Analyzing Distortion Risk Metrics and Weighted Entropy for Unimodal and Symmetric Distributions under Partial Information Constraints

Basic Information

• Paper ID: 2504.19725
• Title: Analyzing distortion risk metrics and weighted entropy for unimodal and symmetric distributions under partial information constraints
• Authors: Baishuai Zuo (Qufu Normal University & Southern University of Science and Technology), Chuancun Yin (Qufu Normal University)
• Classification: q-fin.RM (Quantitative Finance - Risk Management), stat.AP (Statistics - Applications)
• Publication Date: November 21, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2504.19725v2

Abstract

This paper derives worst-case bounds for distortion risk metrics and weighted entropy under unimodal distributions, symmetric distributions, and symmetric unimodal distributions when only mean and variance information is available. Notably, exact upper bounds are provided for symmetric distributions. These theoretical results are applied to various (weighted) entropies, shortfalls, and other risk measures, including cumulative Tsallis past entropy, α-order cumulative residual Tsallis entropy, extended Gini coefficient, fractional generalized cumulative residual entropy, extended Gini shortfall, and n-order expected shortfall.

Research Background and Motivation

Research Problem

The core problem addressed is: How to determine worst-case bounds for distortion risk metrics and weighted entropy of random variables with specific distributional characteristics (unimodality, symmetry) under partial information constraints (knowing only mean and variance)?

Problem Significance

1. Practical Demand: In reality, many datasets are model-free or subject to distributional uncertainty, yet we typically can obtain partial information (e.g., mean, variance) and structural characteristics (e.g., unimodality, symmetry)
2. Risk Management Applications: Distortion risk metrics have widespread applications in behavioral economics and risk management, particularly in premium principle construction and risk aversion modeling
3. Theoretical Value: Entropy plays an important role in life table research, annuity cost calculation, and premium principle construction

Limitations of Existing Methods

1. General Distribution Research: Zuo and Yin (2025) studied worst-case distortion risk metrics and weighted entropy for general distributions but did not consider special distributional structures
2. Lack of Symmetry Constraints: Psarrakos et al. (2024) and Zuo and Yin (2025) provided worst-case results under partial information but did not consider symmetry constraints
3. Insufficient Unimodal Distribution Research: While Bernard et al. (2020, 2025) studied RVaR bounds for unimodal distributions, research on general distortion risk metrics and weighted entropy for unimodal and symmetric unimodal distributions remains incomplete

Research Motivation

1. Distortion risk metrics and weighted entropy have broad application value
2. The practical need to "quickly estimate risk measures given partial information and distributional structure" is common
3. Filling the gap in worst-case risk measure research for special distributional structures

Core Contributions

1. Bounds for Unimodal Distributions: Derives lower and upper bounds for distortion risk metrics and weighted entropy under unimodal distributions (Theorems 1, Corollary 1, Theorem 4)
2. Exact Bounds for Symmetric Distributions: Provides exact upper bounds and worst-case distributions for symmetric distributions (Theorems 2, Corollary 2, Proposition 1, Theorem 5)
3. Bounds for Symmetric Unimodal Distributions: Establishes lower and upper bounds for symmetric unimodal distributions (Theorems 3, Corollary 3, Theorem 6)
4. Broad Applications: Applies theoretical results to multiple specific entropy measures, shortfalls, and risk measures, including:
   • Entropies: cumulative Tsallis past entropy, cumulative residual Tsallis entropy, Gini mean difference, fractional generalized cumulative residual entropy, cumulative residual entropy, etc.
   • Shortfalls: extended Gini shortfall, Gini shortfall, cumulative residual entropy shortfall, etc.
   • Other risk measures: n-order expected shortfall, dual power principle, proportional hazard principle, etc.
5. Numerical Verification: Validates theoretical results using real stock return data

Methodology Details

Problem Formulation

Problem 1: For distortion risk metrics $\rho_g(X) = \int_0^\infty g(P(X > x))dx + \int_{-\infty}^0 [g(P(X > x)) - g(1)]dx$

Solve the optimization problem: $\sup_{X \in \mathcal{V}(\mu, \sigma)} \rho_g(X)$

where $\mathcal{V}(\mu, \sigma) \in \{\mathcal{V}_U(\mu, \sigma), \mathcal{V}_S(\mu, \sigma), \mathcal{V}_{SU}(\mu, \sigma)\}$ denote the sets of unimodal, symmetric, and symmetric unimodal random variables with mean $\mu$ and variance $\sigma^2$, respectively.

Problem 2: For weighted entropy $\int_{-\infty}^{+\infty} \psi(x)g(\bar{F}_X(x))dx$

Solve similar optimization problems under weighted constraints $E[\Psi(F_X^{-1}(U))] = \mu_\Psi$, $\text{Var}[\Psi(F_X^{-1}(U))] = \sigma_\Psi^2$.

Core Methodological Framework

1. Quantile Representation

Using Lemmas 1 and 2, convert distortion risk metrics and weighted entropy to integral forms of quantile functions: $\rho_g(X) = \int_0^1 F_X^{-1}(u)d\hat{g}(u)$$\int_{-\infty}^{+\infty} \psi(x)g(\bar{F}_X(x))dx = \int_0^1 \Psi(F_X^{-1}(u))d\hat{g}(u)$

where $\hat{g}(u) = g(1) - g(1-u)$.

2. Extreme Distribution Construction

Unimodal Distributions: Define two classes of extreme distribution sets $U_R$ and $U_L$ (or weighted versions $U_{\Psi R}$ and $U_{\Psi L}$): $U_R = \left\{X: F_X^{-1}(u) = \begin{cases} a, & u \in [0, b) \\ c(u-b) + a, & u \in [b, 1] \end{cases}\right\}$

These distributions provide lower bounds under given constraints.

Symmetric Unimodal Distributions: Define the set $\mathcal{S}(\alpha)$: $\mathcal{S}(\alpha) = \left\{X: F_X^{-1}(u) = \begin{cases} a + c(u-1+b), & u \in (0, 1-b) \\ a, & u \in [1-b, b] \\ a + c(u-b), & u \in (b, 1) \end{cases}\right\}$

3. Convex Envelope Technique

For upper bounds, utilize the convex envelope $\hat{g}_*$ (the greatest convex minorant of $\hat{g}$) and Lemma 3 (monotonicity lemma): $\rho_g(X) = \int_0^1 F_X^{-1}(u)d\hat{g}(u) \leq \int_0^1 F_X^{-1}(u)d\hat{g}_*(u)$

4. TVaR Connection

Transform the problem into tail conditional expectation (TVaR) form: $\sup_{X \in \mathcal{V}_U(\mu, \sigma)} \rho_g(X) \leq \mu(\hat{g}_*)^\prime(0) + \int_0^1 (1-u)\sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{TVaR}_u(X) d(\hat{g}_*)^\prime(u)$

Utilize known results from Bernard et al. (2020) on TVaR for unimodal distributions: $\sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{TVaR}_\alpha(X) = \begin{cases} \mu + \frac{\sigma}{3}\sqrt{\frac{\alpha(8-9\alpha)}{1-\alpha}}, & \alpha \in (0, \frac{1}{2}) \\ \mu + \sigma\sqrt{\frac{8}{9(1-\alpha)} - 1}, & \alpha \in [\frac{1}{2}, 1) \end{cases}$

Technical Innovations

1. Unified Framework: Incorporates distortion risk metrics and weighted entropy into a unified mathematical framework, systematically addressing them through quantile representation and convex envelope techniques
2. Exact Symmetric Bounds: Provides exact closed-form solutions for symmetric distributions by leveraging the symmetry constraint $F_X^{-1}(u) + F_X^{-1}(1-u) = 2\mu$
3. Piecewise Linear Approximation: Uses sequences of piecewise linear functions to approximate convex envelopes, establishing upper bounds via the monotone convergence theorem
4. Residual and Past Lifetime Extensions: Extends results to residual lifetime $X_t = [X-t|X>t]$ and past lifetime $X_{(t)} = [X|X \leq t]$ (Corollaries 4-9)

Experimental Setup

Dataset

The authors used daily stock return data from three companies:

1. Coca-Cola Company (KO): $\mu_1 = 0.052128514$, $\sigma_1^2 = 0.416460169$
2. Cincinnati Financial (CINF): $\mu_2 = -0.009317269$, $\sigma_2^2 = 3.604281185$
3. eBay Inc. (EBAY): $\mu_3 = 0.09911245$, $\sigma_3^2 = 0.900906979$

Data Period: Business days from May 20, 2024 to May 18, 2024 (from nasdaq.com)

Evaluation Metrics

Compute exact upper bounds for Gini Shortfall ($\text{GS}_p^\tau$) with parameter settings:

• $\tau = 0.25$ (loading parameter)
• $r = 2$
• $p \in \{0.80, 0.90, 0.95, 0.98, 0.99\}$ (confidence levels)

Comparison Methods

Compare upper bounds under four distributional constraints:

1. General distributions ($\mathcal{V}(\mu, \sigma)$)
2. Unimodal distributions ($\mathcal{V}_U(\mu, \sigma)$)
3. Symmetric distributions ($\mathcal{V}_S(\mu, \sigma)$)
4. Symmetric unimodal distributions ($\mathcal{V}_{SU}(\mu, \sigma)$)

Implementation Details

• Verify unimodality and symmetry through histogram analysis
• Compute theoretical bounds using numerical integration and optimization methods
• Use standard mathematical function libraries for cases involving incomplete Gamma functions

Experimental Results

Main Results

Tables 1-3: Gini Shortfall Upper Bounds for Three Stocks

Taking KO stock as an example (Table 1):

Key Findings

1. Monotonicity: All upper bounds increase monotonically with confidence level $p$, consistent with the intuition of risk measures
2. Ordering of Constraint Effects: For the same $p$ value, $\sup_{\mathcal{V}} \text{GS}_p^\tau > \sup_{\mathcal{V}_S} \text{GS}_p^\tau > \sup_{\mathcal{V}_U} \text{GS}_p^\tau > \sup_{\mathcal{V}_{SU}} \text{GS}_p^\tau$
   This indicates that unimodality constraints have a larger impact on upper bounds than symmetry constraints
3. Variance Effects: CINF ($\sigma_2^2 = 3.604$) has the largest upper bounds, while KO ($\sigma_1^2 = 0.416$) has the smallest, reflecting the direct impact of variance on risk measures
4. Distributional Shape Invariance: The relative effects of constraints remain consistent across different stocks, demonstrating the robustness of theoretical results

Specific Application Cases

Example 1: Cumulative Tsallis Past Entropy ($\text{CT}_\alpha(X)$)

For unimodal distributions ($\alpha = 2$, i.e., Gini mean difference): $\sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{Gini}(X) \geq \sigma\sqrt{3}$

For symmetric distributions: $\sup_{X \in \mathcal{V}_S(\mu, \sigma)} \text{Gini}(X) = \sigma\sqrt{3}$ Worst-case distribution: $F_X^{-1}(u) = \mu + \sigma\sqrt{3}(2u-1)$

Example 2: Cumulative Residual Entropy ($\alpha = 1$)

For unimodal distributions: $\sup_{X \in \mathcal{V}_U(\mu, \sigma)} E(X) \geq 0.927341\sigma$

For symmetric distributions: $\sup_{X \in \mathcal{V}_S(\mu, \sigma)} E(X) = \frac{\pi\sigma}{2\sqrt{3}}$

For symmetric unimodal distributions: $\sup_{X \in \mathcal{V}_{SU}(\mu, \sigma)} E(X) \geq 0.878282\sigma$

Example 3: Expected Shortfall ($\text{ES}_p(X)$)

For unimodal distributions: $\sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{ES}_p(X) = \begin{cases} \mu + \frac{\sqrt{p(8-9p)}}{3(1-p)}\sigma, & p \in (0, \frac{1}{2}] \\ \mu + \frac{1}{3}\sqrt{\frac{9p-1}{1-p}}\sigma, & p \in (\frac{1}{2}, 1) \end{cases}$

This result generalizes Lemma 2.6 from Li et al. (2018), which only required $p \in (\frac{5}{6}, 1)$.

Theoretical Verification

Remark 5: Through the inequality chain $\sup_{X \in \mathcal{V}_{SU}(\mu, \sigma)} \text{Gini}(X) \leq \sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{Gini}(X) \leq \sup_{X \in \mathcal{V}(\mu, \sigma)} \text{Gini}(X)$

Combined with the result from Zuo and Yin (2025) that $\sup_{X \in \mathcal{V}(\mu, \sigma)} \text{Gini}(X) = \sigma\sqrt{3}$ and the lower bound from this paper $\sup_{X \in \mathcal{V}_{SU}(\mu, \sigma)} \text{Gini}(X) \geq \sigma\sqrt{3}$, we obtain: $\sup_{X \in \mathcal{V}_{SU}(\mu, \sigma)} \text{Gini}(X) = \sup_{X \in \mathcal{V}_U(\mu, \sigma)} \text{Gini}(X) = \sigma\sqrt{3}$

This shows that for the Gini coefficient, unimodality and symmetric unimodality constraints yield the same exact bound.

Related Work

Risk Measures under Partial Information

1. VaR and TVaR: El Ghaoui et al. (2003) and Chen et al. (2011) provide closed-form solutions for VaR and TVaR when the first two moments are known
2. RVaR: Li et al. (2018) studied worst-case RVaR; Bernard et al. (2020, 2025) further investigated RVaR bounds for unimodal and symmetric unimodal distributions
3. Spectral Risk Measures: Li (2018) studied spectral risk measures (an important subclass of convex distortion risk metrics)
4. Symmetric Information: Zhu and Shao (2018) generalized Li (2018)'s results under symmetric information

Distortion Risk Metrics

1. Theoretical Foundation: Wang et al. (2020a,b) established a general theoretical framework for distortion risk metrics
2. Higher Moments: Liu et al. (2020) and Cai et al. (2025a) studied worst-case scenarios when higher moments are known
3. Wasserstein Ball: Bernard et al. (2024) and Cai et al. (2025b) studied cases where distributions lie in a Wasserstein ball
4. Extreme Cases: Shao and Zhang (2023, 2024) and Zhao et al. (2024) provided closed-form solutions when the first two moments and symmetry are known

Positioning of This Work

This is the first systematic study of distortion risk metrics and weighted entropy under unimodal and symmetric unimodal constraints, filling gaps in:

• Zuo and Yin (2025), which did not consider special distributional structures
• Bernard et al. (2020, 2025), which only addressed RVaR
• Zhao et al. (2024), which did not provide exact bounds for unimodal and symmetric unimodal distributions

Conclusions and Discussion

Main Conclusions

1. Unimodal Distributions: Establishes computable upper and lower bounds for distortion risk metrics and weighted entropy, with lower bounds given by extreme distributions $U_R$ and $U_L$, and upper bounds obtained through convex envelope and TVaR bounds
2. Symmetric Distributions: Obtains exact closed-form upper bounds $\sup_{X \in \mathcal{V}_S(\mu, \sigma)} \rho_g(X) = \mu g(1) + \frac{\sigma}{2}\sqrt{\int_0^1 [(\hat{g}(u) + \hat{g}(1-u))^\prime_*]^2 du}$
3. Symmetric Unimodal Distributions: Establishes upper and lower bounds, with lower bounds given by the set $\mathcal{S}(\alpha)$ and upper bounds obtained through convex envelope techniques
4. Broad Applicability: Theoretical results successfully applied to 15+ specific risk measures and entropy measures

Limitations

1. Computational Complexity:
   • Bounds for unimodal and symmetric unimodal distributions are not closed-form, requiring numerical optimization
   • Involve complex integrals and special functions (e.g., incomplete Gamma functions)
2. Tightness of Bounds:
   • For unimodal and symmetric unimodal distributions, gaps may exist between upper and lower bounds
   • In certain cases (e.g., Gini coefficient), bounds are tight, but tightness is unknown in general
3. Distributional Assumptions:
   • Only considers unimodal distributions (single mode)
   • Does not address multimodal or other complex distributional structures
4. Moment Constraints:
   • Uses only first two moments
   • Higher moments could provide tighter bounds

Future Directions

The authors explicitly propose two research directions:

1. Optimization Problems: Study optimal reinsurance problems for distortion risk metrics of unimodal and symmetric unimodal distributions under partial information
2. Multimodal Distributions: Investigate exact upper bounds for distortion risk metrics and weighted entropy under multimodal distributions
3. Other Distortion Measures: Extend results to other distortion measure families, such as parametric Wang premiums (Fiori and Rosazza Gianin, 2023)

In-Depth Evaluation

Strengths

1. Theoretical Rigor:
   • Complete and rigorous proofs, effectively utilizing convex analysis, quantile function theory, and monotone convergence theorem
   • Provides exact solutions for symmetric distributions, not merely bounds
2. Unified Framework:
   • Incorporates distortion risk metrics and weighted entropy into a unified framework
   • Systematically addresses three classes of distributions: unimodal, symmetric, and symmetric unimodal
3. Broad Applicability:
   • Theoretical results applied to 15+ specific measures
   • Covers three major application categories: entropy, shortfalls, and risk measures
   • Extended to residual and past lifetimes
4. Methodological Innovation:
   • Cleverly combines convex envelope techniques with TVaR bounds
   • Constructed extreme distribution sets have clear geometric interpretation
5. Empirical Validation:
   • Uses real stock data to verify theory
   • Clearly demonstrates relative effects of different constraints

Weaknesses

1. Computability Issues:
   • Bounds for unimodal and symmetric unimodal distributions require solving complex optimization problems
   • Equations like $M_\alpha(b_0) = 0$ in Examples 1 and 2 require numerical solution
   • Lacks algorithmic implementation details and computational complexity analysis
2. Insufficient Tightness Analysis:
   • Except for a few special cases (e.g., Gini coefficient), lacks systematic analysis of gaps between bounds
   • No asymptotic behavior analysis of bounds
3. Weak Experimental Section:
   • Uses only 3 stocks, relatively small sample size
   • Short data period (approximately one year)
   • Lacks quantitative tests for unimodality and symmetry (relies only on histograms)
   • No comparison with other methods (e.g., empirical estimates, parametric methods)
4. Writing Issues:
   • Paper is lengthy (30 pages), structure could be more compact
   • Some proofs are verbose and could be moved to appendix
   • Figure quality is average (e.g., Figures 1-5 have low resolution and labeling)
5. Theoretical Limitations:
   • Unimodality definition based on convexity/concavity of distribution function is restrictive
   • Does not consider cases where mode location is known
   • For weighted entropy, the practical meaning of constraint $E[\Psi(F_X^{-1}(U))] = \mu_\Psi$ is unclear

Impact

1. Academic Contribution:
   • Fills the gap in research on distortion risk metrics under unimodal and symmetric unimodal constraints
   • Provides theoretical foundation and technical tools for subsequent research
   • Expected to be cited in risk management, actuarial science, and decision theory
2. Practical Value:
   • Provides conservative risk assessment tools for insurance and financial institutions
   • Particularly useful when distribution is uncertain but structural information is known
   • Applicable to stress testing and scenario analysis
3. Reproducibility:
   • Detailed theoretical proofs enable reproduction
   • Lacks public code and data; experimental reproducibility is moderate
   • Some numerical results (e.g., $b_0 \approx 0.582812$) lack explanation of calculation methods

Applicable Scenarios

1. Risk Management:
   • Insurance company reserve calculations
   • Bank capital adequacy assessment
   • Portfolio risk limit setting
2. Actuarial Pricing:
   • Robust premium principle design
   • Annuity product pricing
   • Reinsurance contract design
3. Regulatory Compliance:
   • When complete distribution models are unavailable but distributional features are known
   • Cases requiring conservative estimates
   • Risk assessment under Solvency II and similar regulatory frameworks
4. Decision Analysis:
   • Robust decision-making under distributional uncertainty
   • Risk aversion modeling in behavioral economics
5. Inapplicable Scenarios:
   • When data is sufficient and complete distribution can be reliably estimated
   • When distribution is clearly multimodal or violates assumptions
   • Applications requiring exact values rather than bounds

Key References

1. Bernard, C., Kazzi, R., Vanduffel, S. (2020, 2025): RVaR bounds for unimodal and symmetric unimodal distributions; important foundation for this work
2. Zuo, B., Yin, C. (2025): Distortion risk metrics for general distributions; directly extended by this work
3. Wang, Q., Wang, R., Wei, Y. (2020a): General theoretical framework for distortion risk metrics
4. Shao, H., Zhang, Z.G. (2023, 2024): Distortion risk metrics under parameter ambiguity; provides technical methods
5. Zhao, M., et al. (2024): Optimal and worst-case distortion risk metrics under partial information constraints

Overall Assessment

This is a theoretically rigorous and broadly applicable paper in risk management. Main strengths include: (1) establishing a systematic theoretical framework for distortion risk metrics under unimodal and symmetric distributions; (2) providing exact solutions for symmetric distributions; (3) broad application scope. Main weaknesses include: (1) bounds for unimodal and symmetric unimodal distributions require numerical computation, limiting operability; (2) relatively weak experimental section; (3) insufficient tightness analysis.

The paper is suitable for publication in journals in actuarial science, risk management, or applied probability (e.g., Insurance: Mathematics and Economics, ASTIN Bulletin). For practitioners, this work provides practical conservative risk assessment tools; for researchers, it offers rich technical methods and future research directions.

Recommendation Score: 4/5 (Significant theoretical contribution, but room for improvement in experiments and operability)