671X. doi:10.1007/s40304-017-0103-z. URL https://link.springer.com/content/pdf/10.1007/s40304-017-0103-z.pdf

  15. [23]

    A mean-field optimal control formulation of deep learning

    Weinan E, Jiequn Han, and Qianxiao Li. A mean-field optimal control formulation of deep learning. Research in the Mathematical Sciences, 6 0 (1): 0 10, Dec 2018. ISSN 2197-9847. doi:10.1007/s40687-018-0172-y. URL https://link.springer.com/content/pdf/10.1007/s40687-018-0172-y.pdf

  16. [24]

    Dynamical Systems and Optimal Control Approach to Deep Learning, pages 422--438

    Weinan E Jiequn Han and Qianxiao Li. Dynamical Systems and Optimal Control Approach to Deep Learning, pages 422--438. Cambridge University Press, December 2022. ISBN 9781316516782. doi:10.1017/9781009025096.011. URL http://dx.doi.org/10.1017/9781009025096.011

  17. [25]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

  18. [26]

    Royset, Claire Walton, and Isaac Kaminer

    Chris Phelps, Qi Gong, Johannes O. Royset, Claire Walton, and Isaac Kaminer. Consistent approximation of a nonlinear optimal control problem with uncertain parameters. Automatica, 50 0 (12): 0 2987--2997, 2014. ISSN 0005-1098. doi:10.1016/j.automatica.2014.10.025. URL https://...

  19. [27]

    Royset, and Qi Gong

    Chris Phelps, Johannes O. Royset, and Qi Gong. Optimal control of uncertain systems using sample average approximations. SIAM Journal on Control and Optimization, 54 0 (1): 0 1--29, January 2016. ISSN 1095-7138. doi:10.1137/140983161. URL http://dx.doi.org/10.1137/140983161

  20. [28]

    R. W. Brockett. Optimal control of the L iouville equation. In Proceedings of the I nternational C onference on C omplex G eometry and R elated F ields , volume 39 of AMS/IP Stud. Adv. Math., pages 23--35. Amer. Math. Soc., Providence, RI, 2007. doi:10.1090/amsip/039/02. URL h...

  21. [29]

    Notes on the control of the L iouville equation

    Fatiha Alabau-Boussouira, Roger Brockett, Olivier Glass, J \'e r \^o me Le Rousseau, Enrique Zuazua, and Roger Brockett. Notes on the control of the L iouville equation. Control of Partial Differential Equations: Cetraro, Italy 2010, Editors: Piermarco Cannarsa, Jean-Michel Co...

  22. [30]

    A theoretical investigation of B rockett's ensemble optimal control problems

    Jan Bartsch, Alfio Borz \` , Francesco Fanelli, and Souvik Roy. A theoretical investigation of B rockett's ensemble optimal control problems. Calculus of Variations and Partial Differential Equations, 58 0 (5): 0 162, Sep 2019. ISSN 1432-0835. doi:10.1007/s00526-019-1604-2. UR...

  23. [31]

    A F okker- P lanck control framework for stochastic systems

    Mario Annunziato and Alfio Borz\' . A F okker- P lanck control framework for stochastic systems. EMS Surv. Math. Sci., 5 0 (1-2): 0 65--98, 2018. ISSN 2308-2151. doi:10.4171/EMSS/27. URL https://doi.org/10.4171/EMSS/27

  24. [32]

    Optimal Control of the Fokker--Planck Equation with Space-Dependent Controls

    Arthur Fleig and Roberto Guglielmi. Optimal Control of the Fokker--Planck Equation with Space-Dependent Controls . Journal of Optimization Theory and Applications, 174 0 (2): 0 408--427, Aug 2017. ISSN 1573-2878. doi:10.1007/s10957-017-1120-5. URL https://link.springer.com/con...

  25. [33]

    Markov Marginal Problems and Their Applications to Markov Optimal Control

    Toshio Mikami. Markov Marginal Problems and Their Applications to Markov Optimal Control. Birkh \"a user Boston, Boston, MA, 1999. ISBN 978-1-4612-1784-8. doi:10.1007/978-1-4612-1784-8_27. URL https://doi.org/10.1007/978-1-4612-1784-8_27

  26. [34]

    Stochastic Optimal Transportation Problem

    Toshio Mikami. Stochastic Optimal Transportation Problem . In Toshio Mikami, editor, Stochastic Optimal Transportation : Stochastic Control with Fixed Marginals , pages 21--75. Springer, Singapore, 2021. ISBN 9789811617546. doi:10.1007/978-981-16-1754-6_2. URL https://doi.org/...

  27. [35]

    Michael Harrison

    J. Michael Harrison. B rownian Models of Performance and Control . Cambridge University Press, 2013

  28. [36]

    Baris Ata, J. M. Harrison, and L. A. Shepp. Drift rate control of a B rownian processing system . The Annals of Applied Probability, 15 0 (2), May 2005. ISSN 1050-5164. doi:10.1214/105051604000000855. URL http://dx.doi.org/10.1214/105051604000000855

  29. [37]

    Drift control of high-dimensional reflected B rownian Motion: A Computational Method Based on Neural Networks

    Baris Ata, J Michael Harrison, and Nian Si. Drift control of high-dimensional reflected B rownian Motion: A Computational Method Based on Neural Networks . Stochastic Systems, 2024 a

  30. [38]

    Singular control of (reflected) B rownian motion: a computational method suitable for queueing applications

    Baris Ata, J Michael Harrison, and Nian Si. Singular control of (reflected) B rownian motion: a computational method suitable for queueing applications. Queueing Systems, pages 1--37, 2024 b

  31. [39]

    Optimization by vector space methods

    David G Luenberger. Optimization by vector space methods. John Wiley & Sons, 1997

  32. [40]

    Convex Functions and Their Applications, volume 23

    Constantin Niculescu and Lars-Erik Persson. Convex Functions and Their Applications, volume 23. Springer, 2006

  33. [41]

    Fundamentals of Queueing Networks: Performance, Asymptotics, and Optimization, volume 4

    Hong Chen and David D Yao. Fundamentals of Queueing Networks: Performance, Asymptotics, and Optimization, volume 4. Springer, 2001

  34. [42]

    Internet supplement to stochastic-process limits: An introduction to stochastic-process limits and their application to queues

    Ward Whitt. Internet supplement to stochastic-process limits: An introduction to stochastic-process limits and their application to queues. 2002 b . URL https://www.columbia.edu/ ww2040/supplementno.pdf

  35. [43]

    On a theorem of D anskin with an application to a theorem of V on N eumann- S ion

    Pierre Bernhard and Alain Rapaport. On a theorem of D anskin with an application to a theorem of V on N eumann- S ion. Nonlinear Analysis: Theory, Methods & Applications, 24 0 (8): 0 1163--1181, 1995

  36. [44]

    Inverse problems: A B ayesian perspective

    Andrew M Stuart. Inverse problems: A B ayesian perspective. Acta Numerica, 19: 0 451--559, 2010

  37. [45]

    Kreyszig

    E. Kreyszig. Introductory Functional Analysis with Applications . Wiley, New York, W iley classics library edition, 1989. ISBN 0471504599

  38. [46]

    Wong- Z akai approximations for stochastic differential equations

    Krystyna Twardowska. Wong- Z akai approximations for stochastic differential equations. Acta Applicandae Mathematicae, 43 0 (3): 0 317--359, 1996. doi:10.1007/BF00047670. URL https://link.springer.com/article/10.1007/BF00047670

  39. [47]

    Grebenkov, Dmitry Beliaev, and Peter W

    Denis S. Grebenkov, Dmitry Beliaev, and Peter W. Jones. A multiscale guide to B rownian motion. Journal of Physics A: Mathematical and Theoretical, 49 0 (4): 0 043001, 2016. doi:10.1088/1751-8113/49/4/043001. URL https://iopscience.iop.org/article/10.1088/1751-8113/49/4/043001

  40. [48]

    An optimal polynomial approximation of B rownian motion

    James Foster, Terry Lyons, and Harald Oberhauser. An optimal polynomial approximation of B rownian motion. SIAM Journal on Numerical Analysis, 58 0 (3): 0 1393--1421, 2020

  41. [49]

    Nesterov

    Y. Nesterov. Introductory Lectures on Convex Optimization: A Basic Course. Springer Science + Business Media, LLC, New York, NY, 2004

  42. [50]

    S. Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends in Machine Learning, 8 0 (3--4): 0 231--358, 2015

  43. [51]

    Philip E. Protter. Stochastic Integration and Differential Equations. Springer Berlin Heidelberg, 2005. ISBN 9783662100615. doi:10.1007/978-3-662-10061-5. URL http://dx.doi.org/10.1007/978-3-662-10061-5

  44. [52]

    Xiaoying Han and Peter E. Kloeden. Random Ordinary Differential Equations and Their Numerical Solution. Springer Singapore, 2017. ISBN 9789811062650. doi:10.1007/978-981-10-6265-0. URL http://dx.doi.org/10.1007/978-981-10-6265-0

  45. [53]

    Concentration Inequalities: A Nonasymptotic Theory of Independence

    Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, 02 2013. ISBN 9780199535255. doi:10.1093/acprof:oso/9780199535255.001.0001. URL https://doi.org/10.1093/acprof:oso/9780199535255.001.0001

  46. [54]

    Probability in Banach Spaces

    Michel Ledoux and Michel Talagrand. Probability in Banach Spaces. Springer Berlin Heidelberg, 1991. ISBN 9783642202124. doi:10.1007/978-3-642-20212-4. URL http://dx.doi.org/10.1007/978-3-642-20212-4

  47. [55]

    Concentration in unbounded metric spaces and algorithmic stability

    Aryeh Kontorovich. Concentration in unbounded metric spaces and algorithmic stability. In Proceedings of the 31st International Conference on Machine Learning. June 22 nd -24 th , Beijing, China, 28--36, 2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.