Pith. sign in

REVIEW 4 major objections 4 minor 37 references

Fine-Tuning Generative Models for Extreme Events via CVaR-Penalized Wasserstein Gradient Flows

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read CVaR penalty restores the vanishing tail velocity in generative particle flows.

desk verdict A genuinely useful new subgradient calculus for CVaR on empirical measures, packaged as a fine-tuning algorithm that works well on heavy tails but overstates its tail-agnostic robustness: the global radial penalty fails on the nearly Gaussian marginal of their own anisotropic test. read the letter →

arxiv 2608.11544 v1 pith:AFURIH3B submitted 2026-08-12 stat.ML cs.LG

classification stat.MLcs.LG
keywords heavy-taileddistributionsextremeeventsConditionalValue-at-RiskWassersteingradientflowsLipschitz-regularizedKLdivergenceparticlealgorithmsgenerativemodelfine-tuningtail-agnosticlearning
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

The paper proposes CVaR-GPA, a particle fine-tuning algorithm that improves a pre-trained generative model's accuracy on heavy-tailed targets without being told the target's tail index. Its loss is the Lipschitz-regularized KL divergence plus a squared Conditional Value-at-Risk (CVaR) discrepancy, and the Wasserstein gradient flow of this loss has a bounded but non-Lipschitz velocity field that keeps pushing particles outward in the tail after the KL-only velocity has faded. The authors derive the variational subgradients of CVaR from its infimum representation, a step needed because the classical density-based first variation does not exist for empirical measures. On isotropic and anisotropic Student-$t$ targets, a funnel benchmark, and a real 25-dimensional portfolio-return dataset, the fine-tuned particles reduce global $L^1$ and tail log-CCDF errors relative to the Lipschitz-regularized baseline. The paper also reports the method's limitation: on a 5-dimensional Student-$t$ with one nearly Gaussian marginal, that marginal's errors increase.

What carries the argument

The machinery is the CVaR-penalized potential $\Phi^y_Q = \phi^* - \frac{2\lambda}{1-\alpha}\,\Delta C(Q;P_{\mathrm{tar}})\,(g-y)_+$, where $\phi^*$ is the optimal $L$-Lipschitz potential defining the Lipschitz-regularized KL divergence, $g(x)=\|x\|$ is the radial risk function, and $y$ ranges over the quantile interval $T(Q)$. Choosing $-\nabla \Phi^y_Q$ as the velocity gives a bounded, non-Lipschitz field: the KL part contributes $-\nabla\phi^*$ everywhere, and the CVaR part adds an outward radial push of size $\frac{2\lambda}{1-\alpha}|\Delta C|$ exactly where $\|x\|>y$. The subgradient theorem provided in the paper is the essential step: for empirical measures, where the CDF of $g$ is a step function, the classical formula $(g-\mathrm{VaR}_\alpha)_+/(1-\alpha)$ fails because the quantile interval is non-singleton, and every $y\in T(Q)$ yields a valid variational subgradient.

What would settle it

The paper's own 5-dimensional Student-t experiment is the concrete check: with tail indices $\nu=(1,1.5,3,10,30)$, the nearly Gaussian $\nu=30$ marginal's global and tail errors increase after fine-tuning. Repeating that experiment at larger sample sizes and with several seeds, and testing whether a per-coordinate risk function removes the degradation, would settle whether the single-global-CVaR mechanism is the limiting component of the method.

Watch

Extended reading notes

Core claim

The central discovery is that a squared CVaR discrepancy term appended to the Lipschitz-regularized KL divergence repairs the premature vanishing velocity of Lipschitz-regularized Wasserstein gradient flows. The CVaR term contributes a radial velocity component, active only beyond a threshold radius, whose magnitude is proportional to the gap between the target's CVaR and the generated distribution's CVaR; this component is nonzero in the tail even when sample scarcity has made the KL-derived velocity vanish, and it tapers off as the distribution approaches the target. The paper proves that the resulting velocity field is bounded yet not Lipschitz, which is the kind of transport map needed to move a light-tailed pre-trained distribution toward a heavier-tailed target. It also proves that the penalized loss is a divergence whose unique global minimizer is the target distribution, and that the subgradient family indexed by thresholds in the quantile interval $T(Q)$ supplies the correct first-variation objects for empirical measures.

Load-bearing premise

The load-bearing premise is that one global CVaR discrepancy computed from radial distances, with a single activation threshold, can steer all coordinates of an anisotropic target at once; the paper's own $\nu=30$ experiment shows this premise can fail when one marginal is nearly Gaussian while others are heavily tailed.

Editorial extensions

If this is right

  • One fine-tuning run with the same hyperparameters improves both global $L^1$ error and tail log-CCDF error across targets with different tail indices, without estimating the tail decay rate.
  • The kinetic-energy stopping rule makes the effective flow depth depend on the target, not on a preset number of layers or iterations.
  • The algorithm needs only samples from the pre-trained model, so it can fine-tune any black-box generator without architectural access.
  • The bounded velocity keeps particle updates stable even when the target's population distribution has no finite moments, because the empirical target used in the algorithm has finite support.
  • The non-Lipschitz transport map is what allows a light-tailed pre-trained distribution to move toward a heavier-tailed target; Lipschitz maps would preserve the source's light tail.

Reading between the lines

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

  • If the limiting mechanism is the single radial CVaR statistic, replacing $g(x)=\|x\|$ with a per-coordinate risk function should remove the $\nu=30$ degradation while keeping the heavy-marginal gains; this is the paper's own stated future direction.
  • The tail-agnostic claim should be read as holding for the algorithm's empirical objective: the population-level theory needs a finite first moment, but the practical algorithm inherits finite support from its samples, which is why the Cauchy target can be handled.
  • Monitoring the jump in the velocity field across the threshold shell could serve as a live diagnostic of unresolved tail discrepancy, since the jump magnitude is proportional to the current CVaR gap.
  • Comparing inner- versus outer-endpoint choices for the activation threshold $y\in T(Q)$ would probe whether the activation region's size trades global accuracy against tail accuracy; the paper leaves this as a design choice.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a CVaR-penalized loss functional F^CVaR(Q;P_tar) = D^L_KL(Q||P_tar) + λ(ΔC)^2, where ΔC is the difference of CVaR values of a risk function g under Q and the target, and derives the corresponding Wasserstein gradient flow. The main theoretical results are a subgradient calculus for CVaR based on the Rockafellar–Uryasev representation, valid on empirical measures where the classical density-based derivative fails, and an explicit bounded but non-Lipschitz velocity field (Theorem 3.8, Corollary 3.11). The paper then presents the particle algorithm CVaR-GPA, which fine-tunes samples from a pre-trained model, and reports experiments on isotropic and anisotropic Student-t targets, Neal's funnel, and Fama-French 25 portfolio data, comparing global L1 and tail log-CCDF errors against a Lip-KL-GPA baseline.

Significance. If the claims hold, the paper makes a solid theoretical contribution: the first-variation subgradients of CVaR are derived rigorously for empirical measures, the velocity field is explicit and provably bounded, and the proposed algorithm is architecture-agnostic and tail-agnostic in the sense of not requiring an estimated tail index. The authors are also transparent about a known failure mode for a nearly Gaussian marginal in an anisotropic Student-t example. However, the experimental evidence is weakened by the absence of error bars or repeated trials, and the abstract's unqualified 'dramatically improves' and 'tail-agnostic/robust' statements are contradicted by the paper's own ν=30 result in Section 5.2. The central theoretical derivation appears sound, but the empirical scope and the strength of the claims need to be revised.

major comments (4)
  1. [§5.2, Figure 5] The reported results for the 5-dimensional anisotropic Student-t target contradict the abstract's unqualified claim that CVaR-GPA 'dramatically improves global and tail accuracy' on anisotropic Student-t targets. For the ν=30 marginal, E_L1 increases from 0.014 to 0.021 and E_tail increases from 0.078 to 0.14. The text acknowledges this and attributes it to the global scalar CVaR correction being dominated by heavier-tailed directions, but the abstract and introduction are not qualified accordingly. Please revise the central claims to state explicitly that the method is effective for moderate tail heterogeneity and may degrade for near-Gaussian marginals in highly anisotropic targets, or modify the algorithm/risk function so that this failure no longer occurs.
  2. [§5, Figures 3–5] All numerical results are presented as single values with no error bars, no number of random seeds, and no measure of variability. Given that the comparisons of E_L1 and E_tail are the main empirical evidence, please report means and standard deviations over multiple runs, or at least provide box plots or confidence intervals. Without this, the improvements in Figures 3 and 4, and the deterioration in Figure 5, cannot be assessed for statistical significance.
  3. [Assumption 2.1, §5.1] The population-level theory requires E_{P_tar}[g] < ∞, which for the chosen radial risk function g(x)=||x|| means the target must have a finite first moment. This excludes the motivating Cauchy (ν=1) target at the population level; the authors note that the empirical target measure automatically satisfies the condition, but the abstract's 'minimal assumptions on the target distribution' and 'tail-agnostic' statements are therefore too strong. Please state explicitly that the population-level well-posedness of F^CVaR requires finite E_{P_tar}[g], and that the algorithm's use on infinite-mean targets rests on the finite-sample formulation only.
  4. [§5, Eqs. (5.45)–(5.46)] Both error metrics E_L1 and E_tail are defined through the radial CCDF of ||x||, which is the same risk function g used in the CVaR penalty (Eq. (1.3) and Corollary 3.11). Consequently, part of the reported tail improvement is aligned with the penalized quantity by construction. The per-marginal evaluations in Figures 4 and 5 provide some independent evidence, but the paper should either report a metric based on a different risk function or explicitly acknowledge that the tail-fidelity claim is demonstrated for the radial statistic used in the penalty.
minor comments (4)
  1. [Eq. (2.9)] The notation in Eq. (2.9), CVaR^{Q,g}_α := β VaR^{Q,g}_α + (1−β) CVaR^{Q,g}_α, uses the same symbol CVaR^{Q,g}_α on both sides, making the definition self-referential. Please denote the upper conditional expectation, E_Q[g(X)|g(X)>VaR^{Q,g}_α], by a different symbol such as CVaR^+ or CVaR^{Q,g,+}_α.
  2. [Algorithm 2, line 8] The text in Section 5 states that the outer VaR endpoint is used, but Algorithm 2 line 8 allows either 'VaR or VaR'. Please make the choice explicit in the algorithm or its caption so the experiments are exactly reproducible.
  3. [Figure 2] The numbers above the bars in Figure 2 are dense and hard to read at log scale. Consider moving the exact values to a table or using larger fonts and fewer decimal places.
  4. [Section 4] The paper does not state the computational overhead of CVaR-GPA relative to Lip-KL-GPA. Since the added cost is mainly sorting and scalar tail statistics, a short remark on runtime or complexity would help practitioners.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CVaR-penalized flow is derived from an explicit loss, and the radial-tail metric overlap is alignment, not construction.

full rationale

The derivation chain is self-contained. The loss (2.4) is an explicit objective, the subgradient potential (3.31) is constructed from the Rockafellar-Uryasev representation, and the velocity field (3.36)/(3.39) follows by differentiating that potential; nothing in Theorem 3.5, Theorem 3.7, Theorem 3.8, or Corollary 3.11 assumes the conclusion that CVaR-GPA improves tail accuracy. The KL subgradient is imported from the published Lip-KL-GPA theory [9,18]; this is disclosed, parameter-free, and does not depend on the CVaR construction, so it is independent support rather than a load-bearing self-citation. The empirical claims are benchmarked against Lip-KL-GPA with fixed hyperparameters (λ=0.02, α=0.999), not fitted per benchmark, and the results are not forced: Section 5.2 reports that on the nearly Gaussian ν=30 marginal both errors increase, showing the outcome is genuinely empirical. The evaluation metrics (5.45)-(5.46) use the same radial norm g(x)=||x|| as the CVaR penalty, which creates alignment between the objective and the reported tail metric; however, E_tail is a log-CCDF discrepancy averaged over quantiles 0.95-0.999 while the penalty is a squared CVaR discrepancy at α=0.999, so reducing one does not by construction reduce the other. The paper also candidly states it does not claim a rigorous quantitative tail-behavior change (Remark 3.10). Overall, no step in the derivation reduces to its inputs by definition or by fitted-parameter renaming.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard functional-analytic tools, a domain assumption about finite first moments, and a design choice for selecting a subgradient. No new entities are postulated. The main caveat is that the population-level theory excludes Cauchy-like targets without finite first moment, despite the experiments covering them via empirical measures.

free parameters (4)
  • λ = 0.02
    CVaR penalty weight fixed across all experiments in Section 5; chosen by hand, controls the strength of tail velocity.
  • α = 0.999
    Quantile level for CVaR, fixed in Section 5; defines the extreme tail region and activation threshold.
  • L = not reported for experiments; illustrative L=1.0 and L=0.125 in Figure 1
    Lipschitz constant of the KL regularizer; determines the bound on the baseline velocity and is part of the algorithm input.
  • network and optimizer hyperparameters (depth D, width, N_φ, Δt, ε, M, N) = not reported
    Algorithm settings required to reproduce results; only partially described, so the empirical numbers are under-determined.
assumptions (5)
  • standard math Kantorovich-Rubinstein duality characterizes W1 as a supremum over 1-Lipschitz functions
    Used in Lemma 3.1 and Proposition B.1 to prove Lipschitz continuity of CVaR and the KL term.
  • standard math Rademacher's theorem: L-Lipschitz functions are differentiable almost everywhere
    Used in Corollary 3.9 to bound ∇φ* by L.
  • domain assumption Assumption 2.1: Q has finite first moment and E^Q[g] and E^{P_tar}[g] are finite
    Ensures D^L_KL and CVaR are well-defined; paper claims empirical measures satisfy it, but the population-level Cauchy target does not.
  • domain assumption The risk function g is Lipschitz and C^1 off the zero set
    Needed for Theorem 3.8 and Corollary 3.9; satisfied by g(x)=||x||.
  • ad hoc to paper A specific subgradient y in T(Q) (the outer VaR endpoint) is chosen to define the velocity field
    The theory holds for any y in T(Q), but the algorithm's activation region and behavior depend on this design choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Tuning Generative Models for Extreme Events via CVaR-Penalized Wasserstein Gradient Flows." pith.science (2026). https://pith.science/paper/AFURIH3B

@misc{pith2026260811544,
  author       = {Pith},
  title        = {Pith review of: Fine-Tuning Generative Models for Extreme Events via CVaR-Penalized Wasserstein Gradient Flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFURIH3B}},
  note         = {Machine review of arXiv:2608.11544}
}
abstract

We propose CVaR-penalized Generative Particle Algorithm (CVaR-GPA), a robust, tail-agnostic algorithm for fine-tuning generative models to learn heavy-tailed distributions and capture extreme events, requiring no prior knowledge or estimation of the target's tail characteristics. The method is the Wasserstein gradient flow of the Lipschitz-regularized Kullback-Leibler (KL) divergence penalized by a Conditional Value-at-Risk (CVaR) discrepancy term: the Lipschitz-regularized KL divergence enables robust learning under minimal assumptions on the target distribution, while the CVaR penalty restores the velocity that otherwise vanishes prematurely in the under-sampled tails. The penalized flow admits a bounded but non-Lipschitz velocity field. This departs from the Lipschitz transport maps of standard generators, which preserve the tail behavior of a light-tailed source, and enables transport toward heavier-tailed targets. To define this flow on empirical measures, we derive the first-variation subgradients of CVaR from its Rockafellar-Uryasev representation, valid precisely where the classical density-based formula fails. The particle algorithm CVaR-GPA fine-tunes the output samples of any pre-trained model, without access to its architecture, and runs on an adaptive time horizon set by a kinetic-energy stopping criterion rather than a preset depth. On synthetic isotropic and anisotropic Student-$t$ target distributions, Neal's funnel distribution, and the real-world high-dimensional Fama-French 25 portfolio dataset, CVaR-GPA dramatically improves global and tail accuracy on heavy-tailed targets over the pre-trained baseline.

Figures

Figures reproduced from arXiv: 2608.11544 by the authors.

Figure 1
Figure 1. CVaR-penalization restores Lip-KL-GPA’s premature vanishing velocity. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Comparison of Lip-KL-GPA vs. CVaR-GPA on the Fama-French 25 monthly [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the pre-trained model (Lip-KL-GPA) vs. the fine-tuned model [PITH_FULL_IMAGE:figures/full_fig_p032_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of the pre-trained model (Lip-KL-GPA) vs. the fine-tuned model [PITH_FULL_IMAGE:figures/full_fig_p034_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the pre-trained model (Lip-KL-GPA) vs. the fine-tuned model [PITH_FULL_IMAGE:figures/full_fig_p035_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

  1. [1]

    Expected Shortfall as a Tool for Financial Risk Management

    C. Acerbi, C. Nordio, and C. Sirtori. Expected shortfall as a tool for financial risk management.arXiv preprint cond-mat/0102304, 2001

  2. [2]

    Acerbi and P

    C. Acerbi and P. Simonetti. Portfolio optimization with spectral measures of risk. arXiv preprint cond-mat/0203607, 2002

  3. [3]

    Albrecher and S

    H. Albrecher and S. Asmussen. Ruin probabilities and aggregrate claims distributions for shot noise cox processes.Scandinavian Actuarial Journal, 2006(2):86–110, 2006

  4. [4]

    Allouche, S

    M. Allouche, S. Girard, and E. Gobet. EV-GAN: Simulation of extreme events with ReLU neural networks.Journal of Machine Learning Research, 23(150):1–39, 2022

  5. [5]

    Allouche, S

    M. Allouche, S. Girard, and E. Gobet. Exceedgan: simulation above extreme thresh- olds using generative adversarial networks.Extremes, pages 1–23, 2026. 36

  6. [6]

    Bhatia, A

    S. Bhatia, A. Jain, and B. Hooi. Exgan: Adversarial generation of extreme samples. InProceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 6750–6758, 2021

  7. [7]

    Birrell, P

    J. Birrell, P. Dupuis, M. A. Katsoulakis, Y. Pantazis, and L. Rey-Bellet. (f,gamma)- divergences: Interpolating between f-divergences and integral probability metrics. Journal of Machine Learning Research, 23(39):1–70, 2022

  8. [8]

    Chaudhary, U

    S. Chaudhary, U. Dinesha, D. Kalathil, and S. Shakkottai. Risk-averse fine-tuning of large language models.Advances in Neural Information Processing Systems, 37:107003–107038, 2024

Show all 37 references
  1. [9]

    Z. Chen, H. Gu, M. A. Katsoulakis, L. Rey-Bellet, and W. Zhu. Robust generative learning with lipschitz-regularizedα-divergences allows minimal assumptions on target distributions.Information and Inference: A Journal of the IMA, 14(4):iaaf028, 2025

  2. [10]

    Cirillo and N

    P. Cirillo and N. N. Taleb. Tail risk of contagious diseases.Nature Physics, 16(6):606– 613, 2020

  3. [11]

    F. H. Clarke.Optimization and nonsmooth analysis. SIAM, 1990

  4. [12]

    R. Cont, M. Cucuringu, R. Xu, and C. Zhang. Tail-gan: Learning to simulate tail risk scenarios.Management Science, 72(4):2917–2936, 2026

  5. [13]

    De Santi, M

    R. De Santi, M. Vlastelica, Y.-P. Hsieh, Z. Shen, N. He, and A. Krause. Flow density control: Generative optimization beyond entropy-regularized fine-tuning.Advances in neural information processing systems, 38:11056–11088, 2026

  6. [14]

    Dupuis and Y

    P. Dupuis and Y. Mao. Formulation and properties of a divergence used to compare probability measures without absolute continuity.ESAIM: Control, Optimisation and Calculus of Variations, 28:10, 2022

  7. [15]

    Embrechts and N

    P. Embrechts and N. Veraverbeke. Estimates for the probability of ruin with special emphasis on the possibility of large claims.Insurance: Mathematics and Economics, 1(1):55–72, 1982

  8. [16]

    E. F. Fama and K. R. French. Common risk factors in the returns on stocks and bonds.Journal of Financial Economics, 33(1):3–56, 1993. 37

  9. [17]

    Grossi, H

    P. Grossi, H. Kunreuther, and C. C. Patel.Catastrophe modeling: a new approach to managing risk, volume 25. Springer Science & Business Media, 2005

  10. [18]

    H. Gu, P. Birmpa, Y. Pantazis, L. Rey-Bellet, and M. A. Katsoulakis. Lipschitz- regularized gradient flows and generative particle algorithms for high-dimensional scarce data.SIAM Journal on Mathematics of Data Science, 6(4):1205–1235, 2024

  11. [19]

    Y. Guan, K. Balasubramanian, and S. Ma. Mirror flow matching with heavy-tailed priors for generative modeling on convex domains. InInternational Conference on Learning Representations, volume 2026, pages 130098–130124, 2026

  12. [20]

    Hickling and D

    T. Hickling and D. Prangle. Flexible tails for normalizing flows.arXiv preprint arXiv:2406.16971, 2024

  13. [21]

    B. M. Hill. A Simple General Approach to Inference About the Tail of a Distribution. The Annals of Statistics, 3(5):1163 – 1174, 1975

  14. [22]

    Huster, J

    T. Huster, J. Cohen, Z. Lin, K. Chan, C. Kamhoua, N. O. Leslie, C.-Y. J. Chiang, and V. Sekar. Pareto gan: Extending the representational power of gans to heavy-tailed distributions. InInternational Conference on Machine Learning, pages 4523–4532. PMLR, 2021

  15. [23]

    Jaini, I

    P. Jaini, I. Kobyzev, Y. Yu, and M. Brubaker. Tails of lipschitz triangular flows. In International Conference on Machine Learning, pages 4673–4681. PMLR, 2020

  16. [24]

    Jordan, D

    R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker– Planck equation.SIAM Journal on Mathematical Analysis, 29(1):1–17, 1998

  17. [25]

    L. V. Kantorovich and S. Rubinshtein. On a space of totally additive functions.Vestnik of the St. Petersburg University: Mathematics, 13(7):52–59, 1958

  18. [26]

    M. Kishida. Risk-aware stability, ultimate boundedness, and positive invariance.IEEE Transactions on Automatic Control, 69(1):681–688, 2023

  19. [27]

    H. Liu, T. Zhu, N. Jia, J. He, and Z. Zheng. Learning to simulate from heavy-tailed distribution via diffusion model.Available at SSRN 4975931, 2024. 38

  20. [28]

    Miyato, T

    T. Miyato, T. Kataoka, M. Koyama, and Y. Yoshida. Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018

  21. [29]

    R. M. Neal. Slice sampling.The Annals of Statistics, 31(3):705–767, 2003

  22. [30]

    F. Otto. The geometry of dissipative evolution equations: The porous medium equa- tion.Communications in Partial Differential Equations, 26(1-2):101–174, 2001

  23. [31]

    Pandey, J

    K. Pandey, J. Pathak, Y. Xu, S. Mandt, M. Pritchard, A. Vahdat, and M. Mardani. Heavy-tailed diffusion models.arXiv preprint arXiv:2410.14171, 2024

  24. [32]

    R. T. Rockafellar and S. Uryasev. Conditional value-at-risk for general loss distribu- tions.Journal of banking & finance, 26(7):1443–1471, 2002

  25. [33]

    Z. Wang, R. De Santi, X. Mo, M. M. Zavlanos, A. Krause, and K. H. Johansson. Efficient tail-aware generative optimization via flow model fine-tuning.arXiv preprint arXiv:2602.16796, 2026. Appendix A: Clarke’s generalized subgradients LetQ7→J(Q) be a locally Lipschitz continuou...

  26. [34]

    CVaRQ1,g α −CVaRQ2,g α ≤ Lip(g) 1−α W1(Q1,Q 2),(B6) for allQ 1,Q 2∈Q g

    The CVaRQ7→CVaR Q,g α , is Lipschitz continuous with respect to the Wasserstein-1 metric, i.e. CVaRQ1,g α −CVaRQ2,g α ≤ Lip(g) 1−α W1(Q1,Q 2),(B6) for allQ 1,Q 2∈Q g

  27. [35]

    Proof.(1) Fixy∈R

    The functionalQ7→F CVaR(Q;P tar)is locally Lipschitz continuous onQ g with respect toW 1. Proof.(1) Fixy∈R. Sincet7→t + is 1-Lipschitz andgis Lip(g)-Lipschitz,x7→(g(x)−y) + is Lip(g)-Lipschitz, uniformly iny. Thus we have FQ1,g α (y)−F Q2,g α (y) = 1 1−α EQ1[(g−y) +]−E Q2[(g−y...

  28. [36]

    The function(ϵ,y)7→F Qϵ,g α (y)is jointly continuous in[0,ϵ 0]×R. 43

  29. [37]

    Proof.(1) Denote byℓ ρ(y) := 1 1−α R (g(x)−y) +dρ(x) for anyy∈Rand∥ρ∥= R d|ρ|

    The setT:= S ϵ∈[0,ϵ0]T(Qϵ)is compact. Proof.(1) Denote byℓ ρ(y) := 1 1−α R (g(x)−y) +dρ(x) for anyy∈Rand∥ρ∥= R d|ρ|. For ally∈R,|ℓ ρ(y)|≤ 1 1−α R gd|ρ|+|y|∥ρ∥ <∞sinceρis right-admissible and R gd|ρ|<∞ by (3.18). Letϵ 0 be a given threshold and let (ϵ,y)∈[0,ϵ 0]×Rbe arbitrary. ...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.