Pith. sign in

REVIEW 3 major objections 5 minor 49 references

A single weighted model fuses sample data and moment information, with weight C/√N, to get SAA-level 1/√N optimality at every data size.

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 ↗

Weighted SAA and moment-DRO with lambda = C/sqrt(N) combines data and information, is asymptotically optimal at a 1/sqrt(N) rate, and improves scenario reduction.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The HO idea is solid and the numerics are convincing, but the asymptotic CLT in Proposition 8 is wrong as stated, and the finite-sample guarantee does not cover the implemented method. the 3 major comments →

arxiv 2508.18767 v1 pith:GG55FIWX submitted 2025-08-26 math.OC

Harmonizing SAA and DRO

Ziliang Jin , Jianqiang Cheng , Daniel Zhuoyu Long , Kai Pan This is my paper
classification math.OC MSC 90C1590C2290C47
keywords sample average approximationdistributionally robust optimizationmoment ambiguity setasymptotic optimalityfinite-sample guaranteescenario reduction1/sqrt(N) ratestochastic programming
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 reading

SAA uses historical samples but degrades when data are scarce; moment-based DRO uses partial distributional information but stays conservative when data are plentiful. This paper's harmonizing optimization (HO) minimizes a weighted sum of the empirical expectation and the worst-case expectation over a moment ambiguity set, with the information weight λ = C/√N. The weight shrinks automatically as N grows, so data dominate for large samples and information carries small samples. The paper proves that, provided the empirical distribution lies in the projected ambiguity set, the true distribution is contained with high confidence for finite N, and that the optimal value converges to the true optimum with a 1/√N-rate gap. Why care: one would not need to decide whether the available data count as large or small, which normally requires knowing the problem's effective dimension and geometry.

Core claim

The central claim is that there is a single model interpolating SAA and moment-based DRO: Γ(λ) = min_x [(1−λ) E_{P0} f(x,ξ) + λ max_{P∈D} E_P f(x,ξ)], with λ = C/√N. Proposition 8 states Γ(λ) = inf_{x∈X*} F_λ(x) + O(1/√N) and √N(Γ(λ)−V*) converges in distribution to inf_{x∈X*} Y(x) with Y(x)∼N(0,σ²(x)); when the optimum is unique, the gap is asymptotically N(0,σ²(x*)), so HO matches SAA's optimality rate while retaining DRO's protection for small N. The model is equivalent to a DRO with ambiguity set D_H(λ)={(1−λ)P0+λPξ : Pξ∈ΠξD}, whose size shrinks as λ decreases. For common moment sets the paper gives finite-sample containment of the true distribution (Proposition 6) and tractable LP/SDP r

What carries the argument

The load-bearing mechanism is the convex mixture ambiguity set D_H(λ) — every candidate distribution is a (1−λ) share of the empirical distribution P0 plus a λ share of some distribution in the original moment ambiguity set. It converts the choice between SAA and DRO into one scalar λ = C/√N, making the effective ambiguity set shrink as N grows. The Gelbrich distance, a metric on mean-covariance pairs, upper-bounds Wasserstein-2 distance under moment constraints and is used to define λ* by bisection so that the set of distributions within distance ϵ of P0 is contained in D_H(λ*); this containment drives Proposition 6's confidence guarantee. A dual reformulation for moment ambiguity sets yiel

Load-bearing premise

The finite-sample guarantee requires the empirical distribution P0 to lie inside the projected moment ambiguity set ΠξD, which for the equality-mean set used in the experiments holds only if the sample mean coincides with the nominal mean; the constants c1 and c2 needed to compute the associated epsilon are not specified, so the guarantee cannot yet be instantiated numerically.

What would settle it

Take the DD ambiguity set from Appendix B.2 with a nominal mean μ different from the sample mean of N draws from a known P. Compute λ* by the paper's bisection algorithm and check whether the true P is in D_H(λ*). Because P0∉ΠξD, Proposition 6's premise fails exactly when the empirical mean differs from μ, so repeated simulations would show that the claimed ≥1−β containment does not hold; alternatively, once c1 and c2 are fixed on a small synthetic problem, compare the predicted β from (6) with empirical containment frequencies.

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

If this is right

  • A user with N samples does not need to classify N as large or small: λ = C/√N automatically amplifies data as N grows, so HO's out-of-sample performance avoids both SAA's small-data fragility and DRO's large-data conservatism.
  • The constant C is estimated once (cross-validation, confidence-interval tightening, or C=√M0) and then reused for every N, so many instances of the same problem with different sample sizes are solved at no extra calibration cost.
  • For moment ambiguity sets with support, mean-covariance, or mean-absolute-deviation information, HO is a tractable LP or SDP, so the interpolation is directly usable in practice.
  • In scenario reduction, using M<N scenarios with λ=1−√(M/N) preserves the dropped scenarios' moment information; the experiments report approximation errors near 1% with M as small as 10 while total time stays far below solving the full N-scenario SAA.
  • When the original problem has a unique optimum, the asymptotic gap is o(1/√N), faster than the generic O(1/√N); this is a direct corollary of the singleton case in Proposition 8.

Where Pith is reading between the lines

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

  • A testable extension: if a moment ambiguity set is replaced by any convex family of distributions closed under mixtures, the same D_H(λ) construction yields an indexed family of estimators; observing whether the 1/√N rate survives would reveal how much of the theory depends on moment structure rather than convexity.
  • Because Proposition 5 shows that combining SAA with a Wasserstein ball merely rescales the radius, a reader can infer that the qualitative gains of HO over Wasserstein DRO in the experiments come from the moment ambiguity set itself, not from the mixing alone.
  • One consequence of the paper's own formulas (5)–(6) is that numeric use of the finite-sample guarantee requires the constants c1 and c2; since they are not specified, a user cannot instantiate the stated confidence level without additional derivation.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes a 'harmonizing optimization' (HO) model that minimizes a convex combination of the SAA objective and a moment-based DRO objective, with weight λ = C/√N for the DRO/information term. The authors claim that HO adaptively balances data and distributional information, works without classifying sample sizes as large or small, admits a finite-sample performance guarantee, is asymptotically optimal at the 1/√N rate, and can be used for scenario reduction. The paper provides tractable reformulations (LP/SDP), three heuristics for choosing C, and numerical experiments on mean-risk portfolio optimization and lot sizing, where HO is compared with Wasserstein DRO and existing scenario-reduction methods.

Significance. The harmonizing idea is appealing and the application to scenario reduction is a useful practical contribution. The paper contains several correct and useful components: Proposition 5 cleanly shows that linearly combining SAA with Wasserstein DRO is equivalent to a single Wasserstein DRO with scaled radius; the tractable reformulations in Propositions 9 and 10 are concrete; the numerical study demonstrates that HO can outperform Wasserstein DRO in the tested settings; and the 1/√N convergence-rate claim in Proposition 8 is plausible. I found no circular reasoning in the central derivation. However, the two headline theoretical guarantees are not currently established as stated: the distributional limit in Proposition 8 is incorrect because an O(1/√N) remainder cannot be discarded in the CLT, and the finite-sample guarantee in Proposition 6 is not operational because the constants c1,c2 are unspecified and the key assumption P0∈ΠξD fails for the main experiments. If the authors correct these points, the paper would be a solid contribution to the SAA/DRO interface.

major comments (3)
  1. [§3.3, Appendix C.7, Eq. (22), (25)–(26)] The proof of Proposition 8 derives |Γ(λ) − VN| ≤ C1/√N in Eq. (22) and then substitutes Γ(λ) for VN in the SAA CLT (25)–(26) to obtain (7)–(8). This step is invalid: Slutsky-type arguments require the remainder to be o_p(1/√N), not merely O(1/√N). Since λ = C/√N and D is fixed, the remainder is generically of exact order 1/√N. Let h(x) = max_{P∈D} E_P[f(x,ξ)] − F(x). For singleton X* = {x*}, the correct limit is √N(Γ(λ) − V*) ⇒ Y(x*) + C·h(x*), where Y(x*) ∼ N(0, σ²(x*)); with multiple optima it is inf_{x∈X*}[Y(x) + C·h(x)], not inf_{x∈X*}Y(x). Thus (7), (8), and the centered-bias formula (9) are not established and are generally false for the piecewise-linear losses used in Section 5, where h(x*) > 0 unless the worst-case distribution in D has the same expectation as P at x*. The O(1/√N) rate itself survives, but the claimed equivalence to the SAA limiting law does not.
  2. [§3.2, Proposition 6, Eq. (4)–(6); Appendix B.2; §5.1] The finite-sample guarantee requires P0 ∈ ΠξD. For the MAD ambiguity set DD used in the portfolio experiments, ΠξDD includes the exact mean constraint E_{P0}[ξ] = µ; in Section 5.1, µ is the true mean, so the empirical distribution P0 lies in ΠξDD only by chance. Thus the stated guarantee does not apply to the implemented method in the headline experiment. For the lot-sizing experiments, µ is estimated from the training sample, so the mean condition may hold by construction, but the MAD condition E_{P0}[|ξ−µ|] ≤ δ is not verified in the text. In addition, the constants c1 and c2 in Eq. (5) are never specified, so ε in Eq. (6) cannot be computed, and λ* cannot be determined from the paper. The bisection algorithm in Appendix C.5 therefore is not operational. The paper should either supply usable constants, state the guarantee in a form that can be evaluated from data, or clearly separate
  3. [Appendix C.4, final paragraph of proof of Proposition 6] The proof of the claim 'For any λ ∈ [1−β,1], P(P∈DH(λ)) ≥ 1−β' contains a probabilistic confusion. The text states: 'given any λ, we have P(P = (1−λ)P0 + λP) = λ.' Here P on the left-hand side is the true distribution and P0 is the random empirical distribution; the event {P = (1−λ)P0 + λP} has probability 0 for λ<1 (unless P0=P almost surely), not λ. The mixture probability λ is being treated as if it were the probability that a fixed true distribution coincides with a random mixture. This part of Proposition 6 therefore lacks a valid proof. The main inclusion via Fournier–Guillin may be salvageable, but the additional λ∈[1−β,1] statement needs correction or removal.
minor comments (5)
  1. [Eq. (6)] Typo 'otherwie' should be 'otherwise'. Also the regime condition 'N ≥ log(c1β^{-1})/c2' is written in a way that mixes the sample size N with the exponent; clarify the intended threshold.
  2. [Proposition 2] The statement 'FN(x) ∼ N(F(x), σ²(x)/N)' is informal; the CLT is about √N(FN(x)−F(x)). The confidence interval is only approximate. Consider stating this as an asymptotic approximation.
  3. [Figures 1–3 and Tables] The label 'Wassertein' appears in Figures 1–3; should be 'Wasserstein'. Also the paper would benefit from error bars or standard deviations across the 200 runs, since out-of-sample performance differences can be small.
  4. [§3.4 method (ii)] The construction x(C) = (1−C/√N)x_SAA + (C/√N)x_DRO is presented without justification. It is a heuristic interpolation; state explicitly that this is an approximation used to choose C, not an exact equivalence.
  5. [Appendix D.3] In the lot-sizing experiments, the MAD parameter δ for DD is not specified in the text. For reproducibility, give the exact values or estimation procedure used for δ.

Circularity Check

0 steps flagged

No significant circularity: the central HO derivation is self-contained and rests on external concentration and SAA results.

full rationale

The paper's central derivation chain is: (HO) is defined as a convex combination of SAA and moment-DRO objectives; Proposition 3 rewrites it as a DRO against DH(λ); Proposition 6 uses the Fournier–Guillin concentration bound plus Gelbrich/Wasserstein inclusions to show P∈DH(λ*) with confidence; Proposition 7 proves Γ(λ)→V* from uniform convergence of FN plus λ→0; Proposition 8 derives the O(1/√N) gap from λ=C/√N and then attempts to transfer the SAA CLT. None of these steps define the conclusion into the assumptions. The O(1/√N) rate is a direct consequence of the chosen scaling λ=C/√N (a designed rate, not a hidden fit), and the CLT step is a separate mathematical-error issue: (22) only gives Γ(λ)−VN=O(1/√N), so substituting into √N(VN−V*) yields an extra O(1) term; Slutsky would require o_p(1/√N). That is a correctness/technical gap, not a circular reduction. Self-citations (Cheramin et al. 2022 for SDP reformulation; Long et al. 2024 for two-stage solution algorithms) are used as computational tools and are not the source of the asymptotic or finite-sample performance claims. The finite-sample guarantee's reliance on P0∈ΠξD is a stated condition limiting applicability to the DD experiments, not a circularity. Therefore no step satisfies the standard for a circularity finding; score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 9 axioms · 0 invented entities

The central result rests on standard stochastic programming assumptions, on the structural condition P0 in Pi_xi D which is not checked in experiments, and on ad hoc choices for lambda. No new physical or mathematical entity is introduced; DH(lambda) is a convex combination of existing distributions and ambiguity sets.

free parameters (3)
  • weight of information lambda = C/sqrt(N) = C estimated via K-fold cross-validation, confidence-interval tightening, or sqrt(M0)
    Section 3.4; the constant C is chosen from data and reused across N. The theoretical guarantees hold for any fixed C, but the choice affects practical performance.
  • scenario reduction weight lambda = 1 - sqrt(M)/sqrt(N) = Functions of M and N
    Section 4; introduced without derivation. It recovers SAA at M = N and grows as M shrinks, but no error bound is proved for this schedule.
  • epsilon in lambda* definition = Eq. (6), with unspecified constants c1 and c2
    Section 3.2; needed to make the finite-sample guarantee operational, but the constants are not provided, so lambda* cannot be computed from the paper.
axioms (9)
  • standard math Central limit theorem for the SAA estimator
    Used in Proposition 2 and Appendix B.1 to construct the confidence interval for F(x).
  • standard math Fournier-Guillin concentration bound and the inequality G <= W2 between Gelbrich and Wasserstein distances
    Used in the proof of Proposition 6 to show the true distribution is in DH(lambda*) with high confidence.
  • standard math Convexity of the Gelbrich sublevel set BG_epsilon, taken from Nguyen et al. (2021)
    Used inside the proof of Proposition 6 to force a contradiction for distributions outside DH(lambda*).
  • domain assumption f(x, .) is convex and Lipschitz, E|f(x, xi)| < infinity, and X is nonempty, convex, and compact
    Stated in Section 2 as standing assumptions for the stochastic program and the HO derivations.
  • domain assumption P0 in Pi_xi D, the empirical distribution lies in the projected ambiguity set
    Required by Propositions 4 and 6 for the nesting and the finite-sample guarantee. Not satisfied by the equality-moment set DD used in the experiments.
  • domain assumption Light-tailed true distribution P (Assumption 1)
    Assumed in Proposition 6 to apply the Fournier-Guillin concentration result.
  • domain assumption Conditions (i)-(iv) in Section 3.5: bounded top confidence set, Slater-type condition, piecewise linear convex f, and nested conic confidence sets
    Assumed in Theorem 1 to obtain the tractable reformulation (H1).
  • ad hoc to paper lambda = C/sqrt(N) scaling
    Chosen so that the DRO term becomes an O(1/sqrt(N)) perturbation and the asymptotic optimality rate matches SAA. No external principle forces this rate.
  • ad hoc to paper Scenario reduction weight lambda = 1 - sqrt(M)/sqrt(N)
    Introduced in Section 4 without derivation or theoretical error guarantee.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Harmonizing SAA and DRO." pith.science (2026). https://pith.science/paper/GG55FIWX

@misc{pith2026250818767,
  author       = {Pith},
  title        = {Pith review of: Harmonizing SAA and DRO},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GG55FIWX}},
  note         = {Machine review of arXiv:2508.18767}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Decision-makers often encounter uncertainty, and the distribution of uncertain parameters plays a crucial role in making reliable decisions. However, complete information is rarely available. The sample average approximation (SAA) approach utilizes historical data to address this, but struggles with insufficient data. Conversely, moment-based distributionally robust optimization (DRO) effectively employs partial distributional information but can yield conservative solutions even with ample data. To bridge these approaches, we propose a novel method called harmonizing optimization (HO), which integrates SAA and DRO by adaptively adjusting the weights of data and information based on sample size N. This allows HO to amplify data effects in large samples while emphasizing information in smaller ones. More importantly, HO performs well across varying data sizes without needing to classify them as large or small. We provide practical methods for determining these weights and demonstrate that HO offers finite-sample performance guarantees, proving asymptotic optimality when the weight of information follows a 1/\sqrt{N}-rate. In addition, HO can be applied to enhance scenario reduction, improving approximation quality and reducing completion time by retaining critical information from reduced scenarios. Numerical results show significant advantages of HO in solution quality compared to Wasserstein-based DRO, and highlight its effectiveness in scenario reduction.

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

49 extracted references · 46 canonical work pages · 1 internal anchor

  1. [1]

    F., and Li, Y

    Alexander, S., Coleman, T. F., and Li, Y. (2006). Minimizing C V a R and V a R for a portfolio of derivatives. Journal of Banking & Finance , 30(2):583--605

    work page 2006

  2. [2]

    Arya, V., Garg, N., Khandekar, R., Meyerson, A., Munagala, K., and Pandit, V. (2004). Local search heuristics for k -median and facility location problems. SIAM Journal on Computing , 33(3):544--562

    work page 2004

  3. [3]

    Bahari, M., Nejjar, I., and Alahi, A. (2021). Injecting knowledge in data-driven vehicle trajectory predictors. Transportation Research Part C: Emerging Technologies , 128:103010

    work page 2021

  4. [4]

    Basciftci, B., Ahmed, S., and Shen, S. (2021). Distributionally robust facility location problem under decision-dependent stochastic demand. European Journal of Operational Research , 292(2):548--561

    work page 2021

  5. [5]

    and de Ruiter, F

    Bertsimas, D. and de Ruiter, F. J. (2016). Duality in two-stage adaptive linear optimization: Faster computation and stronger bounds. INFORMS Journal on Computing , 28(3):500--511

    work page 2016

  6. [6]

    Birge, J. R. and Louveaux, F. (2011). Introduction to Stochastic Programming . Springer Science & Business Media

    2011

  7. [7]

    Cheramin, M., Cheng, J., Jiang, R., and Pan, K. (2022). Computationally efficient approximations for distributionally robust optimization under moment and W asserstein ambiguity. INFORMS Journal on Computing , 34(3):1768--1794

    work page 2022

  8. [8]

    Cheung, W. C. and Simchi-Levi, D. (2019). Sampling-based approximation schemes for capacitated stochastic inventory control models. Mathematics of Operations Research , 44(2):668--692

    work page 2019

  9. [9]

    and Ye, Y

    Delage, E. and Ye, Y. (2010). Distributionally robust optimization under moment uncertainty with application to data-driven problems. Operations Research , 58(3):595--612

    work page 2010

  10. [10]

    o we-Kuska, N., and R \

    Dupa c ov \'a , J., Gr \"o we-Kuska, N., and R \"o misch, W. (2003). Scenario reduction in stochastic programming. Mathematical Programming , 95:493--511

    work page 2003

  11. [11]

    Esfahani, P. M. and Kuhn, D. (2018). Data-driven distributionally robust optimization using the W asserstein metric: Performance guarantees and tractable reformulations. Mathematical Programming , 171(1-2):115--166

    work page 2018

  12. [12]

    and Guillin, A

    Fournier, N. and Guillin, A. (2015). On the rate of convergence in W asserstein distance of the empirical measure. Probability Theory and Related Fields , 162(3):707--738

    work page 2015

  13. [13]

    Gao, R. (2023). Finite-sample guarantees for W asserstein distributionally robust optimization: Breaking the curse of dimensionality. Operations Research , 71(6):2291--2306

    work page 2023

  14. [14]

    Distributionally Robust Stochastic Optimization with Dependence Structure

    Gao, R. and Kleywegt, A. J. (2017). Distributionally robust stochastic optimization with dependence structure. arXiv preprint arXiv:1701.04200

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  15. [15]

    E., Oks, M., and Oustry, F

    Ghaoui, L. E., Oks, M., and Oustry, F. (2003). Worst-case value-at-risk and robust portfolio optimization: A conic programming approach. Operations Research , 51(4):543--556

    work page 2003

  16. [16]

    and Wiesemann, W

    Ghosal, S. and Wiesemann, W. (2020). The distributionally robust chance-constrained vehicle routing problem. Operations Research , 68(3):716--732

    work page 2020

  17. [17]

    Givens, C. R. and Shortt, R. M. (1984). A class of W asserstein metrics for probability distributions. Michigan Mathematical Journal , 31(2):231--240

    work page 1984

  18. [18]

    J., and Lim, A

    Gotoh, J.-y., Kim, M. J., and Lim, A. E. (2025). A data-driven approach to beating SAA out of sample. Operations Research , 73(2):829--841

    work page 2025

  19. [19]

    Hu, J., Homem-de Mello, T., and Mehrotra, S. (2012). Sample average approximation of stochastic dominance constrained programs. Mathematical Programming , 133(1):171--201

    work page 2012

  20. [20]

    J., Shapiro, A., and Homem-de Mello, T

    Kleywegt, A. J., Shapiro, A., and Homem-de Mello, T. (2002). The sample average approximation method for stochastic discrete optimization. SIAM Journal on Optimization , 12(2):479--502

    work page 2002

  21. [21]

    Kuhn, D., Shafiee, S., and Wiesemann, W. (2025). Distributionally robust optimization

    work page 2025

  22. [22]

    R., Ghaoui, L

    Lanckriet, G. R., Ghaoui, L. E., Bhattacharyya, C., and Jordan, M. I. (2002). A robust minimax approach to classification. Journal of Machine Learning Research , 3:555--582

    work page 2002

  23. [23]

    Li, Y., Saeed, D., Zhang, X., Ziebart, B., and Gimpel, K. (2022). Moment distributionally robust tree structured prediction. Advances in Neural Information Processing Systems , 35:12237--12252

    work page 2022

  24. [24]

    T., Krishnan, H., and Uichanco, J

    Lin, M., Huh, W. T., Krishnan, H., and Uichanco, J. (2022). Data-driven newsvendor problem: Performance of the sample average approximation. Operations Research , 70(4):1996--2012

    work page 2022

  25. [25]

    Liu, Y., Meskarian, R., and Xu, H. (2017). Distributionally robust reward-risk ratio optimization with moment constraints. SIAM Journal on Optimization , 27(2):957--985

    work page 2017

  26. [26]

    Z., Qi, J., and Zhang, A

    Long, D. Z., Qi, J., and Zhang, A. (2024). Supermodularity in two-stage distributionally robust optimization. Management Science , 70(3):1394--1409

    work page 2024

  27. [27]

    Luo, X., Zhang, D., and Zhu, X. (2021). Deep learning based forecasting of photovoltaic power generation by incorporating domain knowledge. Energy , 225:120240

    work page 2021

  28. [28]

    A., Shafiee, S., Filipovi \'c , D., and Kuhn, D

    Nguyen, V. A., Shafiee, S., Filipovi \'c , D., and Kuhn, D. (2021). Mean-covariance robust risk measurement. arXiv preprint arXiv:2112.09959

    arXiv 2021

  29. [29]

    A., Si, N., and Blanchet, J

    Nguyen, V. A., Si, N., and Blanchet, J. (2020). Robust B ayesian classification using an optimistic score ratio. International Conference on Machine Learning , 119:7327--7337

    work page 2020

  30. [30]

    Popescu, I. (2007). Robust mean-covariance solutions for stochastic optimization. Operations Research , 55(1):98--112

    work page 2007

  31. [31]

    M., and Pineda, S

    Porras, \'A ., Dom \' nguez, C., Morales, J. M., and Pineda, S. (2023). Tight and compact sample average approximation for joint chance-constrained problems with applications to optimal power flow. INFORMS Journal on Computing , 35(6):1454--1469

    work page 2023

  32. [32]

    and Mehrotra, S

    Rahimian, H. and Mehrotra, S. (2019). Distributionally robust optimization: A review. arXiv preprint arXiv:1908.05659

    Pith/arXiv arXiv 2019

  33. [33]

    T., Uryasev, S., et al

    Rockafellar, R. T., Uryasev, S., et al. (2000). Optimization of conditional value-at-risk. Journal of Risk , 2:21--42

    work page 2000

  34. [34]

    Rujeerapaiboon, N., Kuhn, D., and Wiesemann, W. (2016). Robust growth-optimal portfolios. Management Science , 62(7):2090--2109

    work page 2016

  35. [35]

    Rujeerapaiboon, N., Schindler, K., Kuhn, D., and Wiesemann, W. (2022). Scenario reduction revisited: Fundamental limits and guarantees. Mathematical Programming , 191(1):207--242

    work page 2022

  36. [36]

    Scarf, H. (1958). A min-max solution of an inventory problem. Studies in the Mathematical Theory of Inventory and Production

    work page 1958

  37. [37]

    Schindler, K., Rujeerapaiboon, N., Kuhn, D., and Wiesemann, W. (2024). A planner-trader decomposition for multimarket hydro scheduling. Operations Research , 72(1):185--202

    work page 2024

  38. [38]

    Sch \"u tz, P., Tomasgard, A., and Ahmed, S. (2009). Supply chain design under uncertainty using sample average approximation and dual decomposition. European Journal of Operational Research , 199(2):409--419

    work page 2009

  39. [39]

    Shapiro, A. (2003). Monte C arlo sampling methods. Handbooks in Operations Research and Management Science , 10:353--425

    work page 2003

  40. [40]

    Shapiro, A., Dentcheva, D., and Ruszczynski, A. (2021). Lectures on Stochastic Programming: Modeling and Theory . SIAM

    work page 2021

  41. [41]

    Shehadeh, K. S. (2023). Distributionally robust optimization approaches for a stochastic mobile facility fleet sizing, routing, and scheduling problem. Transportation Science , 57(1):197--229

    work page 2023

  42. [42]

    Sion, M. (1958). On general minimax theorems. Pacific Journal of Mathematics , 8(1):171--176

    work page 1958

  43. [43]

    Takriti, S., Krasenbrink, B., and Wu, L. S.-Y. (2000). Incorporating fuel constraints and electricity spot prices into the stochastic unit commitment problem. Operations Research , 48(2):268--280

    work page 2000

  44. [44]

    Tsang, M. Y. and Shehadeh, K. S. (2025a). Algorithmic approaches for identifying the trade-off between pessimism and optimism in a stochastic fixed charge facility location problem. Optimization Online

    work page

  45. [45]

    Tsang, M. Y. and Shehadeh, K. S. (2025b). On the tradeoff between distributional belief and ambiguity: Conservatism, finite-sample guarantees, and asymptotic properties. INFORMS Journal on Optimization

    work page

  46. [46]

    Wang, S., Wang, H., Li, X., and Honorio, J. (2025). Learning against distributional uncertainty: On the trade-off between robustness and specificity. IEEE Journal of Selected Topics in Signal Processing

    work page 2025

  47. [47]

    Wiesemann, W., Kuhn, D., and Sim, M. (2014). Distributionally robust convex optimization. Operations Research , 62(6):1358--1376

    work page 2014

  48. [48]

    and Zhang, D

    Xu, H. and Zhang, D. (2009). Smooth sample average approximation of stationary points in nonsmooth stochastic optimization and applications. Mathematical Programming , 119:371--401

    work page 2009

  49. [49]

    Zymler, S., Kuhn, D., and Rustem, B. (2013). Distributionally robust joint chance constraints with second-order moment information. Mathematical Programming , 137:167--198

    work page 2013

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.